Skip to content

Instantly share code, notes, and snippets.

@AmosLewis

AmosLewis/mobilebert-uncased.default.pytorch.linalg.elide.mlir

Created June 24, 2024 23:35
Show Gist options
  • Save AmosLewis/2faceeb13d63ea5574424067b6a12514 to your computer and use it in GitHub Desktop.
Save AmosLewis/2faceeb13d63ea5574424067b6a12514 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
mobilebert-uncased.default.pytorch.linalg.elide.mlir
#map = affine_map<(d0, d1) -> (0, d1)>
#map1 = affine_map<(d0, d1) -> (d0, d1)>
#map2 = affine_map<(d0, d1) -> (d0)>
#map3 = affine_map<(d0, d1) -> (d1, d0)>
#map4 = affine_map<()[s0, s1] -> (s0 + s1 + 128)>
#map5 = affine_map<(d0, d1, d2) -> (0, d1, d2)>
#map6 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map7 = affine_map<(d0, d1, d2) -> (d1, d2)>
#map8 = affine_map<(d0, d1, d2) -> (d2)>
#map9 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map10 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map11 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d3, d2)>
#map12 = affine_map<(d0, d1, d2, d3) -> (0, d1, d2, d3)>
#map13 = affine_map<(d0, d1, d2, d3) -> ()>
#map14 = affine_map<(d0, d1, d2, d3) -> (0, 0, 0, d3)>
#map15 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
#map16 = affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>
#map17 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>
#map18 = affine_map<(d0, d1, d2) -> (d1)>
#map19 = affine_map<(d0, d1) -> (d1)>
module {
ml_program.global private mutable @global_seed(dense<0> : tensor<i64>) : tensor<i64>
func.func @main_graph(%arg0: tensor<1x128xi64>) -> tensor<1x2xf32> {
%cst = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%c0_i64 = arith.constant 0 : i64
%cst_0 = arith.constant 0.000000e+00 : f32
%cst_1 = arith.constant 0xFF800000 : f32
%c512_i64 = arith.constant 512 : i64
%c30522_i64 = arith.constant 30522 : i64
%c384_i64 = arith.constant 384 : i64
%cst_2 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_3 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_4 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_5 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_6 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_7 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_8 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_9 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_10 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_11 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_12 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_13 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_14 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_15 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_16 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_17 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_18 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_19 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_20 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_21 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_22 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_23 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_24 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_25 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_26 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_27 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_28 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_29 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_30 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_31 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_32 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_33 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_34 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_35 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_36 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_37 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_38 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_39 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_40 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_41 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_42 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_43 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_44 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_45 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_46 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_47 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_48 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_49 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_50 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_51 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_52 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_53 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_54 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_55 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_56 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_57 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_58 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_59 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_60 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_61 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_62 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_63 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_64 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_65 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_66 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_67 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_68 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_69 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_70 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_71 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_72 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_73 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_74 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_75 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_76 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_77 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_78 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_79 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_80 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_81 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_82 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_83 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_84 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_85 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_86 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_87 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_88 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_89 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_90 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_91 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_92 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_93 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_94 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_95 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_96 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_97 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_98 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_99 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_100 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_101 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_102 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_103 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_104 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_105 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_106 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_107 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_108 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_109 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_110 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_111 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_112 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_113 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_114 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_115 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_116 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_117 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_118 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_119 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_120 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_121 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_122 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_123 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_124 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_125 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_126 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_127 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_128 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_129 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_130 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_131 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_132 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_133 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_134 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_135 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_136 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_137 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_138 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_139 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_140 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_141 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_142 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_143 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_144 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_145 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_146 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_147 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_148 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_149 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_150 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_151 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_152 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_153 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_154 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_155 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_156 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_157 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_158 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_159 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_160 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_161 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_162 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_163 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_164 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_165 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_166 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_167 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_168 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_169 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_170 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_171 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_172 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_173 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_174 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_175 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_176 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_177 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_178 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_179 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_180 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_181 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_182 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_183 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_184 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_185 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_186 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_187 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_188 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_189 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_190 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_191 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_192 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_193 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_194 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_195 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_196 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_197 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_198 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_199 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_200 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_201 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_202 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_203 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_204 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_205 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_206 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_207 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_208 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_209 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_210 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_211 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_212 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_213 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_214 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_215 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_216 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_217 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_218 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_219 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_220 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_221 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_222 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_223 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_224 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_225 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_226 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_227 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_228 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_229 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_230 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_231 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_232 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_233 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_234 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_235 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_236 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_237 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_238 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_239 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_240 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_241 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_242 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_243 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_244 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_245 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_246 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_247 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_248 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_249 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_250 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_251 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_252 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_253 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_254 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_255 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_256 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_257 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_258 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_259 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_260 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_261 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_262 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_263 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_264 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_265 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_266 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_267 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_268 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_269 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_270 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_271 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_272 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_273 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_274 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_275 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_276 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_277 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_278 