stable_linalg.mlir

#map = affine_map<(d0) -> (d0)>
#map1 = affine_map<(d0) -> ()>
#map2 = affine_map<() -> ()>
#map3 = affine_map<(d0, d1) -> ()>
#map4 = affine_map<(d0, d1) -> (d0, d1)>
#map5 = affine_map<(d0, d1) -> (d0, 0)>
#map6 = affine_map<(d0, d1) -> (0, d1)>
#map7 = affine_map<(d0, d1) -> (d1, d0)>
#map8 = affine_map<(d0, d1) -> (d1)>
#map9 = affine_map<(d0, d1, d2, d3) -> ()>
#map10 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map11 = affine_map<(d0, d1, d2, d3) -> (d1)>
#map12 = affine_map<(d0, d1, d2, d3) -> (d0, d1, 0, 0)>
#map13 = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, 0)>
#map14 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)>
#map15 = affine_map<(d0, d1, d2) -> ()>
#map16 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map17 = affine_map<(d0, d1, d2) -> (d0, d1, 0)>
#map18 = affine_map<(d0, d1, d2) -> (d2)>
#map19 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map20 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d3, d2)>
#map21 = affine_map<(d0, d1, d2, d3) -> (d0, d3, d1, d2)>
#map22 = affine_map<(d0, d1, d2, d3) -> (d0, d3, d2, d1)>
module attributes {torch.debug_module_name = "_lambda"} {
  func.func @forward(%arg0: tensor<2x4x96x96xf32>, %arg1: tensor<2xi64>, %arg2: tensor<2x77x1024xf32>) -> tensor<2x4x96x96xf16> {
    %cst = arith.constant dense<1.250000e-01> : tensor<f64>
    %cst_0 = arith.constant dense<9.9999999999999995E-7> : tensor<f64>
    %cst_1 = arith.constant dense<1.000000e-05> : tensor<f64>
    %cst_2 = arith.constant dense<160> : tensor<i64>
    %cst_3 = arith.constant dense<-9.2103403719761836> : tensor<f64>
    %cst_4 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_5 = arith.constant dense_resource<__elided__> : tensor<1280x320xf32>
    %cst_6 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_7 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_8 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_9 = arith.constant dense_resource<__elided__> : tensor<320x4x3x3xf32>
    %cst_10 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_11 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_12 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_13 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_14 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_15 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_16 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_17 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_18 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_19 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_20 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_21 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_22 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_23 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_24 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_25 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_26 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_27 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_28 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_29 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_30 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_31 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_32 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_33 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_34 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_35 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_36 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_37 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_38 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_39 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_40 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_41 = arith.constant dense_resource<__elided__> : tensor<2560x320xf32>
    %cst_42 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_43 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_44 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_45 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_46 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_47 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_48 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_49 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_50 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_51 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_52 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_53 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_54 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_55 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_56 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_57 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_58 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_59 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_60 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_61 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_62 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_63 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_64 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_65 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_66 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_67 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_68 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_69 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_70 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_71 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_72 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_73 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_74 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_75 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_76 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_77 = arith.constant dense_resource<__elided__> : tensor<2560x320xf32>
    %cst_78 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_79 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_80 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_81 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_82 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_83 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_84 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_85 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_86 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_87 = arith.constant dense_resource<__elided__> : tensor<640x320x3x3xf32>
    %cst_88 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_89 = arith.constant dense_resource<__elided__> : tensor<640x1280xf32>
    %cst_90 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_91 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_92 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_93 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_94 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_95 = arith.constant dense_resource<__elided__> : tensor<640x320x1x1xf32>
    %cst_96 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_97 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_98 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_99 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_100 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_101 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_102 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_103 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_104 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_105 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_106 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_107 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_108 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_109 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_110 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_111 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_112 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_113 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_114 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_115 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_116 = arith.constant dense_resource<__elided__> : tensor<5120xf32>
    %cst_117 = arith.constant dense_resource<__elided__> : tensor<5120x640xf32>
    %cst_118 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_119 = arith.constant dense_resource<__elided__> : tensor<640x2560xf32>
    %cst_120 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_121 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_122 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_123 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_124 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_125 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_126 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_127 = arith.constant dense_resource<__elided__> : tensor<640x1280xf32>
    %cst_128 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_129 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_130 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_131 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_132 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_133 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_134 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_135 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_136 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_137 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_138 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_139 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_140 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_141 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_142 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_143 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_144 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_145 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_146 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_147 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_148 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_149 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_150 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_151 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_152 = arith.constant dense_resource<__elided__> : tensor<5120xf32>
    %cst_153 = arith.constant dense_resource<__elided__> : tensor<5120x640xf32>
    %cst_154 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_155 = arith.constant dense_resource<__elided__> : tensor<640x2560xf32>
    %cst_156 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_157 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_158 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_159 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_160 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_161 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_162 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_163 = arith.constant dense_resource<__elided__> : tensor<1280x640x3x3xf32>
    %cst_164 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_165 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_166 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_167 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_168 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_169 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_170 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_171 = arith.constant dense_resource<__elided__> : tensor<1280x640x1x1xf32>
    %cst_172 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_173 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_174 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_175 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_176 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_177 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_178 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_179 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_180 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_181 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_182 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_183 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_184 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_185 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_186 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_187 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_188 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_189 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_190 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_191 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_192 = arith.constant dense_resource<__elided__> : tensor<10240xf32>
    %cst_193 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf32>
    %cst_194 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_195 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf32>
    %cst_196 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_197 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_198 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_199 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_200 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_201 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_202 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_203 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_204 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_205 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_206 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_207 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_208 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_209 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_210 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_211 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_212 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_213 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_214 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_215 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_216 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_217 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_218 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_219 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_220 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_221 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_222 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_223 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_224 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_225 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_226 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_227 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_228 = arith.constant dense_resource<__elided__> : tensor<10240xf32>
    %cst_229 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf32>
    %cst_230 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_231 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf32>
    %cst_232 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_233 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_234 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_235 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_236 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_237 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_238 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_239 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_240 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_241 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_242 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_243 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_244 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_245 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_246 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_247 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_248 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_249 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_250 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_251 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_252 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_253 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_254 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_255 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_256 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_257 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_258 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_259 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_260 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_261 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_262 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_263 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_264 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_265 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_266 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_267 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_268 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_269 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_270 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_271 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_272 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_273 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_274 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_275 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_276 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_277 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_278 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_279 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_280 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_281 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_282 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_283 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_284 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_285 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_286 = arith.constant dense_resource<__elided__> : tensor<10240xf32>
    %cst_287 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf32>
    %cst_288 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_289 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf32>
    %cst_290 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_291 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_292 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_293 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_294 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_295 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_296 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_297 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_298 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_299 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_300 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_301 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_302 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_303 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_304 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_305 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf32>
    %cst_306 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_307 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_308 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_309 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_310 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_311 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_312 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_313 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf32>
    %cst_314 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_315 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_316 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_317 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf32>
    %cst_318 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_319 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_320 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_321 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_322 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_323 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_324 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_325 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf32>
    %cst_326 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_327 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_328 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_329 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf32>
    %cst_330 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_331 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_332 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_333 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_334 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_335 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_336 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_337 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf32>
    %cst_338 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_339 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_340 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_341 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_342 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_343 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf32>
    %cst_344 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_345 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_346 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_347 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_348 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_349 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_350 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_351 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf32>
    %cst_352 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_353 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_354 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_355 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_356 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_357 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_358 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_359 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_360 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_361 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_362 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_363 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_364 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_365 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_366 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_367 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_368 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_369 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_370 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_371 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_372 = arith.constant dense_resource<__elided__> : tensor<10240xf32>
    %cst_373 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf32>
    %cst_374 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_375 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf32>
    %cst_376 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_377 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_378 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_379 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_380 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_381 = arith.constant dense_resource<__elided__> : tensor<1280x2560x3x3xf32>
    %cst_382 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_383 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_384 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_385 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_386 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_387 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_388 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_389 = arith.constant dense_resource<__elided__> : tensor<1280x2560x1x1xf32>
    %cst_390 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_391 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_392 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_393 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_394 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_395 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_396 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_397 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_398 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_399 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_400 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_401 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_402 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_403 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_404 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_405 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_406 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_407 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_408 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_409 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_410 = arith.constant dense_resource<__elided__> : tensor<10240xf32>
    %cst_411 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf32>
    %cst_412 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_413 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf32>
    %cst_414 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_415 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_416 = arith.constant dense_resource<__elided__> : tensor<1920xf32>
    %cst_417 = arith.constant dense_resource<__elided__> : tensor<1920xf32>
    %cst_418 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_419 = arith.constant dense_resource<__elided__> : tensor<1280x1920x3x3xf32>
    %cst_420 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_421 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_422 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_423 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_424 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_425 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_426 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_427 = arith.constant dense_resource<__elided__> : tensor<1280x1920x1x1xf32>
    %cst_428 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_429 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_430 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_431 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_432 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_433 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_434 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_435 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_436 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_437 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_438 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_439 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_440 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_441 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_442 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_443 = arith.constant dense_resource<__elided__> : tensor<1280x1024xf32>
    %cst_444 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_445 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_446 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_447 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_448 = arith.constant dense_resource<__elided__> : tensor<10240xf32>
    %cst_449 = arith.constant dense_resource<__elided__> : tensor<10240x1280xf32>
    %cst_450 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_451 = arith.constant dense_resource<__elided__> : tensor<1280x5120xf32>
    %cst_452 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_453 = arith.constant dense_resource<__elided__> : tensor<1280x1280xf32>
    %cst_454 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_455 = arith.constant dense_resource<__elided__> : tensor<1280x1280x3x3xf32>
    %cst_456 = arith.constant dense_resource<__elided__> : tensor<1920xf32>
    %cst_457 = arith.constant dense_resource<__elided__> : tensor<1920xf32>
    %cst_458 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_459 = arith.constant dense_resource<__elided__> : tensor<640x1920x3x3xf32>
    %cst_460 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_461 = arith.constant dense_resource<__elided__> : tensor<640x1280xf32>
    %cst_462 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_463 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_464 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_465 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_466 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_467 = arith.constant dense_resource<__elided__> : tensor<640x1920x1x1xf32>
    %cst_468 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_469 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_470 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_471 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_472 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_473 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_474 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_475 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_476 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_477 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_478 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_479 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_480 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_481 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_482 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_483 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_484 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_485 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_486 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_487 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_488 = arith.constant dense_resource<__elided__> : tensor<5120xf32>
    %cst_489 = arith.constant dense_resource<__elided__> : tensor<5120x640xf32>
    %cst_490 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_491 = arith.constant dense_resource<__elided__> : tensor<640x2560xf32>
    %cst_492 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_493 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_494 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_495 = arith.constant dense_resource<__elided__> : tensor<1280xf32>
    %cst_496 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_497 = arith.constant dense_resource<__elided__> : tensor<640x1280x3x3xf32>
    %cst_498 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_499 = arith.constant dense_resource<__elided__> : tensor<640x1280xf32>
    %cst_500 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_501 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_502 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_503 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_504 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_505 = arith.constant dense_resource<__elided__> : tensor<640x1280x1x1xf32>
    %cst_506 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_507 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_508 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_509 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_510 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_511 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_512 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_513 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_514 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_515 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_516 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_517 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_518 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_519 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_520 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_521 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_522 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_523 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_524 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_525 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_526 = arith.constant dense_resource<__elided__> : tensor<5120xf32>
    %cst_527 = arith.constant dense_resource<__elided__> : tensor<5120x640xf32>
    %cst_528 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_529 = arith.constant dense_resource<__elided__> : tensor<640x2560xf32>
    %cst_530 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_531 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_532 = arith.constant dense_resource<__elided__> : tensor<960xf32>
    %cst_533 = arith.constant dense_resource<__elided__> : tensor<960xf32>
    %cst_534 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_535 = arith.constant dense_resource<__elided__> : tensor<640x960x3x3xf32>
    %cst_536 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_537 = arith.constant dense_resource<__elided__> : tensor<640x1280xf32>
    %cst_538 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_539 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_540 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_541 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_542 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_543 = arith.constant dense_resource<__elided__> : tensor<640x960x1x1xf32>
    %cst_544 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_545 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_546 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_547 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_548 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_549 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_550 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_551 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_552 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_553 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_554 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_555 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_556 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_557 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_558 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_559 = arith.constant dense_resource<__elided__> : tensor<640x1024xf32>
    %cst_560 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_561 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_562 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_563 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_564 = arith.constant dense_resource<__elided__> : tensor<5120xf32>
    %cst_565 = arith.constant dense_resource<__elided__> : tensor<5120x640xf32>
    %cst_566 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_567 = arith.constant dense_resource<__elided__> : tensor<640x2560xf32>
    %cst_568 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_569 = arith.constant dense_resource<__elided__> : tensor<640x640xf32>
    %cst_570 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_571 = arith.constant dense_resource<__elided__> : tensor<640x640x3x3xf32>
    %cst_572 = arith.constant dense_resource<__elided__> : tensor<960xf32>
    %cst_573 = arith.constant dense_resource<__elided__> : tensor<960xf32>
    %cst_574 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_575 = arith.constant dense_resource<__elided__> : tensor<320x960x3x3xf32>
    %cst_576 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_577 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_578 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_579 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_580 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_581 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_582 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_583 = arith.constant dense_resource<__elided__> : tensor<320x960x1x1xf32>
    %cst_584 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_585 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_586 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_587 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_588 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_589 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_590 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_591 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_592 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_593 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_594 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_595 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_596 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_597 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_598 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_599 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_600 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_601 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_602 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_603 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_604 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_605 = arith.constant dense_resource<__elided__> : tensor<2560x320xf32>
    %cst_606 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_607 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_608 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_609 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_610 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_611 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_612 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_613 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xf32>
    %cst_614 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_615 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_616 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_617 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_618 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_619 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_620 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_621 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xf32>
    %cst_622 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_623 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_624 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_625 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_626 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_627 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_628 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_629 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_630 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_631 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_632 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_633 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_634 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_635 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_636 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_637 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_638 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_639 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_640 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_641 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_642 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_643 = arith.constant dense_resource<__elided__> : tensor<2560x320xf32>
    %cst_644 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_645 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_646 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_647 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_648 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_649 = arith.constant dense_resource<__elided__> : tensor<640xf32>
    %cst_650 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_651 = arith.constant dense_resource<__elided__> : tensor<320x640x3x3xf32>
    %cst_652 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_653 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_654 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_655 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_656 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_657 = arith.constant dense_resource<__elided__> : tensor<320x320x3x3xf32>
    %cst_658 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_659 = arith.constant dense_resource<__elided__> : tensor<320x640x1x1xf32>
    %cst_660 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_661 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_662 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_663 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_664 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_665 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_666 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_667 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_668 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_669 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_670 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_671 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_672 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_673 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_674 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_675 = arith.constant dense_resource<__elided__> : tensor<320x1024xf32>
    %cst_676 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_677 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_678 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_679 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_680 = arith.constant dense_resource<__elided__> : tensor<2560xf32>
    %cst_681 = arith.constant dense_resource<__elided__> : tensor<2560x320xf32>
    %cst_682 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_683 = arith.constant dense_resource<__elided__> : tensor<320x1280xf32>
    %cst_684 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_685 = arith.constant dense_resource<__elided__> : tensor<320x320xf32>
    %cst_686 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_687 = arith.constant dense_resource<__elided__> : tensor<320xf32>
    %cst_688 = arith.constant dense<[-0.0333971679, 0.0151019702, -0.0098297568, -0.00283672824]> : tensor<4xf32>
    %cst_689 = arith.constant dense_resource<__elided__> : tensor<4x320x3x3xf32>
    %cst_690 = arith.constant 0.000000e+00 : f16
    %cst_691 = arith.constant 1.000000e+00 : f16
    %cst_692 = arith.constant -3.40282347E+38 : f32
    %cst_693 = arith.constant 2.000000e+00 : f16
    %cst_694 = arith.constant 5.000000e-01 : f16
    %cst_695 = arith.constant 0.000000e+00 : f64
    %cst_696 = arith.constant 1.000000e-05 : f64
    %cst_697 = arith.constant 0.000000e+00 : f32
    %cst_698 = arith.constant 1.000000e+00 : f32
    %cst_699 = arith.constant 9.216000e+04 : f64
    %cst_700 = arith.constant 3.200000e+02 : f64
    %cst_701 = arith.constant 2.304000e+04 : f64
    %cst_702 = arith.constant 4.608000e+04 : f64
    %cst_703 = arith.constant 6.400000e+02 : f64
    %cst_704 = arith.constant 1.152000e+04 : f64
    %cst_705 = arith.constant 1.280000e+03 : f64
    %cst_706 = arith.constant 5.760000e+03 : f64
    %cst_707 = arith.constant 3.456000e+04 : f64
    %cst_708 = arith.constant 1.382400e+05 : f64
    %cst_709 = arith.constant 6.912000e+04 : f64
    %cst_710 = arith.constant 2.764800e+05 : f64
    %cst_711 = arith.constant 1.843200e+05 : f64
    %c2 = arith.constant 2 : index
    %0 = tensor.empty() : tensor<160xf32>
    %1 = linalg.generic {indexing_maps = [#map], iterator_types = ["parallel"]} outs(%0 : tensor<160xf32>) {
    ^bb0(%out: f32):
      %7524 = linalg.index 0 : index
      %7525 = arith.index_cast %7524 : index to i64
      %7526 = arith.sitofp %7525 : i64 to f32
      %7527 = arith.addf %7526, %cst_697 : f32
      linalg.yield %7527 : f32
    } -> tensor<160xf32>
    %2 = linalg.generic {indexing_maps = [#map, #map1, #map], iterator_types = ["parallel"]} ins(%1, %cst_3 : tensor<160xf32>, tensor<f64>) outs(%0 : tensor<160xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.mulf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<160xf32>
    %3 = linalg.generic {indexing_maps = [#map, #map1, #map], iterator_types = ["parallel"]} ins(%2, %cst_2 : tensor<160xf32>, tensor<i64>) outs(%0 : tensor<160xf32>) {
    ^bb0(%in: f32, %in_2119: i64, %out: f32):
      %7524 = arith.sitofp %in_2119 : i64 to f32
      %7525 = arith.divf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<160xf32>
    %4 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%3 : tensor<160xf32>) outs(%0 : tensor<160xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<160xf32>
    %expanded = tensor.expand_shape %arg1 [[0, 1]] : tensor<2xi64> into tensor<2x1xi64>
    %5 = tensor.empty() : tensor<f64>
    %6 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %7 = tensor.empty() : tensor<f32>
    %8 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%6 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %9 = tensor.empty() : tensor<2x1xf32>
    %10 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%8 : tensor<f32>) outs(%9 : tensor<2x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1xf32>
    %11 = tensor.empty() : tensor<2x1xi64>
    %12 = linalg.generic {indexing_maps = [#map5, #map4], iterator_types = ["parallel", "parallel"]} ins(%expanded : tensor<2x1xi64>) outs(%11 : tensor<2x1xi64>) {
    ^bb0(%in: i64, %out: i64):
      linalg.yield %in : i64
    } -> tensor<2x1xi64>
    %13 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%12 : tensor<2x1xi64>) outs(%10 : tensor<2x1xf32>) {
    ^bb0(%in: i64, %out: f32):
      %7524 = arith.sitofp %in : i64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1xf32>
    %expanded_712 = tensor.expand_shape %4 [[0, 1]] : tensor<160xf32> into tensor<1x160xf32>
    %14 = tensor.empty() : tensor<2x160xf32>
    %15 = linalg.generic {indexing_maps = [#map5, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%13, %expanded_712 : tensor<2x1xf32>, tensor<1x160xf32>) outs(%14 : tensor<2x160xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x160xf32>
    %16 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%15 : tensor<2x160xf32>) outs(%14 : tensor<2x160xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.cos %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x160xf32>
    %17 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%15 : tensor<2x160xf32>) outs(%14 : tensor<2x160xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.sin %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x160xf32>
    %18 = tensor.empty() : tensor<2x320xf32>
    %inserted_slice = tensor.insert_slice %16 into %18[0, 0] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
    %inserted_slice_713 = tensor.insert_slice %17 into %inserted_slice[0, 160] [2, 160] [1, 1] : tensor<2x160xf32> into tensor<2x320xf32>
    %19 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %20 = tensor.empty() : tensor<f16>
    %21 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%19 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %22 = tensor.empty() : tensor<1280xf16>
    %23 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%21 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %24 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_4 : tensor<1280xf32>) outs(%23 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %25 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %26 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%25 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %27 = tensor.empty() : tensor<1280x320xf16>
    %28 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%26 : tensor<f16>) outs(%27 : tensor<1280x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x320xf16>
    %29 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_5 : tensor<1280x320xf32>) outs(%28 : tensor<1280x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x320xf16>
    %30 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %31 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%30 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %32 = tensor.empty() : tensor<2x320xf16>
    %33 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%31 : tensor<f16>) outs(%32 : tensor<2x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320xf16>
    %34 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%inserted_slice_713 : tensor<2x320xf32>) outs(%33 : tensor<2x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x320xf16>
    %35 = tensor.empty() : tensor<320x1280xf16>
    %36 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%29 : tensor<1280x320xf16>) outs(%35 : tensor<320x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1280xf16>
    %37 = tensor.empty() : tensor<2x1280xf16>
    %38 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %39 = linalg.matmul ins(%34, %36 : tensor<2x320xf16>, tensor<320x1280xf16>) outs(%38 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %40 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%24, %39 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %41 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%40 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %42 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%41, %40 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %43 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %44 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%43 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %45 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%44 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %46 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_6 : tensor<1280xf32>) outs(%45 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %47 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %48 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%47 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %49 = tensor.empty() : tensor<1280x1280xf16>
    %50 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%48 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %51 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_7 : tensor<1280x1280xf32>) outs(%50 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %52 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%51 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %53 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %54 = linalg.matmul ins(%42, %52 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%53 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %55 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%46, %54 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %56 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %57 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%56 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %58 = tensor.empty() : tensor<2x4x96x96xf16>
    %59 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%57 : tensor<f16>) outs(%58 : tensor<2x4x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x4x96x96xf16>
    %60 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<2x4x96x96xf32>) outs(%59 : tensor<2x4x96x96xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x4x96x96xf16>
    %61 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %62 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%61 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %63 = tensor.empty() : tensor<320xf16>
    %64 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%62 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %65 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_8 : tensor<320xf32>) outs(%64 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %66 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %67 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%66 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %68 = tensor.empty() : tensor<320x4x3x3xf16>
    %69 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%67 : tensor<f16>) outs(%68 : tensor<320x4x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x4x3x3xf16>
    %70 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_9 : tensor<320x4x3x3xf32>) outs(%69 : tensor<320x4x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x4x3x3xf16>
    %padded = tensor.pad %60 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x4x96x96xf16> to tensor<2x4x98x98xf16>
    %71 = tensor.empty() : tensor<2x320x96x96xf16>
    %72 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%65 : tensor<320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %73 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded, %70 : tensor<2x4x98x98xf16>, tensor<320x4x3x3xf16>) outs(%72 : tensor<2x320x96x96xf16>) -> tensor<2x320x96x96xf16>
    %74 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %75 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%74 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %76 = tensor.empty() : tensor<2x320x96x96xf32>
    %77 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%75 : tensor<f32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x96x96xf32>
    %78 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%73 : tensor<2x320x96x96xf16>) outs(%77 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %collapsed = tensor.collapse_shape %78 [[0], [1], [2, 3]] : tensor<2x320x96x96xf32> into tensor<2x320x9216xf32>
    %expanded_714 = tensor.expand_shape %collapsed [[0], [1, 2], [3]] : tensor<2x320x9216xf32> into tensor<2x32x10x9216xf32>
    %79 = tensor.empty() : tensor<2x32x10x9216xf64>
    %80 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_714 : tensor<2x32x10x9216xf32>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %81 = tensor.empty() : tensor<2x32x1x1xf64>
    %82 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %83 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%80 : tensor<2x32x10x9216xf64>) outs(%82 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %84 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%83 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %85 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%80, %84 : tensor<2x32x10x9216xf64>, tensor<2x32x1x1xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %86 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%85 : tensor<2x32x10x9216xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %87 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %88 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%86 : tensor<2x32x10x9216xf64>) outs(%87 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %89 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%88 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %90 = tensor.empty() : tensor<2x32x1x1xf32>
    %91 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%89 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %92 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %93 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_714 : tensor<2x32x10x9216xf32>) outs(%92 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %94 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%93 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %95 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%94 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %96 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%91, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %97 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%96 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %98 = tensor.empty() : tensor<2x32x10x9216xf32>
    %99 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_714, %95 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %100 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%99, %97 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %collapsed_715 = tensor.collapse_shape %100 [[0], [1, 2], [3]] : tensor<2x32x10x9216xf32> into tensor<2x320x9216xf32>
    %expanded_716 = tensor.expand_shape %collapsed_715 [[0], [1], [2, 3]] : tensor<2x320x9216xf32> into tensor<2x320x96x96xf32>
    %expanded_717 = tensor.expand_shape %cst_10 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_718 = tensor.expand_shape %cst_11 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %101 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_716, %expanded_718 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %102 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%101, %expanded_717 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %103 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %104 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%103 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %105 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%104 : tensor<f16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %106 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%102 : tensor<2x320x96x96xf32>) outs(%105 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %107 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%106 : tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x320x96x96xf16>
    %108 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%107, %106 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %109 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %110 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%109 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %111 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%110 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %112 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_12 : tensor<320xf32>) outs(%111 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %113 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %114 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%113 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %115 = tensor.empty() : tensor<320x320x3x3xf16>
    %116 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%114 : tensor<f16>) outs(%115 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320x3x3xf16>
    %117 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_13 : tensor<320x320x3x3xf32>) outs(%116 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320x3x3xf16>
    %padded_719 = tensor.pad %108 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x320x96x96xf16> to tensor<2x320x98x98xf16>
    %118 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%112 : tensor<320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %119 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_719, %117 : tensor<2x320x98x98xf16>, tensor<320x320x3x3xf16>) outs(%118 : tensor<2x320x96x96xf16>) -> tensor<2x320x96x96xf16>
    %120 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %121 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%120, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %122 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %123 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%122 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %124 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%123 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %125 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_14 : tensor<320xf32>) outs(%124 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %126 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %127 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%126 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %128 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%127 : tensor<f16>) outs(%35 : tensor<320x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1280xf16>
    %129 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_15 : tensor<320x1280xf32>) outs(%128 : tensor<320x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1280xf16>
    %130 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%129 : tensor<320x1280xf16>) outs(%27 : tensor<1280x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x320xf16>
    %131 = linalg.fill ins(%cst_690 : f16) outs(%32 : tensor<2x320xf16>) -> tensor<2x320xf16>
    %132 = linalg.matmul ins(%121, %130 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%131 : tensor<2x320xf16>) -> tensor<2x320xf16>
    %133 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%125, %132 : tensor<320xf16>, tensor<2x320xf16>) outs(%32 : tensor<2x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320xf16>
    %expanded_720 = tensor.expand_shape %133 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16>
    %134 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%119, %expanded_720 : tensor<2x320x96x96xf16>, tensor<2x320x1x1xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %135 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %136 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%135 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %137 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%136 : tensor<f32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x96x96xf32>
    %138 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%134 : tensor<2x320x96x96xf16>) outs(%137 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %collapsed_721 = tensor.collapse_shape %138 [[0], [1], [2, 3]] : tensor<2x320x96x96xf32> into tensor<2x320x9216xf32>