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_279 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_280 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_281 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_282 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_283 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_284 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_285 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_286 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_287 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_288 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_289 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_290 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_291 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_292 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_293 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_294 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_295 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_296 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_297 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_298 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_299 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_300 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_301 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_302 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_303 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_304 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_305 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_306 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_307 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_308 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_309 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_310 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_311 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_312 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_313 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_314 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_315 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_316 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_317 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_318 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_319 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_320 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_321 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_322 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_323 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_324 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_325 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_326 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_327 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_328 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_329 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_330 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_331 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_332 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_333 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_334 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_335 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_336 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_337 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_338 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_339 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_340 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_341 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_342 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_343 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_344 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_345 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_346 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_347 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_348 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_349 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_350 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_351 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_352 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_353 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_354 = arith.constant dense_resource<__elided__> : tensor<128x512xf32>
%cst_355 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_356 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_357 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_358 = arith.constant dense_resource<__elided__> : tensor<128x128xf32>
%cst_359 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_360 = arith.constant dense_resource<__elided__> : tensor<512x128xf32>
%cst_361 = arith.constant dense_resource<__elided__> : tensor<384x512xf32>
%cst_362 = arith.constant dense<0.000000e+00> : tensor<2xf32>
%cst_363 = arith.constant dense_resource<__elided__> : tensor<2x512xf32>
%cst_364 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_365 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_366 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_367 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_368 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_369 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_370 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_371 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_372 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_373 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_374 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_375 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_376 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_377 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_378 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_379 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_380 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_381 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_382 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_383 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_384 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_385 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_386 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_387 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_388 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_389 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_390 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_391 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_392 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_393 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_394 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_395 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_396 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_397 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_398 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_399 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_400 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_401 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_402 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_403 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_404 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_405 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_406 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_407 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_408 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_409 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_410 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_411 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_412 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_413 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_414 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_415 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_416 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_417 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_418 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_419 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_420 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_421 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_422 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_423 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_424 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_425 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_426 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_427 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_428 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_429 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_430 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_431 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_432 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_433 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_434 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_435 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_436 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_437 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_438 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_439 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_440 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_441 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_442 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_443 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_444 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_445 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_446 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_447 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_448 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_449 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_450 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_451 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_452 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_453 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_454 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_455 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_456 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_457 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_458 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_459 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_460 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_461 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_462 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_463 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_464 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_465 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_466 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_467 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_468 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_469 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_470 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_471 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_472 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_473 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_474 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_475 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_476 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_477 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_478 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_479 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_480 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_481 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_482 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_483 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_484 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_485 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_486 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_487 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_488 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_489 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_490 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_491 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_492 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_493 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_494 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_495 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_496 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_497 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_498 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_499 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_500 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_501 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_502 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_503 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_504 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_505 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_506 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_507 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_508 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_509 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_510 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_511 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_512 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_513 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_514 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_515 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_516 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_517 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_518 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_519 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_520 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_521 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_522 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_523 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_524 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_525 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_526 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_527 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_528 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_529 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_530 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_531 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_532 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_533 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_534 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_535 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_536 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_537 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_538 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_539 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_540 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_541 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_542 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_543 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_544 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_545 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_546 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_547 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_548 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_549 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_550 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_551 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_552 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_553 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_554 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_555 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_556 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_557 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_558 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_559 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_560 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_561 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_562 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_563 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_564 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_565 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_566 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_567 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_568 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_569 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_570 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_571 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_572 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_573 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_574 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_575 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_576 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_577 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_578 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_579 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_580 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_581 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_582 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_583 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_584 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_585 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_586 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_587 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_588 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_589 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_590 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_591 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_592 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_593 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_594 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_595 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_596 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_597 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_598 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_599 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_600 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_601 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_602 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_603 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_604 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_605 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_606 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_607 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_608 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_609 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_610 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_611 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_612 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_613 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_614 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_615 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_616 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_617 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_618 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_619 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_620 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_621 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_622 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_623 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_624 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_625 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_626 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_627 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_628 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_629 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_630 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_631 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_632 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_633 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_634 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_635 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_636 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_637 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_638 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_639 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_640 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_641 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_642 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_643 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_644 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_645 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_646 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_647 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_648 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_649 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_650 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_651 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_652 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_653 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_654 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_655 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_656 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_657 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_658 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_659 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_660 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_661 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_662 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_663 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_664 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_665 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_666 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_667 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_668 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_669 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_670 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_671 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_672 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_673 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_674 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_675 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_676 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_677 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_678 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_679 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_680 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_681 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_682 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_683 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_684 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_685 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_686 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_687 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_688 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_689 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_690 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_691 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_692 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_693 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_694 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_695 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_696 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_697 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_698 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_699 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_700 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_701 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_702 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_703 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_704 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_705 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_706 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_707 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_708 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_709 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_710 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_711 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_712 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_713 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_714 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_715 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_716 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_717 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_718 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_719 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_720 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_721 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_722 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_723 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_724 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_725 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_726 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_727 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_728 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_729 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_730 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_731 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_732 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_733 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_734 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_735 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_736 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_737 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_738 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_739 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_740 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_741 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_742 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_743 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_744 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_745 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_746 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_747 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_748 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_749 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_750 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_751 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_752 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_753 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_754 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_755 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_756 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_757 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_758 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_759 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_760 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_761 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_762 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_763 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_764 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_765 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_766 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_767 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_768 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_769 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_770 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_771 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_772 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_773 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_774 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_775 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_776 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_777 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_778 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_779 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_780 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_781 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_782 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_783 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_784 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_785 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_786 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_787 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_788 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_789 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_790 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_791 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_792 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_793 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_794 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_795 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_796 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_797 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_798 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_799 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_800 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_801 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_802 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_803 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_804 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_805 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_806 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_807 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_808 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_809 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_810 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_811 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_812 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_813 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_814 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_815 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_816 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_817 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_818 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_819 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_820 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_821 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_822 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_823 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_824 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_825 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_826 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_827 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_828 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_829 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_830 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_831 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_832 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_833 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_834 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_835 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_836 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_837 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_838 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_839 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_840 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_841 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_842 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_843 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_844 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_845 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_846 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_847 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_848 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_849 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_850 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_851 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_852 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_853 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_854 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_855 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_856 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_857 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_858 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_859 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_860 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_861 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_862 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_863 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_864 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_865 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_866 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_867 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_868 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_869 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_870 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_871 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_872 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_873 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_874 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_875 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_876 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_877 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_878 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_879 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_880 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_881 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_882 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_883 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_884 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_885 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_886 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_887 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_888 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_889 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_890 