    %expanded_722 = tensor.expand_shape %collapsed_721 [[0], [1, 2], [3]] : tensor<2x320x9216xf32> into tensor<2x32x10x9216xf32>
    %139 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_722 : tensor<2x32x10x9216xf32>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %140 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %141 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%139 : tensor<2x32x10x9216xf64>) outs(%140 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %142 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%141 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %143 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%139, %142 : tensor<2x32x10x9216xf64>, tensor<2x32x1x1xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %144 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%143 : tensor<2x32x10x9216xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %145 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %146 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%144 : tensor<2x32x10x9216xf64>) outs(%145 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %147 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%146 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %148 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%147 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %149 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %150 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_722 : tensor<2x32x10x9216xf32>) outs(%149 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %151 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%150 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %152 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%151 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %153 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%148, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %154 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%153 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %155 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_722, %152 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %156 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%155, %154 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %collapsed_723 = tensor.collapse_shape %156 [[0], [1, 2], [3]] : tensor<2x32x10x9216xf32> into tensor<2x320x9216xf32>
    %expanded_724 = tensor.expand_shape %collapsed_723 [[0], [1], [2, 3]] : tensor<2x320x9216xf32> into tensor<2x320x96x96xf32>
    %expanded_725 = tensor.expand_shape %cst_16 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_726 = tensor.expand_shape %cst_17 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %157 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_724, %expanded_726 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %158 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%157, %expanded_725 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %159 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %160 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%159 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %161 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%160 : tensor<f16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %162 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%158 : tensor<2x320x96x96xf32>) outs(%161 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %163 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%162 : tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x320x96x96xf16>
    %164 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%163, %162 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %165 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %166 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%165 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %167 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%166 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %168 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_18 : tensor<320xf32>) outs(%167 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %169 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %170 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%169 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %171 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%170 : tensor<f16>) outs(%115 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320x3x3xf16>
    %172 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_19 : tensor<320x320x3x3xf32>) outs(%171 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320x3x3xf16>
    %padded_727 = tensor.pad %164 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x320x96x96xf16> to tensor<2x320x98x98xf16>
    %173 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%168 : tensor<320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %174 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_727, %172 : tensor<2x320x98x98xf16>, tensor<320x320x3x3xf16>) outs(%173 : tensor<2x320x96x96xf16>) -> tensor<2x320x96x96xf16>
    %175 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%73, %174 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %176 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %177 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%176 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %178 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%177 : tensor<f32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x96x96xf32>
    %179 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%175 : tensor<2x320x96x96xf16>) outs(%178 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %collapsed_728 = tensor.collapse_shape %179 [[0], [1], [2, 3]] : tensor<2x320x96x96xf32> into tensor<2x320x9216xf32>
    %expanded_729 = tensor.expand_shape %collapsed_728 [[0], [1, 2], [3]] : tensor<2x320x9216xf32> into tensor<2x32x10x9216xf32>
    %180 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_729 : tensor<2x32x10x9216xf32>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %181 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %182 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%180 : tensor<2x32x10x9216xf64>) outs(%181 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %183 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%182 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %184 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%180, %183 : tensor<2x32x10x9216xf64>, tensor<2x32x1x1xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %185 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%184 : tensor<2x32x10x9216xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %186 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %187 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%185 : tensor<2x32x10x9216xf64>) outs(%186 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %188 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%187 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %189 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%188 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %190 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %191 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_729 : tensor<2x32x10x9216xf32>) outs(%190 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %192 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%191 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %193 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%192 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %194 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%189, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %195 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%194 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %196 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_729, %193 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %197 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%196, %195 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %collapsed_730 = tensor.collapse_shape %197 [[0], [1, 2], [3]] : tensor<2x32x10x9216xf32> into tensor<2x320x9216xf32>
    %expanded_731 = tensor.expand_shape %collapsed_730 [[0], [1], [2, 3]] : tensor<2x320x9216xf32> into tensor<2x320x96x96xf32>
    %expanded_732 = tensor.expand_shape %cst_20 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_733 = tensor.expand_shape %cst_21 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %198 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_731, %expanded_733 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %199 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%198, %expanded_732 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %200 = tensor.empty() : tensor<2x96x96x320xf32>
    %201 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%199 : tensor<2x320x96x96xf32>) outs(%200 : tensor<2x96x96x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x96x96x320xf32>
    %collapsed_734 = tensor.collapse_shape %201 [[0], [1, 2], [3]] : tensor<2x96x96x320xf32> into tensor<2x9216x320xf32>
    %202 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %203 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%202 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %204 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%203 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %205 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_22 : tensor<320xf32>) outs(%204 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %206 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %207 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%206 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %208 = tensor.empty() : tensor<320x320xf16>
    %209 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%207 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %210 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_23 : tensor<320x320xf32>) outs(%209 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %211 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %212 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%211 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %213 = tensor.empty() : tensor<2x9216x320xf16>
    %214 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%212 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %215 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_734 : tensor<2x9216x320xf32>) outs(%214 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %216 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%210 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_735 = tensor.collapse_shape %215 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %217 = tensor.empty() : tensor<18432x320xf16>
    %218 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %219 = linalg.matmul ins(%collapsed_735, %216 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%218 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %220 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%205, %219 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_736 = tensor.expand_shape %220 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %221 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %222 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%221 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %223 = tensor.empty() : tensor<2x9216x320xf32>
    %224 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%222 : tensor<f32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %225 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_736 : tensor<2x9216x320xf16>) outs(%224 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %226 = tensor.empty() : tensor<2x9216x1xf64>
    %227 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %228 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%225 : tensor<2x9216x320xf32>) outs(%227 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %229 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%228 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %230 = tensor.empty() : tensor<2x9216x1xf32>
    %231 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%229 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %232 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%231 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %233 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%225, %232 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %234 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%233 : tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %235 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %236 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%234 : tensor<2x9216x320xf32>) outs(%235 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %237 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%236 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %238 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%237 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %239 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%238 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x9216x1xf32>
    %240 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%239 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %241 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%240 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %242 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%233, %241 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %243 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%242, %cst_24 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %244 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%243, %cst_25 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %245 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %246 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%245 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %247 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%246 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %248 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_26 : tensor<320x320xf32>) outs(%247 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %249 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %250 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%249 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %251 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%250 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %252 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%244 : tensor<2x9216x320xf32>) outs(%251 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %253 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%248 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_737 = tensor.collapse_shape %252 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %254 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %255 = linalg.matmul ins(%collapsed_737, %253 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%254 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %256 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %257 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%256 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %258 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%257 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %259 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_27 : tensor<320x320xf32>) outs(%258 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %260 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %261 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%260 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %262 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%261 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %263 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%244 : tensor<2x9216x320xf32>) outs(%262 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %264 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%259 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_738 = tensor.collapse_shape %263 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %265 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %266 = linalg.matmul ins(%collapsed_738, %264 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%265 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %267 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %268 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%267 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %269 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%268 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %270 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_28 : tensor<320x320xf32>) outs(%269 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %271 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %272 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%271 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %273 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%272 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %274 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%244 : tensor<2x9216x320xf32>) outs(%273 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %275 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%270 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_739 = tensor.collapse_shape %274 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %276 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %277 = linalg.matmul ins(%collapsed_739, %275 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%276 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %expanded_740 = tensor.expand_shape %255 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %278 = tensor.empty() : tensor<2x5x9216x64xf16>
    %279 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_740 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_741 = tensor.collapse_shape %279 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_742 = tensor.expand_shape %266 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %280 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_742 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_743 = tensor.collapse_shape %280 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_744 = tensor.expand_shape %277 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %281 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_744 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_745 = tensor.collapse_shape %281 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_746 = tensor.expand_shape %collapsed_741 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %282 = tensor.empty() : tensor<10x9216x1x64xf16>
    %283 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_746 : tensor<10x9216x64x1xf16>) outs(%282 : tensor<10x9216x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x64xf16>
    %expanded_747 = tensor.expand_shape %collapsed_743 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %284 = tensor.empty() : tensor<10x1x9216x64xf16>
    %285 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_747 : tensor<10x9216x64x1xf16>) outs(%284 : tensor<10x1x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x9216x64xf16>
    %286 = tensor.empty() : tensor<10x9216x64x1xf16>
    %287 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%283 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_748 = tensor.collapse_shape %287 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %288 = tensor.empty() : tensor<10x64x9216x1xf16>
    %289 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%285 : tensor<10x1x9216x64xf16>) outs(%288 : tensor<10x64x9216x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x64x9216x1xf16>
    %collapsed_749 = tensor.collapse_shape %289 [[0], [1], [2, 3]] : tensor<10x64x9216x1xf16> into tensor<10x64x9216xf16>
    %290 = tensor.empty() : tensor<10x9216x9216xf16>
    %291 = linalg.fill ins(%cst_690 : f16) outs(%290 : tensor<10x9216x9216xf16>) -> tensor<10x9216x9216xf16>
    %292 = linalg.batch_matmul ins(%collapsed_748, %collapsed_749 : tensor<10x9216x64xf16>, tensor<10x64x9216xf16>) outs(%291 : tensor<10x9216x9216xf16>) -> tensor<10x9216x9216xf16>
    %expanded_750 = tensor.expand_shape %292 [[0], [1], [2, 3]] : tensor<10x9216x9216xf16> into tensor<10x9216x1x9216xf16>
    %293 = tensor.empty() : tensor<10x9216x9216x1xf16>
    %294 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_750 : tensor<10x9216x1x9216xf16>) outs(%293 : tensor<10x9216x9216x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x9216x1xf16>
    %collapsed_751 = tensor.collapse_shape %294 [[0], [1], [2, 3]] : tensor<10x9216x9216x1xf16> into tensor<10x9216x9216xf16>
    %295 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_751, %cst : tensor<10x9216x9216xf16>, tensor<f64>) outs(%290 : tensor<10x9216x9216xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<10x9216x9216xf16>
    %296 = tensor.empty() : tensor<10x9216x9216xf32>
    %297 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%295 : tensor<10x9216x9216xf16>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %298 = tensor.empty() : tensor<10x9216x1xf32>
    %299 = linalg.fill ins(%cst_692 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %300 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%297 : tensor<10x9216x9216xf32>) outs(%299 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %301 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%297, %300 : tensor<10x9216x9216xf32>, tensor<10x9216x1xf32>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %302 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%301 : tensor<10x9216x9216xf32>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %303 = linalg.fill ins(%cst_697 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %304 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%302 : tensor<10x9216x9216xf32>) outs(%303 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %305 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%302, %304 : tensor<10x9216x9216xf32>, tensor<10x9216x1xf32>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %306 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %307 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%306 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %308 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%307 : tensor<f16>) outs(%290 : tensor<10x9216x9216xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x9216xf16>
    %309 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%305 : tensor<10x9216x9216xf32>) outs(%308 : tensor<10x9216x9216xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10x9216x9216xf16>
    %expanded_752 = tensor.expand_shape %309 [[0], [1], [2, 3]] : tensor<10x9216x9216xf16> into tensor<10x9216x9216x1xf16>
    %310 = tensor.empty() : tensor<10x9216x1x9216xf16>
    %311 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_752 : tensor<10x9216x9216x1xf16>) outs(%310 : tensor<10x9216x1x9216xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x9216xf16>
    %expanded_753 = tensor.expand_shape %collapsed_745 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %312 = tensor.empty() : tensor<10x1x64x9216xf16>
    %313 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_753 : tensor<10x9216x64x1xf16>) outs(%312 : tensor<10x1x64x9216xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x64x9216xf16>
    %314 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%311 : tensor<10x9216x1x9216xf16>) outs(%293 : tensor<10x9216x9216x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x9216x1xf16>
    %collapsed_754 = tensor.collapse_shape %314 [[0], [1], [2, 3]] : tensor<10x9216x9216x1xf16> into tensor<10x9216x9216xf16>
    %315 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%313 : tensor<10x1x64x9216xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_755 = tensor.collapse_shape %315 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %316 = tensor.empty() : tensor<10x9216x64xf16>
    %317 = linalg.fill ins(%cst_690 : f16) outs(%316 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %318 = linalg.batch_matmul ins(%collapsed_754, %collapsed_755 : tensor<10x9216x9216xf16>, tensor<10x9216x64xf16>) outs(%317 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %expanded_756 = tensor.expand_shape %318 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x1x64xf16>
    %319 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_756 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_757 = tensor.collapse_shape %319 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %expanded_758 = tensor.expand_shape %collapsed_757 [[0, 1], [2], [3]] : tensor<10x9216x64xf16> into tensor<2x5x9216x64xf16>
    %320 = tensor.empty() : tensor<2x9216x5x64xf16>
    %321 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_758 : tensor<2x5x9216x64xf16>) outs(%320 : tensor<2x9216x5x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x5x64xf16>
    %322 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %323 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%322 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %324 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%323 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %325 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_29 : tensor<320xf32>) outs(%324 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %326 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %327 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%326 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %328 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%327 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %329 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_30 : tensor<320x320xf32>) outs(%328 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %330 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%329 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_759 = tensor.collapse_shape %321 [[0, 1], [2, 3]] : tensor<2x9216x5x64xf16> into tensor<18432x320xf16>
    %331 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %332 = linalg.matmul ins(%collapsed_759, %330 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%331 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %333 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%325, %332 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_760 = tensor.expand_shape %333 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %334 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_760, %expanded_736 : tensor<2x9216x320xf16>, tensor<2x9216x320xf16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %335 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %336 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%335 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %337 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%336 : tensor<f32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %338 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%334 : tensor<2x9216x320xf16>) outs(%337 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %339 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %340 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%338 : tensor<2x9216x320xf32>) outs(%339 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %341 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%340 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %342 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%341 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %343 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%342 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %344 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%338, %343 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %345 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%344 : tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %346 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %347 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%345 : tensor<2x9216x320xf32>) outs(%346 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %348 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%347 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %349 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%348 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %350 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%349 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x9216x1xf32>
    %351 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%350 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %352 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%351 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %353 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%344, %352 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %354 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%353, %cst_31 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %355 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%354, %cst_32 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %356 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %357 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%356 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %358 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%357 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %359 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_33 : tensor<320x320xf32>) outs(%358 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %360 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %361 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%360 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %362 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%361 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %363 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%355 : tensor<2x9216x320xf32>) outs(%362 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %364 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%359 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_761 = tensor.collapse_shape %363 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %365 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %366 = linalg.matmul ins(%collapsed_761, %364 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%365 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %367 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %368 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%367 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %369 = tensor.empty() : tensor<320x1024xf16>
    %370 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%368 : tensor<f16>) outs(%369 : tensor<320x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1024xf16>
    %371 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_34 : tensor<320x1024xf32>) outs(%370 : tensor<320x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1024xf16>
    %372 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %373 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%372 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %374 = tensor.empty() : tensor<2x77x1024xf16>
    %375 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%373 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %376 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%375 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %377 = tensor.empty() : tensor<1024x320xf16>
    %378 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%371 : tensor<320x1024xf16>) outs(%377 : tensor<1024x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x320xf16>
    %collapsed_762 = tensor.collapse_shape %376 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %379 = tensor.empty() : tensor<154x320xf16>
    %380 = linalg.fill ins(%cst_690 : f16) outs(%379 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %381 = linalg.matmul ins(%collapsed_762, %378 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%380 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %382 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %383 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%382 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %384 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%383 : tensor<f16>) outs(%369 : tensor<320x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1024xf16>
    %385 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_35 : tensor<320x1024xf32>) outs(%384 : tensor<320x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1024xf16>
    %386 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %387 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%386 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %388 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%387 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %389 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%388 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %390 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%385 : tensor<320x1024xf16>) outs(%377 : tensor<1024x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x320xf16>
    %collapsed_763 = tensor.collapse_shape %389 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %391 = linalg.fill ins(%cst_690 : f16) outs(%379 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %392 = linalg.matmul ins(%collapsed_763, %390 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%391 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %expanded_764 = tensor.expand_shape %366 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %393 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_764 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_765 = tensor.collapse_shape %393 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_766 = tensor.expand_shape %381 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
    %394 = tensor.empty() : tensor<2x5x77x64xf16>
    %395 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_766 : tensor<2x77x5x64xf16>) outs(%394 : tensor<2x5x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x77x64xf16>
    %collapsed_767 = tensor.collapse_shape %395 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
    %expanded_768 = tensor.expand_shape %392 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
    %396 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_768 : tensor<2x77x5x64xf16>) outs(%394 : tensor<2x5x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x77x64xf16>
    %collapsed_769 = tensor.collapse_shape %396 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
    %expanded_770 = tensor.expand_shape %collapsed_765 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %397 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_770 : tensor<10x9216x64x1xf16>) outs(%282 : tensor<10x9216x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x64xf16>
    %expanded_771 = tensor.expand_shape %collapsed_767 [[0], [1], [2, 3]] : tensor<10x77x64xf16> into tensor<10x77x64x1xf16>
    %398 = tensor.empty() : tensor<10x1x77x64xf16>
    %399 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_771 : tensor<10x77x64x1xf16>) outs(%398 : tensor<10x1x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x77x64xf16>
    %400 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%397 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_772 = tensor.collapse_shape %400 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %401 = tensor.empty() : tensor<10x64x77x1xf16>
    %402 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%399 : tensor<10x1x77x64xf16>) outs(%401 : tensor<10x64x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x64x77x1xf16>
    %collapsed_773 = tensor.collapse_shape %402 [[0], [1], [2, 3]] : tensor<10x64x77x1xf16> into tensor<10x64x77xf16>
    %403 = tensor.empty() : tensor<10x9216x77xf16>
    %404 = linalg.fill ins(%cst_690 : f16) outs(%403 : tensor<10x9216x77xf16>) -> tensor<10x9216x77xf16>
    %405 = linalg.batch_matmul ins(%collapsed_772, %collapsed_773 : tensor<10x9216x64xf16>, tensor<10x64x77xf16>) outs(%404 : tensor<10x9216x77xf16>) -> tensor<10x9216x77xf16>
    %expanded_774 = tensor.expand_shape %405 [[0], [1], [2, 3]] : tensor<10x9216x77xf16> into tensor<10x9216x1x77xf16>
    %406 = tensor.empty() : tensor<10x9216x77x1xf16>
    %407 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_774 : tensor<10x9216x1x77xf16>) outs(%406 : tensor<10x9216x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x77x1xf16>
    %collapsed_775 = tensor.collapse_shape %407 [[0], [1], [2, 3]] : tensor<10x9216x77x1xf16> into tensor<10x9216x77xf16>
    %408 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_775, %cst : tensor<10x9216x77xf16>, tensor<f64>) outs(%403 : tensor<10x9216x77xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<10x9216x77xf16>
    %409 = tensor.empty() : tensor<10x9216x77xf32>
    %410 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%408 : tensor<10x9216x77xf16>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %411 = linalg.fill ins(%cst_692 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %412 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%410 : tensor<10x9216x77xf32>) outs(%411 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %413 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%410, %412 : tensor<10x9216x77xf32>, tensor<10x9216x1xf32>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %414 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%413 : tensor<10x9216x77xf32>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %415 = linalg.fill ins(%cst_697 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %416 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%414 : tensor<10x9216x77xf32>) outs(%415 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %417 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%414, %416 : tensor<10x9216x77xf32>, tensor<10x9216x1xf32>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %418 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %419 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%418 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %420 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%419 : tensor<f16>) outs(%403 : tensor<10x9216x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x77xf16>
    %421 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%417 : tensor<10x9216x77xf32>) outs(%420 : tensor<10x9216x77xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10x9216x77xf16>
    %expanded_776 = tensor.expand_shape %421 [[0], [1], [2, 3]] : tensor<10x9216x77xf16> into tensor<10x9216x77x1xf16>
    %422 = tensor.empty() : tensor<10x9216x1x77xf16>
    %423 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_776 : tensor<10x9216x77x1xf16>) outs(%422 : tensor<10x9216x1x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x77xf16>
    %expanded_777 = tensor.expand_shape %collapsed_769 [[0], [1], [2, 3]] : tensor<10x77x64xf16> into tensor<10x77x64x1xf16>
    %424 = tensor.empty() : tensor<10x1x64x77xf16>
    %425 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_777 : tensor<10x77x64x1xf16>) outs(%424 : tensor<10x1x64x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x64x77xf16>
    %426 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%423 : tensor<10x9216x1x77xf16>) outs(%406 : tensor<10x9216x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x77x1xf16>
    %collapsed_778 = tensor.collapse_shape %426 [[0], [1], [2, 3]] : tensor<10x9216x77x1xf16> into tensor<10x9216x77xf16>
    %427 = tensor.empty() : tensor<10x77x64x1xf16>
    %428 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%425 : tensor<10x1x64x77xf16>) outs(%427 : tensor<10x77x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x77x64x1xf16>
    %collapsed_779 = tensor.collapse_shape %428 [[0], [1], [2, 3]] : tensor<10x77x64x1xf16> into tensor<10x77x64xf16>
    %429 = linalg.fill ins(%cst_690 : f16) outs(%316 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %430 = linalg.batch_matmul ins(%collapsed_778, %collapsed_779 : tensor<10x9216x77xf16>, tensor<10x77x64xf16>) outs(%429 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %expanded_780 = tensor.expand_shape %430 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x1x64xf16>
    %431 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_780 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_781 = tensor.collapse_shape %431 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %expanded_782 = tensor.expand_shape %collapsed_781 [[0, 1], [2], [3]] : tensor<10x9216x64xf16> into tensor<2x5x9216x64xf16>
    %432 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_782 : tensor<2x5x9216x64xf16>) outs(%320 : tensor<2x9216x5x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x5x64xf16>
    %433 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %434 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%433 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %435 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%434 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %436 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_36 : tensor<320xf32>) outs(%435 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %437 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %438 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%437 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %439 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%438 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %440 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_37 : tensor<320x320xf32>) outs(%439 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %441 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%440 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_783 = tensor.collapse_shape %432 [[0, 1], [2, 3]] : tensor<2x9216x5x64xf16> into tensor<18432x320xf16>
    %442 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %443 = linalg.matmul ins(%collapsed_783, %441 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%442 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %444 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%436, %443 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_784 = tensor.expand_shape %444 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %445 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_784, %334 : tensor<2x9216x320xf16>, tensor<2x9216x320xf16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %446 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %447 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%446 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %448 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%447 : tensor<f32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %449 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%445 : tensor<2x9216x320xf16>) outs(%448 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %450 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %451 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%449 : tensor<2x9216x320xf32>) outs(%450 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %452 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%451 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %453 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%452 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %454 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%453 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %455 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%449, %454 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %456 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%455 : tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %457 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %458 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%456 : tensor<2x9216x320xf32>) outs(%457 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %459 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%458 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %460 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%459 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %461 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%460 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x9216x1xf32>
    %462 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%461 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %463 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%462 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %464 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%455, %463 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %465 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%464, %cst_38 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %466 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%465, %cst_39 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %467 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %468 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%467 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %469 = tensor.empty() : tensor<2560xf16>
    %470 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%468 : tensor<f16>) outs(%469 : tensor<2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2560xf16>
    %471 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_40 : tensor<2560xf32>) outs(%470 : tensor<2560xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2560xf16>
    %472 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %473 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%472 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %474 = tensor.empty() : tensor<2560x320xf16>
    %475 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%473 : tensor<f16>) outs(%474 : tensor<2560x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2560x320xf16>
    %476 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_41 : tensor<2560x320xf32>) outs(%475 : tensor<2560x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2560x320xf16>
    %477 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %478 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%477 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %479 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%478 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %480 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%466 : tensor<2x9216x320xf32>) outs(%479 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %481 = tensor.empty() : tensor<320x2560xf16>
    %482 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%476 : tensor<2560x320xf16>) outs(%481 : tensor<320x2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x2560xf16>
    %collapsed_785 = tensor.collapse_shape %480 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %483 = tensor.empty() : tensor<18432x2560xf16>
    %484 = linalg.fill ins(%cst_690 : f16) outs(%483 : tensor<18432x2560xf16>) -> tensor<18432x2560xf16>
    %485 = linalg.matmul ins(%collapsed_785, %482 : tensor<18432x320xf16>, tensor<320x2560xf16>) outs(%484 : tensor<18432x2560xf16>) -> tensor<18432x2560xf16>
    %486 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%471, %485 : tensor<2560xf16>, tensor<18432x2560xf16>) outs(%483 : tensor<18432x2560xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x2560xf16>
    %expanded_786 = tensor.expand_shape %486 [[0, 1], [2]] : tensor<18432x2560xf16> into tensor<2x9216x2560xf16>
    %extracted_slice = tensor.extract_slice %expanded_786[0, 0, 0] [2, 9216, 1280] [1, 1, 1] : tensor<2x9216x2560xf16> to tensor<2x9216x1280xf16>
    %extracted_slice_787 = tensor.extract_slice %expanded_786[0, 0, 1280] [2, 9216, 1280] [1, 1, 1] : tensor<2x9216x2560xf16> to tensor<2x9216x1280xf16>
    %487 = tensor.empty() : tensor<2x9216x1280xf16>
    %488 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_787 : tensor<2x9216x1280xf16>) outs(%487 : tensor<2x9216x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = math.sqrt %cst_693 : f16
      %7525 = arith.divf %in, %7524 : f16
      %7526 = math.erf %7525 : f16
      %7527 = arith.addf %7526, %cst_691 : f16
      %7528 = arith.mulf %7527, %cst_694 : f16
      %7529 = arith.mulf %in, %7528 : f16
      linalg.yield %7529 : f16
    } -> tensor<2x9216x1280xf16>
    %489 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice, %488 : tensor<2x9216x1280xf16>, tensor<2x9216x1280xf16>) outs(%487 : tensor<2x9216x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x1280xf16>
    %490 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %491 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%490 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %492 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%491 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %493 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_42 : tensor<320xf32>) outs(%492 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %494 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %495 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%494 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %496 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%495 : tensor<f16>) outs(%35 : tensor<320x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1280xf16>
    %497 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_43 : tensor<320x1280xf32>) outs(%496 : tensor<320x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1280xf16>
    %498 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%497 : tensor<320x1280xf16>) outs(%27 : tensor<1280x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x320xf16>
    %collapsed_788 = tensor.collapse_shape %489 [[0, 1], [2]] : tensor<2x9216x1280xf16> into tensor<18432x1280xf16>
    %499 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %500 = linalg.matmul ins(%collapsed_788, %498 : tensor<18432x1280xf16>, tensor<1280x320xf16>) outs(%499 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %501 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%493, %500 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_789 = tensor.expand_shape %501 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %502 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_789, %445 : tensor<2x9216x320xf16>, tensor<2x9216x320xf16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %503 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %504 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%503 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %505 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%504 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %506 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_44 : tensor<320xf32>) outs(%505 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %507 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %508 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%507 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %509 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%508 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %510 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_45 : tensor<320x320xf32>) outs(%509 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %511 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%510 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_790 = tensor.collapse_shape %502 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %512 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %513 = linalg.matmul ins(%collapsed_790, %511 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%512 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %514 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%506, %513 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_791 = tensor.expand_shape %514 [[0, 1, 2], [3]] : tensor<18432x320xf16> into tensor<2x96x96x320xf16>
    %515 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_791 : tensor<2x96x96x320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %516 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%515, %175 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %517 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %518 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%517 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %519 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%518 : tensor<f32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x96x96xf32>
    %520 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%516 : tensor<2x320x96x96xf16>) outs(%519 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %collapsed_792 = tensor.collapse_shape %520 [[0], [1], [2, 3]] : tensor<2x320x96x96xf32> into tensor<2x320x9216xf32>
    %expanded_793 = tensor.expand_shape %collapsed_792 [[0], [1, 2], [3]] : tensor<2x320x9216xf32> into tensor<2x32x10x9216xf32>
    %521 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_793 : tensor<2x32x10x9216xf32>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %522 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %523 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%521 : tensor<2x32x10x9216xf64>) outs(%522 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %524 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%523 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %525 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%521, %524 : tensor<2x32x10x9216xf64>, tensor<2x32x1x1xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %526 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%525 : tensor<2x32x10x9216xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %527 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %528 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%526 : tensor<2x32x10x9216xf64>) outs(%527 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %529 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%528 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %530 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%529 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %531 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %532 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_793 : tensor<2x32x10x9216xf32>) outs(%531 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %533 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%532 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %534 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%533 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %535 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%530, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %536 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%535 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %537 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_793, %534 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %538 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%537, %536 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %collapsed_794 = tensor.collapse_shape %538 [[0], [1, 2], [3]] : tensor<2x32x10x9216xf32> into tensor<2x320x9216xf32>
    %expanded_795 = tensor.expand_shape %collapsed_794 [[0], [1], [2, 3]] : tensor<2x320x9216xf32> into tensor<2x320x96x96xf32>
    %expanded_796 = tensor.expand_shape %cst_46 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_797 = tensor.expand_shape %cst_47 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %539 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_795, %expanded_797 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %540 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%539, %expanded_796 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %541 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %542 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%541 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %543 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%542 : tensor<f16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %544 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%540 : tensor<2x320x96x96xf32>) outs(%543 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %545 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%544 : tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x320x96x96xf16>
    %546 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%545, %544 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %547 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %548 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%547 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %549 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%548 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %550 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_48 : tensor<320xf32>) outs(%549 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %551 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %552 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%551 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %553 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%552 : tensor<f16>) outs(%115 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320x3x3xf16>
    %554 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_49 : tensor<320x320x3x3xf32>) outs(%553 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320x3x3xf16>
    %padded_798 = tensor.pad %546 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x320x96x96xf16> to tensor<2x320x98x98xf16>
    %555 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%550 : tensor<320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %556 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_798, %554 : tensor<2x320x98x98xf16>, tensor<320x320x3x3xf16>) outs(%555 : tensor<2x320x96x96xf16>) -> tensor<2x320x96x96xf16>
    %557 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %558 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%557, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %559 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %560 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%559 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %561 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%560 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %562 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_50 : tensor<320xf32>) outs(%561 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %563 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %564 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%563 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %565 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%564 : tensor<f16>) outs(%35 : tensor<320x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1280xf16>
    %566 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_51 : tensor<320x1280xf32>) outs(%565 : tensor<320x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1280xf16>
    %567 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%566 : tensor<320x1280xf16>) outs(%27 : tensor<1280x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x320xf16>
    %568 = linalg.fill ins(%cst_690 : f16) outs(%32 : tensor<2x320xf16>) -> tensor<2x320xf16>
    %569 = linalg.matmul ins(%558, %567 : tensor<2x1280xf16>, tensor<1280x320xf16>) outs(%568 : tensor<2x320xf16>) -> tensor<2x320xf16>
    %570 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%562, %569 : tensor<320xf16>, tensor<2x320xf16>) outs(%32 : tensor<2x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320xf16>
    %expanded_799 = tensor.expand_shape %570 [[0], [1, 2, 3]] : tensor<2x320xf16> into tensor<2x320x1x1xf16>
    %571 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%556, %expanded_799 : tensor<2x320x96x96xf16>, tensor<2x320x1x1xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %572 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %573 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%572 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %574 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%573 : tensor<f32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x96x96xf32>
    %575 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%571 : tensor<2x320x96x96xf16>) outs(%574 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %collapsed_800 = tensor.collapse_shape %575 [[0], [1], [2, 3]] : tensor<2x320x96x96xf32> into tensor<2x320x9216xf32>
    %expanded_801 = tensor.expand_shape %collapsed_800 [[0], [1, 2], [3]] : tensor<2x320x9216xf32> into tensor<2x32x10x9216xf32>
    %576 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_801 : tensor<2x32x10x9216xf32>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %577 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %578 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%576 : tensor<2x32x10x9216xf64>) outs(%577 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %579 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%578 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %580 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%576, %579 : tensor<2x32x10x9216xf64>, tensor<2x32x1x1xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %581 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%580 : tensor<2x32x10x9216xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %582 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %583 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%581 : tensor<2x32x10x9216xf64>) outs(%582 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %584 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%583 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %585 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%584 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %586 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %587 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_801 : tensor<2x32x10x9216xf32>) outs(%586 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %588 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%587 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %589 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%588 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %590 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%585, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %591 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%590 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %592 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_801, %589 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %593 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%592, %591 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %collapsed_802 = tensor.collapse_shape %593 [[0], [1, 2], [3]] : tensor<2x32x10x9216xf32> into tensor<2x320x9216xf32>