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_891 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_892 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_893 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_894 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_895 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_896 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_897 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_898 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_899 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_900 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_901 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_902 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_903 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_904 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_905 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_906 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_907 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_908 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_909 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_910 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_911 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_912 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_913 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_914 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_915 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_916 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_917 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_918 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_919 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_920 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_921 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_922 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_923 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_924 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_925 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_926 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_927 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_928 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_929 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_930 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_931 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_932 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_933 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_934 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_935 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_936 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_937 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_938 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_939 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_940 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_941 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_942 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_943 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_944 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_945 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_946 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_947 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_948 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_949 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_950 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_951 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_952 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_953 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_954 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_955 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_956 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_957 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_958 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_959 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_960 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_961 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_962 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_963 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_964 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_965 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_966 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_967 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_968 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_969 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_970 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_971 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_972 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_973 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_974 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_975 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_976 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_977 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_978 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_979 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_980 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_981 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_982 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_983 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_984 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_985 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_986 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_987 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_988 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_989 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_990 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_991 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_992 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_993 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_994 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_995 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_996 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_997 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_998 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_999 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1000 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1001 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1002 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1003 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1004 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1005 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1006 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1007 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1008 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1009 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1010 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1011 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1012 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1013 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1014 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1015 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1016 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1017 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1018 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1019 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1020 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1021 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1022 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1023 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1024 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1025 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1026 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1027 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1028 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1029 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1030 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1031 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1032 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1033 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1034 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1035 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1036 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1037 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1038 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1039 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1040 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1041 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1042 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1043 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1044 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1045 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1046 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1047 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1048 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1049 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1050 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1051 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1052 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1053 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1054 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1055 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1056 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1057 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1058 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1059 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1060 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1061 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1062 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1063 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1064 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1065 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1066 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1067 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1068 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1069 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1070 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1071 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1072 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1073 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1074 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1075 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1076 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1077 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1078 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1079 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1080 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1081 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1082 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1083 = arith.constant dense_resource<__elided__> : tensor<128xf32>
%cst_1084 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1085 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1086 = arith.constant dense_resource<__elided__> : tensor<512xf32>
%cst_1087 = arith.constant dense_resource<__elided__> : tensor<2x512xf32>
%cst_1088 = arith.constant dense_resource<__elided__> : tensor<512x512xf32>
%cst_1089 = arith.constant dense_resource<__elided__> : tensor<30522x128xf32>
%cst_1090 = arith.constant dense<0> : tensor<1x128xi64>
%cst_1091 = arith.constant dense_resource<__elided__> : tensor<1x128xi64>
%cst_1092 = arith.constant dense<5.65685415> : tensor<f32>
%cst_1093 = arith.constant dense<-0.000000e+00> : tensor<1x1x1x128xf32>
%cst_1094 = arith.constant dense_resource<__elided__> : tensor<6xi64>
%cst_1095 = arith.constant dense_resource<__elided__> : tensor<6xi64>
%cst_1096 = arith.constant dense<0> : tensor<1xi64>
%c2 = arith.constant 2 : index
%0 = tensor.empty() : tensor<1x128xi1>
%1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel"]} ins(%arg0 : tensor<1x128xi64>) outs(%0 : tensor<1x128xi1>) {
^bb0(%in: i64, %out: i1):
%2492 = arith.cmpi slt, %in, %c0_i64 : i64
linalg.yield %2492 : i1
} -> tensor<1x128xi1>
%2 = tensor.empty() : tensor<1x128xi64>
%3 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel"]} ins(%arg0 : tensor<1x128xi64>) outs(%2 : tensor<1x128xi64>) {
^bb0(%in: i64, %out: i64):
%2492 = arith.addi %in, %c30522_i64 : i64
linalg.yield %2492 : i64
} -> tensor<1x128xi64>
%4 = linalg.generic {indexing_maps = [#map, #map, #map, #map1], iterator_types = ["parallel", "parallel"]} ins(%1, %3, %arg0 : tensor<1x128xi1>, tensor<1x128xi64>, tensor<1x128xi64>) outs(%2 : tensor<1x128xi64>) {
^bb0(%in: i1, %in_1408: i64, %in_1409: i64, %out: i64):
%2492 = arith.select %in, %in_1408, %in_1409 : i64
linalg.yield %2492 : i64
} -> tensor<1x128xi64>
%collapsed = tensor.collapse_shape %4 [[0, 1]] : tensor<1x128xi64> into tensor<128xi64>
%5 = tensor.empty() : tensor<128x128xf32>
%6 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel"]} ins(%collapsed : tensor<128xi64>) outs(%5 : tensor<128x128xf32>) {
^bb0(%in: i64, %out: f32):
%2492 = arith.index_cast %in : i64 to index
%2493 = linalg.index 1 : index
%extracted_1408 = tensor.extract %cst_1089[%2492, %2493] : tensor<30522x128xf32>
linalg.yield %extracted_1408 : f32
} -> tensor<128x128xf32>
%expanded = tensor.expand_shape %6 [[0, 1], [2]] output_shape [1, 128, 128] : tensor<128x128xf32> into tensor<1x128x128xf32>
%extracted_slice = tensor.extract_slice %expanded[0, 1, 0] [1, 127, 128] [1, 1, 1] : tensor<1x128x128xf32> to tensor<1x127x128xf32>
%expanded_1097 = tensor.expand_shape %cst_1094 [[0, 1]] output_shape [3, 2] : tensor<6xi64> into tensor<3x2xi64>
%7 = tensor.empty() : tensor<2x3xi64>
%8 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel"]} ins(%expanded_1097 : tensor<3x2xi64>) outs(%7 : tensor<2x3xi64>) {
^bb0(%in: i64, %out: i64):
linalg.yield %in : i64
} -> tensor<2x3xi64>
%collapsed_1098 = tensor.collapse_shape %8 [[0, 1]] : tensor<2x3xi64> into tensor<6xi64>
%extracted_slice_1099 = tensor.extract_slice %collapsed_1098[0] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1100 = tensor.collapse_shape %extracted_slice_1099 [] : tensor<1xi64> into tensor<i64>
%extracted = tensor.extract %collapsed_1100[] : tensor<i64>
%extracted_slice_1101 = tensor.extract_slice %collapsed_1098[1] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1102 = tensor.collapse_shape %extracted_slice_1101 [] : tensor<1xi64> into tensor<i64>
%extracted_1103 = tensor.extract %collapsed_1102[] : tensor<i64>
%extracted_slice_1104 = tensor.extract_slice %collapsed_1098[2] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1105 = tensor.collapse_shape %extracted_slice_1104 [] : tensor<1xi64> into tensor<i64>
%extracted_1106 = tensor.extract %collapsed_1105[] : tensor<i64>
%extracted_slice_1107 = tensor.extract_slice %collapsed_1098[3] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1108 = tensor.collapse_shape %extracted_slice_1107 [] : tensor<1xi64> into tensor<i64>
%extracted_1109 = tensor.extract %collapsed_1108[] : tensor<i64>
%extracted_slice_1110 = tensor.extract_slice %collapsed_1098[4] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1111 = tensor.collapse_shape %extracted_slice_1110 [] : tensor<1xi64> into tensor<i64>
%extracted_1112 = tensor.extract %collapsed_1111[] : tensor<i64>
%extracted_slice_1113 = tensor.extract_slice %collapsed_1098[5] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1114 = tensor.collapse_shape %extracted_slice_1113 [] : tensor<1xi64> into tensor<i64>
%extracted_1115 = tensor.extract %collapsed_1114[] : tensor<i64>
%9 = arith.index_cast %extracted_1106 : i64 to index
%10 = arith.index_cast %extracted_1115 : i64 to index
%11 = arith.index_cast %extracted_1103 : i64 to index
%12 = arith.index_cast %extracted_1112 : i64 to index
%13 = arith.index_cast %extracted : i64 to index
%14 = arith.index_cast %extracted_1109 : i64 to index
%padded = tensor.pad %extracted_slice low[%9, %11, %13] high[%10, %12, %14] {
^bb0(%arg1: index, %arg2: index, %arg3: index):
tensor.yield %cst_0 : f32
} : tensor<1x127x128xf32> to tensor<?x?x?xf32>
%extracted_slice_1116 = tensor.extract_slice %expanded[0, 0, 0] [1, 127, 128] [1, 1, 1] : tensor<1x128x128xf32> to tensor<1x127x128xf32>
%expanded_1117 = tensor.expand_shape %cst_1095 [[0, 1]] output_shape [3, 2] : tensor<6xi64> into tensor<3x2xi64>
%15 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel"]} ins(%expanded_1117 : tensor<3x2xi64>) outs(%7 : tensor<2x3xi64>) {
^bb0(%in: i64, %out: i64):
linalg.yield %in : i64
} -> tensor<2x3xi64>
%collapsed_1118 = tensor.collapse_shape %15 [[0, 1]] : tensor<2x3xi64> into tensor<6xi64>
%extracted_slice_1119 = tensor.extract_slice %collapsed_1118[0] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1120 = tensor.collapse_shape %extracted_slice_1119 [] : tensor<1xi64> into tensor<i64>
%extracted_1121 = tensor.extract %collapsed_1120[] : tensor<i64>
%extracted_slice_1122 = tensor.extract_slice %collapsed_1118[1] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1123 = tensor.collapse_shape %extracted_slice_1122 [] : tensor<1xi64> into tensor<i64>
%extracted_1124 = tensor.extract %collapsed_1123[] : tensor<i64>
%extracted_slice_1125 = tensor.extract_slice %collapsed_1118[2] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1126 = tensor.collapse_shape %extracted_slice_1125 [] : tensor<1xi64> into tensor<i64>
%extracted_1127 = tensor.extract %collapsed_1126[] : tensor<i64>
%extracted_slice_1128 = tensor.extract_slice %collapsed_1118[3] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1129 = tensor.collapse_shape %extracted_slice_1128 [] : tensor<1xi64> into tensor<i64>
%extracted_1130 = tensor.extract %collapsed_1129[] : tensor<i64>
%extracted_slice_1131 = tensor.extract_slice %collapsed_1118[4] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1132 = tensor.collapse_shape %extracted_slice_1131 [] : tensor<1xi64> into tensor<i64>
%extracted_1133 = tensor.extract %collapsed_1132[] : tensor<i64>
%extracted_slice_1134 = tensor.extract_slice %collapsed_1118[5] [1] [1] : tensor<6xi64> to tensor<1xi64>
%collapsed_1135 = tensor.collapse_shape %extracted_slice_1134 [] : tensor<1xi64> into tensor<i64>
%extracted_1136 = tensor.extract %collapsed_1135[] : tensor<i64>
%16 = arith.index_cast %extracted_1127 : i64 to index
%17 = arith.index_cast %extracted_1136 : i64 to index
%18 = arith.index_cast %extracted_1124 : i64 to index
%19 = arith.index_cast %extracted_1133 : i64 to index
%20 = arith.index_cast %extracted_1121 : i64 to index
%21 = arith.index_cast %extracted_1130 : i64 to index
%padded_1137 = tensor.pad %extracted_slice_1116 low[%16, %18, %20] high[%17, %19, %21] {
^bb0(%arg1: index, %arg2: index, %arg3: index):
tensor.yield %cst_0 : f32
} : tensor<1x127x128xf32> to tensor<?x?x?xf32>
%concat = tensor.concat dim(2) %padded, %expanded, %padded_1137 : (tensor<?x?x?xf32>, tensor<1x128x128xf32>, tensor<?x?x?xf32>) -> tensor<1x128x?xf32>
%dim = tensor.dim %padded, %c2 : tensor<?x?x?xf32>
%dim_1138 = tensor.dim %padded_1137, %c2 : tensor<?x?x?xf32>
%22 = affine.apply #map4()[%dim, %dim_1138]
%23 = arith.index_cast %22 : index to i64
%24 = arith.cmpi eq, %23, %c384_i64 : i64
cf.assert %24, "mismatching contracting dimension"
%25 = tensor.empty() : tensor<1x128x384xf32>
%26 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%concat : tensor<1x128x?xf32>) outs(%25 : tensor<1x128x384xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x384xf32>
%27 = tensor.empty() : tensor<1x384x512xf32>
%28 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_361 : tensor<384x512xf32>) outs(%27 : tensor<1x384x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x384x512xf32>
%29 = tensor.empty() : tensor<1x128x512xf32>
%30 = linalg.fill ins(%cst_0 : f32) outs(%29 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%31 = linalg.batch_matmul ins(%26, %28 : tensor<1x128x384xf32>, tensor<1x384x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%32 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1086, %31 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%33 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel"]} ins(%cst_1091 : tensor<1x128xi64>) outs(%0 : tensor<1x128xi1>) {
^bb0(%in: i64, %out: i1):
%2492 = arith.cmpi slt, %in, %c0_i64 : i64
linalg.yield %2492 : i1
} -> tensor<1x128xi1>
%34 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel"]} ins(%cst_1091 : tensor<1x128xi64>) outs(%2 : tensor<1x128xi64>) {
^bb0(%in: i64, %out: i64):
%2492 = arith.addi %in, %c512_i64 : i64
linalg.yield %2492 : i64
} -> tensor<1x128xi64>
%35 = linalg.generic {indexing_maps = [#map, #map, #map, #map1], iterator_types = ["parallel", "parallel"]} ins(%33, %34, %cst_1091 : tensor<1x128xi1>, tensor<1x128xi64>, tensor<1x128xi64>) outs(%2 : tensor<1x128xi64>) {
^bb0(%in: i1, %in_1408: i64, %in_1409: i64, %out: i64):
%2492 = arith.select %in, %in_1408, %in_1409 : i64
linalg.yield %2492 : i64
} -> tensor<1x128xi64>
%collapsed_1139 = tensor.collapse_shape %35 [[0, 1]] : tensor<1x128xi64> into tensor<128xi64>
%36 = tensor.empty() : tensor<128x512xf32>
%37 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel"]} ins(%collapsed_1139 : tensor<128xi64>) outs(%36 : tensor<128x512xf32>) {
^bb0(%in: i64, %out: f32):
%2492 = arith.index_cast %in : i64 to index
%2493 = linalg.index 1 : index
%extracted_1408 = tensor.extract %cst_1088[%2492, %2493] : tensor<512x512xf32>
linalg.yield %extracted_1408 : f32
} -> tensor<128x512xf32>
%expanded_1140 = tensor.expand_shape %37 [[0, 1], [2]] output_shape [1, 128, 512] : tensor<128x512xf32> into tensor<1x128x512xf32>
%collapsed_1141 = tensor.collapse_shape %cst_1090 [[0, 1]] : tensor<1x128xi64> into tensor<128xi64>
%38 = linalg.generic {indexing_maps = [#map2, #map1], iterator_types = ["parallel", "parallel"]} ins(%collapsed_1141 : tensor<128xi64>) outs(%36 : tensor<128x512xf32>) {
^bb0(%in: i64, %out: f32):
%2492 = arith.index_cast %in : i64 to index
%2493 = linalg.index 1 : index
%extracted_1408 = tensor.extract %cst_1087[%2492, %2493] : tensor<2x512xf32>
linalg.yield %extracted_1408 : f32
} -> tensor<128x512xf32>
%expanded_1142 = tensor.expand_shape %38 [[0, 1], [2]] output_shape [1, 128, 512] : tensor<128x512xf32> into tensor<1x128x512xf32>
%39 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%32, %expanded_1140 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%40 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%39, %expanded_1142 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%41 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%40, %cst_1084 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%42 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%41, %cst_1085 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%43 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%42 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%44 = tensor.empty() : tensor<1x512x128xf32>
%45 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_360 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%46 = tensor.empty() : tensor<1x128x128xf32>
%47 = linalg.fill ins(%cst_0 : f32) outs(%46 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%48 = linalg.batch_matmul ins(%43, %45 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%49 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1070, %48 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%50 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%49, %cst_1069 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%51 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%50, %cst_1080 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%52 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_359 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%53 = linalg.batch_matmul ins(%43, %52 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%54 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1068, %53 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%55 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%54, %cst_1066 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%56 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%55, %cst_1067 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%57 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%56 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%58 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_358 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%59 = linalg.batch_matmul ins(%57, %58 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%60 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1083, %59 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%61 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_357 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%62 = linalg.batch_matmul ins(%57, %61 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%63 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1082, %62 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%64 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_356 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%65 = linalg.batch_matmul ins(%43, %64 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%66 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1081, %65 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1143 = tensor.expand_shape %60 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%67 = tensor.empty() : tensor<1x4x128x32xf32>
%68 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1143 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1144 = tensor.expand_shape %63 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1145 = tensor.expand_shape %66 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%69 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1145 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%70 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1144 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%71 = tensor.empty() : tensor<1x4x32x128xf32>
%72 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%70 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%73 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%68 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%74 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%72 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1146 = tensor.collapse_shape %73 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1147 = tensor.collapse_shape %74 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%75 = tensor.empty() : tensor<4x128x128xf32>
%76 = linalg.fill ins(%cst_0 : f32) outs(%75 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%77 = linalg.batch_matmul ins(%collapsed_1146, %collapsed_1147 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1148 = tensor.expand_shape %77 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%78 = tensor.empty() : tensor<1x4x128x128xf32>
%79 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1148, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%80 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%79, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%81 = tensor.empty() : tensor<1x4x128xi64>
%82 = linalg.fill ins(%c0_i64 : i64) outs(%81 : tensor<1x4x128xi64>) -> tensor<1x4x128xi64>
%83 = tensor.empty() : tensor<1x4x128xf32>
%84 = linalg.fill ins(%cst_1 : f32) outs(%83 : tensor<1x4x128xf32>) -> tensor<1x4x128xf32>
%85:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%80 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1149 = tensor.expand_shape %85#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%86 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%80, %expanded_1149 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%87 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%86 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%88 = tensor.empty() : tensor<1x4x128x1xf32>
%89 = linalg.fill ins(%cst_0 : f32) outs(%88 : tensor<1x4x128x1xf32>) -> tensor<1x4x128x1xf32>
%90 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%87 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%91 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%87, %90 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%92 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%91 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%93 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%69 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1150 = tensor.collapse_shape %92 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1151 = tensor.collapse_shape %93 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%94 = tensor.empty() : tensor<4x128x32xf32>
%95 = linalg.fill ins(%cst_0 : f32) outs(%94 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%96 = linalg.batch_matmul ins(%collapsed_1150, %collapsed_1151 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1152 = tensor.expand_shape %96 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%97 = tensor.empty() : tensor<1x128x4x32xf32>
%98 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1152 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1153 = tensor.collapse_shape %98 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%99 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1153 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%100 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_355 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%101 = linalg.batch_matmul ins(%99, %100 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%102 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1080, %101 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%103 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%102, %51 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%104 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%103, %cst_1078 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%105 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%104, %cst_1079 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%106 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%105 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%107 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_354 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%108 = linalg.batch_matmul ins(%106, %107 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%109 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1065, %108 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%110 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%109 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%111 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%110 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%112 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_353 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%113 = linalg.batch_matmul ins(%111, %112 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%114 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1064, %113 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%115 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%114, %105 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%116 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%115, %cst_1062 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%117 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%116, %cst_1063 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%118 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%117 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%119 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_352 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%120 = linalg.batch_matmul ins(%118, %119 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%121 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1061, %120 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%122 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%121 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%123 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%122 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%124 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_351 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%125 = linalg.batch_matmul ins(%123, %124 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%126 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1060, %125 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%127 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%126, %117 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%128 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%127, %cst_1058 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%129 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%128, %cst_1059 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%130 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%129 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%131 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_350 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%132 = linalg.batch_matmul ins(%130, %131 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%133 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1057, %132 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%134 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%133 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%135 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%134 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%136 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_349 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%137 = linalg.batch_matmul ins(%135, %136 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%138 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1056, %137 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%139 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%138, %129 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%140 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%139, %cst_1054 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%141 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%140, %cst_1055 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%142 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%141 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%143 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_348 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%144 = linalg.batch_matmul ins(%142, %143 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%145 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1077, %144 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%146 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%145 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%147 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%146 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%148 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_347 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%149 = linalg.batch_matmul ins(%147, %148 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%150 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1076, %149 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%151 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%150, %141 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%152 