    %expanded_803 = tensor.expand_shape %collapsed_802 [[0], [1], [2, 3]] : tensor<2x320x9216xf32> into tensor<2x320x96x96xf32>
    %expanded_804 = tensor.expand_shape %cst_52 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_805 = tensor.expand_shape %cst_53 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %594 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_803, %expanded_805 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %595 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%594, %expanded_804 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %596 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %597 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%596 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %598 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%597 : tensor<f16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %599 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%595 : tensor<2x320x96x96xf32>) outs(%598 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %600 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%599 : tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x320x96x96xf16>
    %601 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%600, %599 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %602 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %603 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%602 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %604 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%603 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %605 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_54 : tensor<320xf32>) outs(%604 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %606 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %607 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%606 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %608 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%607 : tensor<f16>) outs(%115 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320x3x3xf16>
    %609 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_55 : tensor<320x320x3x3xf32>) outs(%608 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320x3x3xf16>
    %padded_806 = tensor.pad %601 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x320x96x96xf16> to tensor<2x320x98x98xf16>
    %610 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%605 : tensor<320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %611 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_806, %609 : tensor<2x320x98x98xf16>, tensor<320x320x3x3xf16>) outs(%610 : tensor<2x320x96x96xf16>) -> tensor<2x320x96x96xf16>
    %612 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%516, %611 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %613 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %614 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%613 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %615 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%614 : tensor<f32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x96x96xf32>
    %616 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%612 : tensor<2x320x96x96xf16>) outs(%615 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %collapsed_807 = tensor.collapse_shape %616 [[0], [1], [2, 3]] : tensor<2x320x96x96xf32> into tensor<2x320x9216xf32>
    %expanded_808 = tensor.expand_shape %collapsed_807 [[0], [1, 2], [3]] : tensor<2x320x9216xf32> into tensor<2x32x10x9216xf32>
    %617 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_808 : tensor<2x32x10x9216xf32>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %618 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %619 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%617 : tensor<2x32x10x9216xf64>) outs(%618 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %620 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%619 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %621 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%617, %620 : tensor<2x32x10x9216xf64>, tensor<2x32x1x1xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %622 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%621 : tensor<2x32x10x9216xf64>) outs(%79 : tensor<2x32x10x9216xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x9216xf64>
    %623 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %624 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%622 : tensor<2x32x10x9216xf64>) outs(%623 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %625 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%624 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %626 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%625 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %627 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %628 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_808 : tensor<2x32x10x9216xf32>) outs(%627 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %629 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%628 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_699 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %630 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%629 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %631 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%626, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %632 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%631 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %633 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_808, %630 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %634 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%633, %632 : tensor<2x32x10x9216xf32>, tensor<2x32x1x1xf32>) outs(%98 : tensor<2x32x10x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x9216xf32>
    %collapsed_809 = tensor.collapse_shape %634 [[0], [1, 2], [3]] : tensor<2x32x10x9216xf32> into tensor<2x320x9216xf32>
    %expanded_810 = tensor.expand_shape %collapsed_809 [[0], [1], [2, 3]] : tensor<2x320x9216xf32> into tensor<2x320x96x96xf32>
    %expanded_811 = tensor.expand_shape %cst_56 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_812 = tensor.expand_shape %cst_57 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %635 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_810, %expanded_812 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %636 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%635, %expanded_811 : tensor<2x320x96x96xf32>, tensor<1x320x1x1xf32>) outs(%76 : tensor<2x320x96x96xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x96x96xf32>
    %637 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%636 : tensor<2x320x96x96xf32>) outs(%200 : tensor<2x96x96x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x96x96x320xf32>
    %collapsed_813 = tensor.collapse_shape %637 [[0], [1, 2], [3]] : tensor<2x96x96x320xf32> into tensor<2x9216x320xf32>
    %638 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %639 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%638 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %640 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%639 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %641 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_58 : tensor<320xf32>) outs(%640 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %642 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %643 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%642 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %644 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%643 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %645 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_59 : tensor<320x320xf32>) outs(%644 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %646 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %647 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%646 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %648 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%647 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %649 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_813 : tensor<2x9216x320xf32>) outs(%648 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %650 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%645 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_814 = tensor.collapse_shape %649 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %651 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %652 = linalg.matmul ins(%collapsed_814, %650 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%651 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %653 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%641, %652 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_815 = tensor.expand_shape %653 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %654 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %655 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%654 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %656 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%655 : tensor<f32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %657 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_815 : tensor<2x9216x320xf16>) outs(%656 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %658 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %659 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%657 : tensor<2x9216x320xf32>) outs(%658 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %660 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%659 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %661 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%660 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %662 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%661 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %663 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%657, %662 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %664 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%663 : tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %665 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %666 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%664 : tensor<2x9216x320xf32>) outs(%665 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %667 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%666 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %668 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%667 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %669 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%668 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x9216x1xf32>
    %670 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%669 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %671 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%670 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %672 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%663, %671 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %673 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%672, %cst_60 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %674 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%673, %cst_61 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %675 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %676 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%675 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %677 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%676 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %678 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_62 : tensor<320x320xf32>) outs(%677 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %679 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %680 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%679 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %681 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%680 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %682 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%674 : tensor<2x9216x320xf32>) outs(%681 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %683 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%678 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_816 = tensor.collapse_shape %682 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %684 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %685 = linalg.matmul ins(%collapsed_816, %683 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%684 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %686 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %687 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%686 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %688 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%687 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %689 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_63 : tensor<320x320xf32>) outs(%688 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %690 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %691 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%690 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %692 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%691 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %693 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%674 : tensor<2x9216x320xf32>) outs(%692 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %694 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%689 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_817 = tensor.collapse_shape %693 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %695 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %696 = linalg.matmul ins(%collapsed_817, %694 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%695 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %697 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %698 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%697 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %699 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%698 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %700 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_64 : tensor<320x320xf32>) outs(%699 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %701 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %702 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%701 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %703 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%702 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %704 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%674 : tensor<2x9216x320xf32>) outs(%703 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %705 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%700 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_818 = tensor.collapse_shape %704 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %706 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %707 = linalg.matmul ins(%collapsed_818, %705 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%706 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %expanded_819 = tensor.expand_shape %685 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %708 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_819 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_820 = tensor.collapse_shape %708 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_821 = tensor.expand_shape %696 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %709 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_821 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_822 = tensor.collapse_shape %709 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_823 = tensor.expand_shape %707 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %710 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_823 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_824 = tensor.collapse_shape %710 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_825 = tensor.expand_shape %collapsed_820 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %711 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_825 : tensor<10x9216x64x1xf16>) outs(%282 : tensor<10x9216x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x64xf16>
    %expanded_826 = tensor.expand_shape %collapsed_822 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %712 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_826 : tensor<10x9216x64x1xf16>) outs(%284 : tensor<10x1x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x9216x64xf16>
    %713 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%711 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_827 = tensor.collapse_shape %713 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %714 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%712 : tensor<10x1x9216x64xf16>) outs(%288 : tensor<10x64x9216x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x64x9216x1xf16>
    %collapsed_828 = tensor.collapse_shape %714 [[0], [1], [2, 3]] : tensor<10x64x9216x1xf16> into tensor<10x64x9216xf16>
    %715 = linalg.fill ins(%cst_690 : f16) outs(%290 : tensor<10x9216x9216xf16>) -> tensor<10x9216x9216xf16>
    %716 = linalg.batch_matmul ins(%collapsed_827, %collapsed_828 : tensor<10x9216x64xf16>, tensor<10x64x9216xf16>) outs(%715 : tensor<10x9216x9216xf16>) -> tensor<10x9216x9216xf16>
    %expanded_829 = tensor.expand_shape %716 [[0], [1], [2, 3]] : tensor<10x9216x9216xf16> into tensor<10x9216x1x9216xf16>
    %717 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_829 : tensor<10x9216x1x9216xf16>) outs(%293 : tensor<10x9216x9216x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x9216x1xf16>
    %collapsed_830 = tensor.collapse_shape %717 [[0], [1], [2, 3]] : tensor<10x9216x9216x1xf16> into tensor<10x9216x9216xf16>
    %718 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_830, %cst : tensor<10x9216x9216xf16>, tensor<f64>) outs(%290 : tensor<10x9216x9216xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<10x9216x9216xf16>
    %719 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%718 : tensor<10x9216x9216xf16>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %720 = linalg.fill ins(%cst_692 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %721 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%719 : tensor<10x9216x9216xf32>) outs(%720 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %722 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%719, %721 : tensor<10x9216x9216xf32>, tensor<10x9216x1xf32>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %723 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%722 : tensor<10x9216x9216xf32>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %724 = linalg.fill ins(%cst_697 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %725 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%723 : tensor<10x9216x9216xf32>) outs(%724 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %726 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%723, %725 : tensor<10x9216x9216xf32>, tensor<10x9216x1xf32>) outs(%296 : tensor<10x9216x9216xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x9216xf32>
    %727 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %728 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%727 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %729 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%728 : tensor<f16>) outs(%290 : tensor<10x9216x9216xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x9216xf16>
    %730 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%726 : tensor<10x9216x9216xf32>) outs(%729 : tensor<10x9216x9216xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10x9216x9216xf16>
    %expanded_831 = tensor.expand_shape %730 [[0], [1], [2, 3]] : tensor<10x9216x9216xf16> into tensor<10x9216x9216x1xf16>
    %731 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_831 : tensor<10x9216x9216x1xf16>) outs(%310 : tensor<10x9216x1x9216xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x9216xf16>
    %expanded_832 = tensor.expand_shape %collapsed_824 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %732 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_832 : tensor<10x9216x64x1xf16>) outs(%312 : tensor<10x1x64x9216xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x64x9216xf16>
    %733 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%731 : tensor<10x9216x1x9216xf16>) outs(%293 : tensor<10x9216x9216x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x9216x1xf16>
    %collapsed_833 = tensor.collapse_shape %733 [[0], [1], [2, 3]] : tensor<10x9216x9216x1xf16> into tensor<10x9216x9216xf16>
    %734 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%732 : tensor<10x1x64x9216xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_834 = tensor.collapse_shape %734 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %735 = linalg.fill ins(%cst_690 : f16) outs(%316 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %736 = linalg.batch_matmul ins(%collapsed_833, %collapsed_834 : tensor<10x9216x9216xf16>, tensor<10x9216x64xf16>) outs(%735 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %expanded_835 = tensor.expand_shape %736 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x1x64xf16>
    %737 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_835 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_836 = tensor.collapse_shape %737 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %expanded_837 = tensor.expand_shape %collapsed_836 [[0, 1], [2], [3]] : tensor<10x9216x64xf16> into tensor<2x5x9216x64xf16>
    %738 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_837 : tensor<2x5x9216x64xf16>) outs(%320 : tensor<2x9216x5x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x5x64xf16>
    %739 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %740 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%739 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %741 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%740 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %742 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_65 : tensor<320xf32>) outs(%741 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %743 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %744 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%743 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %745 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%744 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %746 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_66 : tensor<320x320xf32>) outs(%745 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %747 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%746 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_838 = tensor.collapse_shape %738 [[0, 1], [2, 3]] : tensor<2x9216x5x64xf16> into tensor<18432x320xf16>
    %748 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %749 = linalg.matmul ins(%collapsed_838, %747 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%748 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %750 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%742, %749 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_839 = tensor.expand_shape %750 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %751 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_839, %expanded_815 : tensor<2x9216x320xf16>, tensor<2x9216x320xf16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %752 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %753 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%752 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %754 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%753 : tensor<f32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %755 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%751 : tensor<2x9216x320xf16>) outs(%754 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %756 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %757 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%755 : tensor<2x9216x320xf32>) outs(%756 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %758 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%757 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %759 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%758 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %760 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%759 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %761 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%755, %760 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %762 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%761 : tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %763 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %764 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%762 : tensor<2x9216x320xf32>) outs(%763 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %765 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%764 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %766 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%765 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %767 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%766 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x9216x1xf32>
    %768 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%767 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %769 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%768 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %770 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%761, %769 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %771 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%770, %cst_67 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %772 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%771, %cst_68 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %773 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %774 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%773 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %775 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%774 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %776 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_69 : tensor<320x320xf32>) outs(%775 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %777 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %778 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%777 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %779 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%778 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %780 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%772 : tensor<2x9216x320xf32>) outs(%779 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %781 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%776 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_840 = tensor.collapse_shape %780 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %782 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %783 = linalg.matmul ins(%collapsed_840, %781 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%782 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %784 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %785 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%784 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %786 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%785 : tensor<f16>) outs(%369 : tensor<320x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1024xf16>
    %787 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_70 : tensor<320x1024xf32>) outs(%786 : tensor<320x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1024xf16>
    %788 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %789 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%788 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %790 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%789 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %791 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%790 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %792 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%787 : tensor<320x1024xf16>) outs(%377 : tensor<1024x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x320xf16>
    %collapsed_841 = tensor.collapse_shape %791 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %793 = linalg.fill ins(%cst_690 : f16) outs(%379 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %794 = linalg.matmul ins(%collapsed_841, %792 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%793 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %795 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %796 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%795 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %797 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%796 : tensor<f16>) outs(%369 : tensor<320x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1024xf16>
    %798 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_71 : tensor<320x1024xf32>) outs(%797 : tensor<320x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1024xf16>
    %799 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %800 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%799 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %801 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%800 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %802 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%801 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %803 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%798 : tensor<320x1024xf16>) outs(%377 : tensor<1024x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x320xf16>
    %collapsed_842 = tensor.collapse_shape %802 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %804 = linalg.fill ins(%cst_690 : f16) outs(%379 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %805 = linalg.matmul ins(%collapsed_842, %803 : tensor<154x1024xf16>, tensor<1024x320xf16>) outs(%804 : tensor<154x320xf16>) -> tensor<154x320xf16>
    %expanded_843 = tensor.expand_shape %783 [[0, 1], [2, 3]] : tensor<18432x320xf16> into tensor<2x9216x5x64xf16>
    %806 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_843 : tensor<2x9216x5x64xf16>) outs(%278 : tensor<2x5x9216x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x9216x64xf16>
    %collapsed_844 = tensor.collapse_shape %806 [[0, 1], [2], [3]] : tensor<2x5x9216x64xf16> into tensor<10x9216x64xf16>
    %expanded_845 = tensor.expand_shape %794 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
    %807 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_845 : tensor<2x77x5x64xf16>) outs(%394 : tensor<2x5x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x77x64xf16>
    %collapsed_846 = tensor.collapse_shape %807 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
    %expanded_847 = tensor.expand_shape %805 [[0, 1], [2, 3]] : tensor<154x320xf16> into tensor<2x77x5x64xf16>
    %808 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_847 : tensor<2x77x5x64xf16>) outs(%394 : tensor<2x5x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x5x77x64xf16>
    %collapsed_848 = tensor.collapse_shape %808 [[0, 1], [2], [3]] : tensor<2x5x77x64xf16> into tensor<10x77x64xf16>
    %expanded_849 = tensor.expand_shape %collapsed_844 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x64x1xf16>
    %809 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_849 : tensor<10x9216x64x1xf16>) outs(%282 : tensor<10x9216x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x64xf16>
    %expanded_850 = tensor.expand_shape %collapsed_846 [[0], [1], [2, 3]] : tensor<10x77x64xf16> into tensor<10x77x64x1xf16>
    %810 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_850 : tensor<10x77x64x1xf16>) outs(%398 : tensor<10x1x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x77x64xf16>
    %811 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%809 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_851 = tensor.collapse_shape %811 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %812 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%810 : tensor<10x1x77x64xf16>) outs(%401 : tensor<10x64x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x64x77x1xf16>
    %collapsed_852 = tensor.collapse_shape %812 [[0], [1], [2, 3]] : tensor<10x64x77x1xf16> into tensor<10x64x77xf16>
    %813 = linalg.fill ins(%cst_690 : f16) outs(%403 : tensor<10x9216x77xf16>) -> tensor<10x9216x77xf16>
    %814 = linalg.batch_matmul ins(%collapsed_851, %collapsed_852 : tensor<10x9216x64xf16>, tensor<10x64x77xf16>) outs(%813 : tensor<10x9216x77xf16>) -> tensor<10x9216x77xf16>
    %expanded_853 = tensor.expand_shape %814 [[0], [1], [2, 3]] : tensor<10x9216x77xf16> into tensor<10x9216x1x77xf16>
    %815 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_853 : tensor<10x9216x1x77xf16>) outs(%406 : tensor<10x9216x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x77x1xf16>
    %collapsed_854 = tensor.collapse_shape %815 [[0], [1], [2, 3]] : tensor<10x9216x77x1xf16> into tensor<10x9216x77xf16>
    %816 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_854, %cst : tensor<10x9216x77xf16>, tensor<f64>) outs(%403 : tensor<10x9216x77xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<10x9216x77xf16>
    %817 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%816 : tensor<10x9216x77xf16>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %818 = linalg.fill ins(%cst_692 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %819 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%817 : tensor<10x9216x77xf32>) outs(%818 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %820 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%817, %819 : tensor<10x9216x77xf32>, tensor<10x9216x1xf32>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %821 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%820 : tensor<10x9216x77xf32>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %822 = linalg.fill ins(%cst_697 : f32) outs(%298 : tensor<10x9216x1xf32>) -> tensor<10x9216x1xf32>
    %823 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%821 : tensor<10x9216x77xf32>) outs(%822 : tensor<10x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x1xf32>
    %824 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%821, %823 : tensor<10x9216x77xf32>, tensor<10x9216x1xf32>) outs(%409 : tensor<10x9216x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<10x9216x77xf32>
    %825 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %826 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%825 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %827 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%826 : tensor<f16>) outs(%403 : tensor<10x9216x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x77xf16>
    %828 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%824 : tensor<10x9216x77xf32>) outs(%827 : tensor<10x9216x77xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10x9216x77xf16>
    %expanded_855 = tensor.expand_shape %828 [[0], [1], [2, 3]] : tensor<10x9216x77xf16> into tensor<10x9216x77x1xf16>
    %829 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_855 : tensor<10x9216x77x1xf16>) outs(%422 : tensor<10x9216x1x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x1x77xf16>
    %expanded_856 = tensor.expand_shape %collapsed_848 [[0], [1], [2, 3]] : tensor<10x77x64xf16> into tensor<10x77x64x1xf16>
    %830 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_856 : tensor<10x77x64x1xf16>) outs(%424 : tensor<10x1x64x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x1x64x77xf16>
    %831 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%829 : tensor<10x9216x1x77xf16>) outs(%406 : tensor<10x9216x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x77x1xf16>
    %collapsed_857 = tensor.collapse_shape %831 [[0], [1], [2, 3]] : tensor<10x9216x77x1xf16> into tensor<10x9216x77xf16>
    %832 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%830 : tensor<10x1x64x77xf16>) outs(%427 : tensor<10x77x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x77x64x1xf16>
    %collapsed_858 = tensor.collapse_shape %832 [[0], [1], [2, 3]] : tensor<10x77x64x1xf16> into tensor<10x77x64xf16>
    %833 = linalg.fill ins(%cst_690 : f16) outs(%316 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %834 = linalg.batch_matmul ins(%collapsed_857, %collapsed_858 : tensor<10x9216x77xf16>, tensor<10x77x64xf16>) outs(%833 : tensor<10x9216x64xf16>) -> tensor<10x9216x64xf16>
    %expanded_859 = tensor.expand_shape %834 [[0], [1], [2, 3]] : tensor<10x9216x64xf16> into tensor<10x9216x1x64xf16>
    %835 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_859 : tensor<10x9216x1x64xf16>) outs(%286 : tensor<10x9216x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10x9216x64x1xf16>
    %collapsed_860 = tensor.collapse_shape %835 [[0], [1], [2, 3]] : tensor<10x9216x64x1xf16> into tensor<10x9216x64xf16>
    %expanded_861 = tensor.expand_shape %collapsed_860 [[0, 1], [2], [3]] : tensor<10x9216x64xf16> into tensor<2x5x9216x64xf16>
    %836 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_861 : tensor<2x5x9216x64xf16>) outs(%320 : tensor<2x9216x5x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x5x64xf16>
    %837 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %838 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%837 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %839 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%838 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %840 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_72 : tensor<320xf32>) outs(%839 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %841 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %842 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%841 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %843 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%842 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %844 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_73 : tensor<320x320xf32>) outs(%843 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %845 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%844 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_862 = tensor.collapse_shape %836 [[0, 1], [2, 3]] : tensor<2x9216x5x64xf16> into tensor<18432x320xf16>
    %846 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %847 = linalg.matmul ins(%collapsed_862, %845 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%846 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %848 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%840, %847 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_863 = tensor.expand_shape %848 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %849 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_863, %751 : tensor<2x9216x320xf16>, tensor<2x9216x320xf16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %850 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %851 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%850 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %852 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%851 : tensor<f32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %853 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%849 : tensor<2x9216x320xf16>) outs(%852 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %854 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %855 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%853 : tensor<2x9216x320xf32>) outs(%854 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %856 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%855 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %857 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%856 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %858 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%857 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %859 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%853, %858 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %860 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%859 : tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %861 = linalg.fill ins(%cst_695 : f64) outs(%226 : tensor<2x9216x1xf64>) -> tensor<2x9216x1xf64>
    %862 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%860 : tensor<2x9216x320xf32>) outs(%861 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x9216x1xf64>
    %863 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%862 : tensor<2x9216x1xf64>) outs(%226 : tensor<2x9216x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_700 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x9216x1xf64>
    %864 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%863 : tensor<2x9216x1xf64>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %865 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%864 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x9216x1xf32>
    %866 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%865 : tensor<2x9216x1xf32>) outs(%230 : tensor<2x9216x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x1xf32>
    %867 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%866 : tensor<2x9216x1xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x9216x320xf32>
    %868 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%859, %867 : tensor<2x9216x320xf32>, tensor<2x9216x320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %869 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%868, %cst_74 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %870 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%869, %cst_75 : tensor<2x9216x320xf32>, tensor<320xf32>) outs(%223 : tensor<2x9216x320xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x9216x320xf32>
    %871 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %872 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%871 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %873 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%872 : tensor<f16>) outs(%469 : tensor<2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2560xf16>
    %874 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_76 : tensor<2560xf32>) outs(%873 : tensor<2560xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2560xf16>
    %875 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %876 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%875 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %877 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%876 : tensor<f16>) outs(%474 : tensor<2560x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2560x320xf16>
    %878 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_77 : tensor<2560x320xf32>) outs(%877 : tensor<2560x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2560x320xf16>
    %879 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %880 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%879 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %881 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%880 : tensor<f16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x9216x320xf16>
    %882 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%870 : tensor<2x9216x320xf32>) outs(%881 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %883 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%878 : tensor<2560x320xf16>) outs(%481 : tensor<320x2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x2560xf16>
    %collapsed_864 = tensor.collapse_shape %882 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %884 = linalg.fill ins(%cst_690 : f16) outs(%483 : tensor<18432x2560xf16>) -> tensor<18432x2560xf16>
    %885 = linalg.matmul ins(%collapsed_864, %883 : tensor<18432x320xf16>, tensor<320x2560xf16>) outs(%884 : tensor<18432x2560xf16>) -> tensor<18432x2560xf16>
    %886 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%874, %885 : tensor<2560xf16>, tensor<18432x2560xf16>) outs(%483 : tensor<18432x2560xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x2560xf16>
    %expanded_865 = tensor.expand_shape %886 [[0, 1], [2]] : tensor<18432x2560xf16> into tensor<2x9216x2560xf16>
    %extracted_slice_866 = tensor.extract_slice %expanded_865[0, 0, 0] [2, 9216, 1280] [1, 1, 1] : tensor<2x9216x2560xf16> to tensor<2x9216x1280xf16>
    %extracted_slice_867 = tensor.extract_slice %expanded_865[0, 0, 1280] [2, 9216, 1280] [1, 1, 1] : tensor<2x9216x2560xf16> to tensor<2x9216x1280xf16>
    %887 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_867 : tensor<2x9216x1280xf16>) outs(%487 : tensor<2x9216x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = math.sqrt %cst_693 : f16
      %7525 = arith.divf %in, %7524 : f16
      %7526 = math.erf %7525 : f16
      %7527 = arith.addf %7526, %cst_691 : f16
      %7528 = arith.mulf %7527, %cst_694 : f16
      %7529 = arith.mulf %in, %7528 : f16
      linalg.yield %7529 : f16
    } -> tensor<2x9216x1280xf16>
    %888 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_866, %887 : tensor<2x9216x1280xf16>, tensor<2x9216x1280xf16>) outs(%487 : tensor<2x9216x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x1280xf16>
    %889 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %890 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%889 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %891 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%890 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %892 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_78 : tensor<320xf32>) outs(%891 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %893 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %894 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%893 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %895 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%894 : tensor<f16>) outs(%35 : tensor<320x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x1280xf16>
    %896 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_79 : tensor<320x1280xf32>) outs(%895 : tensor<320x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x1280xf16>
    %897 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%896 : tensor<320x1280xf16>) outs(%27 : tensor<1280x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x320xf16>
    %collapsed_868 = tensor.collapse_shape %888 [[0, 1], [2]] : tensor<2x9216x1280xf16> into tensor<18432x1280xf16>
    %898 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %899 = linalg.matmul ins(%collapsed_868, %897 : tensor<18432x1280xf16>, tensor<1280x320xf16>) outs(%898 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %900 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%892, %899 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_869 = tensor.expand_shape %900 [[0, 1], [2]] : tensor<18432x320xf16> into tensor<2x9216x320xf16>
    %901 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_869, %849 : tensor<2x9216x320xf16>, tensor<2x9216x320xf16>) outs(%213 : tensor<2x9216x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x9216x320xf16>
    %902 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %903 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%902 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %904 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%903 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %905 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_80 : tensor<320xf32>) outs(%904 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %906 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %907 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%906 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %908 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%907 : tensor<f16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %909 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_81 : tensor<320x320xf32>) outs(%908 : tensor<320x320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320xf16>
    %910 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%909 : tensor<320x320xf16>) outs(%208 : tensor<320x320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320xf16>
    %collapsed_870 = tensor.collapse_shape %901 [[0, 1], [2]] : tensor<2x9216x320xf16> into tensor<18432x320xf16>
    %911 = linalg.fill ins(%cst_690 : f16) outs(%217 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %912 = linalg.matmul ins(%collapsed_870, %910 : tensor<18432x320xf16>, tensor<320x320xf16>) outs(%911 : tensor<18432x320xf16>) -> tensor<18432x320xf16>
    %913 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%905, %912 : tensor<320xf16>, tensor<18432x320xf16>) outs(%217 : tensor<18432x320xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<18432x320xf16>
    %expanded_871 = tensor.expand_shape %913 [[0, 1, 2], [3]] : tensor<18432x320xf16> into tensor<2x96x96x320xf16>
    %914 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_871 : tensor<2x96x96x320xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x96x96xf16>
    %915 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%914, %612 : tensor<2x320x96x96xf16>, tensor<2x320x96x96xf16>) outs(%71 : tensor<2x320x96x96xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x96x96xf16>
    %916 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %917 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%916 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %918 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%917 : tensor<f16>) outs(%63 : tensor<320xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320xf16>
    %919 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_82 : tensor<320xf32>) outs(%918 : tensor<320xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320xf16>
    %920 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %921 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%920 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %922 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%921 : tensor<f16>) outs(%115 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<320x320x3x3xf16>
    %923 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_83 : tensor<320x320x3x3xf32>) outs(%922 : tensor<320x320x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<320x320x3x3xf16>
    %padded_872 = tensor.pad %915 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x320x96x96xf16> to tensor<2x320x98x98xf16>
    %924 = tensor.empty() : tensor<2x320x48x48xf16>
    %925 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%919 : tensor<320xf16>) outs(%924 : tensor<2x320x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x48x48xf16>
    %926 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_872, %923 : tensor<2x320x98x98xf16>, tensor<320x320x3x3xf16>) outs(%925 : tensor<2x320x48x48xf16>) -> tensor<2x320x48x48xf16>
    %927 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %928 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%927 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %929 = tensor.empty() : tensor<2x320x48x48xf32>
    %930 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%928 : tensor<f32>) outs(%929 : tensor<2x320x48x48xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x320x48x48xf32>
    %931 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%926 : tensor<2x320x48x48xf16>) outs(%930 : tensor<2x320x48x48xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x48x48xf32>
    %collapsed_873 = tensor.collapse_shape %931 [[0], [1], [2, 3]] : tensor<2x320x48x48xf32> into tensor<2x320x2304xf32>
    %expanded_874 = tensor.expand_shape %collapsed_873 [[0], [1, 2], [3]] : tensor<2x320x2304xf32> into tensor<2x32x10x2304xf32>
    %932 = tensor.empty() : tensor<2x32x10x2304xf64>
    %933 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_874 : tensor<2x32x10x2304xf32>) outs(%932 : tensor<2x32x10x2304xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x2304xf64>
    %934 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %935 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%933 : tensor<2x32x10x2304xf64>) outs(%934 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %936 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%935 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %937 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%933, %936 : tensor<2x32x10x2304xf64>, tensor<2x32x1x1xf64>) outs(%932 : tensor<2x32x10x2304xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x2304xf64>
    %938 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%937 : tensor<2x32x10x2304xf64>) outs(%932 : tensor<2x32x10x2304xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x10x2304xf64>
    %939 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %940 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%938 : tensor<2x32x10x2304xf64>) outs(%939 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %941 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%940 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %942 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%941 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %943 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %944 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_874 : tensor<2x32x10x2304xf32>) outs(%943 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %945 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%944 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %946 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%945 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %947 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%942, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %948 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%947 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %949 = tensor.empty() : tensor<2x32x10x2304xf32>
    %950 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_874, %946 : tensor<2x32x10x2304xf32>, tensor<2x32x1x1xf32>) outs(%949 : tensor<2x32x10x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x2304xf32>
    %951 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%950, %948 : tensor<2x32x10x2304xf32>, tensor<2x32x1x1xf32>) outs(%949 : tensor<2x32x10x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x10x2304xf32>
    %collapsed_875 = tensor.collapse_shape %951 [[0], [1, 2], [3]] : tensor<2x32x10x2304xf32> into tensor<2x320x2304xf32>
    %expanded_876 = tensor.expand_shape %collapsed_875 [[0], [1], [2, 3]] : tensor<2x320x2304xf32> into tensor<2x320x48x48xf32>
    %expanded_877 = tensor.expand_shape %cst_84 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %expanded_878 = tensor.expand_shape %cst_85 [[0, 1, 2, 3]] : tensor<320xf32> into tensor<1x320x1x1xf32>
    %952 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_876, %expanded_878 : tensor<2x320x48x48xf32>, tensor<1x320x1x1xf32>) outs(%929 : tensor<2x320x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x48x48xf32>
    %953 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%952, %expanded_877 : tensor<2x320x48x48xf32>, tensor<1x320x1x1xf32>) outs(%929 : tensor<2x320x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x320x48x48xf32>
    %954 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %955 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%954 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %956 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%955 : tensor<f16>) outs(%924 : tensor<2x320x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x320x48x48xf16>
    %957 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%953 : tensor<2x320x48x48xf32>) outs(%956 : tensor<2x320x48x48xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x48x48xf16>
    %958 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%957 : tensor<2x320x48x48xf16>) outs(%924 : tensor<2x320x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x320x48x48xf16>
    %959 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%958, %957 : tensor<2x320x48x48xf16>, tensor<2x320x48x48xf16>) outs(%924 : tensor<2x320x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x320x48x48xf16>
    %960 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %961 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%960 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %962 = tensor.empty() : tensor<640xf16>
    %963 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%961 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %964 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_86 : tensor<640xf32>) outs(%963 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %965 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %966 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%965 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %967 = tensor.empty() : tensor<640x320x3x3xf16>
    %968 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%966 : tensor<f16>) outs(%967 : tensor<640x320x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x320x3x3xf16>
    %969 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_87 : tensor<640x320x3x3xf32>) outs(%968 : tensor<640x320x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x320x3x3xf16>
    %padded_879 = tensor.pad %959 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x320x48x48xf16> to tensor<2x320x50x50xf16>
    %970 = tensor.empty() : tensor<2x640x48x48xf16>
    %971 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%964 : tensor<640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %972 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_879, %969 : tensor<2x320x50x50xf16>, tensor<640x320x3x3xf16>) outs(%971 : tensor<2x640x48x48xf16>) -> tensor<2x640x48x48xf16>
    %973 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %974 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%973, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %975 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %976 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%975 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %977 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%976 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %978 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_88 : tensor<640xf32>) outs(%977 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %979 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %980 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%979 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %981 = tensor.empty() : tensor<640x1280xf16>
    %982 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%980 : tensor<f16>) outs(%981 : tensor<640x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x1280xf16>
    %983 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_89 : tensor<640x1280xf32>) outs(%982 : tensor<640x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x1280xf16>
    %984 = tensor.empty() : tensor<1280x640xf16>
    %985 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%983 : tensor<640x1280xf16>) outs(%984 : tensor<1280x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x640xf16>
    %986 = tensor.empty() : tensor<2x640xf16>
    %987 = linalg.fill ins(%cst_690 : f16) outs(%986 : tensor<2x640xf16>) -> tensor<2x640xf16>
    %988 = linalg.matmul ins(%974, %985 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%987 : tensor<2x640xf16>) -> tensor<2x640xf16>
    %989 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%978, %988 : tensor<640xf16>, tensor<2x640xf16>) outs(%986 : tensor<2x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640xf16>
    %expanded_880 = tensor.expand_shape %989 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16>
    %990 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%972, %expanded_880 : tensor<2x640x48x48xf16>, tensor<2x640x1x1xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %991 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %992 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%991 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %993 = tensor.empty() : tensor<2x640x48x48xf32>
    %994 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%992 : tensor<f32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x640x48x48xf32>
    %995 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%990 : tensor<2x640x48x48xf16>) outs(%994 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %collapsed_881 = tensor.collapse_shape %995 [[0], [1], [2, 3]] : tensor<2x640x48x48xf32> into tensor<2x640x2304xf32>
    %expanded_882 = tensor.expand_shape %collapsed_881 [[0], [1, 2], [3]] : tensor<2x640x2304xf32> into tensor<2x32x20x2304xf32>
    %996 = tensor.empty() : tensor<2x32x20x2304xf64>
    %997 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_882 : tensor<2x32x20x2304xf32>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %998 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %999 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%997 : tensor<2x32x20x2304xf64>) outs(%998 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1000 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%999 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1001 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%997, %1000 : tensor<2x32x20x2304xf64>, tensor<2x32x1x1xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1002 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1001 : tensor<2x32x20x2304xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1003 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1004 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1002 : tensor<2x32x20x2304xf64>) outs(%1003 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1005 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1004 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1006 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1005 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1007 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1008 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_882 : tensor<2x32x20x2304xf32>) outs(%1007 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1009 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1008 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1010 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1009 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1011 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1006, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1012 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1011 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1013 = tensor.empty() : tensor<2x32x20x2304xf32>
    %1014 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_882, %1010 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %1015 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1014, %1012 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %collapsed_883 = tensor.collapse_shape %1015 [[0], [1, 2], [3]] : tensor<2x32x20x2304xf32> into tensor<2x640x2304xf32>
    %expanded_884 = tensor.expand_shape %collapsed_883 [[0], [1], [2, 3]] : tensor<2x640x2304xf32> into tensor<2x640x48x48xf32>
    %expanded_885 = tensor.expand_shape %cst_90 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %expanded_886 = tensor.expand_shape %cst_91 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %1016 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_884, %expanded_886 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1017 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1016, %expanded_885 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1018 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1019 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1018 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1020 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1019 : tensor<f16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1021 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1017 : tensor<2x640x48x48xf32>) outs(%1020 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1022 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1021 : tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x640x48x48xf16>
    %1023 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1022, %1021 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1024 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1025 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1024 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1026 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1025 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1027 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_92 : tensor<640xf32>) outs(%1026 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1028 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1029 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1028 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1030 = tensor.empty() : tensor<640x640x3x3xf16>
    %1031 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1029 : tensor<f16>) outs(%1030 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640x3x3xf16>
    %1032 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_93 : tensor<640x640x3x3xf32>) outs(%1031 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640x3x3xf16>
    %padded_887 = tensor.pad %1023 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x640x48x48xf16> to tensor<2x640x50x50xf16>
    %1033 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1027 : tensor<640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1034 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_887, %1032 : tensor<2x640x50x50xf16>, tensor<640x640x3x3xf16>) outs(%1033 : tensor<2x640x48x48xf16>) -> tensor<2x640x48x48xf16>
    %1035 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1036 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1035 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1037 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1036 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1038 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_94 : tensor<640xf32>) outs(%1037 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1039 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1040 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1039 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1041 = tensor.empty() : tensor<640x320x1x1xf16>
    %1042 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1040 : tensor<f16>) outs(%1041 : tensor<640x320x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x320x1x1xf16>
    %1043 = tensor.empty() : tensor<640x320x1x1xf32>
    %1044 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_95 : tensor<640x320x1x1xf32>) outs(%1043 : tensor<640x320x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<640x320x1x1xf32>
    %1045 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1044 : tensor<640x320x1x1xf32>) outs(%1042 : tensor<640x320x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x320x1x1xf16>
    %1046 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1038 : tensor<640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1047 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%926, %1045 : tensor<2x320x48x48xf16>, tensor<640x320x1x1xf16>) outs(%1046 : tensor<2x640x48x48xf16>) -> tensor<2x640x48x48xf16>
    %1048 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1047, %1034 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1049 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1050 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1049 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1051 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1050 : tensor<f32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x640x48x48xf32>
    %1052 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1048 : tensor<2x640x48x48xf16>) outs(%1051 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %collapsed_888 = tensor.collapse_shape %1052 [[0], [1], [2, 3]] : tensor<2x640x48x48xf32> into tensor<2x640x2304xf32>
    %expanded_889 = tensor.expand_shape %collapsed_888 [[0], [1, 2], [3]] : tensor<2x640x2304xf32> into tensor<2x32x20x2304xf32>
    %1053 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_889 : tensor<2x32x20x2304xf32>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1054 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1055 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1053 : tensor<2x32x20x2304xf64>) outs(%1054 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1056 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1055 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1057 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1053, %1056 : tensor<2x32x20x2304xf64>, tensor<2x32x1x1xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1058 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1057 : tensor<2x32x20x2304xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1059 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1060 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1058 : tensor<2x32x20x2304xf64>) outs(%1059 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1061 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1060 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1062 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1061 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1063 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1064 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_889 : tensor<2x32x20x2304xf32>) outs(%1063 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1065 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1064 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1066 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1065 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1067 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1062, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1068 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1067 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1069 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_889, %1066 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %1070 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1069, %1068 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %collapsed_890 = tensor.collapse_shape %1070 [[0], [1, 2], [3]] : tensor<2x32x20x2304xf32> into tensor<2x640x2304xf32>
    %expanded_891 = tensor.expand_shape %collapsed_890 [[0], [1], [2, 3]] : tensor<2x640x2304xf32> into tensor<2x640x48x48xf32>
    %expanded_892 = tensor.expand_shape %cst_96 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %expanded_893 = tensor.expand_shape %cst_97 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %1071 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_891, %expanded_893 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1072 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1071, %expanded_892 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1073 = tensor.empty() : tensor<2x48x48x640xf32>
    %1074 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1072 : tensor<2x640x48x48xf32>) outs(%1073 : tensor<2x48x48x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x48x48x640xf32>
    %collapsed_894 = tensor.collapse_shape %1074 [[0], [1, 2], [3]] : tensor<2x48x48x640xf32> into tensor<2x2304x640xf32>
    %1075 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1076 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1075 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1077 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1076 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1078 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_98 : tensor<640xf32>) outs(%1077 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1079 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1080 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1079 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1081 = tensor.empty() : tensor<640x640xf16>
    %1082 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1080 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1083 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_99 : tensor<640x640xf32>) outs(%1082 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1084 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1085 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1084 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1086 = tensor.empty() : tensor<2x2304x640xf16>
    %1087 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1085 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1088 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_894 : tensor<2x2304x640xf32>) outs(%1087 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1089 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1083 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_895 = tensor.collapse_shape %1088 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1090 = tensor.empty() : tensor<4608x640xf16>
    %1091 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1092 = linalg.matmul ins(%collapsed_895, %1089 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1091 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1093 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1078, %1092 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_896 = tensor.expand_shape %1093 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1094 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1095 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1094 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1096 = tensor.empty() : tensor<2x2304x640xf32>
    %1097 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1095 : tensor<f32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1098 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_896 : tensor<2x2304x640xf16>) outs(%1097 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1099 = tensor.empty() : tensor<2x2304x1xf64>
    %1100 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1101 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1098 : tensor<2x2304x640xf32>) outs(%1100 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1102 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1101 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1103 = tensor.empty() : tensor<2x2304x1xf32>
    %1104 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1102 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1105 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1104 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1106 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1098, %1105 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1107 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1106 : tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1108 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1109 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1107 : tensor<2x2304x640xf32>) outs(%1108 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1110 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1109 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1111 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1110 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1112 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1111 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x2304x1xf32>
    %1113 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1112 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1114 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1113 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1115 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1106, %1114 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1116 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1115, %cst_100 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1117 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1116, %cst_101 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1118 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1119 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1118 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1120 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1119 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1121 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_102 : tensor<640x640xf32>) outs(%1120 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1122 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1123 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1122 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1124 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1123 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1125 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1117 : tensor<2x2304x640xf32>) outs(%1124 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1126 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1121 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_897 = tensor.collapse_shape %1125 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1127 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1128 = linalg.matmul ins(%collapsed_897, %1126 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1127 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1129 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1130 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1129 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1131 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1130 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1132 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_103 : tensor<640x640xf32>) outs(%1131 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1133 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1134 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1133 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1135 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1134 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1136 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1117 : tensor<2x2304x640xf32>) outs(%1135 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1137 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1132 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_898 = tensor.collapse_shape %1136 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1138 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1139 = linalg.matmul ins(%collapsed_898, %1137 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1138 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1140 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1141 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1140 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1142 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1141 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1143 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_104 : tensor<640x640xf32>) outs(%1142 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1144 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1145 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1144 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1146 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1145 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1147 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1117 : tensor<2x2304x640xf32>) outs(%1146 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1148 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1143 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_899 = tensor.collapse_shape %1147 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1149 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1150 = linalg.matmul ins(%collapsed_899, %1148 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1149 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %expanded_900 = tensor.expand_shape %1128 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1151 = tensor.empty() : tensor<2x10x2304x64xf16>
    %1152 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_900 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_901 = tensor.collapse_shape %1152 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_902 = tensor.expand_shape %1139 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1153 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_902 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_903 = tensor.collapse_shape %1153 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_904 = tensor.expand_shape %1150 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1154 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_904 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_905 = tensor.collapse_shape %1154 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_906 = tensor.expand_shape %collapsed_901 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1155 = tensor.empty() : tensor<20x2304x1x64xf16>
    %1156 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_906 : tensor<20x2304x64x1xf16>) outs(%1155 : tensor<20x2304x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x64xf16>
    %expanded_907 = tensor.expand_shape %collapsed_903 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1157 = tensor.empty() : tensor<20x1x2304x64xf16>
    %1158 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_907 : tensor<20x2304x64x1xf16>) outs(%1157 : tensor<20x1x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x2304x64xf16>
    %1159 = tensor.empty() : tensor<20x2304x64x1xf16>
    %1160 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1156 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_908 = tensor.collapse_shape %1160 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %1161 = tensor.empty() : tensor<20x64x2304x1xf16>
    %1162 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1158 : tensor<20x1x2304x64xf16>) outs(%1161 : tensor<20x64x2304x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x64x2304x1xf16>
    %collapsed_909 = tensor.collapse_shape %1162 [[0], [1], [2, 3]] : tensor<20x64x2304x1xf16> into tensor<20x64x2304xf16>
    %1163 = tensor.empty() : tensor<20x2304x2304xf16>
    %1164 = linalg.fill ins(%cst_690 : f16) outs(%1163 : tensor<20x2304x2304xf16>) -> tensor<20x2304x2304xf16>
    %1165 = linalg.batch_matmul ins(%collapsed_908, %collapsed_909 : tensor<20x2304x64xf16>, tensor<20x64x2304xf16>) outs(%1164 : tensor<20x2304x2304xf16>) -> tensor<20x2304x2304xf16>
    %expanded_910 = tensor.expand_shape %1165 [[0], [1], [2, 3]] : tensor<20x2304x2304xf16> into tensor<20x2304x1x2304xf16>
    %1166 = tensor.empty() : tensor<20x2304x2304x1xf16>
    %1167 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_910 : tensor<20x2304x1x2304xf16>) outs(%1166 : tensor<20x2304x2304x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x2304x1xf16>
    %collapsed_911 = tensor.collapse_shape %1167 [[0], [1], [2, 3]] : tensor<20x2304x2304x1xf16> into tensor<20x2304x2304xf16>
    %1168 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_911, %cst : tensor<20x2304x2304xf16>, tensor<f64>) outs(%1163 : tensor<20x2304x2304xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<20x2304x2304xf16>
    %1169 = tensor.empty() : tensor<20x2304x2304xf32>
    %1170 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1168 : tensor<20x2304x2304xf16>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1171 = tensor.empty() : tensor<20x2304x1xf32>
    %1172 = linalg.fill ins(%cst_692 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1173 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1170 : tensor<20x2304x2304xf32>) outs(%1172 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1174 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1170, %1173 : tensor<20x2304x2304xf32>, tensor<20x2304x1xf32>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1175 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1174 : tensor<20x2304x2304xf32>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1176 = linalg.fill ins(%cst_697 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1177 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1175 : tensor<20x2304x2304xf32>) outs(%1176 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1178 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1175, %1177 : tensor<20x2304x2304xf32>, tensor<20x2304x1xf32>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1179 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1180 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1179 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1181 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1180 : tensor<f16>) outs(%1163 : tensor<20x2304x2304xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x2304xf16>
    %1182 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1178 : tensor<20x2304x2304xf32>) outs(%1181 : tensor<20x2304x2304xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<20x2304x2304xf16>
    %expanded_912 = tensor.expand_shape %1182 [[0], [1], [2, 3]] : tensor<20x2304x2304xf16> into tensor<20x2304x2304x1xf16>
    %1183 = tensor.empty() : tensor<20x2304x1x2304xf16>
    %1184 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_912 : tensor<20x2304x2304x1xf16>) outs(%1183 : tensor<20x2304x1x2304xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x2304xf16>
    %expanded_913 = tensor.expand_shape %collapsed_905 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1185 = tensor.empty() : tensor<20x1x64x2304xf16>
    %1186 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_913 : tensor<20x2304x64x1xf16>) outs(%1185 : tensor<20x1x64x2304xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x64x2304xf16>
    %1187 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1184 : tensor<20x2304x1x2304xf16>) outs(%1166 : tensor<20x2304x2304x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x2304x1xf16>
    %collapsed_914 = tensor.collapse_shape %1187 [[0], [1], [2, 3]] : tensor<20x2304x2304x1xf16> into tensor<20x2304x2304xf16>
    %1188 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1186 : tensor<20x1x64x2304xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_915 = tensor.collapse_shape %1188 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %1189 = tensor.empty() : tensor<20x2304x64xf16>
    %1190 = linalg.fill ins(%cst_690 : f16) outs(%1189 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %1191 = linalg.batch_matmul ins(%collapsed_914, %collapsed_915 : tensor<20x2304x2304xf16>, tensor<20x2304x64xf16>) outs(%1190 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %expanded_916 = tensor.expand_shape %1191 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x1x64xf16>
    %1192 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_916 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_917 = tensor.collapse_shape %1192 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %expanded_918 = tensor.expand_shape %collapsed_917 [[0, 1], [2], [3]] : tensor<20x2304x64xf16> into tensor<2x10x2304x64xf16>
    %1193 = tensor.empty() : tensor<2x2304x10x64xf16>
    %1194 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_918 : tensor<2x10x2304x64xf16>) outs(%1193 : tensor<2x2304x10x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x10x64xf16>
    %1195 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1196 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1195 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1197 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1196 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1198 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_105 : tensor<640xf32>) outs(%1197 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1199 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1200 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1199 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1201 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1200 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1202 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_106 : tensor<640x640xf32>) outs(%1201 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1203 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1202 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_919 = tensor.collapse_shape %1194 [[0, 1], [2, 3]] : tensor<2x2304x10x64xf16> into tensor<4608x640xf16>
    %1204 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1205 = linalg.matmul ins(%collapsed_919, %1203 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1204 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1206 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1198, %1205 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_920 = tensor.expand_shape %1206 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1207 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_920, %expanded_896 : tensor<2x2304x640xf16>, tensor<2x2304x640xf16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1208 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1209 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1208 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1210 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1209 : tensor<f32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1211 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1207 : tensor<2x2304x640xf16>) outs(%1210 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1212 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1213 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1211 : tensor<2x2304x640xf32>) outs(%1212 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1214 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1213 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1215 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1214 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1216 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1215 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1217 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1211, %1216 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1218 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1217 : tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1219 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1220 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1218 : tensor<2x2304x640xf32>) outs(%1219 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1221 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1220 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1222 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1221 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1223 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1222 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x2304x1xf32>
    %1224 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1223 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1225 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1224 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1226 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1217, %1225 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1227 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1226, %cst_107 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1228 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1227, %cst_108 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1229 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1230 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1229 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1231 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1230 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1232 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_109 : tensor<640x640xf32>) outs(%1231 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1233 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1234 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1233 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1235 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1234 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1236 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1228 : tensor<2x2304x640xf32>) outs(%1235 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1237 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1232 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_921 = tensor.collapse_shape %1236 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1238 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1239 = linalg.matmul ins(%collapsed_921, %1237 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1238 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1240 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1241 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1240 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1242 = tensor.empty() : tensor<640x1024xf16>
    %1243 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1241 : tensor<f16>) outs(%1242 : tensor<640x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x1024xf16>
    %1244 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_110 : tensor<640x1024xf32>) outs(%1243 : tensor<640x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x1024xf16>
    %1245 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1246 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1245 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1247 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1246 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %1248 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%1247 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %1249 = tensor.empty() : tensor<1024x640xf16>
    %1250 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1244 : tensor<640x1024xf16>) outs(%1249 : tensor<1024x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x640xf16>
    %collapsed_922 = tensor.collapse_shape %1248 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %1251 = tensor.empty() : tensor<154x640xf16>
    %1252 = linalg.fill ins(%cst_690 : f16) outs(%1251 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %1253 = linalg.matmul ins(%collapsed_922, %1250 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1252 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %1254 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1255 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1254 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1256 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1255 : tensor<f16>) outs(%1242 : tensor<640x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x1024xf16>
    %1257 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_111 : tensor<640x1024xf32>) outs(%1256 : tensor<640x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x1024xf16>
    %1258 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1259 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1258 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1260 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1259 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %1261 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%1260 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %1262 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1257 : tensor<640x1024xf16>) outs(%1249 : tensor<1024x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x640xf16>
    %collapsed_923 = tensor.collapse_shape %1261 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %1263 = linalg.fill ins(%cst_690 : f16) outs(%1251 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %1264 = linalg.matmul ins(%collapsed_923, %1262 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1263 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %expanded_924 = tensor.expand_shape %1239 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1265 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_924 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_925 = tensor.collapse_shape %1265 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_926 = tensor.expand_shape %1253 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
    %1266 = tensor.empty() : tensor<2x10x77x64xf16>
    %1267 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_926 : tensor<2x77x10x64xf16>) outs(%1266 : tensor<2x10x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x77x64xf16>
    %collapsed_927 = tensor.collapse_shape %1267 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
    %expanded_928 = tensor.expand_shape %1264 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
    %1268 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_928 : tensor<2x77x10x64xf16>) outs(%1266 : tensor<2x10x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x77x64xf16>
    %collapsed_929 = tensor.collapse_shape %1268 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
    %expanded_930 = tensor.expand_shape %collapsed_925 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1269 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_930 : tensor<20x2304x64x1xf16>) outs(%1155 : tensor<20x2304x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x64xf16>
    %expanded_931 = tensor.expand_shape %collapsed_927 [[0], [1], [2, 3]] : tensor<20x77x64xf16> into tensor<20x77x64x1xf16>
    %1270 = tensor.empty() : tensor<20x1x77x64xf16>
    %1271 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_931 : tensor<20x77x64x1xf16>) outs(%1270 : tensor<20x1x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x77x64xf16>
    %1272 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1269 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_932 = tensor.collapse_shape %1272 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %1273 = tensor.empty() : tensor<20x64x77x1xf16>
    %1274 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1271 : tensor<20x1x77x64xf16>) outs(%1273 : tensor<20x64x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x64x77x1xf16>
    %collapsed_933 = tensor.collapse_shape %1274 [[0], [1], [2, 3]] : tensor<20x64x77x1xf16> into tensor<20x64x77xf16>
    %1275 = tensor.empty() : tensor<20x2304x77xf16>
    %1276 = linalg.fill ins(%cst_690 : f16) outs(%1275 : tensor<20x2304x77xf16>) -> tensor<20x2304x77xf16>
    %1277 = linalg.batch_matmul ins(%collapsed_932, %collapsed_933 : tensor<20x2304x64xf16>, tensor<20x64x77xf16>) outs(%1276 : tensor<20x2304x77xf16>) -> tensor<20x2304x77xf16>
    %expanded_934 = tensor.expand_shape %1277 [[0], [1], [2, 3]] : tensor<20x2304x77xf16> into tensor<20x2304x1x77xf16>
    %1278 = tensor.empty() : tensor<20x2304x77x1xf16>
    %1279 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_934 : tensor<20x2304x1x77xf16>) outs(%1278 : tensor<20x2304x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x77x1xf16>
    %collapsed_935 = tensor.collapse_shape %1279 [[0], [1], [2, 3]] : tensor<20x2304x77x1xf16> into tensor<20x2304x77xf16>
    %1280 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_935, %cst : tensor<20x2304x77xf16>, tensor<f64>) outs(%1275 : tensor<20x2304x77xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<20x2304x77xf16>
    %1281 = tensor.empty() : tensor<20x2304x77xf32>
    %1282 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1280 : tensor<20x2304x77xf16>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1283 = linalg.fill ins(%cst_692 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1284 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1282 : tensor<20x2304x77xf32>) outs(%1283 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1285 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1282, %1284 : tensor<20x2304x77xf32>, tensor<20x2304x1xf32>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1286 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1285 : tensor<20x2304x77xf32>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1287 = linalg.fill ins(%cst_697 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1288 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1286 : tensor<20x2304x77xf32>) outs(%1287 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1289 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1286, %1288 : tensor<20x2304x77xf32>, tensor<20x2304x1xf32>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1290 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1291 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1290 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1292 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1291 : tensor<f16>) outs(%1275 : tensor<20x2304x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x77xf16>
    %1293 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1289 : tensor<20x2304x77xf32>) outs(%1292 : tensor<20x2304x77xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<20x2304x77xf16>
    %expanded_936 = tensor.expand_shape %1293 [[0], [1], [2, 3]] : tensor<20x2304x77xf16> into tensor<20x2304x77x1xf16>
    %1294 = tensor.empty() : tensor<20x2304x1x77xf16>
    %1295 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_936 : tensor<20x2304x77x1xf16>) outs(%1294 : tensor<20x2304x1x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x77xf16>
    %expanded_937 = tensor.expand_shape %collapsed_929 [[0], [1], [2, 3]] : tensor<20x77x64xf16> into tensor<20x77x64x1xf16>
    %1296 = tensor.empty() : tensor<20x1x64x77xf16>
    %1297 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_937 : tensor<20x77x64x1xf16>) outs(%1296 : tensor<20x1x64x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x64x77xf16>
    %1298 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1295 : tensor<20x2304x1x77xf16>) outs(%1278 : tensor<20x2304x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x77x1xf16>
    %collapsed_938 = tensor.collapse_shape %1298 [[0], [1], [2, 3]] : tensor<20x2304x77x1xf16> into tensor<20x2304x77xf16>
    %1299 = tensor.empty() : tensor<20x77x64x1xf16>
    %1300 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1297 : tensor<20x1x64x77xf16>) outs(%1299 : tensor<20x77x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x77x64x1xf16>
    %collapsed_939 = tensor.collapse_shape %1300 [[0], [1], [2, 3]] : tensor<20x77x64x1xf16> into tensor<20x77x64xf16>
    %1301 = linalg.fill ins(%cst_690 : f16) outs(%1189 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %1302 = linalg.batch_matmul ins(%collapsed_938, %collapsed_939 : tensor<20x2304x77xf16>, tensor<20x77x64xf16>) outs(%1301 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %expanded_940 = tensor.expand_shape %1302 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x1x64xf16>
    %1303 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_940 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_941 = tensor.collapse_shape %1303 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %expanded_942 = tensor.expand_shape %collapsed_941 [[0, 1], [2], [3]] : tensor<20x2304x64xf16> into tensor<2x10x2304x64xf16>
    %1304 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_942 : tensor<2x10x2304x64xf16>) outs(%1193 : tensor<2x2304x10x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x10x64xf16>
    %1305 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1306 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1305 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1307 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1306 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1308 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_112 : tensor<640xf32>) outs(%1307 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1309 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1310 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1309 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1311 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1310 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1312 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_113 : tensor<640x640xf32>) outs(%1311 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1313 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1312 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_943 = tensor.collapse_shape %1304 [[0, 1], [2, 3]] : tensor<2x2304x10x64xf16> into tensor<4608x640xf16>
    %1314 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1315 = linalg.matmul ins(%collapsed_943, %1313 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1314 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1316 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1308, %1315 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_944 = tensor.expand_shape %1316 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1317 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_944, %1207 : tensor<2x2304x640xf16>, tensor<2x2304x640xf16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1318 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1319 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1318 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1320 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1319 : tensor<f32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1321 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1317 : tensor<2x2304x640xf16>) outs(%1320 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1322 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1323 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1321 : tensor<2x2304x640xf32>) outs(%1322 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1324 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1323 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1325 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1324 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1326 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1325 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1327 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1321, %1326 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1328 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1327 : tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1329 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1330 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1328 : tensor<2x2304x640xf32>) outs(%1329 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1331 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1330 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1332 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1331 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1333 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1332 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x2304x1xf32>
    %1334 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1333 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1335 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1334 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1336 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1327, %1335 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1337 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1336, %cst_114 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1338 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1337, %cst_115 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1339 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1340 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1339 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1341 = tensor.empty() : tensor<5120xf16>
    %1342 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1340 : tensor<f16>) outs(%1341 : tensor<5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<5120xf16>
    %1343 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_116 : tensor<5120xf32>) outs(%1342 : tensor<5120xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<5120xf16>
    %1344 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1345 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1344 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1346 = tensor.empty() : tensor<5120x640xf16>
    %1347 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1345 : tensor<f16>) outs(%1346 : tensor<5120x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<5120x640xf16>
    %1348 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_117 : tensor<5120x640xf32>) outs(%1347 : tensor<5120x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<5120x640xf16>
    %1349 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1350 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1349 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1351 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1350 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1352 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1338 : tensor<2x2304x640xf32>) outs(%1351 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1353 = tensor.empty() : tensor<640x5120xf16>
    %1354 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1348 : tensor<5120x640xf16>) outs(%1353 : tensor<640x5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x5120xf16>
    %collapsed_945 = tensor.collapse_shape %1352 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1355 = tensor.empty() : tensor<4608x5120xf16>
    %1356 = linalg.fill ins(%cst_690 : f16) outs(%1355 : tensor<4608x5120xf16>) -> tensor<4608x5120xf16>
    %1357 = linalg.matmul ins(%collapsed_945, %1354 : tensor<4608x640xf16>, tensor<640x5120xf16>) outs(%1356 : tensor<4608x5120xf16>) -> tensor<4608x5120xf16>
    %1358 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1343, %1357 : tensor<5120xf16>, tensor<4608x5120xf16>) outs(%1355 : tensor<4608x5120xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x5120xf16>
    %expanded_946 = tensor.expand_shape %1358 [[0, 1], [2]] : tensor<4608x5120xf16> into tensor<2x2304x5120xf16>
    %extracted_slice_947 = tensor.extract_slice %expanded_946[0, 0, 0] [2, 2304, 2560] [1, 1, 1] : tensor<2x2304x5120xf16> to tensor<2x2304x2560xf16>
    %extracted_slice_948 = tensor.extract_slice %expanded_946[0, 0, 2560] [2, 2304, 2560] [1, 1, 1] : tensor<2x2304x5120xf16> to tensor<2x2304x2560xf16>
    %1359 = tensor.empty() : tensor<2x2304x2560xf16>
    %1360 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_948 : tensor<2x2304x2560xf16>) outs(%1359 : tensor<2x2304x2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = math.sqrt %cst_693 : f16
      %7525 = arith.divf %in, %7524 : f16
      %7526 = math.erf %7525 : f16
      %7527 = arith.addf %7526, %cst_691 : f16
      %7528 = arith.mulf %7527, %cst_694 : f16
      %7529 = arith.mulf %in, %7528 : f16
      linalg.yield %7529 : f16
    } -> tensor<2x2304x2560xf16>
    %1361 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_947, %1360 : tensor<2x2304x2560xf16>, tensor<2x2304x2560xf16>) outs(%1359 : tensor<2x2304x2560xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x2560xf16>
    %1362 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1363 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1362 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1364 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1363 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1365 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_118 : tensor<640xf32>) outs(%1364 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1366 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1367 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1366 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1368 = tensor.empty() : tensor<640x2560xf16>
    %1369 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1367 : tensor<f16>) outs(%1368 : tensor<640x2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x2560xf16>
    %1370 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_119 : tensor<640x2560xf32>) outs(%1369 : tensor<640x2560xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x2560xf16>
    %1371 = tensor.empty() : tensor<2560x640xf16>
    %1372 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1370 : tensor<640x2560xf16>) outs(%1371 : tensor<2560x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2560x640xf16>
    %collapsed_949 = tensor.collapse_shape %1361 [[0, 1], [2]] : tensor<2x2304x2560xf16> into tensor<4608x2560xf16>
    %1373 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1374 = linalg.matmul ins(%collapsed_949, %1372 : tensor<4608x2560xf16>, tensor<2560x640xf16>) outs(%1373 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1375 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1365, %1374 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_950 = tensor.expand_shape %1375 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1376 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_950, %1317 : tensor<2x2304x640xf16>, tensor<2x2304x640xf16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1377 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1378 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1377 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1379 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1378 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1380 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_120 : tensor<640xf32>) outs(%1379 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1381 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1382 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1381 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1383 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1382 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1384 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_121 : tensor<640x640xf32>) outs(%1383 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1385 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1384 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_951 = tensor.collapse_shape %1376 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1386 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1387 = linalg.matmul ins(%collapsed_951, %1385 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1386 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1388 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1380, %1387 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_952 = tensor.expand_shape %1388 [[0, 1, 2], [3]] : tensor<4608x640xf16> into tensor<2x48x48x640xf16>
    %1389 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_952 : tensor<2x48x48x640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1390 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1389, %1048 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1391 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1392 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1391 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1393 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1392 : tensor<f32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x640x48x48xf32>
    %1394 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1390 : tensor<2x640x48x48xf16>) outs(%1393 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %collapsed_953 = tensor.collapse_shape %1394 [[0], [1], [2, 3]] : tensor<2x640x48x48xf32> into tensor<2x640x2304xf32>
    %expanded_954 = tensor.expand_shape %collapsed_953 [[0], [1, 2], [3]] : tensor<2x640x2304xf32> into tensor<2x32x20x2304xf32>
    %1395 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_954 : tensor<2x32x20x2304xf32>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1396 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1397 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1395 : tensor<2x32x20x2304xf64>) outs(%1396 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1398 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1397 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1399 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1395, %1398 : tensor<2x32x20x2304xf64>, tensor<2x32x1x1xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1400 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1399 : tensor<2x32x20x2304xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1401 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1402 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1400 : tensor<2x32x20x2304xf64>) outs(%1401 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1403 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1402 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1404 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1403 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1405 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1406 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_954 : tensor<2x32x20x2304xf32>) outs(%1405 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1407 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1406 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1408 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1407 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1409 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1404, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1410 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1409 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1411 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_954, %1408 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %1412 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1411, %1410 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %collapsed_955 = tensor.collapse_shape %1412 [[0], [1, 2], [3]] : tensor<2x32x20x2304xf32> into tensor<2x640x2304xf32>
    %expanded_956 = tensor.expand_shape %collapsed_955 [[0], [1], [2, 3]] : tensor<2x640x2304xf32> into tensor<2x640x48x48xf32>
    %expanded_957 = tensor.expand_shape %cst_122 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %expanded_958 = tensor.expand_shape %cst_123 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %1413 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_956, %expanded_958 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1414 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1413, %expanded_957 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1415 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1416 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1415 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1417 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1416 : tensor<f16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1418 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1414 : tensor<2x640x48x48xf32>) outs(%1417 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1419 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1418 : tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x640x48x48xf16>
    %1420 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1419, %1418 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1421 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1422 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1421 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1423 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1422 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1424 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_124 : tensor<640xf32>) outs(%1423 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1425 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1426 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1425 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1427 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1426 : tensor<f16>) outs(%1030 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640x3x3xf16>
    %1428 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_125 : tensor<640x640x3x3xf32>) outs(%1427 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640x3x3xf16>
    %padded_959 = tensor.pad %1420 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x640x48x48xf16> to tensor<2x640x50x50xf16>
    %1429 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1424 : tensor<640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1430 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_959, %1428 : tensor<2x640x50x50xf16>, tensor<640x640x3x3xf16>) outs(%1429 : tensor<2x640x48x48xf16>) -> tensor<2x640x48x48xf16>
    %1431 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %1432 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1431, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %1433 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1434 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1433 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1435 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1434 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1436 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_126 : tensor<640xf32>) outs(%1435 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1437 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1438 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1437 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1439 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1438 : tensor<f16>) outs(%981 : tensor<640x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x1280xf16>
    %1440 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_127 : tensor<640x1280xf32>) outs(%1439 : tensor<640x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x1280xf16>
    %1441 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1440 : tensor<640x1280xf16>) outs(%984 : tensor<1280x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x640xf16>
    %1442 = linalg.fill ins(%cst_690 : f16) outs(%986 : tensor<2x640xf16>) -> tensor<2x640xf16>
    %1443 = linalg.matmul ins(%1432, %1441 : tensor<2x1280xf16>, tensor<1280x640xf16>) outs(%1442 : tensor<2x640xf16>) -> tensor<2x640xf16>
    %1444 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1436, %1443 : tensor<640xf16>, tensor<2x640xf16>) outs(%986 : tensor<2x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640xf16>
    %expanded_960 = tensor.expand_shape %1444 [[0], [1, 2, 3]] : tensor<2x640xf16> into tensor<2x640x1x1xf16>
    %1445 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1430, %expanded_960 : tensor<2x640x48x48xf16>, tensor<2x640x1x1xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1446 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1447 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1446 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1448 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1447 : tensor<f32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x640x48x48xf32>
    %1449 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1445 : tensor<2x640x48x48xf16>) outs(%1448 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %collapsed_961 = tensor.collapse_shape %1449 [[0], [1], [2, 3]] : tensor<2x640x48x48xf32> into tensor<2x640x2304xf32>
    %expanded_962 = tensor.expand_shape %collapsed_961 [[0], [1, 2], [3]] : tensor<2x640x2304xf32> into tensor<2x32x20x2304xf32>