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%151, %cst_1074 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%153 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%152, %cst_1075 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%154 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%153 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%155 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_346 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%156 = linalg.batch_matmul ins(%154, %155 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%157 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1073, %156 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%158 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%157, %42 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%159 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%158, %cst_1071 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%160 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%159, %cst_1072 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%161 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%160 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%162 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_345 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%163 = linalg.batch_matmul ins(%161, %162 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%164 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1040, %163 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%165 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%164, %cst_1039 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%166 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%165, %cst_1050 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%167 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_344 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%168 = linalg.batch_matmul ins(%161, %167 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%169 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1038, %168 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%170 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%169, %cst_1036 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%171 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%170, %cst_1037 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%172 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%171 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%173 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_343 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%174 = linalg.batch_matmul ins(%172, %173 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%175 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1053, %174 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%176 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_342 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%177 = linalg.batch_matmul ins(%172, %176 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%178 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1052, %177 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%179 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_341 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%180 = linalg.batch_matmul ins(%161, %179 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%181 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1051, %180 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1154 = tensor.expand_shape %175 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%182 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1154 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1155 = tensor.expand_shape %178 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1156 = tensor.expand_shape %181 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%183 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1156 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%184 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1155 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%185 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%184 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%186 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%182 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%187 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%185 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1157 = tensor.collapse_shape %186 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1158 = tensor.collapse_shape %187 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%188 = linalg.batch_matmul ins(%collapsed_1157, %collapsed_1158 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1159 = tensor.expand_shape %188 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%189 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1159, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%190 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%189, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%191:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%190 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1160 = tensor.expand_shape %191#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%192 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%190, %expanded_1160 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%193 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%192 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%194 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%193 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%195 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%193, %194 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%196 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%195 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%197 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%183 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1161 = tensor.collapse_shape %196 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1162 = tensor.collapse_shape %197 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%198 = linalg.batch_matmul ins(%collapsed_1161, %collapsed_1162 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1163 = tensor.expand_shape %198 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%199 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1163 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1164 = tensor.collapse_shape %199 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%200 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1164 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%201 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_340 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%202 = linalg.batch_matmul ins(%200, %201 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%203 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1050, %202 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%204 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%203, %166 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%205 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%204, %cst_1048 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%206 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%205, %cst_1049 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%207 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%206 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%208 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_339 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%209 = linalg.batch_matmul ins(%207, %208 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%210 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1035, %209 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%211 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%210 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%212 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%211 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%213 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_338 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%214 = linalg.batch_matmul ins(%212, %213 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%215 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1034, %214 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%216 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%215, %206 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%217 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%216, %cst_1032 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%218 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%217, %cst_1033 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%219 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%218 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%220 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_337 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%221 = linalg.batch_matmul ins(%219, %220 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%222 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1031, %221 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%223 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%222 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%224 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%223 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%225 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_336 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%226 = linalg.batch_matmul ins(%224, %225 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%227 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1030, %226 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%228 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%227, %218 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%229 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%228, %cst_1028 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%230 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%229, %cst_1029 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%231 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%230 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%232 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_335 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%233 = linalg.batch_matmul ins(%231, %232 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%234 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1027, %233 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%235 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%234 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%236 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%235 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%237 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_334 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%238 = linalg.batch_matmul ins(%236, %237 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%239 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1026, %238 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%240 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%239, %230 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%241 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%240, %cst_1024 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%242 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%241, %cst_1025 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%243 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%242 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%244 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_333 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%245 = linalg.batch_matmul ins(%243, %244 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%246 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1047, %245 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%247 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%246 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%248 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%247 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%249 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_332 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%250 = linalg.batch_matmul ins(%248, %249 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%251 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1046, %250 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%252 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%251, %242 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%253 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%252, %cst_1044 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%254 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%253, %cst_1045 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%255 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%254 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%256 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_331 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%257 = linalg.batch_matmul ins(%255, %256 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%258 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1043, %257 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%259 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%258, %160 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%260 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%259, %cst_1041 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%261 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%260, %cst_1042 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%262 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%261 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%263 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_330 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%264 = linalg.batch_matmul ins(%262, %263 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%265 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1010, %264 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%266 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%265, %cst_1009 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%267 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%266, %cst_1020 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%268 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_329 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%269 = linalg.batch_matmul ins(%262, %268 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%270 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1008, %269 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%271 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%270, %cst_1006 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%272 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%271, %cst_1007 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%273 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%272 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%274 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_328 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%275 = linalg.batch_matmul ins(%273, %274 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%276 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1023, %275 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%277 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_327 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%278 = linalg.batch_matmul ins(%273, %277 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%279 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1022, %278 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%280 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_326 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%281 = linalg.batch_matmul ins(%262, %280 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%282 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1021, %281 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1165 = tensor.expand_shape %276 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%283 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1165 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1166 = tensor.expand_shape %279 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1167 = tensor.expand_shape %282 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%284 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1167 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%285 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1166 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%286 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%285 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%287 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%283 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%288 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%286 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1168 = tensor.collapse_shape %287 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1169 = tensor.collapse_shape %288 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%289 = linalg.batch_matmul ins(%collapsed_1168, %collapsed_1169 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1170 = tensor.expand_shape %289 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%290 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1170, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%291 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%290, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%292:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%291 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1171 = tensor.expand_shape %292#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%293 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%291, %expanded_1171 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%294 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%293 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%295 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%294 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%296 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%294, %295 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%297 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%296 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%298 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%284 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1172 = tensor.collapse_shape %297 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1173 = tensor.collapse_shape %298 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%299 = linalg.batch_matmul ins(%collapsed_1172, %collapsed_1173 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1174 = tensor.expand_shape %299 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%300 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1174 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1175 = tensor.collapse_shape %300 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%301 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1175 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%302 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_325 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%303 = linalg.batch_matmul ins(%301, %302 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%304 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1020, %303 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%305 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%304, %267 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%306 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%305, %cst_1018 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%307 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%306, %cst_1019 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%308 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%307 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%309 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_324 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%310 = linalg.batch_matmul ins(%308, %309 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%311 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1005, %310 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%312 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%311 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%313 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%312 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%314 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_323 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%315 = linalg.batch_matmul ins(%313, %314 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%316 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1004, %315 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%317 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%316, %307 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%318 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%317, %cst_1002 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%319 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%318, %cst_1003 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%320 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%319 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%321 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_322 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%322 = linalg.batch_matmul ins(%320, %321 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%323 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1001, %322 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%324 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%323 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%325 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%324 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%326 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_321 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%327 = linalg.batch_matmul ins(%325, %326 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%328 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1000, %327 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%329 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%328, %319 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%330 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%329, %cst_998 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%331 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%330, %cst_999 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%332 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%331 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%333 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_320 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%334 = linalg.batch_matmul ins(%332, %333 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%335 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_997, %334 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%336 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%335 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%337 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%336 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%338 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_319 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%339 = linalg.batch_matmul ins(%337, %338 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%340 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_996, %339 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%341 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%340, %331 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%342 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%341, %cst_994 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%343 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%342, %cst_995 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%344 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%343 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%345 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_318 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%346 = linalg.batch_matmul ins(%344, %345 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%347 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1017, %346 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%348 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%347 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%349 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%348 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%350 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_317 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%351 = linalg.batch_matmul ins(%349, %350 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%352 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1016, %351 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%353 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%352, %343 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%354 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%353, %cst_1014 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%355 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%354, %cst_1015 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%356 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%355 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%357 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_316 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%358 = linalg.batch_matmul ins(%356, %357 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%359 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1013, %358 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%360 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%359, %261 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%361 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%360, %cst_1011 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%362 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%361, %cst_1012 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%363 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%362 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%364 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_315 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%365 = linalg.batch_matmul ins(%363, %364 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%366 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_980, %365 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%367 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%366, %cst_979 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%368 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%367, %cst_990 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%369 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_314 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%370 = linalg.batch_matmul ins(%363, %369 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%371 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_978, %370 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%372 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%371, %cst_976 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%373 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%372, %cst_977 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%374 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%373 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%375 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_313 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%376 = linalg.batch_matmul ins(%374, %375 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%377 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_993, %376 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%378 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_312 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%379 = linalg.batch_matmul ins(%374, %378 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%380 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_992, %379 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%381 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_311 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%382 = linalg.batch_matmul ins(%363, %381 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%383 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_991, %382 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1176 = tensor.expand_shape %377 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%384 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1176 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1177 = tensor.expand_shape %380 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1178 = tensor.expand_shape %383 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%385 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1178 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%386 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1177 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%387 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%386 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%388 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%384 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%389 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%387 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1179 = tensor.collapse_shape %388 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1180 = tensor.collapse_shape %389 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%390 = linalg.batch_matmul ins(%collapsed_1179, %collapsed_1180 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1181 = tensor.expand_shape %390 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%391 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1181, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%392 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%391, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%393:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%392 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1182 = tensor.expand_shape %393#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%394 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%392, %expanded_1182 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%395 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%394 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%396 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%395 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%397 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%395, %396 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%398 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%397 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%399 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%385 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1183 = tensor.collapse_shape %398 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1184 = tensor.collapse_shape %399 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%400 = linalg.batch_matmul ins(%collapsed_1183, %collapsed_1184 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1185 = tensor.expand_shape %400 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%401 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1185 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1186 = tensor.collapse_shape %401 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%402 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1186 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%403 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_310 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%404 = linalg.batch_matmul ins(%402, %403 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%405 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_990, %404 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%406 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%405, %368 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%407 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%406, %cst_988 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%408 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%407, %cst_989 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%409 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%408 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%410 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_309 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%411 = linalg.batch_matmul ins(%409, %410 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%412 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_975, %411 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%413 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%412 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%414 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%413 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%415 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_308 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%416 = linalg.batch_matmul ins(%414, %415 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%417 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_974, %416 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%418 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%417, %408 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%419 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%418, %cst_972 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%420 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%419, %cst_973 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%421 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%420 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%422 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_307 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%423 = linalg.batch_matmul ins(%421, %422 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%424 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_971, %423 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%425 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%424 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%426 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%425 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%427 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_306 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%428 = linalg.batch_matmul ins(%426, %427 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%429 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_970, %428 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%430 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%429, %420 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%431 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%430, %cst_968 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%432 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%431, %cst_969 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%433 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%432 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%434 