    %1450 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_962 : tensor<2x32x20x2304xf32>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1451 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1452 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1450 : tensor<2x32x20x2304xf64>) outs(%1451 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1453 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1452 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1454 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1450, %1453 : tensor<2x32x20x2304xf64>, tensor<2x32x1x1xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1455 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1454 : tensor<2x32x20x2304xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1456 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1457 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1455 : tensor<2x32x20x2304xf64>) outs(%1456 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1458 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1457 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1459 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1458 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1460 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1461 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_962 : tensor<2x32x20x2304xf32>) outs(%1460 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1462 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1461 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1463 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1462 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1464 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1459, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1465 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1464 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1466 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_962, %1463 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %1467 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1466, %1465 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %collapsed_963 = tensor.collapse_shape %1467 [[0], [1, 2], [3]] : tensor<2x32x20x2304xf32> into tensor<2x640x2304xf32>
    %expanded_964 = tensor.expand_shape %collapsed_963 [[0], [1], [2, 3]] : tensor<2x640x2304xf32> into tensor<2x640x48x48xf32>
    %expanded_965 = tensor.expand_shape %cst_128 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %expanded_966 = tensor.expand_shape %cst_129 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %1468 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_964, %expanded_966 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1469 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1468, %expanded_965 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1470 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1471 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1470 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1472 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1471 : tensor<f16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1473 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1469 : tensor<2x640x48x48xf32>) outs(%1472 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1474 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1473 : tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x640x48x48xf16>
    %1475 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1474, %1473 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1476 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1477 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1476 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1478 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1477 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1479 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_130 : tensor<640xf32>) outs(%1478 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1480 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1481 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1480 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1482 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1481 : tensor<f16>) outs(%1030 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640x3x3xf16>
    %1483 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_131 : tensor<640x640x3x3xf32>) outs(%1482 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640x3x3xf16>
    %padded_967 = tensor.pad %1475 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x640x48x48xf16> to tensor<2x640x50x50xf16>
    %1484 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1479 : tensor<640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1485 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_967, %1483 : tensor<2x640x50x50xf16>, tensor<640x640x3x3xf16>) outs(%1484 : tensor<2x640x48x48xf16>) -> tensor<2x640x48x48xf16>
    %1486 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1390, %1485 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1487 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1488 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1487 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1489 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1488 : tensor<f32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x640x48x48xf32>
    %1490 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1486 : tensor<2x640x48x48xf16>) outs(%1489 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %collapsed_968 = tensor.collapse_shape %1490 [[0], [1], [2, 3]] : tensor<2x640x48x48xf32> into tensor<2x640x2304xf32>
    %expanded_969 = tensor.expand_shape %collapsed_968 [[0], [1, 2], [3]] : tensor<2x640x2304xf32> into tensor<2x32x20x2304xf32>
    %1491 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_969 : tensor<2x32x20x2304xf32>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1492 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1493 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1491 : tensor<2x32x20x2304xf64>) outs(%1492 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1494 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1493 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1495 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1491, %1494 : tensor<2x32x20x2304xf64>, tensor<2x32x1x1xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1496 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1495 : tensor<2x32x20x2304xf64>) outs(%996 : tensor<2x32x20x2304xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x2304xf64>
    %1497 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1498 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1496 : tensor<2x32x20x2304xf64>) outs(%1497 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1499 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1498 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1500 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1499 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1501 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1502 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_969 : tensor<2x32x20x2304xf32>) outs(%1501 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1503 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1502 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_702 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1504 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1503 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1505 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1500, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1506 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1505 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1507 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_969, %1504 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %1508 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1507, %1506 : tensor<2x32x20x2304xf32>, tensor<2x32x1x1xf32>) outs(%1013 : tensor<2x32x20x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x2304xf32>
    %collapsed_970 = tensor.collapse_shape %1508 [[0], [1, 2], [3]] : tensor<2x32x20x2304xf32> into tensor<2x640x2304xf32>
    %expanded_971 = tensor.expand_shape %collapsed_970 [[0], [1], [2, 3]] : tensor<2x640x2304xf32> into tensor<2x640x48x48xf32>
    %expanded_972 = tensor.expand_shape %cst_132 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %expanded_973 = tensor.expand_shape %cst_133 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %1509 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_971, %expanded_973 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1510 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1509, %expanded_972 : tensor<2x640x48x48xf32>, tensor<1x640x1x1xf32>) outs(%993 : tensor<2x640x48x48xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x48x48xf32>
    %1511 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1510 : tensor<2x640x48x48xf32>) outs(%1073 : tensor<2x48x48x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x48x48x640xf32>
    %collapsed_974 = tensor.collapse_shape %1511 [[0], [1, 2], [3]] : tensor<2x48x48x640xf32> into tensor<2x2304x640xf32>
    %1512 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1513 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1512 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1514 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1513 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1515 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_134 : tensor<640xf32>) outs(%1514 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1516 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1517 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1516 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1518 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1517 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1519 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_135 : tensor<640x640xf32>) outs(%1518 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1520 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1521 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1520 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1522 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1521 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1523 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_974 : tensor<2x2304x640xf32>) outs(%1522 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1524 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1519 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_975 = tensor.collapse_shape %1523 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1525 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1526 = linalg.matmul ins(%collapsed_975, %1524 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1525 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1527 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1515, %1526 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_976 = tensor.expand_shape %1527 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1528 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1529 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1528 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1530 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1529 : tensor<f32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1531 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_976 : tensor<2x2304x640xf16>) outs(%1530 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1532 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1533 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1531 : tensor<2x2304x640xf32>) outs(%1532 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1534 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1533 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1535 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1534 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1536 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1535 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1537 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1531, %1536 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1538 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1537 : tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1539 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1540 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1538 : tensor<2x2304x640xf32>) outs(%1539 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1541 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1540 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1542 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1541 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1543 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1542 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x2304x1xf32>
    %1544 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1543 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1545 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1544 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1546 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1537, %1545 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1547 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1546, %cst_136 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1548 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1547, %cst_137 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1549 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1550 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1549 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1551 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1550 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1552 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_138 : tensor<640x640xf32>) outs(%1551 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1553 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1554 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1553 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1555 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1554 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1556 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1548 : tensor<2x2304x640xf32>) outs(%1555 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1557 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1552 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_977 = tensor.collapse_shape %1556 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1558 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1559 = linalg.matmul ins(%collapsed_977, %1557 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1558 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1560 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1561 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1560 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1562 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1561 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1563 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_139 : tensor<640x640xf32>) outs(%1562 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1564 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1565 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1564 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1566 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1565 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1567 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1548 : tensor<2x2304x640xf32>) outs(%1566 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1568 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1563 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_978 = tensor.collapse_shape %1567 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1569 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1570 = linalg.matmul ins(%collapsed_978, %1568 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1569 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1571 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1572 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1571 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1573 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1572 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1574 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_140 : tensor<640x640xf32>) outs(%1573 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1575 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1576 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1575 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1577 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1576 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1578 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1548 : tensor<2x2304x640xf32>) outs(%1577 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1579 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1574 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_979 = tensor.collapse_shape %1578 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1580 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1581 = linalg.matmul ins(%collapsed_979, %1579 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1580 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %expanded_980 = tensor.expand_shape %1559 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1582 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_980 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_981 = tensor.collapse_shape %1582 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_982 = tensor.expand_shape %1570 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1583 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_982 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_983 = tensor.collapse_shape %1583 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_984 = tensor.expand_shape %1581 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1584 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_984 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_985 = tensor.collapse_shape %1584 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_986 = tensor.expand_shape %collapsed_981 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1585 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_986 : tensor<20x2304x64x1xf16>) outs(%1155 : tensor<20x2304x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x64xf16>
    %expanded_987 = tensor.expand_shape %collapsed_983 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1586 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_987 : tensor<20x2304x64x1xf16>) outs(%1157 : tensor<20x1x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x2304x64xf16>
    %1587 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1585 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_988 = tensor.collapse_shape %1587 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %1588 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1586 : tensor<20x1x2304x64xf16>) outs(%1161 : tensor<20x64x2304x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x64x2304x1xf16>
    %collapsed_989 = tensor.collapse_shape %1588 [[0], [1], [2, 3]] : tensor<20x64x2304x1xf16> into tensor<20x64x2304xf16>
    %1589 = linalg.fill ins(%cst_690 : f16) outs(%1163 : tensor<20x2304x2304xf16>) -> tensor<20x2304x2304xf16>
    %1590 = linalg.batch_matmul ins(%collapsed_988, %collapsed_989 : tensor<20x2304x64xf16>, tensor<20x64x2304xf16>) outs(%1589 : tensor<20x2304x2304xf16>) -> tensor<20x2304x2304xf16>
    %expanded_990 = tensor.expand_shape %1590 [[0], [1], [2, 3]] : tensor<20x2304x2304xf16> into tensor<20x2304x1x2304xf16>
    %1591 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_990 : tensor<20x2304x1x2304xf16>) outs(%1166 : tensor<20x2304x2304x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x2304x1xf16>
    %collapsed_991 = tensor.collapse_shape %1591 [[0], [1], [2, 3]] : tensor<20x2304x2304x1xf16> into tensor<20x2304x2304xf16>
    %1592 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_991, %cst : tensor<20x2304x2304xf16>, tensor<f64>) outs(%1163 : tensor<20x2304x2304xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<20x2304x2304xf16>
    %1593 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1592 : tensor<20x2304x2304xf16>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1594 = linalg.fill ins(%cst_692 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1595 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1593 : tensor<20x2304x2304xf32>) outs(%1594 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1596 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1593, %1595 : tensor<20x2304x2304xf32>, tensor<20x2304x1xf32>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1597 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1596 : tensor<20x2304x2304xf32>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1598 = linalg.fill ins(%cst_697 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1599 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1597 : tensor<20x2304x2304xf32>) outs(%1598 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1600 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1597, %1599 : tensor<20x2304x2304xf32>, tensor<20x2304x1xf32>) outs(%1169 : tensor<20x2304x2304xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x2304xf32>
    %1601 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1602 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1601 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1603 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1602 : tensor<f16>) outs(%1163 : tensor<20x2304x2304xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x2304xf16>
    %1604 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1600 : tensor<20x2304x2304xf32>) outs(%1603 : tensor<20x2304x2304xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<20x2304x2304xf16>
    %expanded_992 = tensor.expand_shape %1604 [[0], [1], [2, 3]] : tensor<20x2304x2304xf16> into tensor<20x2304x2304x1xf16>
    %1605 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_992 : tensor<20x2304x2304x1xf16>) outs(%1183 : tensor<20x2304x1x2304xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x2304xf16>
    %expanded_993 = tensor.expand_shape %collapsed_985 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1606 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_993 : tensor<20x2304x64x1xf16>) outs(%1185 : tensor<20x1x64x2304xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x64x2304xf16>
    %1607 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1605 : tensor<20x2304x1x2304xf16>) outs(%1166 : tensor<20x2304x2304x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x2304x1xf16>
    %collapsed_994 = tensor.collapse_shape %1607 [[0], [1], [2, 3]] : tensor<20x2304x2304x1xf16> into tensor<20x2304x2304xf16>
    %1608 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1606 : tensor<20x1x64x2304xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_995 = tensor.collapse_shape %1608 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %1609 = linalg.fill ins(%cst_690 : f16) outs(%1189 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %1610 = linalg.batch_matmul ins(%collapsed_994, %collapsed_995 : tensor<20x2304x2304xf16>, tensor<20x2304x64xf16>) outs(%1609 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %expanded_996 = tensor.expand_shape %1610 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x1x64xf16>
    %1611 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_996 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_997 = tensor.collapse_shape %1611 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %expanded_998 = tensor.expand_shape %collapsed_997 [[0, 1], [2], [3]] : tensor<20x2304x64xf16> into tensor<2x10x2304x64xf16>
    %1612 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_998 : tensor<2x10x2304x64xf16>) outs(%1193 : tensor<2x2304x10x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x10x64xf16>
    %1613 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1614 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1613 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1615 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1614 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1616 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_141 : tensor<640xf32>) outs(%1615 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1617 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1618 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1617 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1619 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1618 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1620 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_142 : tensor<640x640xf32>) outs(%1619 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1621 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1620 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_999 = tensor.collapse_shape %1612 [[0, 1], [2, 3]] : tensor<2x2304x10x64xf16> into tensor<4608x640xf16>
    %1622 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1623 = linalg.matmul ins(%collapsed_999, %1621 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1622 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1624 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1616, %1623 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_1000 = tensor.expand_shape %1624 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1625 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1000, %expanded_976 : tensor<2x2304x640xf16>, tensor<2x2304x640xf16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1626 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1627 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1626 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1628 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1627 : tensor<f32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1629 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1625 : tensor<2x2304x640xf16>) outs(%1628 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1630 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1631 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1629 : tensor<2x2304x640xf32>) outs(%1630 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1632 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1631 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1633 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1632 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1634 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1633 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1635 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1629, %1634 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1636 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1635 : tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1637 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1638 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1636 : tensor<2x2304x640xf32>) outs(%1637 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1639 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1638 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1640 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1639 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1641 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1640 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x2304x1xf32>
    %1642 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1641 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1643 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1642 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1644 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1635, %1643 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1645 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1644, %cst_143 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1646 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1645, %cst_144 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1647 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1648 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1647 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1649 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1648 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1650 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_145 : tensor<640x640xf32>) outs(%1649 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1651 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1652 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1651 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1653 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1652 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1654 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1646 : tensor<2x2304x640xf32>) outs(%1653 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1655 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1650 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_1001 = tensor.collapse_shape %1654 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1656 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1657 = linalg.matmul ins(%collapsed_1001, %1655 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1656 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1658 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1659 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1658 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1660 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1659 : tensor<f16>) outs(%1242 : tensor<640x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x1024xf16>
    %1661 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_146 : tensor<640x1024xf32>) outs(%1660 : tensor<640x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x1024xf16>
    %1662 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1663 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1662 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1664 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1663 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %1665 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%1664 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %1666 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1661 : tensor<640x1024xf16>) outs(%1249 : tensor<1024x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x640xf16>
    %collapsed_1002 = tensor.collapse_shape %1665 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %1667 = linalg.fill ins(%cst_690 : f16) outs(%1251 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %1668 = linalg.matmul ins(%collapsed_1002, %1666 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1667 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %1669 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1670 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1669 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1671 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1670 : tensor<f16>) outs(%1242 : tensor<640x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x1024xf16>
    %1672 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_147 : tensor<640x1024xf32>) outs(%1671 : tensor<640x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x1024xf16>
    %1673 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1674 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1673 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1675 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1674 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %1676 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%1675 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %1677 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1672 : tensor<640x1024xf16>) outs(%1249 : tensor<1024x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x640xf16>
    %collapsed_1003 = tensor.collapse_shape %1676 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %1678 = linalg.fill ins(%cst_690 : f16) outs(%1251 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %1679 = linalg.matmul ins(%collapsed_1003, %1677 : tensor<154x1024xf16>, tensor<1024x640xf16>) outs(%1678 : tensor<154x640xf16>) -> tensor<154x640xf16>
    %expanded_1004 = tensor.expand_shape %1657 [[0, 1], [2, 3]] : tensor<4608x640xf16> into tensor<2x2304x10x64xf16>
    %1680 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1004 : tensor<2x2304x10x64xf16>) outs(%1151 : tensor<2x10x2304x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x2304x64xf16>
    %collapsed_1005 = tensor.collapse_shape %1680 [[0, 1], [2], [3]] : tensor<2x10x2304x64xf16> into tensor<20x2304x64xf16>
    %expanded_1006 = tensor.expand_shape %1668 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
    %1681 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1006 : tensor<2x77x10x64xf16>) outs(%1266 : tensor<2x10x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x77x64xf16>
    %collapsed_1007 = tensor.collapse_shape %1681 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
    %expanded_1008 = tensor.expand_shape %1679 [[0, 1], [2, 3]] : tensor<154x640xf16> into tensor<2x77x10x64xf16>
    %1682 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1008 : tensor<2x77x10x64xf16>) outs(%1266 : tensor<2x10x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x10x77x64xf16>
    %collapsed_1009 = tensor.collapse_shape %1682 [[0, 1], [2], [3]] : tensor<2x10x77x64xf16> into tensor<20x77x64xf16>
    %expanded_1010 = tensor.expand_shape %collapsed_1005 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x64x1xf16>
    %1683 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1010 : tensor<20x2304x64x1xf16>) outs(%1155 : tensor<20x2304x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x64xf16>
    %expanded_1011 = tensor.expand_shape %collapsed_1007 [[0], [1], [2, 3]] : tensor<20x77x64xf16> into tensor<20x77x64x1xf16>
    %1684 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1011 : tensor<20x77x64x1xf16>) outs(%1270 : tensor<20x1x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x77x64xf16>
    %1685 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1683 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_1012 = tensor.collapse_shape %1685 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %1686 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1684 : tensor<20x1x77x64xf16>) outs(%1273 : tensor<20x64x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x64x77x1xf16>
    %collapsed_1013 = tensor.collapse_shape %1686 [[0], [1], [2, 3]] : tensor<20x64x77x1xf16> into tensor<20x64x77xf16>
    %1687 = linalg.fill ins(%cst_690 : f16) outs(%1275 : tensor<20x2304x77xf16>) -> tensor<20x2304x77xf16>
    %1688 = linalg.batch_matmul ins(%collapsed_1012, %collapsed_1013 : tensor<20x2304x64xf16>, tensor<20x64x77xf16>) outs(%1687 : tensor<20x2304x77xf16>) -> tensor<20x2304x77xf16>
    %expanded_1014 = tensor.expand_shape %1688 [[0], [1], [2, 3]] : tensor<20x2304x77xf16> into tensor<20x2304x1x77xf16>
    %1689 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1014 : tensor<20x2304x1x77xf16>) outs(%1278 : tensor<20x2304x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x77x1xf16>
    %collapsed_1015 = tensor.collapse_shape %1689 [[0], [1], [2, 3]] : tensor<20x2304x77x1xf16> into tensor<20x2304x77xf16>
    %1690 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1015, %cst : tensor<20x2304x77xf16>, tensor<f64>) outs(%1275 : tensor<20x2304x77xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<20x2304x77xf16>
    %1691 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1690 : tensor<20x2304x77xf16>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1692 = linalg.fill ins(%cst_692 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1693 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1691 : tensor<20x2304x77xf32>) outs(%1692 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1694 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1691, %1693 : tensor<20x2304x77xf32>, tensor<20x2304x1xf32>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1695 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1694 : tensor<20x2304x77xf32>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1696 = linalg.fill ins(%cst_697 : f32) outs(%1171 : tensor<20x2304x1xf32>) -> tensor<20x2304x1xf32>
    %1697 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1695 : tensor<20x2304x77xf32>) outs(%1696 : tensor<20x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x1xf32>
    %1698 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1695, %1697 : tensor<20x2304x77xf32>, tensor<20x2304x1xf32>) outs(%1281 : tensor<20x2304x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<20x2304x77xf32>
    %1699 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1700 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1699 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1701 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1700 : tensor<f16>) outs(%1275 : tensor<20x2304x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x77xf16>
    %1702 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1698 : tensor<20x2304x77xf32>) outs(%1701 : tensor<20x2304x77xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<20x2304x77xf16>
    %expanded_1016 = tensor.expand_shape %1702 [[0], [1], [2, 3]] : tensor<20x2304x77xf16> into tensor<20x2304x77x1xf16>
    %1703 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1016 : tensor<20x2304x77x1xf16>) outs(%1294 : tensor<20x2304x1x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x1x77xf16>
    %expanded_1017 = tensor.expand_shape %collapsed_1009 [[0], [1], [2, 3]] : tensor<20x77x64xf16> into tensor<20x77x64x1xf16>
    %1704 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1017 : tensor<20x77x64x1xf16>) outs(%1296 : tensor<20x1x64x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x1x64x77xf16>
    %1705 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1703 : tensor<20x2304x1x77xf16>) outs(%1278 : tensor<20x2304x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x77x1xf16>
    %collapsed_1018 = tensor.collapse_shape %1705 [[0], [1], [2, 3]] : tensor<20x2304x77x1xf16> into tensor<20x2304x77xf16>
    %1706 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1704 : tensor<20x1x64x77xf16>) outs(%1299 : tensor<20x77x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x77x64x1xf16>
    %collapsed_1019 = tensor.collapse_shape %1706 [[0], [1], [2, 3]] : tensor<20x77x64x1xf16> into tensor<20x77x64xf16>
    %1707 = linalg.fill ins(%cst_690 : f16) outs(%1189 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %1708 = linalg.batch_matmul ins(%collapsed_1018, %collapsed_1019 : tensor<20x2304x77xf16>, tensor<20x77x64xf16>) outs(%1707 : tensor<20x2304x64xf16>) -> tensor<20x2304x64xf16>
    %expanded_1020 = tensor.expand_shape %1708 [[0], [1], [2, 3]] : tensor<20x2304x64xf16> into tensor<20x2304x1x64xf16>
    %1709 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1020 : tensor<20x2304x1x64xf16>) outs(%1159 : tensor<20x2304x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<20x2304x64x1xf16>
    %collapsed_1021 = tensor.collapse_shape %1709 [[0], [1], [2, 3]] : tensor<20x2304x64x1xf16> into tensor<20x2304x64xf16>
    %expanded_1022 = tensor.expand_shape %collapsed_1021 [[0, 1], [2], [3]] : tensor<20x2304x64xf16> into tensor<2x10x2304x64xf16>
    %1710 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1022 : tensor<2x10x2304x64xf16>) outs(%1193 : tensor<2x2304x10x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x10x64xf16>
    %1711 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1712 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1711 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1713 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1712 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1714 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_148 : tensor<640xf32>) outs(%1713 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1715 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1716 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1715 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1717 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1716 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1718 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_149 : tensor<640x640xf32>) outs(%1717 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1719 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1718 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_1023 = tensor.collapse_shape %1710 [[0, 1], [2, 3]] : tensor<2x2304x10x64xf16> into tensor<4608x640xf16>
    %1720 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1721 = linalg.matmul ins(%collapsed_1023, %1719 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1720 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1722 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1714, %1721 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_1024 = tensor.expand_shape %1722 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1723 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1024, %1625 : tensor<2x2304x640xf16>, tensor<2x2304x640xf16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1724 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1725 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1724 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1726 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1725 : tensor<f32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1727 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1723 : tensor<2x2304x640xf16>) outs(%1726 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1728 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1729 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1727 : tensor<2x2304x640xf32>) outs(%1728 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1730 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1729 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1731 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1730 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1732 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1731 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1733 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1727, %1732 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1734 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1733 : tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1735 = linalg.fill ins(%cst_695 : f64) outs(%1099 : tensor<2x2304x1xf64>) -> tensor<2x2304x1xf64>
    %1736 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1734 : tensor<2x2304x640xf32>) outs(%1735 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x2304x1xf64>
    %1737 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1736 : tensor<2x2304x1xf64>) outs(%1099 : tensor<2x2304x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_703 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x2304x1xf64>
    %1738 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1737 : tensor<2x2304x1xf64>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1739 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1738 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x2304x1xf32>
    %1740 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1739 : tensor<2x2304x1xf32>) outs(%1103 : tensor<2x2304x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x1xf32>
    %1741 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1740 : tensor<2x2304x1xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x2304x640xf32>
    %1742 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1733, %1741 : tensor<2x2304x640xf32>, tensor<2x2304x640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1743 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1742, %cst_150 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1744 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1743, %cst_151 : tensor<2x2304x640xf32>, tensor<640xf32>) outs(%1096 : tensor<2x2304x640xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x2304x640xf32>
    %1745 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1746 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1745 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1747 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1746 : tensor<f16>) outs(%1341 : tensor<5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<5120xf16>
    %1748 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_152 : tensor<5120xf32>) outs(%1747 : tensor<5120xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<5120xf16>
    %1749 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1750 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1749 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1751 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1750 : tensor<f16>) outs(%1346 : tensor<5120x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<5120x640xf16>
    %1752 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_153 : tensor<5120x640xf32>) outs(%1751 : tensor<5120x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<5120x640xf16>
    %1753 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1754 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1753 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1755 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1754 : tensor<f16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x2304x640xf16>
    %1756 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1744 : tensor<2x2304x640xf32>) outs(%1755 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1757 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1752 : tensor<5120x640xf16>) outs(%1353 : tensor<640x5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x5120xf16>
    %collapsed_1025 = tensor.collapse_shape %1756 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1758 = linalg.fill ins(%cst_690 : f16) outs(%1355 : tensor<4608x5120xf16>) -> tensor<4608x5120xf16>
    %1759 = linalg.matmul ins(%collapsed_1025, %1757 : tensor<4608x640xf16>, tensor<640x5120xf16>) outs(%1758 : tensor<4608x5120xf16>) -> tensor<4608x5120xf16>
    %1760 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1748, %1759 : tensor<5120xf16>, tensor<4608x5120xf16>) outs(%1355 : tensor<4608x5120xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x5120xf16>
    %expanded_1026 = tensor.expand_shape %1760 [[0, 1], [2]] : tensor<4608x5120xf16> into tensor<2x2304x5120xf16>
    %extracted_slice_1027 = tensor.extract_slice %expanded_1026[0, 0, 0] [2, 2304, 2560] [1, 1, 1] : tensor<2x2304x5120xf16> to tensor<2x2304x2560xf16>
    %extracted_slice_1028 = tensor.extract_slice %expanded_1026[0, 0, 2560] [2, 2304, 2560] [1, 1, 1] : tensor<2x2304x5120xf16> to tensor<2x2304x2560xf16>
    %1761 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1028 : tensor<2x2304x2560xf16>) outs(%1359 : tensor<2x2304x2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = math.sqrt %cst_693 : f16
      %7525 = arith.divf %in, %7524 : f16
      %7526 = math.erf %7525 : f16
      %7527 = arith.addf %7526, %cst_691 : f16
      %7528 = arith.mulf %7527, %cst_694 : f16
      %7529 = arith.mulf %in, %7528 : f16
      linalg.yield %7529 : f16
    } -> tensor<2x2304x2560xf16>
    %1762 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1027, %1761 : tensor<2x2304x2560xf16>, tensor<2x2304x2560xf16>) outs(%1359 : tensor<2x2304x2560xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x2560xf16>
    %1763 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1764 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1763 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1765 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1764 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1766 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_154 : tensor<640xf32>) outs(%1765 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1767 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1768 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1767 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1769 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1768 : tensor<f16>) outs(%1368 : tensor<640x2560xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x2560xf16>
    %1770 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_155 : tensor<640x2560xf32>) outs(%1769 : tensor<640x2560xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x2560xf16>
    %1771 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1770 : tensor<640x2560xf16>) outs(%1371 : tensor<2560x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2560x640xf16>
    %collapsed_1029 = tensor.collapse_shape %1762 [[0, 1], [2]] : tensor<2x2304x2560xf16> into tensor<4608x2560xf16>
    %1772 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1773 = linalg.matmul ins(%collapsed_1029, %1771 : tensor<4608x2560xf16>, tensor<2560x640xf16>) outs(%1772 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1774 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1766, %1773 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_1030 = tensor.expand_shape %1774 [[0, 1], [2]] : tensor<4608x640xf16> into tensor<2x2304x640xf16>
    %1775 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1030, %1723 : tensor<2x2304x640xf16>, tensor<2x2304x640xf16>) outs(%1086 : tensor<2x2304x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x2304x640xf16>
    %1776 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1777 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1776 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1778 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1777 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1779 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_156 : tensor<640xf32>) outs(%1778 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1780 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1781 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1780 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1782 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1781 : tensor<f16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %1783 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_157 : tensor<640x640xf32>) outs(%1782 : tensor<640x640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640xf16>
    %1784 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1783 : tensor<640x640xf16>) outs(%1081 : tensor<640x640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640xf16>
    %collapsed_1031 = tensor.collapse_shape %1775 [[0, 1], [2]] : tensor<2x2304x640xf16> into tensor<4608x640xf16>
    %1785 = linalg.fill ins(%cst_690 : f16) outs(%1090 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1786 = linalg.matmul ins(%collapsed_1031, %1784 : tensor<4608x640xf16>, tensor<640x640xf16>) outs(%1785 : tensor<4608x640xf16>) -> tensor<4608x640xf16>
    %1787 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1779, %1786 : tensor<640xf16>, tensor<4608x640xf16>) outs(%1090 : tensor<4608x640xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<4608x640xf16>
    %expanded_1032 = tensor.expand_shape %1787 [[0, 1, 2], [3]] : tensor<4608x640xf16> into tensor<2x48x48x640xf16>
    %1788 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1032 : tensor<2x48x48x640xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x48x48xf16>
    %1789 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1788, %1486 : tensor<2x640x48x48xf16>, tensor<2x640x48x48xf16>) outs(%970 : tensor<2x640x48x48xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x48x48xf16>
    %1790 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1791 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1790 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1792 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1791 : tensor<f16>) outs(%962 : tensor<640xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640xf16>
    %1793 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_158 : tensor<640xf32>) outs(%1792 : tensor<640xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640xf16>
    %1794 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1795 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1794 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1796 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1795 : tensor<f16>) outs(%1030 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<640x640x3x3xf16>
    %1797 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_159 : tensor<640x640x3x3xf32>) outs(%1796 : tensor<640x640x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<640x640x3x3xf16>
    %padded_1033 = tensor.pad %1789 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x640x48x48xf16> to tensor<2x640x50x50xf16>
    %1798 = tensor.empty() : tensor<2x640x24x24xf16>
    %1799 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1793 : tensor<640xf16>) outs(%1798 : tensor<2x640x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x24x24xf16>
    %1800 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1033, %1797 : tensor<2x640x50x50xf16>, tensor<640x640x3x3xf16>) outs(%1799 : tensor<2x640x24x24xf16>) -> tensor<2x640x24x24xf16>
    %1801 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1802 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1801 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1803 = tensor.empty() : tensor<2x640x24x24xf32>
    %1804 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1802 : tensor<f32>) outs(%1803 : tensor<2x640x24x24xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x640x24x24xf32>
    %1805 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1800 : tensor<2x640x24x24xf16>) outs(%1804 : tensor<2x640x24x24xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x24x24xf32>
    %collapsed_1034 = tensor.collapse_shape %1805 [[0], [1], [2, 3]] : tensor<2x640x24x24xf32> into tensor<2x640x576xf32>
    %expanded_1035 = tensor.expand_shape %collapsed_1034 [[0], [1, 2], [3]] : tensor<2x640x576xf32> into tensor<2x32x20x576xf32>
    %1806 = tensor.empty() : tensor<2x32x20x576xf64>
    %1807 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1035 : tensor<2x32x20x576xf32>) outs(%1806 : tensor<2x32x20x576xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x576xf64>
    %1808 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1809 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1807 : tensor<2x32x20x576xf64>) outs(%1808 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1810 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1809 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_704 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1811 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1807, %1810 : tensor<2x32x20x576xf64>, tensor<2x32x1x1xf64>) outs(%1806 : tensor<2x32x20x576xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x576xf64>
    %1812 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1811 : tensor<2x32x20x576xf64>) outs(%1806 : tensor<2x32x20x576xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x20x576xf64>
    %1813 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1814 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1812 : tensor<2x32x20x576xf64>) outs(%1813 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1815 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1814 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_704 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1816 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1815 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1817 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1818 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1035 : tensor<2x32x20x576xf32>) outs(%1817 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1819 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1818 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_704 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1820 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1819 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1821 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1816, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1822 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1821 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1823 = tensor.empty() : tensor<2x32x20x576xf32>
    %1824 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1035, %1820 : tensor<2x32x20x576xf32>, tensor<2x32x1x1xf32>) outs(%1823 : tensor<2x32x20x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x576xf32>
    %1825 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1824, %1822 : tensor<2x32x20x576xf32>, tensor<2x32x1x1xf32>) outs(%1823 : tensor<2x32x20x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x20x576xf32>
    %collapsed_1036 = tensor.collapse_shape %1825 [[0], [1, 2], [3]] : tensor<2x32x20x576xf32> into tensor<2x640x576xf32>
    %expanded_1037 = tensor.expand_shape %collapsed_1036 [[0], [1], [2, 3]] : tensor<2x640x576xf32> into tensor<2x640x24x24xf32>
    %expanded_1038 = tensor.expand_shape %cst_160 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %expanded_1039 = tensor.expand_shape %cst_161 [[0, 1, 2, 3]] : tensor<640xf32> into tensor<1x640x1x1xf32>
    %1826 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1037, %expanded_1039 : tensor<2x640x24x24xf32>, tensor<1x640x1x1xf32>) outs(%1803 : tensor<2x640x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x24x24xf32>
    %1827 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1826, %expanded_1038 : tensor<2x640x24x24xf32>, tensor<1x640x1x1xf32>) outs(%1803 : tensor<2x640x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x640x24x24xf32>
    %1828 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1829 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1828 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1830 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1829 : tensor<f16>) outs(%1798 : tensor<2x640x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x640x24x24xf16>
    %1831 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1827 : tensor<2x640x24x24xf32>) outs(%1830 : tensor<2x640x24x24xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x24x24xf16>
    %1832 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1831 : tensor<2x640x24x24xf16>) outs(%1798 : tensor<2x640x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x640x24x24xf16>
    %1833 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1832, %1831 : tensor<2x640x24x24xf16>, tensor<2x640x24x24xf16>) outs(%1798 : tensor<2x640x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x640x24x24xf16>
    %1834 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1835 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1834 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1836 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1835 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %1837 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_162 : tensor<1280xf32>) outs(%1836 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %1838 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1839 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1838 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1840 = tensor.empty() : tensor<1280x640x3x3xf16>
    %1841 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1839 : tensor<f16>) outs(%1840 : tensor<1280x640x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x640x3x3xf16>
    %1842 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_163 : tensor<1280x640x3x3xf32>) outs(%1841 : tensor<1280x640x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x640x3x3xf16>
    %padded_1040 = tensor.pad %1833 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x640x24x24xf16> to tensor<2x640x26x26xf16>
    %1843 = tensor.empty() : tensor<2x1280x24x24xf16>
    %1844 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1837 : tensor<1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %1845 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1040, %1842 : tensor<2x640x26x26xf16>, tensor<1280x640x3x3xf16>) outs(%1844 : tensor<2x1280x24x24xf16>) -> tensor<2x1280x24x24xf16>
    %1846 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %1847 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1846, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %1848 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1849 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1848 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1850 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1849 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %1851 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_164 : tensor<1280xf32>) outs(%1850 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %1852 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1853 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1852 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1854 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1853 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %1855 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_165 : tensor<1280x1280xf32>) outs(%1854 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %1856 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1855 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %1857 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %1858 = linalg.matmul ins(%1847, %1856 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%1857 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %1859 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1851, %1858 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %expanded_1041 = tensor.expand_shape %1859 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
    %1860 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1845, %expanded_1041 : tensor<2x1280x24x24xf16>, tensor<2x1280x1x1xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %1861 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1862 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1861 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1863 = tensor.empty() : tensor<2x1280x24x24xf32>
    %1864 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1862 : tensor<f32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x24x24xf32>
    %1865 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1860 : tensor<2x1280x24x24xf16>) outs(%1864 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %collapsed_1042 = tensor.collapse_shape %1865 [[0], [1], [2, 3]] : tensor<2x1280x24x24xf32> into tensor<2x1280x576xf32>
    %expanded_1043 = tensor.expand_shape %collapsed_1042 [[0], [1, 2], [3]] : tensor<2x1280x576xf32> into tensor<2x32x40x576xf32>
    %1866 = tensor.empty() : tensor<2x32x40x576xf64>
    %1867 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1043 : tensor<2x32x40x576xf32>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %1868 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1869 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1867 : tensor<2x32x40x576xf64>) outs(%1868 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1870 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1869 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1871 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1867, %1870 : tensor<2x32x40x576xf64>, tensor<2x32x1x1xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %1872 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1871 : tensor<2x32x40x576xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %1873 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1874 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1872 : tensor<2x32x40x576xf64>) outs(%1873 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1875 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1874 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1876 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1875 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1877 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1878 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1043 : tensor<2x32x40x576xf32>) outs(%1877 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1879 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1878 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1880 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1879 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1881 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1876, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1882 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1881 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1883 = tensor.empty() : tensor<2x32x40x576xf32>
    %1884 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1043, %1880 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %1885 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1884, %1882 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %collapsed_1044 = tensor.collapse_shape %1885 [[0], [1, 2], [3]] : tensor<2x32x40x576xf32> into tensor<2x1280x576xf32>
    %expanded_1045 = tensor.expand_shape %collapsed_1044 [[0], [1], [2, 3]] : tensor<2x1280x576xf32> into tensor<2x1280x24x24xf32>
    %expanded_1046 = tensor.expand_shape %cst_166 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1047 = tensor.expand_shape %cst_167 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %1886 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1045, %expanded_1047 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %1887 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1886, %expanded_1046 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %1888 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1889 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1888 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1890 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1889 : tensor<f16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %1891 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1887 : tensor<2x1280x24x24xf32>) outs(%1890 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %1892 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1891 : tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x24x24xf16>
    %1893 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1892, %1891 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %1894 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1895 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1894 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1896 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1895 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %1897 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_168 : tensor<1280xf32>) outs(%1896 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %1898 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1899 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1898 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1900 = tensor.empty() : tensor<1280x1280x3x3xf16>
    %1901 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1899 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %1902 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_169 : tensor<1280x1280x3x3xf32>) outs(%1901 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1048 = tensor.pad %1893 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x24x24xf16> to tensor<2x1280x26x26xf16>
    %1903 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1897 : tensor<1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %1904 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1048, %1902 : tensor<2x1280x26x26xf16>, tensor<1280x1280x3x3xf16>) outs(%1903 : tensor<2x1280x24x24xf16>) -> tensor<2x1280x24x24xf16>
    %1905 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1906 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1905 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1907 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1906 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %1908 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_170 : tensor<1280xf32>) outs(%1907 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %1909 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1910 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1909 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1911 = tensor.empty() : tensor<1280x640x1x1xf16>
    %1912 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1910 : tensor<f16>) outs(%1911 : tensor<1280x640x1x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x640x1x1xf16>
    %1913 = tensor.empty() : tensor<1280x640x1x1xf32>
    %1914 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_171 : tensor<1280x640x1x1xf32>) outs(%1913 : tensor<1280x640x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1280x640x1x1xf32>
    %1915 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1914 : tensor<1280x640x1x1xf32>) outs(%1912 : tensor<1280x640x1x1xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x640x1x1xf16>
    %1916 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1908 : tensor<1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %1917 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1800, %1915 : tensor<2x640x24x24xf16>, tensor<1280x640x1x1xf16>) outs(%1916 : tensor<2x1280x24x24xf16>) -> tensor<2x1280x24x24xf16>
    %1918 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1917, %1904 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %1919 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1920 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1919 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1921 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1920 : tensor<f32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x24x24xf32>
    %1922 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1918 : tensor<2x1280x24x24xf16>) outs(%1921 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %collapsed_1049 = tensor.collapse_shape %1922 [[0], [1], [2, 3]] : tensor<2x1280x24x24xf32> into tensor<2x1280x576xf32>
    %expanded_1050 = tensor.expand_shape %collapsed_1049 [[0], [1, 2], [3]] : tensor<2x1280x576xf32> into tensor<2x32x40x576xf32>
    %1923 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1050 : tensor<2x32x40x576xf32>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %1924 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1925 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1923 : tensor<2x32x40x576xf64>) outs(%1924 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1926 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1925 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1927 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1923, %1926 : tensor<2x32x40x576xf64>, tensor<2x32x1x1xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %1928 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1927 : tensor<2x32x40x576xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %1929 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1930 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1928 : tensor<2x32x40x576xf64>) outs(%1929 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1931 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1930 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1932 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1931 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1933 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %1934 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1050 : tensor<2x32x40x576xf32>) outs(%1933 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %1935 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1934 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %1936 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1935 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1937 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1932, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %1938 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1937 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %1939 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1050, %1936 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %1940 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1939, %1938 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %collapsed_1051 = tensor.collapse_shape %1940 [[0], [1, 2], [3]] : tensor<2x32x40x576xf32> into tensor<2x1280x576xf32>
    %expanded_1052 = tensor.expand_shape %collapsed_1051 [[0], [1], [2, 3]] : tensor<2x1280x576xf32> into tensor<2x1280x24x24xf32>
    %expanded_1053 = tensor.expand_shape %cst_172 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1054 = tensor.expand_shape %cst_173 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %1941 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1052, %expanded_1054 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %1942 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1941, %expanded_1053 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %1943 = tensor.empty() : tensor<2x24x24x1280xf32>
    %1944 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1942 : tensor<2x1280x24x24xf32>) outs(%1943 : tensor<2x24x24x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x24x24x1280xf32>
    %collapsed_1055 = tensor.collapse_shape %1944 [[0], [1, 2], [3]] : tensor<2x24x24x1280xf32> into tensor<2x576x1280xf32>
    %1945 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1946 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1945 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1947 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%1946 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %1948 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_174 : tensor<1280xf32>) outs(%1947 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %1949 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1950 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1949 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1951 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1950 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %1952 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_175 : tensor<1280x1280xf32>) outs(%1951 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %1953 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1954 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1953 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1955 = tensor.empty() : tensor<2x576x1280xf16>
    %1956 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1954 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %1957 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1055 : tensor<2x576x1280xf32>) outs(%1956 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %1958 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1952 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1056 = tensor.collapse_shape %1957 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %1959 = tensor.empty() : tensor<1152x1280xf16>
    %1960 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %1961 = linalg.matmul ins(%collapsed_1056, %1958 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%1960 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %1962 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1948, %1961 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1057 = tensor.expand_shape %1962 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %1963 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1964 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1963 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %1965 = tensor.empty() : tensor<2x576x1280xf32>
    %1966 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1964 : tensor<f32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %1967 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1057 : tensor<2x576x1280xf16>) outs(%1966 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %1968 = tensor.empty() : tensor<2x576x1xf64>
    %1969 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %1970 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1967 : tensor<2x576x1280xf32>) outs(%1969 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %1971 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1970 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %1972 = tensor.empty() : tensor<2x576x1xf32>
    %1973 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1971 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %1974 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1973 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %1975 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1967, %1974 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %1976 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1975 : tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %1977 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %1978 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1976 : tensor<2x576x1280xf32>) outs(%1977 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %1979 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1978 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %1980 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1979 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %1981 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1980 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x576x1xf32>
    %1982 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1981 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %1983 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1982 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %1984 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1975, %1983 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %1985 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1984, %cst_176 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %1986 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1985, %cst_177 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %1987 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1988 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1987 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1989 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1988 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %1990 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_178 : tensor<1280x1280xf32>) outs(%1989 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %1991 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1992 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1991 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %1993 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1992 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %1994 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1986 : tensor<2x576x1280xf32>) outs(%1993 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %1995 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%1990 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1058 = tensor.collapse_shape %1994 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %1996 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %1997 = linalg.matmul ins(%collapsed_1058, %1995 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%1996 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %1998 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %1999 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%1998 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2000 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%1999 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2001 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_179 : tensor<1280x1280xf32>) outs(%2000 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2002 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2003 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2002 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2004 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2003 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2005 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1986 : tensor<2x576x1280xf32>) outs(%2004 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2006 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2001 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1059 = tensor.collapse_shape %2005 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2007 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2008 = linalg.matmul ins(%collapsed_1059, %2006 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2007 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2009 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2010 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2009 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2011 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2010 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2012 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_180 : tensor<1280x1280xf32>) outs(%2011 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2013 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2014 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2013 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2015 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2014 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2016 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1986 : tensor<2x576x1280xf32>) outs(%2015 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2017 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2012 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1060 = tensor.collapse_shape %2016 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2018 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2019 = linalg.matmul ins(%collapsed_1060, %2017 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2018 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %expanded_1061 = tensor.expand_shape %1997 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2020 = tensor.empty() : tensor<2x20x576x64xf16>
    %2021 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1061 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1062 = tensor.collapse_shape %2021 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1063 = tensor.expand_shape %2008 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2022 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1063 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1064 = tensor.collapse_shape %2022 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1065 = tensor.expand_shape %2019 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2023 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1065 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1066 = tensor.collapse_shape %2023 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1067 = tensor.expand_shape %collapsed_1062 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2024 = tensor.empty() : tensor<40x576x1x64xf16>
    %2025 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1067 : tensor<40x576x64x1xf16>) outs(%2024 : tensor<40x576x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x64xf16>
    %expanded_1068 = tensor.expand_shape %collapsed_1064 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2026 = tensor.empty() : tensor<40x1x576x64xf16>
    %2027 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1068 : tensor<40x576x64x1xf16>) outs(%2026 : tensor<40x1x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x576x64xf16>
    %2028 = tensor.empty() : tensor<40x576x64x1xf16>
    %2029 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2025 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1069 = tensor.collapse_shape %2029 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %2030 = tensor.empty() : tensor<40x64x576x1xf16>
    %2031 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2027 : tensor<40x1x576x64xf16>) outs(%2030 : tensor<40x64x576x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x64x576x1xf16>
    %collapsed_1070 = tensor.collapse_shape %2031 [[0], [1], [2, 3]] : tensor<40x64x576x1xf16> into tensor<40x64x576xf16>
    %2032 = tensor.empty() : tensor<40x576x576xf16>
    %2033 = linalg.fill ins(%cst_690 : f16) outs(%2032 : tensor<40x576x576xf16>) -> tensor<40x576x576xf16>
    %2034 = linalg.batch_matmul ins(%collapsed_1069, %collapsed_1070 : tensor<40x576x64xf16>, tensor<40x64x576xf16>) outs(%2033 : tensor<40x576x576xf16>) -> tensor<40x576x576xf16>
    %expanded_1071 = tensor.expand_shape %2034 [[0], [1], [2, 3]] : tensor<40x576x576xf16> into tensor<40x576x1x576xf16>
    %2035 = tensor.empty() : tensor<40x576x576x1xf16>
    %2036 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1071 : tensor<40x576x1x576xf16>) outs(%2035 : tensor<40x576x576x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x576x1xf16>
    %collapsed_1072 = tensor.collapse_shape %2036 [[0], [1], [2, 3]] : tensor<40x576x576x1xf16> into tensor<40x576x576xf16>
    %2037 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1072, %cst : tensor<40x576x576xf16>, tensor<f64>) outs(%2032 : tensor<40x576x576xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<40x576x576xf16>
    %2038 = tensor.empty() : tensor<40x576x576xf32>
    %2039 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2037 : tensor<40x576x576xf16>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2040 = tensor.empty() : tensor<40x576x1xf32>
    %2041 = linalg.fill ins(%cst_692 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2042 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2039 : tensor<40x576x576xf32>) outs(%2041 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2043 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2039, %2042 : tensor<40x576x576xf32>, tensor<40x576x1xf32>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2044 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2043 : tensor<40x576x576xf32>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2045 = linalg.fill ins(%cst_697 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2046 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2044 : tensor<40x576x576xf32>) outs(%2045 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2047 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2044, %2046 : tensor<40x576x576xf32>, tensor<40x576x1xf32>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2048 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2049 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2048 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2050 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2049 : tensor<f16>) outs(%2032 : tensor<40x576x576xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x576xf16>
    %2051 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2047 : tensor<40x576x576xf32>) outs(%2050 : tensor<40x576x576xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<40x576x576xf16>
    %expanded_1073 = tensor.expand_shape %2051 [[0], [1], [2, 3]] : tensor<40x576x576xf16> into tensor<40x576x576x1xf16>
    %2052 = tensor.empty() : tensor<40x576x1x576xf16>
    %2053 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1073 : tensor<40x576x576x1xf16>) outs(%2052 : tensor<40x576x1x576xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x576xf16>
    %expanded_1074 = tensor.expand_shape %collapsed_1066 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2054 = tensor.empty() : tensor<40x1x64x576xf16>
    %2055 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1074 : tensor<40x576x64x1xf16>) outs(%2054 : tensor<40x1x64x576xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x64x576xf16>
    %2056 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2053 : tensor<40x576x1x576xf16>) outs(%2035 : tensor<40x576x576x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x576x1xf16>
    %collapsed_1075 = tensor.collapse_shape %2056 [[0], [1], [2, 3]] : tensor<40x576x576x1xf16> into tensor<40x576x576xf16>
    %2057 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2055 : tensor<40x1x64x576xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1076 = tensor.collapse_shape %2057 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %2058 = tensor.empty() : tensor<40x576x64xf16>
    %2059 = linalg.fill ins(%cst_690 : f16) outs(%2058 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %2060 = linalg.batch_matmul ins(%collapsed_1075, %collapsed_1076 : tensor<40x576x576xf16>, tensor<40x576x64xf16>) outs(%2059 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %expanded_1077 = tensor.expand_shape %2060 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x1x64xf16>
    %2061 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1077 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1078 = tensor.collapse_shape %2061 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %expanded_1079 = tensor.expand_shape %collapsed_1078 [[0, 1], [2], [3]] : tensor<40x576x64xf16> into tensor<2x20x576x64xf16>
    %2062 = tensor.empty() : tensor<2x576x20x64xf16>
    %2063 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1079 : tensor<2x20x576x64xf16>) outs(%2062 : tensor<2x576x20x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x20x64xf16>
    %2064 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2065 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2064 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2066 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2065 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2067 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_181 : tensor<1280xf32>) outs(%2066 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2068 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2069 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2068 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2070 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2069 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2071 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_182 : tensor<1280x1280xf32>) outs(%2070 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2072 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2071 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1080 = tensor.collapse_shape %2063 [[0, 1], [2, 3]] : tensor<2x576x20x64xf16> into tensor<1152x1280xf16>
    %2073 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2074 = linalg.matmul ins(%collapsed_1080, %2072 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2073 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2075 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2067, %2074 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1081 = tensor.expand_shape %2075 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2076 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1081, %expanded_1057 : tensor<2x576x1280xf16>, tensor<2x576x1280xf16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2077 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2078 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2077 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2079 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2078 : tensor<f32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2080 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2076 : tensor<2x576x1280xf16>) outs(%2079 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2081 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2082 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2080 : tensor<2x576x1280xf32>) outs(%2081 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2083 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2082 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2084 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2083 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2085 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2084 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2086 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2080, %2085 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2087 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2086 : tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2088 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2089 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2087 : tensor<2x576x1280xf32>) outs(%2088 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2090 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2089 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2091 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2090 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2092 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2091 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x576x1xf32>
    %2093 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2092 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2094 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2093 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2095 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2086, %2094 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2096 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2095, %cst_183 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2097 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2096, %cst_184 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2098 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2099 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2098 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2100 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2099 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2101 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_185 : tensor<1280x1280xf32>) outs(%2100 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2102 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2103 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2102 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2104 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2103 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2105 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2097 : tensor<2x576x1280xf32>) outs(%2104 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2106 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2101 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1082 = tensor.collapse_shape %2105 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2107 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2108 = linalg.matmul ins(%collapsed_1082, %2106 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2107 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2109 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2110 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2109 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2111 = tensor.empty() : tensor<1280x1024xf16>
    %2112 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2110 : tensor<f16>) outs(%2111 : tensor<1280x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1024xf16>
    %2113 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_186 : tensor<1280x1024xf32>) outs(%2112 : tensor<1280x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1024xf16>
    %2114 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2115 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2114 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2116 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2115 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %2117 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%2116 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %2118 = tensor.empty() : tensor<1024x1280xf16>
    %2119 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2113 : tensor<1280x1024xf16>) outs(%2118 : tensor<1024x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x1280xf16>
    %collapsed_1083 = tensor.collapse_shape %2117 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %2120 = tensor.empty() : tensor<154x1280xf16>
    %2121 = linalg.fill ins(%cst_690 : f16) outs(%2120 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %2122 = linalg.matmul ins(%collapsed_1083, %2119 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%2121 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %2123 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2124 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2123 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2125 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2124 : tensor<f16>) outs(%2111 : tensor<1280x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1024xf16>
    %2126 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_187 : tensor<1280x1024xf32>) outs(%2125 : tensor<1280x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1024xf16>
    %2127 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2128 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2127 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2129 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2128 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %2130 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%2129 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %2131 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2126 : tensor<1280x1024xf16>) outs(%2118 : tensor<1024x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x1280xf16>
    %collapsed_1084 = tensor.collapse_shape %2130 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %2132 = linalg.fill ins(%cst_690 : f16) outs(%2120 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %2133 = linalg.matmul ins(%collapsed_1084, %2131 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%2132 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %expanded_1085 = tensor.expand_shape %2108 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2134 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1085 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1086 = tensor.collapse_shape %2134 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1087 = tensor.expand_shape %2122 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
    %2135 = tensor.empty() : tensor<2x20x77x64xf16>
    %2136 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1087 : tensor<2x77x20x64xf16>) outs(%2135 : tensor<2x20x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x77x64xf16>
    %collapsed_1088 = tensor.collapse_shape %2136 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
    %expanded_1089 = tensor.expand_shape %2133 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
    %2137 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1089 : tensor<2x77x20x64xf16>) outs(%2135 : tensor<2x20x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x77x64xf16>
    %collapsed_1090 = tensor.collapse_shape %2137 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
    %expanded_1091 = tensor.expand_shape %collapsed_1086 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2138 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1091 : tensor<40x576x64x1xf16>) outs(%2024 : tensor<40x576x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x64xf16>
    %expanded_1092 = tensor.expand_shape %collapsed_1088 [[0], [1], [2, 3]] : tensor<40x77x64xf16> into tensor<40x77x64x1xf16>
    %2139 = tensor.empty() : tensor<40x1x77x64xf16>
    %2140 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1092 : tensor<40x77x64x1xf16>) outs(%2139 : tensor<40x1x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x77x64xf16>
    %2141 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2138 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1093 = tensor.collapse_shape %2141 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %2142 = tensor.empty() : tensor<40x64x77x1xf16>
    %2143 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2140 : tensor<40x1x77x64xf16>) outs(%2142 : tensor<40x64x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x64x77x1xf16>
    %collapsed_1094 = tensor.collapse_shape %2143 [[0], [1], [2, 3]] : tensor<40x64x77x1xf16> into tensor<40x64x77xf16>
    %2144 = tensor.empty() : tensor<40x576x77xf16>
    %2145 = linalg.fill ins(%cst_690 : f16) outs(%2144 : tensor<40x576x77xf16>) -> tensor<40x576x77xf16>
    %2146 = linalg.batch_matmul ins(%collapsed_1093, %collapsed_1094 : tensor<40x576x64xf16>, tensor<40x64x77xf16>) outs(%2145 : tensor<40x576x77xf16>) -> tensor<40x576x77xf16>
    %expanded_1095 = tensor.expand_shape %2146 [[0], [1], [2, 3]] : tensor<40x576x77xf16> into tensor<40x576x1x77xf16>
    %2147 = tensor.empty() : tensor<40x576x77x1xf16>
    %2148 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1095 : tensor<40x576x1x77xf16>) outs(%2147 : tensor<40x576x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x77x1xf16>
    %collapsed_1096 = tensor.collapse_shape %2148 [[0], [1], [2, 3]] : tensor<40x576x77x1xf16> into tensor<40x576x77xf16>
    %2149 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1096, %cst : tensor<40x576x77xf16>, tensor<f64>) outs(%2144 : tensor<40x576x77xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<40x576x77xf16>
    %2150 = tensor.empty() : tensor<40x576x77xf32>
    %2151 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2149 : tensor<40x576x77xf16>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2152 = linalg.fill ins(%cst_692 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2153 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2151 : tensor<40x576x77xf32>) outs(%2152 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2154 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2151, %2153 : tensor<40x576x77xf32>, tensor<40x576x1xf32>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2155 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2154 : tensor<40x576x77xf32>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2156 = linalg.fill ins(%cst_697 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2157 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2155 : tensor<40x576x77xf32>) outs(%2156 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2158 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2155, %2157 : tensor<40x576x77xf32>, tensor<40x576x1xf32>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2159 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2160 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2159 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2161 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2160 : tensor<f16>) outs(%2144 : tensor<40x576x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x77xf16>
    %2162 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2158 : tensor<40x576x77xf32>) outs(%2161 : tensor<40x576x77xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<40x576x77xf16>
    %expanded_1097 = tensor.expand_shape %2162 [[0], [1], [2, 3]] : tensor<40x576x77xf16> into tensor<40x576x77x1xf16>
    %2163 = tensor.empty() : tensor<40x576x1x77xf16>
    %2164 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1097 : tensor<40x576x77x1xf16>) outs(%2163 : tensor<40x576x1x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x77xf16>
    %expanded_1098 = tensor.expand_shape %collapsed_1090 [[0], [1], [2, 3]] : tensor<40x77x64xf16> into tensor<40x77x64x1xf16>
    %2165 = tensor.empty() : tensor<40x1x64x77xf16>
    %2166 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1098 : tensor<40x77x64x1xf16>) outs(%2165 : tensor<40x1x64x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x64x77xf16>
    %2167 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2164 : tensor<40x576x1x77xf16>) outs(%2147 : tensor<40x576x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x77x1xf16>
    %collapsed_1099 = tensor.collapse_shape %2167 [[0], [1], [2, 3]] : tensor<40x576x77x1xf16> into tensor<40x576x77xf16>
    %2168 = tensor.empty() : tensor<40x77x64x1xf16>
    %2169 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2166 : tensor<40x1x64x77xf16>) outs(%2168 : tensor<40x77x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x77x64x1xf16>
    %collapsed_1100 = tensor.collapse_shape %2169 [[0], [1], [2, 3]] : tensor<40x77x64x1xf16> into tensor<40x77x64xf16>
    %2170 = linalg.fill ins(%cst_690 : f16) outs(%2058 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %2171 = linalg.batch_matmul ins(%collapsed_1099, %collapsed_1100 : tensor<40x576x77xf16>, tensor<40x77x64xf16>) outs(%2170 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %expanded_1101 = tensor.expand_shape %2171 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x1x64xf16>
    %2172 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1101 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1102 = tensor.collapse_shape %2172 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %expanded_1103 = tensor.expand_shape %collapsed_1102 [[0, 1], [2], [3]] : tensor<40x576x64xf16> into tensor<2x20x576x64xf16>
    %2173 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1103 : tensor<2x20x576x64xf16>) outs(%2062 : tensor<2x576x20x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x20x64xf16>
    %2174 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2175 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2174 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2176 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2175 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2177 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_188 : tensor<1280xf32>) outs(%2176 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2178 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2179 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2178 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2180 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2179 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2181 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_189 : tensor<1280x1280xf32>) outs(%2180 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2182 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2181 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1104 = tensor.collapse_shape %2173 [[0, 1], [2, 3]] : tensor<2x576x20x64xf16> into tensor<1152x1280xf16>
    %2183 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2184 = linalg.matmul ins(%collapsed_1104, %2182 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2183 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2185 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2177, %2184 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1105 = tensor.expand_shape %2185 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2186 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1105, %2076 : tensor<2x576x1280xf16>, tensor<2x576x1280xf16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2187 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2188 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2187 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2189 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2188 : tensor<f32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2190 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2186 : tensor<2x576x1280xf16>) outs(%2189 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2191 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2192 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2190 : tensor<2x576x1280xf32>) outs(%2191 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2193 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2192 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2194 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2193 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2195 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2194 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2196 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2190, %2195 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2197 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2196 : tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2198 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2199 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2197 : tensor<2x576x1280xf32>) outs(%2198 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2200 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2199 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2201 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2200 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2202 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2201 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x576x1xf32>
    %2203 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2202 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2204 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2203 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2205 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2196, %2204 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2206 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2205, %cst_190 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2207 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2206, %cst_191 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2208 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2209 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2208 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2210 = tensor.empty() : tensor<10240xf16>
    %2211 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2209 : tensor<f16>) outs(%2210 : tensor<10240xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10240xf16>
    %2212 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_192 : tensor<10240xf32>) outs(%2211 : tensor<10240xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10240xf16>
    %2213 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2214 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2213 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2215 = tensor.empty() : tensor<10240x1280xf16>
    %2216 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2214 : tensor<f16>) outs(%2215 : tensor<10240x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10240x1280xf16>
    %2217 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_193 : tensor<10240x1280xf32>) outs(%2216 : tensor<10240x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10240x1280xf16>
    %2218 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2219 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2218 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2220 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2219 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2221 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2207 : tensor<2x576x1280xf32>) outs(%2220 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2222 = tensor.empty() : tensor<1280x10240xf16>
    %2223 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2217 : tensor<10240x1280xf16>) outs(%2222 : tensor<1280x10240xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x10240xf16>
    %collapsed_1106 = tensor.collapse_shape %2221 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2224 = tensor.empty() : tensor<1152x10240xf16>
    %2225 = linalg.fill ins(%cst_690 : f16) outs(%2224 : tensor<1152x10240xf16>) -> tensor<1152x10240xf16>
    %2226 = linalg.matmul ins(%collapsed_1106, %2223 : tensor<1152x1280xf16>, tensor<1280x10240xf16>) outs(%2225 : tensor<1152x10240xf16>) -> tensor<1152x10240xf16>
    %2227 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2212, %2226 : tensor<10240xf16>, tensor<1152x10240xf16>) outs(%2224 : tensor<1152x10240xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x10240xf16>
    %expanded_1107 = tensor.expand_shape %2227 [[0, 1], [2]] : tensor<1152x10240xf16> into tensor<2x576x10240xf16>
    %extracted_slice_1108 = tensor.extract_slice %expanded_1107[0, 0, 0] [2, 576, 5120] [1, 1, 1] : tensor<2x576x10240xf16> to tensor<2x576x5120xf16>
    %extracted_slice_1109 = tensor.extract_slice %expanded_1107[0, 0, 5120] [2, 576, 5120] [1, 1, 1] : tensor<2x576x10240xf16> to tensor<2x576x5120xf16>
    %2228 = tensor.empty() : tensor<2x576x5120xf16>
    %2229 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1109 : tensor<2x576x5120xf16>) outs(%2228 : tensor<2x576x5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = math.sqrt %cst_693 : f16
      %7525 = arith.divf %in, %7524 : f16
      %7526 = math.erf %7525 : f16
      %7527 = arith.addf %7526, %cst_691 : f16
      %7528 = arith.mulf %7527, %cst_694 : f16
      %7529 = arith.mulf %in, %7528 : f16
      linalg.yield %7529 : f16
    } -> tensor<2x576x5120xf16>
    %2230 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1108, %2229 : tensor<2x576x5120xf16>, tensor<2x576x5120xf16>) outs(%2228 : tensor<2x576x5120xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x5120xf16>
    %2231 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2232 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2231 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2233 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2232 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2234 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_194 : tensor<1280xf32>) outs(%2233 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2235 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2236 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2235 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2237 = tensor.empty() : tensor<1280x5120xf16>
    %2238 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2236 : tensor<f16>) outs(%2237 : tensor<1280x5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x5120xf16>
    %2239 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_195 : tensor<1280x5120xf32>) outs(%2238 : tensor<1280x5120xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x5120xf16>
    %2240 = tensor.empty() : tensor<5120x1280xf16>
    %2241 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2239 : tensor<1280x5120xf16>) outs(%2240 : tensor<5120x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<5120x1280xf16>
    %collapsed_1110 = tensor.collapse_shape %2230 [[0, 1], [2]] : tensor<2x576x5120xf16> into tensor<1152x5120xf16>
    %2242 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2243 = linalg.matmul ins(%collapsed_1110, %2241 : tensor<1152x5120xf16>, tensor<5120x1280xf16>) outs(%2242 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2244 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2234, %2243 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1111 = tensor.expand_shape %2244 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2245 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1111, %2186 : tensor<2x576x1280xf16>, tensor<2x576x1280xf16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2246 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2247 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2246 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2248 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2247 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2249 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_196 : tensor<1280xf32>) outs(%2248 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2250 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2251 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2250 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2252 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2251 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2253 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_197 : tensor<1280x1280xf32>) outs(%2252 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2254 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2253 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1112 = tensor.collapse_shape %2245 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2255 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2256 = linalg.matmul ins(%collapsed_1112, %2254 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2255 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2257 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2249, %2256 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1113 = tensor.expand_shape %2257 [[0, 1, 2], [3]] : tensor<1152x1280xf16> into tensor<2x24x24x1280xf16>
    %2258 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1113 : tensor<2x24x24x1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %2259 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2258, %1918 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2260 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2261 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2260 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2262 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2261 : tensor<f32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x24x24xf32>
    %2263 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2259 : tensor<2x1280x24x24xf16>) outs(%2262 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %collapsed_1114 = tensor.collapse_shape %2263 [[0], [1], [2, 3]] : tensor<2x1280x24x24xf32> into tensor<2x1280x576xf32>
    %expanded_1115 = tensor.expand_shape %collapsed_1114 [[0], [1, 2], [3]] : tensor<2x1280x576xf32> into tensor<2x32x40x576xf32>
    %2264 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1115 : tensor<2x32x40x576xf32>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2265 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2266 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2264 : tensor<2x32x40x576xf64>) outs(%2265 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2267 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2266 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2268 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2264, %2267 : tensor<2x32x40x576xf64>, tensor<2x32x1x1xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2269 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2268 : tensor<2x32x40x576xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2270 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2271 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2269 : tensor<2x32x40x576xf64>) outs(%2270 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2272 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2271 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2273 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2272 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2274 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2275 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1115 : tensor<2x32x40x576xf32>) outs(%2274 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2276 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2275 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2277 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2276 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2278 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2273, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2279 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2278 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2280 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1115, %2277 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %2281 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2280, %2279 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %collapsed_1116 = tensor.collapse_shape %2281 [[0], [1, 2], [3]] : tensor<2x32x40x576xf32> into tensor<2x1280x576xf32>
    %expanded_1117 = tensor.expand_shape %collapsed_1116 [[0], [1], [2, 3]] : tensor<2x1280x576xf32> into tensor<2x1280x24x24xf32>
    %expanded_1118 = tensor.expand_shape %cst_198 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1119 = tensor.expand_shape %cst_199 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2282 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1117, %expanded_1119 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %2283 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2282, %expanded_1118 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %2284 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2285 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2284 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2286 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2285 : tensor<f16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %2287 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2283 : tensor<2x1280x24x24xf32>) outs(%2286 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2288 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2287 : tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x24x24xf16>
    %2289 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2288, %2287 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2290 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2291 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2290 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2292 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2291 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2293 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_200 : tensor<1280xf32>) outs(%2292 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2294 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2295 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2294 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2296 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2295 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2297 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_201 : tensor<1280x1280x3x3xf32>) outs(%2296 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1120 = tensor.pad %2289 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x24x24xf16> to tensor<2x1280x26x26xf16>
    %2298 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2293 : tensor<1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %2299 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1120, %2297 : tensor<2x1280x26x26xf16>, tensor<1280x1280x3x3xf16>) outs(%2298 : tensor<2x1280x24x24xf16>) -> tensor<2x1280x24x24xf16>
    %2300 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %2301 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2300, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %2302 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2303 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2302 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2304 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2303 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2305 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_202 : tensor<1280xf32>) outs(%2304 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2306 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2307 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2306 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2308 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2307 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2309 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_203 : tensor<1280x1280xf32>) outs(%2308 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2310 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2309 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2311 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2312 = linalg.matmul ins(%2301, %2310 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2311 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2313 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2305, %2312 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %expanded_1121 = tensor.expand_shape %2313 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
    %2314 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2299, %expanded_1121 : tensor<2x1280x24x24xf16>, tensor<2x1280x1x1xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2315 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2316 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2315 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2317 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2316 : tensor<f32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x24x24xf32>
    %2318 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2314 : tensor<2x1280x24x24xf16>) outs(%2317 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %collapsed_1122 = tensor.collapse_shape %2318 [[0], [1], [2, 3]] : tensor<2x1280x24x24xf32> into tensor<2x1280x576xf32>
    %expanded_1123 = tensor.expand_shape %collapsed_1122 [[0], [1, 2], [3]] : tensor<2x1280x576xf32> into tensor<2x32x40x576xf32>
    %2319 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1123 : tensor<2x32x40x576xf32>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2320 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2321 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2319 : tensor<2x32x40x576xf64>) outs(%2320 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2322 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2321 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2323 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2319, %2322 : tensor<2x32x40x576xf64>, tensor<2x32x1x1xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2324 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2323 : tensor<2x32x40x576xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2325 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2326 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2324 : tensor<2x32x40x576xf64>) outs(%2325 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2327 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2326 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2328 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2327 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2329 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2330 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1123 : tensor<2x32x40x576xf32>) outs(%2329 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2331 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2330 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2332 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2331 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2333 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2328, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2334 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2333 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2335 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1123, %2332 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %2336 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2335, %2334 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %collapsed_1124 = tensor.collapse_shape %2336 [[0], [1, 2], [3]] : tensor<2x32x40x576xf32> into tensor<2x1280x576xf32>
    %expanded_1125 = tensor.expand_shape %collapsed_1124 [[0], [1], [2, 3]] : tensor<2x1280x576xf32> into tensor<2x1280x24x24xf32>
    %expanded_1126 = tensor.expand_shape %cst_204 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1127 = tensor.expand_shape %cst_205 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2337 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1125, %expanded_1127 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %2338 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2337, %expanded_1126 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %2339 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2340 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2339 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2341 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2340 : tensor<f16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %2342 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2338 : tensor<2x1280x24x24xf32>) outs(%2341 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2343 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2342 : tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x24x24xf16>
    %2344 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2343, %2342 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2345 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2346 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2345 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2347 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2346 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2348 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_206 : tensor<1280xf32>) outs(%2347 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2349 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2350 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2349 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2351 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2350 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2352 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_207 : tensor<1280x1280x3x3xf32>) outs(%2351 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1128 = tensor.pad %2344 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x24x24xf16> to tensor<2x1280x26x26xf16>
    %2353 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2348 : tensor<1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %2354 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1128, %2352 : tensor<2x1280x26x26xf16>, tensor<1280x1280x3x3xf16>) outs(%2353 : tensor<2x1280x24x24xf16>) -> tensor<2x1280x24x24xf16>
    %2355 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2259, %2354 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2356 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2357 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2356 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2358 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2357 : tensor<f32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x24x24xf32>
    %2359 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2355 : tensor<2x1280x24x24xf16>) outs(%2358 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %collapsed_1129 = tensor.collapse_shape %2359 [[0], [1], [2, 3]] : tensor<2x1280x24x24xf32> into tensor<2x1280x576xf32>
    %expanded_1130 = tensor.expand_shape %collapsed_1129 [[0], [1, 2], [3]] : tensor<2x1280x576xf32> into tensor<2x32x40x576xf32>
    %2360 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1130 : tensor<2x32x40x576xf32>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2361 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2362 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2360 : tensor<2x32x40x576xf64>) outs(%2361 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2363 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2362 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2364 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2360, %2363 : tensor<2x32x40x576xf64>, tensor<2x32x1x1xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2365 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2364 : tensor<2x32x40x576xf64>) outs(%1866 : tensor<2x32x40x576xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x576xf64>
    %2366 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2367 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2365 : tensor<2x32x40x576xf64>) outs(%2366 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2368 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2367 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2369 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2368 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2370 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2371 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1130 : tensor<2x32x40x576xf32>) outs(%2370 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2372 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2371 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_701 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2373 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2372 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2374 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2369, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2375 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2374 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2376 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1130, %2373 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %2377 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2376, %2375 : tensor<2x32x40x576xf32>, tensor<2x32x1x1xf32>) outs(%1883 : tensor<2x32x40x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x576xf32>
    %collapsed_1131 = tensor.collapse_shape %2377 [[0], [1, 2], [3]] : tensor<2x32x40x576xf32> into tensor<2x1280x576xf32>
    %expanded_1132 = tensor.expand_shape %collapsed_1131 [[0], [1], [2, 3]] : tensor<2x1280x576xf32> into tensor<2x1280x24x24xf32>
    %expanded_1133 = tensor.expand_shape %cst_208 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1134 = tensor.expand_shape %cst_209 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2378 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1132, %expanded_1134 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %2379 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2378, %expanded_1133 : tensor<2x1280x24x24xf32>, tensor<1x1280x1x1xf32>) outs(%1863 : tensor<2x1280x24x24xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x24x24xf32>
    %2380 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2379 : tensor<2x1280x24x24xf32>) outs(%1943 : tensor<2x24x24x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x24x24x1280xf32>
    %collapsed_1135 = tensor.collapse_shape %2380 [[0], [1, 2], [3]] : tensor<2x24x24x1280xf32> into tensor<2x576x1280xf32>
    %2381 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2382 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2381 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2383 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2382 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2384 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_210 : tensor<1280xf32>) outs(%2383 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2385 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2386 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2385 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2387 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2386 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2388 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_211 : tensor<1280x1280xf32>) outs(%2387 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2389 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2390 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2389 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2391 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2390 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2392 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1135 : tensor<2x576x1280xf32>) outs(%2391 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2393 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2388 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1136 = tensor.collapse_shape %2392 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2394 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2395 = linalg.matmul ins(%collapsed_1136, %2393 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2394 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2396 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2384, %2395 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1137 = tensor.expand_shape %2396 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2397 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2398 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2397 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2399 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2398 : tensor<f32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2400 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1137 : tensor<2x576x1280xf16>) outs(%2399 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2401 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2402 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2400 : tensor<2x576x1280xf32>) outs(%2401 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2403 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2402 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2404 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2403 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2405 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2404 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2406 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2400, %2405 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2407 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2406 : tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2408 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2409 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2407 : tensor<2x576x1280xf32>) outs(%2408 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2410 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2409 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2411 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2410 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2412 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2411 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x576x1xf32>
    %2413 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2412 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2414 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2413 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2415 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2406, %2414 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2416 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2415, %cst_212 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2417 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2416, %cst_213 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2418 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2419 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2418 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2420 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2419 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2421 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_214 : tensor<1280x1280xf32>) outs(%2420 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2422 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2423 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2422 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2424 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2423 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2425 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2417 : tensor<2x576x1280xf32>) outs(%2424 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2426 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2421 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1138 = tensor.collapse_shape %2425 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2427 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2428 = linalg.matmul ins(%collapsed_1138, %2426 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2427 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2429 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2430 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2429 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2431 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2430 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2432 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_215 : tensor<1280x1280xf32>) outs(%2431 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2433 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2434 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2433 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2435 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2434 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2436 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2417 : tensor<2x576x1280xf32>) outs(%2435 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2437 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2432 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1139 = tensor.collapse_shape %2436 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2438 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2439 = linalg.matmul ins(%collapsed_1139, %2437 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2438 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2440 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2441 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2440 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2442 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2441 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2443 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_216 : tensor<1280x1280xf32>) outs(%2442 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2444 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2445 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2444 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2446 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2445 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2447 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2417 : tensor<2x576x1280xf32>) outs(%2446 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2448 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2443 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1140 = tensor.collapse_shape %2447 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2449 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2450 = linalg.matmul ins(%collapsed_1140, %2448 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2449 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %expanded_1141 = tensor.expand_shape %2428 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2451 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1141 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1142 = tensor.collapse_shape %2451 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1143 = tensor.expand_shape %2439 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2452 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1143 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1144 = tensor.collapse_shape %2452 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1145 = tensor.expand_shape %2450 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2453 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1145 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1146 = tensor.collapse_shape %2453 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1147 = tensor.expand_shape %collapsed_1142 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2454 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1147 : tensor<40x576x64x1xf16>) outs(%2024 : tensor<40x576x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x64xf16>
    %expanded_1148 = tensor.expand_shape %collapsed_1144 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2455 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1148 : tensor<40x576x64x1xf16>) outs(%2026 : tensor<40x1x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x576x64xf16>
    %2456 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2454 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1149 = tensor.collapse_shape %2456 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %2457 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2455 : tensor<40x1x576x64xf16>) outs(%2030 : tensor<40x64x576x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x64x576x1xf16>
    %collapsed_1150 = tensor.collapse_shape %2457 [[0], [1], [2, 3]] : tensor<40x64x576x1xf16> into tensor<40x64x576xf16>
    %2458 = linalg.fill ins(%cst_690 : f16) outs(%2032 : tensor<40x576x576xf16>) -> tensor<40x576x576xf16>
    %2459 = linalg.batch_matmul ins(%collapsed_1149, %collapsed_1150 : tensor<40x576x64xf16>, tensor<40x64x576xf16>) outs(%2458 : tensor<40x576x576xf16>) -> tensor<40x576x576xf16>
    %expanded_1151 = tensor.expand_shape %2459 [[0], [1], [2, 3]] : tensor<40x576x576xf16> into tensor<40x576x1x576xf16>
    %2460 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1151 : tensor<40x576x1x576xf16>) outs(%2035 : tensor<40x576x576x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x576x1xf16>
    %collapsed_1152 = tensor.collapse_shape %2460 [[0], [1], [2, 3]] : tensor<40x576x576x1xf16> into tensor<40x576x576xf16>
    %2461 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1152, %cst : tensor<40x576x576xf16>, tensor<f64>) outs(%2032 : tensor<40x576x576xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<40x576x576xf16>
    %2462 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2461 : tensor<40x576x576xf16>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2463 = linalg.fill ins(%cst_692 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2464 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2462 : tensor<40x576x576xf32>) outs(%2463 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2465 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2462, %2464 : tensor<40x576x576xf32>, tensor<40x576x1xf32>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2466 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2465 : tensor<40x576x576xf32>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2467 = linalg.fill ins(%cst_697 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2468 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2466 : tensor<40x576x576xf32>) outs(%2467 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2469 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2466, %2468 : tensor<40x576x576xf32>, tensor<40x576x1xf32>) outs(%2038 : tensor<40x576x576xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x576xf32>
    %2470 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2471 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2470 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2472 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2471 : tensor<f16>) outs(%2032 : tensor<40x576x576xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x576xf16>
    %2473 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2469 : tensor<40x576x576xf32>) outs(%2472 : tensor<40x576x576xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<40x576x576xf16>
    %expanded_1153 = tensor.expand_shape %2473 [[0], [1], [2, 3]] : tensor<40x576x576xf16> into tensor<40x576x576x1xf16>
    %2474 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1153 : tensor<40x576x576x1xf16>) outs(%2052 : tensor<40x576x1x576xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x576xf16>
    %expanded_1154 = tensor.expand_shape %collapsed_1146 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2475 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1154 : tensor<40x576x64x1xf16>) outs(%2054 : tensor<40x1x64x576xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x64x576xf16>
    %2476 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2474 : tensor<40x576x1x576xf16>) outs(%2035 : tensor<40x576x576x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x576x1xf16>
    %collapsed_1155 = tensor.collapse_shape %2476 [[0], [1], [2, 3]] : tensor<40x576x576x1xf16> into tensor<40x576x576xf16>
    %2477 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2475 : tensor<40x1x64x576xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1156 = tensor.collapse_shape %2477 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %2478 = linalg.fill ins(%cst_690 : f16) outs(%2058 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %2479 = linalg.batch_matmul ins(%collapsed_1155, %collapsed_1156 : tensor<40x576x576xf16>, tensor<40x576x64xf16>) outs(%2478 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %expanded_1157 = tensor.expand_shape %2479 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x1x64xf16>
    %2480 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1157 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1158 = tensor.collapse_shape %2480 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %expanded_1159 = tensor.expand_shape %collapsed_1158 [[0, 1], [2], [3]] : tensor<40x576x64xf16> into tensor<2x20x576x64xf16>
    %2481 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1159 : tensor<2x20x576x64xf16>) outs(%2062 : tensor<2x576x20x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x20x64xf16>
    %2482 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2483 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2482 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2484 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2483 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2485 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_217 : tensor<1280xf32>) outs(%2484 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2486 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2487 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2486 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2488 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2487 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2489 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_218 : tensor<1280x1280xf32>) outs(%2488 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2490 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2489 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1160 = tensor.collapse_shape %2481 [[0, 1], [2, 3]] : tensor<2x576x20x64xf16> into tensor<1152x1280xf16>
    %2491 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2492 = linalg.matmul ins(%collapsed_1160, %2490 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2491 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2493 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2485, %2492 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1161 = tensor.expand_shape %2493 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2494 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1161, %expanded_1137 : tensor<2x576x1280xf16>, tensor<2x576x1280xf16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2495 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2496 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2495 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2497 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2496 : tensor<f32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2498 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2494 : tensor<2x576x1280xf16>) outs(%2497 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2499 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2500 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2498 : tensor<2x576x1280xf32>) outs(%2499 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2501 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2500 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2502 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2501 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2503 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2502 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2504 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2498, %2503 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2505 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2504 : tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2506 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2507 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2505 : tensor<2x576x1280xf32>) outs(%2506 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2508 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2507 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2509 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2508 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2510 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2509 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x576x1xf32>
    %2511 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2510 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2512 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2511 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2513 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2504, %2512 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2514 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2513, %cst_219 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2515 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2514, %cst_220 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2516 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2517 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2516 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2518 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2517 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2519 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_221 : tensor<1280x1280xf32>) outs(%2518 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2520 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2521 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2520 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2522 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2521 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2523 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2515 : tensor<2x576x1280xf32>) outs(%2522 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2524 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2519 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1162 = tensor.collapse_shape %2523 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2525 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2526 = linalg.matmul ins(%collapsed_1162, %2524 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2525 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2527 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2528 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2527 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2529 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2528 : tensor<f16>) outs(%2111 : tensor<1280x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1024xf16>
    %2530 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_222 : tensor<1280x1024xf32>) outs(%2529 : tensor<1280x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1024xf16>
    %2531 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2532 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2531 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2533 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2532 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %2534 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%2533 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %2535 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2530 : tensor<1280x1024xf16>) outs(%2118 : tensor<1024x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x1280xf16>
    %collapsed_1163 = tensor.collapse_shape %2534 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %2536 = linalg.fill ins(%cst_690 : f16) outs(%2120 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %2537 = linalg.matmul ins(%collapsed_1163, %2535 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%2536 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %2538 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2539 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2538 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2540 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2539 : tensor<f16>) outs(%2111 : tensor<1280x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1024xf16>
    %2541 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_223 : tensor<1280x1024xf32>) outs(%2540 : tensor<1280x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1024xf16>
    %2542 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2543 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2542 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2544 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2543 : tensor<f16>) outs(%374 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x77x1024xf16>
    %2545 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg2 : tensor<2x77x1024xf32>) outs(%2544 : tensor<2x77x1024xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x77x1024xf16>
    %2546 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2541 : tensor<1280x1024xf16>) outs(%2118 : tensor<1024x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1024x1280xf16>
    %collapsed_1164 = tensor.collapse_shape %2545 [[0, 1], [2]] : tensor<2x77x1024xf16> into tensor<154x1024xf16>
    %2547 = linalg.fill ins(%cst_690 : f16) outs(%2120 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %2548 = linalg.matmul ins(%collapsed_1164, %2546 : tensor<154x1024xf16>, tensor<1024x1280xf16>) outs(%2547 : tensor<154x1280xf16>) -> tensor<154x1280xf16>
    %expanded_1165 = tensor.expand_shape %2526 [[0, 1], [2, 3]] : tensor<1152x1280xf16> into tensor<2x576x20x64xf16>
    %2549 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1165 : tensor<2x576x20x64xf16>) outs(%2020 : tensor<2x20x576x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x576x64xf16>
    %collapsed_1166 = tensor.collapse_shape %2549 [[0, 1], [2], [3]] : tensor<2x20x576x64xf16> into tensor<40x576x64xf16>
    %expanded_1167 = tensor.expand_shape %2537 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
    %2550 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1167 : tensor<2x77x20x64xf16>) outs(%2135 : tensor<2x20x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x77x64xf16>
    %collapsed_1168 = tensor.collapse_shape %2550 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
    %expanded_1169 = tensor.expand_shape %2548 [[0, 1], [2, 3]] : tensor<154x1280xf16> into tensor<2x77x20x64xf16>
    %2551 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1169 : tensor<2x77x20x64xf16>) outs(%2135 : tensor<2x20x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x77x64xf16>
    %collapsed_1170 = tensor.collapse_shape %2551 [[0, 1], [2], [3]] : tensor<2x20x77x64xf16> into tensor<40x77x64xf16>
    %expanded_1171 = tensor.expand_shape %collapsed_1166 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x64x1xf16>
    %2552 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1171 : tensor<40x576x64x1xf16>) outs(%2024 : tensor<40x576x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x64xf16>
    %expanded_1172 = tensor.expand_shape %collapsed_1168 [[0], [1], [2, 3]] : tensor<40x77x64xf16> into tensor<40x77x64x1xf16>
    %2553 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1172 : tensor<40x77x64x1xf16>) outs(%2139 : tensor<40x1x77x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x77x64xf16>
    %2554 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2552 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1173 = tensor.collapse_shape %2554 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %2555 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2553 : tensor<40x1x77x64xf16>) outs(%2142 : tensor<40x64x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x64x77x1xf16>
    %collapsed_1174 = tensor.collapse_shape %2555 [[0], [1], [2, 3]] : tensor<40x64x77x1xf16> into tensor<40x64x77xf16>
    %2556 = linalg.fill ins(%cst_690 : f16) outs(%2144 : tensor<40x576x77xf16>) -> tensor<40x576x77xf16>
    %2557 = linalg.batch_matmul ins(%collapsed_1173, %collapsed_1174 : tensor<40x576x64xf16>, tensor<40x64x77xf16>) outs(%2556 : tensor<40x576x77xf16>) -> tensor<40x576x77xf16>
    %expanded_1175 = tensor.expand_shape %2557 [[0], [1], [2, 3]] : tensor<40x576x77xf16> into tensor<40x576x1x77xf16>
    %2558 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1175 : tensor<40x576x1x77xf16>) outs(%2147 : tensor<40x576x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x77x1xf16>
    %collapsed_1176 = tensor.collapse_shape %2558 [[0], [1], [2, 3]] : tensor<40x576x77x1xf16> into tensor<40x576x77xf16>
    %2559 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1176, %cst : tensor<40x576x77xf16>, tensor<f64>) outs(%2144 : tensor<40x576x77xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<40x576x77xf16>
    %2560 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2559 : tensor<40x576x77xf16>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2561 = linalg.fill ins(%cst_692 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2562 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2560 : tensor<40x576x77xf32>) outs(%2561 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2563 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2560, %2562 : tensor<40x576x77xf32>, tensor<40x576x1xf32>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2564 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2563 : tensor<40x576x77xf32>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2565 = linalg.fill ins(%cst_697 : f32) outs(%2040 : tensor<40x576x1xf32>) -> tensor<40x576x1xf32>
    %2566 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2564 : tensor<40x576x77xf32>) outs(%2565 : tensor<40x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x1xf32>
    %2567 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2564, %2566 : tensor<40x576x77xf32>, tensor<40x576x1xf32>) outs(%2150 : tensor<40x576x77xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x576x77xf32>
    %2568 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2569 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2568 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2570 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2569 : tensor<f16>) outs(%2144 : tensor<40x576x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x77xf16>
    %2571 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2567 : tensor<40x576x77xf32>) outs(%2570 : tensor<40x576x77xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<40x576x77xf16>
    %expanded_1177 = tensor.expand_shape %2571 [[0], [1], [2, 3]] : tensor<40x576x77xf16> into tensor<40x576x77x1xf16>
    %2572 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1177 : tensor<40x576x77x1xf16>) outs(%2163 : tensor<40x576x1x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x1x77xf16>
    %expanded_1178 = tensor.expand_shape %collapsed_1170 [[0], [1], [2, 3]] : tensor<40x77x64xf16> into tensor<40x77x64x1xf16>
    %2573 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1178 : tensor<40x77x64x1xf16>) outs(%2165 : tensor<40x1x64x77xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x64x77xf16>
    %2574 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2572 : tensor<40x576x1x77xf16>) outs(%2147 : tensor<40x576x77x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x77x1xf16>
    %collapsed_1179 = tensor.collapse_shape %2574 [[0], [1], [2, 3]] : tensor<40x576x77x1xf16> into tensor<40x576x77xf16>
    %2575 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2573 : tensor<40x1x64x77xf16>) outs(%2168 : tensor<40x77x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x77x64x1xf16>
    %collapsed_1180 = tensor.collapse_shape %2575 [[0], [1], [2, 3]] : tensor<40x77x64x1xf16> into tensor<40x77x64xf16>
    %2576 = linalg.fill ins(%cst_690 : f16) outs(%2058 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %2577 = linalg.batch_matmul ins(%collapsed_1179, %collapsed_1180 : tensor<40x576x77xf16>, tensor<40x77x64xf16>) outs(%2576 : tensor<40x576x64xf16>) -> tensor<40x576x64xf16>
    %expanded_1181 = tensor.expand_shape %2577 [[0], [1], [2, 3]] : tensor<40x576x64xf16> into tensor<40x576x1x64xf16>
    %2578 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1181 : tensor<40x576x1x64xf16>) outs(%2028 : tensor<40x576x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x576x64x1xf16>
    %collapsed_1182 = tensor.collapse_shape %2578 [[0], [1], [2, 3]] : tensor<40x576x64x1xf16> into tensor<40x576x64xf16>
    %expanded_1183 = tensor.expand_shape %collapsed_1182 [[0, 1], [2], [3]] : tensor<40x576x64xf16> into tensor<2x20x576x64xf16>
    %2579 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1183 : tensor<2x20x576x64xf16>) outs(%2062 : tensor<2x576x20x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x20x64xf16>
    %2580 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2581 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2580 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2582 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2581 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2583 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_224 : tensor<1280xf32>) outs(%2582 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2584 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2585 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2584 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2586 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2585 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2587 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_225 : tensor<1280x1280xf32>) outs(%2586 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2588 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2587 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1184 = tensor.collapse_shape %2579 [[0, 1], [2, 3]] : tensor<2x576x20x64xf16> into tensor<1152x1280xf16>
    %2589 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2590 = linalg.matmul ins(%collapsed_1184, %2588 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2589 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2591 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2583, %2590 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1185 = tensor.expand_shape %2591 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2592 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1185, %2494 : tensor<2x576x1280xf16>, tensor<2x576x1280xf16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2593 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2594 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2593 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2595 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2594 : tensor<f32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2596 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2592 : tensor<2x576x1280xf16>) outs(%2595 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2597 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2598 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2596 : tensor<2x576x1280xf32>) outs(%2597 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2599 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2598 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2600 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2599 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2601 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2600 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2602 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2596, %2601 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2603 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2602 : tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2604 = linalg.fill ins(%cst_695 : f64) outs(%1968 : tensor<2x576x1xf64>) -> tensor<2x576x1xf64>
    %2605 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%2603 : tensor<2x576x1280xf32>) outs(%2604 : tensor<2x576x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x576x1xf64>
    %2606 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2605 : tensor<2x576x1xf64>) outs(%1968 : tensor<2x576x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x576x1xf64>
    %2607 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2606 : tensor<2x576x1xf64>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2608 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2607 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x576x1xf32>
    %2609 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2608 : tensor<2x576x1xf32>) outs(%1972 : tensor<2x576x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1xf32>
    %2610 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2609 : tensor<2x576x1xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x576x1280xf32>
    %2611 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2602, %2610 : tensor<2x576x1280xf32>, tensor<2x576x1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2612 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2611, %cst_226 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2613 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2612, %cst_227 : tensor<2x576x1280xf32>, tensor<1280xf32>) outs(%1965 : tensor<2x576x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x576x1280xf32>
    %2614 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2615 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2614 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2616 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2615 : tensor<f16>) outs(%2210 : tensor<10240xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10240xf16>
    %2617 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_228 : tensor<10240xf32>) outs(%2616 : tensor<10240xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10240xf16>
    %2618 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2619 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2618 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2620 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2619 : tensor<f16>) outs(%2215 : tensor<10240x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<10240x1280xf16>
    %2621 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_229 : tensor<10240x1280xf32>) outs(%2620 : tensor<10240x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<10240x1280xf16>
    %2622 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2623 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2622 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2624 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2623 : tensor<f16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x576x1280xf16>
    %2625 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2613 : tensor<2x576x1280xf32>) outs(%2624 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2626 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2621 : tensor<10240x1280xf16>) outs(%2222 : tensor<1280x10240xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x10240xf16>
    %collapsed_1186 = tensor.collapse_shape %2625 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2627 = linalg.fill ins(%cst_690 : f16) outs(%2224 : tensor<1152x10240xf16>) -> tensor<1152x10240xf16>
    %2628 = linalg.matmul ins(%collapsed_1186, %2626 : tensor<1152x1280xf16>, tensor<1280x10240xf16>) outs(%2627 : tensor<1152x10240xf16>) -> tensor<1152x10240xf16>
    %2629 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2617, %2628 : tensor<10240xf16>, tensor<1152x10240xf16>) outs(%2224 : tensor<1152x10240xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x10240xf16>
    %expanded_1187 = tensor.expand_shape %2629 [[0, 1], [2]] : tensor<1152x10240xf16> into tensor<2x576x10240xf16>
    %extracted_slice_1188 = tensor.extract_slice %expanded_1187[0, 0, 0] [2, 576, 5120] [1, 1, 1] : tensor<2x576x10240xf16> to tensor<2x576x5120xf16>
    %extracted_slice_1189 = tensor.extract_slice %expanded_1187[0, 0, 5120] [2, 576, 5120] [1, 1, 1] : tensor<2x576x10240xf16> to tensor<2x576x5120xf16>
    %2630 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1189 : tensor<2x576x5120xf16>) outs(%2228 : tensor<2x576x5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = math.sqrt %cst_693 : f16
      %7525 = arith.divf %in, %7524 : f16
      %7526 = math.erf %7525 : f16
      %7527 = arith.addf %7526, %cst_691 : f16
      %7528 = arith.mulf %7527, %cst_694 : f16
      %7529 = arith.mulf %in, %7528 : f16
      linalg.yield %7529 : f16
    } -> tensor<2x576x5120xf16>
    %2631 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%extracted_slice_1188, %2630 : tensor<2x576x5120xf16>, tensor<2x576x5120xf16>) outs(%2228 : tensor<2x576x5120xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x5120xf16>
    %2632 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2633 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2632 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2634 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2633 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2635 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_230 : tensor<1280xf32>) outs(%2634 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2636 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2637 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2636 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2638 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2637 : tensor<f16>) outs(%2237 : tensor<1280x5120xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x5120xf16>
    %2639 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_231 : tensor<1280x5120xf32>) outs(%2638 : tensor<1280x5120xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x5120xf16>
    %2640 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2639 : tensor<1280x5120xf16>) outs(%2240 : tensor<5120x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<5120x1280xf16>
    %collapsed_1190 = tensor.collapse_shape %2631 [[0, 1], [2]] : tensor<2x576x5120xf16> into tensor<1152x5120xf16>
    %2641 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2642 = linalg.matmul ins(%collapsed_1190, %2640 : tensor<1152x5120xf16>, tensor<5120x1280xf16>) outs(%2641 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2643 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2635, %2642 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1191 = tensor.expand_shape %2643 [[0, 1], [2]] : tensor<1152x1280xf16> into tensor<2x576x1280xf16>
    %2644 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1191, %2592 : tensor<2x576x1280xf16>, tensor<2x576x1280xf16>) outs(%1955 : tensor<2x576x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x576x1280xf16>
    %2645 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2646 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2645 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2647 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2646 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2648 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_232 : tensor<1280xf32>) outs(%2647 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2649 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2650 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2649 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2651 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2650 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2652 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_233 : tensor<1280x1280xf32>) outs(%2651 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2653 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2652 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1192 = tensor.collapse_shape %2644 [[0, 1], [2]] : tensor<2x576x1280xf16> into tensor<1152x1280xf16>
    %2654 = linalg.fill ins(%cst_690 : f16) outs(%1959 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2655 = linalg.matmul ins(%collapsed_1192, %2653 : tensor<1152x1280xf16>, tensor<1280x1280xf16>) outs(%2654 : tensor<1152x1280xf16>) -> tensor<1152x1280xf16>
    %2656 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2648, %2655 : tensor<1280xf16>, tensor<1152x1280xf16>) outs(%1959 : tensor<1152x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<1152x1280xf16>
    %expanded_1193 = tensor.expand_shape %2656 [[0, 1, 2], [3]] : tensor<1152x1280xf16> into tensor<2x24x24x1280xf16>
    %2657 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1193 : tensor<2x24x24x1280xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x24x24xf16>
    %2658 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2657, %2355 : tensor<2x1280x24x24xf16>, tensor<2x1280x24x24xf16>) outs(%1843 : tensor<2x1280x24x24xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x24x24xf16>
    %2659 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2660 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2659 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2661 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2660 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2662 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_234 : tensor<1280xf32>) outs(%2661 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2663 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2664 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2663 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2665 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2664 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2666 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_235 : tensor<1280x1280x3x3xf32>) outs(%2665 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1194 = tensor.pad %2658 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x24x24xf16> to tensor<2x1280x26x26xf16>
    %2667 = tensor.empty() : tensor<2x1280x12x12xf16>
    %2668 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2662 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2669 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%padded_1194, %2666 : tensor<2x1280x26x26xf16>, tensor<1280x1280x3x3xf16>) outs(%2668 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2670 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2671 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2670 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2672 = tensor.empty() : tensor<2x1280x12x12xf32>
    %2673 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2671 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2674 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2669 : tensor<2x1280x12x12xf16>) outs(%2673 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1195 = tensor.collapse_shape %2674 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1196 = tensor.expand_shape %collapsed_1195 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2675 = tensor.empty() : tensor<2x32x40x144xf64>
    %2676 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1196 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2677 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2678 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2676 : tensor<2x32x40x144xf64>) outs(%2677 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2679 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2678 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2680 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2676, %2679 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2681 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2680 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2682 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2683 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2681 : tensor<2x32x40x144xf64>) outs(%2682 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2684 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2683 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2685 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2684 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2686 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2687 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1196 : tensor<2x32x40x144xf32>) outs(%2686 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2688 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2687 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2689 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2688 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2690 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2685, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2691 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2690 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2692 = tensor.empty() : tensor<2x32x40x144xf32>
    %2693 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1196, %2689 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2694 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2693, %2691 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1197 = tensor.collapse_shape %2694 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1198 = tensor.expand_shape %collapsed_1197 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1199 = tensor.expand_shape %cst_236 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1200 = tensor.expand_shape %cst_237 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2695 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1198, %expanded_1200 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2696 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2695, %expanded_1199 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2697 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2698 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2697 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2699 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2698 : tensor<f16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2700 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2696 : tensor<2x1280x12x12xf32>) outs(%2699 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2701 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2700 : tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x12x12xf16>
    %2702 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2701, %2700 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2703 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2704 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2703 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2705 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2704 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2706 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_238 : tensor<1280xf32>) outs(%2705 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2707 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2708 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2707 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2709 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2708 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2710 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_239 : tensor<1280x1280x3x3xf32>) outs(%2709 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1201 = tensor.pad %2702 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x12x12xf16> to tensor<2x1280x14x14xf16>
    %2711 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2706 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2712 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1201, %2710 : tensor<2x1280x14x14xf16>, tensor<1280x1280x3x3xf16>) outs(%2711 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2713 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %2714 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2713, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %2715 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2716 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2715 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2717 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2716 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2718 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_240 : tensor<1280xf32>) outs(%2717 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2719 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2720 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2719 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2721 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2720 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2722 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_241 : tensor<1280x1280xf32>) outs(%2721 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2723 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2722 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2724 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2725 = linalg.matmul ins(%2714, %2723 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2724 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2726 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2718, %2725 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %expanded_1202 = tensor.expand_shape %2726 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
    %2727 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2712, %expanded_1202 : tensor<2x1280x12x12xf16>, tensor<2x1280x1x1xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2728 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2729 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2728 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2730 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2729 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2731 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2727 : tensor<2x1280x12x12xf16>) outs(%2730 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1203 = tensor.collapse_shape %2731 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1204 = tensor.expand_shape %collapsed_1203 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2732 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1204 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2733 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2734 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2732 : tensor<2x32x40x144xf64>) outs(%2733 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2735 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2734 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2736 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2732, %2735 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2737 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2736 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2738 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2739 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2737 : tensor<2x32x40x144xf64>) outs(%2738 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2740 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2739 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2741 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2740 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2742 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2743 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1204 : tensor<2x32x40x144xf32>) outs(%2742 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2744 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2743 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2745 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2744 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2746 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2741, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2747 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2746 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2748 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1204, %2745 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2749 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2748, %2747 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1205 = tensor.collapse_shape %2749 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1206 = tensor.expand_shape %collapsed_1205 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1207 = tensor.expand_shape %cst_242 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1208 = tensor.expand_shape %cst_243 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2750 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1206, %expanded_1208 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2751 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2750, %expanded_1207 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2752 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2753 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2752 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2754 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2753 : tensor<f16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2755 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2751 : tensor<2x1280x12x12xf32>) outs(%2754 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2756 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2755 : tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x12x12xf16>
    %2757 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2756, %2755 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2758 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2759 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2758 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2760 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2759 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2761 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_244 : tensor<1280xf32>) outs(%2760 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2762 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2763 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2762 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2764 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2763 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2765 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_245 : tensor<1280x1280x3x3xf32>) outs(%2764 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1209 = tensor.pad %2757 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x12x12xf16> to tensor<2x1280x14x14xf16>
    %2766 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2761 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2767 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1209, %2765 : tensor<2x1280x14x14xf16>, tensor<1280x1280x3x3xf16>) outs(%2766 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2768 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2669, %2767 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2769 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2770 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2769 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2771 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2770 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2772 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2768 : tensor<2x1280x12x12xf16>) outs(%2771 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1210 = tensor.collapse_shape %2772 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1211 = tensor.expand_shape %collapsed_1210 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2773 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1211 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2774 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2775 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2773 : tensor<2x32x40x144xf64>) outs(%2774 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2776 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2775 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2777 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2773, %2776 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2778 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2777 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2779 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2780 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2778 : tensor<2x32x40x144xf64>) outs(%2779 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2781 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2780 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2782 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2781 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2783 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2784 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1211 : tensor<2x32x40x144xf32>) outs(%2783 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2785 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2784 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2786 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2785 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2787 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2782, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2788 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2787 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2789 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1211, %2786 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2790 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2789, %2788 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1212 = tensor.collapse_shape %2790 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1213 = tensor.expand_shape %collapsed_1212 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1214 = tensor.expand_shape %cst_246 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1215 = tensor.expand_shape %cst_247 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2791 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1213, %expanded_1215 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2792 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2791, %expanded_1214 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2793 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2794 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2793 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2795 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2794 : tensor<f16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2796 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2792 : tensor<2x1280x12x12xf32>) outs(%2795 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2797 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2796 : tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x12x12xf16>
    %2798 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2797, %2796 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2799 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2800 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2799 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2801 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2800 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2802 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_248 : tensor<1280xf32>) outs(%2801 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2803 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2804 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2803 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2805 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2804 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2806 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_249 : tensor<1280x1280x3x3xf32>) outs(%2805 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1216 = tensor.pad %2798 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x12x12xf16> to tensor<2x1280x14x14xf16>
    %2807 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2802 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2808 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1216, %2806 : tensor<2x1280x14x14xf16>, tensor<1280x1280x3x3xf16>) outs(%2807 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2809 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %2810 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2809, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %2811 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2812 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2811 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2813 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2812 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2814 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_250 : tensor<1280xf32>) outs(%2813 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2815 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2816 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2815 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2817 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2816 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2818 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_251 : tensor<1280x1280xf32>) outs(%2817 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2819 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2818 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2820 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2821 = linalg.matmul ins(%2810, %2819 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2820 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2822 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2814, %2821 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %expanded_1217 = tensor.expand_shape %2822 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
    %2823 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2808, %expanded_1217 : tensor<2x1280x12x12xf16>, tensor<2x1280x1x1xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2824 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2825 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2824 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2826 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2825 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2827 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2823 : tensor<2x1280x12x12xf16>) outs(%2826 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1218 = tensor.collapse_shape %2827 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1219 = tensor.expand_shape %collapsed_1218 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2828 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1219 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2829 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2830 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2828 : tensor<2x32x40x144xf64>) outs(%2829 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2831 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2830 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2832 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2828, %2831 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2833 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2832 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2834 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2835 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2833 : tensor<2x32x40x144xf64>) outs(%2834 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2836 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2835 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2837 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2836 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2838 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2839 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1219 : tensor<2x32x40x144xf32>) outs(%2838 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2840 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2839 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2841 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2840 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2842 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2837, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2843 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2842 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2844 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1219, %2841 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2845 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2844, %2843 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1220 = tensor.collapse_shape %2845 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1221 = tensor.expand_shape %collapsed_1220 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1222 = tensor.expand_shape %cst_252 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1223 = tensor.expand_shape %cst_253 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2846 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1221, %expanded_1223 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2847 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2846, %expanded_1222 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2848 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2849 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2848 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2850 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2849 : tensor<f16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2851 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2847 : tensor<2x1280x12x12xf32>) outs(%2850 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2852 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2851 : tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x12x12xf16>
    %2853 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2852, %2851 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2854 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2855 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2854 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2856 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2855 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2857 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_254 : tensor<1280xf32>) outs(%2856 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2858 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2859 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2858 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2860 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2859 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2861 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_255 : tensor<1280x1280x3x3xf32>) outs(%2860 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1224 = tensor.pad %2853 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x12x12xf16> to tensor<2x1280x14x14xf16>
    %2862 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2857 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2863 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1224, %2861 : tensor<2x1280x14x14xf16>, tensor<1280x1280x3x3xf16>) outs(%2862 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2864 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2768, %2863 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2865 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2866 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2865 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2867 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2866 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2868 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2864 : tensor<2x1280x12x12xf16>) outs(%2867 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1225 = tensor.collapse_shape %2868 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1226 = tensor.expand_shape %collapsed_1225 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2869 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1226 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2870 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2871 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2869 : tensor<2x32x40x144xf64>) outs(%2870 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2872 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2871 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2873 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2869, %2872 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2874 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2873 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2875 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2876 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2874 : tensor<2x32x40x144xf64>) outs(%2875 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2877 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2876 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2878 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2877 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2879 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2880 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1226 : tensor<2x32x40x144xf32>) outs(%2879 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2881 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2880 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2882 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2881 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2883 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2878, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2884 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2883 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2885 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1226, %2882 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2886 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2885, %2884 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1227 = tensor.collapse_shape %2886 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1228 = tensor.expand_shape %collapsed_1227 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1229 = tensor.expand_shape %cst_256 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1230 = tensor.expand_shape %cst_257 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2887 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1228, %expanded_1230 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2888 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2887, %expanded_1229 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2889 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2890 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2889 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2891 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2890 : tensor<f16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2892 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2888 : tensor<2x1280x12x12xf32>) outs(%2891 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2893 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2892 : tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x12x12xf16>
    %2894 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2893, %2892 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2895 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2896 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2895 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2897 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2896 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2898 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_258 : tensor<1280xf32>) outs(%2897 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2899 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2900 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2899 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2901 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2900 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2902 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_259 : tensor<1280x1280x3x3xf32>) outs(%2901 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1231 = tensor.pad %2894 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x12x12xf16> to tensor<2x1280x14x14xf16>
    %2903 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2898 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2904 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1231, %2902 : tensor<2x1280x14x14xf16>, tensor<1280x1280x3x3xf16>) outs(%2903 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2905 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%55 : tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280xf16>
    %2906 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2905, %55 : tensor<2x1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %2907 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2908 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2907 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2909 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2908 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2910 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_260 : tensor<1280xf32>) outs(%2909 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2911 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2912 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2911 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2913 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2912 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2914 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_261 : tensor<1280x1280xf32>) outs(%2913 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2915 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2914 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2916 = linalg.fill ins(%cst_690 : f16) outs(%37 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2917 = linalg.matmul ins(%2906, %2915 : tensor<2x1280xf16>, tensor<1280x1280xf16>) outs(%2916 : tensor<2x1280xf16>) -> tensor<2x1280xf16>
    %2918 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2910, %2917 : tensor<1280xf16>, tensor<2x1280xf16>) outs(%37 : tensor<2x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280xf16>
    %expanded_1232 = tensor.expand_shape %2918 [[0], [1, 2, 3]] : tensor<2x1280xf16> into tensor<2x1280x1x1xf16>
    %2919 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2904, %expanded_1232 : tensor<2x1280x12x12xf16>, tensor<2x1280x1x1xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2920 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2921 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2920 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2922 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2921 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2923 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2919 : tensor<2x1280x12x12xf16>) outs(%2922 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1233 = tensor.collapse_shape %2923 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1234 = tensor.expand_shape %collapsed_1233 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2924 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1234 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2925 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2926 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2924 : tensor<2x32x40x144xf64>) outs(%2925 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2927 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2926 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2928 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2924, %2927 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2929 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2928 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2930 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2931 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2929 : tensor<2x32x40x144xf64>) outs(%2930 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2932 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2931 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2933 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2932 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2934 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2935 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1234 : tensor<2x32x40x144xf32>) outs(%2934 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2936 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2935 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2937 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2936 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2938 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2933, %cst_1 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2939 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2938 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2940 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1234, %2937 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2941 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2940, %2939 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1235 = tensor.collapse_shape %2941 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1236 = tensor.expand_shape %collapsed_1235 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1237 = tensor.expand_shape %cst_262 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1238 = tensor.expand_shape %cst_263 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2942 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1236, %expanded_1238 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2943 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2942, %expanded_1237 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2944 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2945 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2944 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2946 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2945 : tensor<f16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2947 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2943 : tensor<2x1280x12x12xf32>) outs(%2946 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2948 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2947 : tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      %7524 = arith.negf %in : f16
      %7525 = math.exp %7524 : f16
      %7526 = arith.addf %7525, %cst_691 : f16
      %7527 = arith.divf %cst_691, %7526 : f16
      linalg.yield %7527 : f16
    } -> tensor<2x1280x12x12xf16>
    %2949 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2948, %2947 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.mulf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2950 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2951 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2950 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2952 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2951 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2953 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_264 : tensor<1280xf32>) outs(%2952 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2954 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2955 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2954 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2956 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2955 : tensor<f16>) outs(%1900 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280x3x3xf16>
    %2957 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_265 : tensor<1280x1280x3x3xf32>) outs(%2956 : tensor<1280x1280x3x3xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280x3x3xf16>
    %padded_1239 = tensor.pad %2949 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_690 : f16
    } : tensor<2x1280x12x12xf16> to tensor<2x1280x14x14xf16>
    %2958 = linalg.generic {indexing_maps = [#map11, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2953 : tensor<1280xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x1280x12x12xf16>
    %2959 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%padded_1239, %2957 : tensor<2x1280x14x14xf16>, tensor<1280x1280x3x3xf16>) outs(%2958 : tensor<2x1280x12x12xf16>) -> tensor<2x1280x12x12xf16>
    %2960 = linalg.generic {indexing_maps = [#map10, #map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2864, %2959 : tensor<2x1280x12x12xf16>, tensor<2x1280x12x12xf16>) outs(%2667 : tensor<2x1280x12x12xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x1280x12x12xf16>
    %2961 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2962 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2961 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %2963 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2962 : tensor<f32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x1280x12x12xf32>
    %2964 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2960 : tensor<2x1280x12x12xf16>) outs(%2963 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %collapsed_1240 = tensor.collapse_shape %2964 [[0], [1], [2, 3]] : tensor<2x1280x12x12xf32> into tensor<2x1280x144xf32>
    %expanded_1241 = tensor.expand_shape %collapsed_1240 [[0], [1, 2], [3]] : tensor<2x1280x144xf32> into tensor<2x32x40x144xf32>
    %2965 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1241 : tensor<2x32x40x144xf32>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2966 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2967 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2965 : tensor<2x32x40x144xf64>) outs(%2966 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2968 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2967 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2969 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2965, %2968 : tensor<2x32x40x144xf64>, tensor<2x32x1x1xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %in_2119: f64, %out: f64):
      %7524 = arith.subf %in, %in_2119 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2970 = linalg.generic {indexing_maps = [#map10, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2969 : tensor<2x32x40x144xf64>) outs(%2675 : tensor<2x32x40x144xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.mulf %in, %in : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x40x144xf64>
    %2971 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2972 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2970 : tensor<2x32x40x144xf64>) outs(%2971 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.addf %in, %out : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2973 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2972 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2974 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2973 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2975 = linalg.fill ins(%cst_695 : f64) outs(%81 : tensor<2x32x1x1xf64>) -> tensor<2x32x1x1xf64>
    %2976 = linalg.generic {indexing_maps = [#map10, #map12], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%expanded_1241 : tensor<2x32x40x144xf32>) outs(%2975 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x32x1x1xf64>
    %2977 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2976 : tensor<2x32x1x1xf64>) outs(%81 : tensor<2x32x1x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_706 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x32x1x1xf64>
    %2978 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2977 : tensor<2x32x1x1xf64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2979 = linalg.generic {indexing_maps = [#map12, #map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2974, %cst_0 : tensor<2x32x1x1xf32>, tensor<f64>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %in_2119: f64, %out: f32):
      %7524 = arith.truncf %in_2119 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x32x1x1xf32>
    %2980 = linalg.generic {indexing_maps = [#map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2979 : tensor<2x32x1x1xf32>) outs(%90 : tensor<2x32x1x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x1x1xf32>
    %2981 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1241, %2978 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %2982 = linalg.generic {indexing_maps = [#map10, #map12, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2981, %2980 : tensor<2x32x40x144xf32>, tensor<2x32x1x1xf32>) outs(%2692 : tensor<2x32x40x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x32x40x144xf32>
    %collapsed_1242 = tensor.collapse_shape %2982 [[0], [1, 2], [3]] : tensor<2x32x40x144xf32> into tensor<2x1280x144xf32>
    %expanded_1243 = tensor.expand_shape %collapsed_1242 [[0], [1], [2, 3]] : tensor<2x1280x144xf32> into tensor<2x1280x12x12xf32>
    %expanded_1244 = tensor.expand_shape %cst_266 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %expanded_1245 = tensor.expand_shape %cst_267 [[0, 1, 2, 3]] : tensor<1280xf32> into tensor<1x1280x1x1xf32>
    %2983 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1243, %expanded_1245 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2984 = linalg.generic {indexing_maps = [#map10, #map13, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2983, %expanded_1244 : tensor<2x1280x12x12xf32>, tensor<1x1280x1x1xf32>) outs(%2672 : tensor<2x1280x12x12xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x1280x12x12xf32>
    %2985 = tensor.empty() : tensor<2x12x12x1280xf32>
    %2986 = linalg.generic {indexing_maps = [#map10, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2984 : tensor<2x1280x12x12xf32>) outs(%2985 : tensor<2x12x12x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x12x12x1280xf32>
    %collapsed_1246 = tensor.collapse_shape %2986 [[0], [1, 2], [3]] : tensor<2x12x12x1280xf32> into tensor<2x144x1280xf32>
    %2987 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2988 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2987 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2989 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%2988 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %2990 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_268 : tensor<1280xf32>) outs(%2989 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %2991 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2992 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2991 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2993 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%2992 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %2994 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_269 : tensor<1280x1280xf32>) outs(%2993 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %2995 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %2996 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%2995 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %2997 = tensor.empty() : tensor<2x144x1280xf16>
    %2998 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2996 : tensor<f16>) outs(%2997 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x144x1280xf16>
    %2999 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1246 : tensor<2x144x1280xf32>) outs(%2998 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x144x1280xf16>
    %3000 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%2994 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1247 = tensor.collapse_shape %2999 [[0, 1], [2]] : tensor<2x144x1280xf16> into tensor<288x1280xf16>
    %3001 = tensor.empty() : tensor<288x1280xf16>
    %3002 = linalg.fill ins(%cst_690 : f16) outs(%3001 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3003 = linalg.matmul ins(%collapsed_1247, %3000 : tensor<288x1280xf16>, tensor<1280x1280xf16>) outs(%3002 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3004 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%2990, %3003 : tensor<1280xf16>, tensor<288x1280xf16>) outs(%3001 : tensor<288x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<288x1280xf16>
    %expanded_1248 = tensor.expand_shape %3004 [[0, 1], [2]] : tensor<288x1280xf16> into tensor<2x144x1280xf16>
    %3005 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3006 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3005 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %3007 = tensor.empty() : tensor<2x144x1280xf32>
    %3008 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3006 : tensor<f32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x144x1280xf32>
    %3009 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1248 : tensor<2x144x1280xf16>) outs(%3008 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3010 = tensor.empty() : tensor<2x144x1xf64>
    %3011 = linalg.fill ins(%cst_695 : f64) outs(%3010 : tensor<2x144x1xf64>) -> tensor<2x144x1xf64>
    %3012 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3009 : tensor<2x144x1280xf32>) outs(%3011 : tensor<2x144x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x144x1xf64>
    %3013 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3012 : tensor<2x144x1xf64>) outs(%3010 : tensor<2x144x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x144x1xf64>
    %3014 = tensor.empty() : tensor<2x144x1xf32>
    %3015 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3013 : tensor<2x144x1xf64>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1xf32>
    %3016 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3015 : tensor<2x144x1xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x144x1280xf32>
    %3017 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3009, %3016 : tensor<2x144x1280xf32>, tensor<2x144x1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3018 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3017 : tensor<2x144x1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3019 = linalg.fill ins(%cst_695 : f64) outs(%3010 : tensor<2x144x1xf64>) -> tensor<2x144x1xf64>
    %3020 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3018 : tensor<2x144x1280xf32>) outs(%3019 : tensor<2x144x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x144x1xf64>
    %3021 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3020 : tensor<2x144x1xf64>) outs(%3010 : tensor<2x144x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x144x1xf64>
    %3022 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3021 : tensor<2x144x1xf64>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1xf32>
    %3023 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3022 : tensor<2x144x1xf32>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x144x1xf32>
    %3024 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3023 : tensor<2x144x1xf32>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1xf32>
    %3025 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3024 : tensor<2x144x1xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x144x1280xf32>
    %3026 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3017, %3025 : tensor<2x144x1280xf32>, tensor<2x144x1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3027 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3026, %cst_270 : tensor<2x144x1280xf32>, tensor<1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.mulf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3028 = linalg.generic {indexing_maps = [#map16, #map18, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3027, %cst_271 : tensor<2x144x1280xf32>, tensor<1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.addf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3029 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3030 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3029 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3031 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%3030 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %3032 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_272 : tensor<1280x1280xf32>) outs(%3031 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %3033 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3034 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3033 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3035 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3034 : tensor<f16>) outs(%2997 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x144x1280xf16>
    %3036 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3028 : tensor<2x144x1280xf32>) outs(%3035 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x144x1280xf16>
    %3037 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%3032 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1249 = tensor.collapse_shape %3036 [[0, 1], [2]] : tensor<2x144x1280xf16> into tensor<288x1280xf16>
    %3038 = linalg.fill ins(%cst_690 : f16) outs(%3001 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3039 = linalg.matmul ins(%collapsed_1249, %3037 : tensor<288x1280xf16>, tensor<1280x1280xf16>) outs(%3038 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3040 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3041 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3040 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3042 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%3041 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %3043 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_273 : tensor<1280x1280xf32>) outs(%3042 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %3044 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3045 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3044 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3046 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3045 : tensor<f16>) outs(%2997 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x144x1280xf16>
    %3047 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3028 : tensor<2x144x1280xf32>) outs(%3046 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x144x1280xf16>
    %3048 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%3043 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1250 = tensor.collapse_shape %3047 [[0, 1], [2]] : tensor<2x144x1280xf16> into tensor<288x1280xf16>
    %3049 = linalg.fill ins(%cst_690 : f16) outs(%3001 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3050 = linalg.matmul ins(%collapsed_1250, %3048 : tensor<288x1280xf16>, tensor<1280x1280xf16>) outs(%3049 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3051 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3052 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3051 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3053 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%3052 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %3054 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_274 : tensor<1280x1280xf32>) outs(%3053 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %3055 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3056 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3055 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3057 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3056 : tensor<f16>) outs(%2997 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x144x1280xf16>
    %3058 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3028 : tensor<2x144x1280xf32>) outs(%3057 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<2x144x1280xf16>
    %3059 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%3054 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1251 = tensor.collapse_shape %3058 [[0, 1], [2]] : tensor<2x144x1280xf16> into tensor<288x1280xf16>
    %3060 = linalg.fill ins(%cst_690 : f16) outs(%3001 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3061 = linalg.matmul ins(%collapsed_1251, %3059 : tensor<288x1280xf16>, tensor<1280x1280xf16>) outs(%3060 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %expanded_1252 = tensor.expand_shape %3039 [[0, 1], [2, 3]] : tensor<288x1280xf16> into tensor<2x144x20x64xf16>
    %3062 = tensor.empty() : tensor<2x20x144x64xf16>
    %3063 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1252 : tensor<2x144x20x64xf16>) outs(%3062 : tensor<2x20x144x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x144x64xf16>
    %collapsed_1253 = tensor.collapse_shape %3063 [[0, 1], [2], [3]] : tensor<2x20x144x64xf16> into tensor<40x144x64xf16>
    %expanded_1254 = tensor.expand_shape %3050 [[0, 1], [2, 3]] : tensor<288x1280xf16> into tensor<2x144x20x64xf16>
    %3064 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1254 : tensor<2x144x20x64xf16>) outs(%3062 : tensor<2x20x144x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x144x64xf16>
    %collapsed_1255 = tensor.collapse_shape %3064 [[0, 1], [2], [3]] : tensor<2x20x144x64xf16> into tensor<40x144x64xf16>
    %expanded_1256 = tensor.expand_shape %3061 [[0, 1], [2, 3]] : tensor<288x1280xf16> into tensor<2x144x20x64xf16>
    %3065 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1256 : tensor<2x144x20x64xf16>) outs(%3062 : tensor<2x20x144x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x20x144x64xf16>
    %collapsed_1257 = tensor.collapse_shape %3065 [[0, 1], [2], [3]] : tensor<2x20x144x64xf16> into tensor<40x144x64xf16>
    %expanded_1258 = tensor.expand_shape %collapsed_1253 [[0], [1], [2, 3]] : tensor<40x144x64xf16> into tensor<40x144x64x1xf16>
    %3066 = tensor.empty() : tensor<40x144x1x64xf16>
    %3067 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1258 : tensor<40x144x64x1xf16>) outs(%3066 : tensor<40x144x1x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x1x64xf16>
    %expanded_1259 = tensor.expand_shape %collapsed_1255 [[0], [1], [2, 3]] : tensor<40x144x64xf16> into tensor<40x144x64x1xf16>
    %3068 = tensor.empty() : tensor<40x1x144x64xf16>
    %3069 = linalg.generic {indexing_maps = [#map10, #map21], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1259 : tensor<40x144x64x1xf16>) outs(%3068 : tensor<40x1x144x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x144x64xf16>
    %3070 = tensor.empty() : tensor<40x144x64x1xf16>
    %3071 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3067 : tensor<40x144x1x64xf16>) outs(%3070 : tensor<40x144x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x64x1xf16>
    %collapsed_1260 = tensor.collapse_shape %3071 [[0], [1], [2, 3]] : tensor<40x144x64x1xf16> into tensor<40x144x64xf16>
    %3072 = tensor.empty() : tensor<40x64x144x1xf16>
    %3073 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3069 : tensor<40x1x144x64xf16>) outs(%3072 : tensor<40x64x144x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x64x144x1xf16>
    %collapsed_1261 = tensor.collapse_shape %3073 [[0], [1], [2, 3]] : tensor<40x64x144x1xf16> into tensor<40x64x144xf16>
    %3074 = tensor.empty() : tensor<40x144x144xf16>
    %3075 = linalg.fill ins(%cst_690 : f16) outs(%3074 : tensor<40x144x144xf16>) -> tensor<40x144x144xf16>
    %3076 = linalg.batch_matmul ins(%collapsed_1260, %collapsed_1261 : tensor<40x144x64xf16>, tensor<40x64x144xf16>) outs(%3075 : tensor<40x144x144xf16>) -> tensor<40x144x144xf16>
    %expanded_1262 = tensor.expand_shape %3076 [[0], [1], [2, 3]] : tensor<40x144x144xf16> into tensor<40x144x1x144xf16>
    %3077 = tensor.empty() : tensor<40x144x144x1xf16>
    %3078 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1262 : tensor<40x144x1x144xf16>) outs(%3077 : tensor<40x144x144x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x144x1xf16>
    %collapsed_1263 = tensor.collapse_shape %3078 [[0], [1], [2, 3]] : tensor<40x144x144x1xf16> into tensor<40x144x144xf16>
    %3079 = linalg.generic {indexing_maps = [#map16, #map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1263, %cst : tensor<40x144x144xf16>, tensor<f64>) outs(%3074 : tensor<40x144x144xf16>) {
    ^bb0(%in: f16, %in_2119: f64, %out: f16):
      %7524 = arith.truncf %in_2119 : f64 to f16
      %7525 = arith.mulf %in, %7524 : f16
      linalg.yield %7525 : f16
    } -> tensor<40x144x144xf16>
    %3080 = tensor.empty() : tensor<40x144x144xf32>
    %3081 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3079 : tensor<40x144x144xf16>) outs(%3080 : tensor<40x144x144xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<40x144x144xf32>
    %3082 = tensor.empty() : tensor<40x144x1xf32>
    %3083 = linalg.fill ins(%cst_692 : f32) outs(%3082 : tensor<40x144x1xf32>) -> tensor<40x144x1xf32>
    %3084 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3081 : tensor<40x144x144xf32>) outs(%3083 : tensor<40x144x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.maxf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x144x1xf32>
    %3085 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3081, %3084 : tensor<40x144x144xf32>, tensor<40x144x1xf32>) outs(%3080 : tensor<40x144x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x144x144xf32>
    %3086 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3085 : tensor<40x144x144xf32>) outs(%3080 : tensor<40x144x144xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.exp %in : f32
      linalg.yield %7524 : f32
    } -> tensor<40x144x144xf32>
    %3087 = linalg.fill ins(%cst_697 : f32) outs(%3082 : tensor<40x144x1xf32>) -> tensor<40x144x1xf32>
    %3088 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3086 : tensor<40x144x144xf32>) outs(%3087 : tensor<40x144x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.addf %in, %out : f32
      linalg.yield %7524 : f32
    } -> tensor<40x144x1xf32>
    %3089 = linalg.generic {indexing_maps = [#map16, #map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3086, %3088 : tensor<40x144x144xf32>, tensor<40x144x1xf32>) outs(%3080 : tensor<40x144x144xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.divf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<40x144x144xf32>
    %3090 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3091 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3090 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3092 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3091 : tensor<f16>) outs(%3074 : tensor<40x144x144xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x144xf16>
    %3093 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3089 : tensor<40x144x144xf32>) outs(%3092 : tensor<40x144x144xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<40x144x144xf16>
    %expanded_1264 = tensor.expand_shape %3093 [[0], [1], [2, 3]] : tensor<40x144x144xf16> into tensor<40x144x144x1xf16>
    %3094 = tensor.empty() : tensor<40x144x1x144xf16>
    %3095 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1264 : tensor<40x144x144x1xf16>) outs(%3094 : tensor<40x144x1x144xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x1x144xf16>
    %expanded_1265 = tensor.expand_shape %collapsed_1257 [[0], [1], [2, 3]] : tensor<40x144x64xf16> into tensor<40x144x64x1xf16>
    %3096 = tensor.empty() : tensor<40x1x64x144xf16>
    %3097 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1265 : tensor<40x144x64x1xf16>) outs(%3096 : tensor<40x1x64x144xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x1x64x144xf16>
    %3098 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3095 : tensor<40x144x1x144xf16>) outs(%3077 : tensor<40x144x144x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x144x1xf16>
    %collapsed_1266 = tensor.collapse_shape %3098 [[0], [1], [2, 3]] : tensor<40x144x144x1xf16> into tensor<40x144x144xf16>
    %3099 = linalg.generic {indexing_maps = [#map10, #map22], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3097 : tensor<40x1x64x144xf16>) outs(%3070 : tensor<40x144x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x64x1xf16>
    %collapsed_1267 = tensor.collapse_shape %3099 [[0], [1], [2, 3]] : tensor<40x144x64x1xf16> into tensor<40x144x64xf16>
    %3100 = tensor.empty() : tensor<40x144x64xf16>
    %3101 = linalg.fill ins(%cst_690 : f16) outs(%3100 : tensor<40x144x64xf16>) -> tensor<40x144x64xf16>
    %3102 = linalg.batch_matmul ins(%collapsed_1266, %collapsed_1267 : tensor<40x144x144xf16>, tensor<40x144x64xf16>) outs(%3101 : tensor<40x144x64xf16>) -> tensor<40x144x64xf16>
    %expanded_1268 = tensor.expand_shape %3102 [[0], [1], [2, 3]] : tensor<40x144x64xf16> into tensor<40x144x1x64xf16>
    %3103 = linalg.generic {indexing_maps = [#map10, #map20], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1268 : tensor<40x144x1x64xf16>) outs(%3070 : tensor<40x144x64x1xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<40x144x64x1xf16>
    %collapsed_1269 = tensor.collapse_shape %3103 [[0], [1], [2, 3]] : tensor<40x144x64x1xf16> into tensor<40x144x64xf16>
    %expanded_1270 = tensor.expand_shape %collapsed_1269 [[0, 1], [2], [3]] : tensor<40x144x64xf16> into tensor<2x20x144x64xf16>
    %3104 = tensor.empty() : tensor<2x144x20x64xf16>
    %3105 = linalg.generic {indexing_maps = [#map10, #map19], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1270 : tensor<2x20x144x64xf16>) outs(%3104 : tensor<2x144x20x64xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<2x144x20x64xf16>
    %3106 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3107 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3106 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3108 = linalg.generic {indexing_maps = [#map1, #map], iterator_types = ["parallel"]} ins(%3107 : tensor<f16>) outs(%22 : tensor<1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280xf16>
    %3109 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%cst_275 : tensor<1280xf32>) outs(%3108 : tensor<1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280xf16>
    %3110 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3111 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3110 : tensor<f64>) outs(%20 : tensor<f16>) {
    ^bb0(%in: f64, %out: f16):
      %7524 = arith.truncf %in : f64 to f16
      linalg.yield %7524 : f16
    } -> tensor<f16>
    %3112 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%3111 : tensor<f16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %3113 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_276 : tensor<1280x1280xf32>) outs(%3112 : tensor<1280x1280xf16>) {
    ^bb0(%in: f32, %out: f16):
      %7524 = arith.truncf %in : f32 to f16
      linalg.yield %7524 : f16
    } -> tensor<1280x1280xf16>
    %3114 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "parallel"]} ins(%3113 : tensor<1280x1280xf16>) outs(%49 : tensor<1280x1280xf16>) {
    ^bb0(%in: f16, %out: f16):
      linalg.yield %in : f16
    } -> tensor<1280x1280xf16>
    %collapsed_1271 = tensor.collapse_shape %3105 [[0, 1], [2, 3]] : tensor<2x144x20x64xf16> into tensor<288x1280xf16>
    %3115 = linalg.fill ins(%cst_690 : f16) outs(%3001 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3116 = linalg.matmul ins(%collapsed_1271, %3114 : tensor<288x1280xf16>, tensor<1280x1280xf16>) outs(%3115 : tensor<288x1280xf16>) -> tensor<288x1280xf16>
    %3117 = linalg.generic {indexing_maps = [#map8, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%3109, %3116 : tensor<1280xf16>, tensor<288x1280xf16>) outs(%3001 : tensor<288x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<288x1280xf16>
    %expanded_1272 = tensor.expand_shape %3117 [[0, 1], [2]] : tensor<288x1280xf16> into tensor<2x144x1280xf16>
    %3118 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_1272, %expanded_1248 : tensor<2x144x1280xf16>, tensor<2x144x1280xf16>) outs(%2997 : tensor<2x144x1280xf16>) {
    ^bb0(%in: f16, %in_2119: f16, %out: f16):
      %7524 = arith.addf %in, %in_2119 : f16
      linalg.yield %7524 : f16
    } -> tensor<2x144x1280xf16>
    %3119 = linalg.fill ins(%cst_695 : f64) outs(%5 : tensor<f64>) -> tensor<f64>
    %3120 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3119 : tensor<f64>) outs(%7 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<f32>
    %3121 = linalg.generic {indexing_maps = [#map15, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3120 : tensor<f32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x144x1280xf32>
    %3122 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3118 : tensor<2x144x1280xf16>) outs(%3121 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f16, %out: f32):
      %7524 = arith.extf %in : f16 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3123 = linalg.fill ins(%cst_695 : f64) outs(%3010 : tensor<2x144x1xf64>) -> tensor<2x144x1xf64>
    %3124 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3122 : tensor<2x144x1280xf32>) outs(%3123 : tensor<2x144x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x144x1xf64>
    %3125 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3124 : tensor<2x144x1xf64>) outs(%3010 : tensor<2x144x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x144x1xf64>
    %3126 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3125 : tensor<2x144x1xf64>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1xf32>
    %3127 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3126 : tensor<2x144x1xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2x144x1280xf32>
    %3128 = linalg.generic {indexing_maps = [#map16, #map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3122, %3127 : tensor<2x144x1280xf32>, tensor<2x144x1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %in_2119: f32, %out: f32):
      %7524 = arith.subf %in, %in_2119 : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3129 = linalg.generic {indexing_maps = [#map16, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3128 : tensor<2x144x1280xf32>) outs(%3007 : tensor<2x144x1280xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.mulf %in, %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1280xf32>
    %3130 = linalg.fill ins(%cst_695 : f64) outs(%3010 : tensor<2x144x1xf64>) -> tensor<2x144x1xf64>
    %3131 = linalg.generic {indexing_maps = [#map16, #map17], iterator_types = ["parallel", "parallel", "reduction"]} ins(%3129 : tensor<2x144x1280xf32>) outs(%3130 : tensor<2x144x1xf64>) {
    ^bb0(%in: f32, %out: f64):
      %7524 = arith.extf %in : f32 to f64
      %7525 = arith.addf %7524, %out : f64
      linalg.yield %7525 : f64
    } -> tensor<2x144x1xf64>
    %3132 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3131 : tensor<2x144x1xf64>) outs(%3010 : tensor<2x144x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %7524 = arith.divf %in, %cst_705 : f64
      linalg.yield %7524 : f64
    } -> tensor<2x144x1xf64>
    %3133 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3132 : tensor<2x144x1xf64>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %7524 = arith.truncf %in : f64 to f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1xf32>
    %3134 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3133 : tensor<2x144x1xf32>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = arith.truncf %cst_696 : f64 to f32
      %7525 = arith.addf %in, %7524 : f32
      linalg.yield %7525 : f32
    } -> tensor<2x144x1xf32>
    %3135 = linalg.generic {indexing_maps = [#map17, #map16], iterator_types = ["parallel", "parallel", "parallel"]} ins(%3134 : tensor<2x144x1xf32>) outs(%3014 : tensor<2x144x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %7524 = math.rsqrt %in : f32
      linalg.yield %7524 : f32
    } -> tensor<2x144x1xf32>
    %3136 = linalg.generic {i