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_305 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%435 = linalg.batch_matmul ins(%433, %434 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%436 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_967, %435 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%437 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%436 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%438 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%437 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%439 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_304 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%440 = linalg.batch_matmul ins(%438, %439 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%441 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_966, %440 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%442 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%441, %432 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%443 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%442, %cst_964 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%444 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%443, %cst_965 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%445 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%444 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%446 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_303 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%447 = linalg.batch_matmul ins(%445, %446 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%448 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_987, %447 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%449 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%448 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%450 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%449 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%451 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_302 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%452 = linalg.batch_matmul ins(%450, %451 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%453 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_986, %452 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%454 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%453, %444 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%455 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%454, %cst_984 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%456 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%455, %cst_985 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%457 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%456 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%458 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_301 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%459 = linalg.batch_matmul ins(%457, %458 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%460 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_983, %459 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%461 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%460, %362 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%462 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%461, %cst_981 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%463 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%462, %cst_982 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%464 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%463 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%465 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_300 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%466 = linalg.batch_matmul ins(%464, %465 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%467 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_950, %466 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%468 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%467, %cst_949 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%469 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%468, %cst_960 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%470 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_299 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%471 = linalg.batch_matmul ins(%464, %470 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%472 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_948, %471 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%473 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%472, %cst_946 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%474 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%473, %cst_947 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%475 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%474 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%476 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_298 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%477 = linalg.batch_matmul ins(%475, %476 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%478 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_963, %477 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%479 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_297 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%480 = linalg.batch_matmul ins(%475, %479 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%481 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_962, %480 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%482 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_296 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%483 = linalg.batch_matmul ins(%464, %482 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%484 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_961, %483 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1187 = tensor.expand_shape %478 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%485 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1187 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1188 = tensor.expand_shape %481 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1189 = tensor.expand_shape %484 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%486 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1189 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%487 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1188 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%488 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%487 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%489 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%485 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%490 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%488 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1190 = tensor.collapse_shape %489 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1191 = tensor.collapse_shape %490 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%491 = linalg.batch_matmul ins(%collapsed_1190, %collapsed_1191 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1192 = tensor.expand_shape %491 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%492 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1192, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%493 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%492, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%494:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%493 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1193 = tensor.expand_shape %494#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%495 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%493, %expanded_1193 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%496 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%495 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%497 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%496 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%498 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%496, %497 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%499 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%498 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%500 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%486 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1194 = tensor.collapse_shape %499 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1195 = tensor.collapse_shape %500 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%501 = linalg.batch_matmul ins(%collapsed_1194, %collapsed_1195 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1196 = tensor.expand_shape %501 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%502 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1196 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1197 = tensor.collapse_shape %502 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%503 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1197 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%504 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_295 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%505 = linalg.batch_matmul ins(%503, %504 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%506 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_960, %505 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%507 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%506, %469 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%508 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%507, %cst_958 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%509 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%508, %cst_959 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%510 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%509 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%511 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_294 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%512 = linalg.batch_matmul ins(%510, %511 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%513 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_945, %512 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%514 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%513 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%515 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%514 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%516 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_293 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%517 = linalg.batch_matmul ins(%515, %516 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%518 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_944, %517 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%519 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%518, %509 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%520 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%519, %cst_942 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%521 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%520, %cst_943 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%522 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%521 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%523 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_292 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%524 = linalg.batch_matmul ins(%522, %523 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%525 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_941, %524 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%526 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%525 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%527 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%526 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%528 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_291 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%529 = linalg.batch_matmul ins(%527, %528 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%530 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_940, %529 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%531 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%530, %521 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%532 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%531, %cst_938 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%533 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%532, %cst_939 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%534 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%533 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%535 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_290 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%536 = linalg.batch_matmul ins(%534, %535 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%537 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_937, %536 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%538 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%537 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%539 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%538 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%540 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_289 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%541 = linalg.batch_matmul ins(%539, %540 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%542 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_936, %541 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%543 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%542, %533 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%544 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%543, %cst_934 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%545 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%544, %cst_935 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%546 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%545 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%547 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_288 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%548 = linalg.batch_matmul ins(%546, %547 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%549 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_957, %548 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%550 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%549 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%551 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%550 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%552 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_287 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%553 = linalg.batch_matmul ins(%551, %552 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%554 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_956, %553 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%555 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%554, %545 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%556 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%555, %cst_954 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%557 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%556, %cst_955 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%558 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%557 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%559 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_286 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%560 = linalg.batch_matmul ins(%558, %559 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%561 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_953, %560 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%562 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%561, %463 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%563 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%562, %cst_951 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%564 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%563, %cst_952 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%565 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%564 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%566 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_285 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%567 = linalg.batch_matmul ins(%565, %566 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%568 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_920, %567 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%569 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%568, %cst_919 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%570 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%569, %cst_930 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%571 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_284 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%572 = linalg.batch_matmul ins(%565, %571 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%573 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_918, %572 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%574 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%573, %cst_916 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%575 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%574, %cst_917 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%576 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%575 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%577 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_283 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%578 = linalg.batch_matmul ins(%576, %577 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%579 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_933, %578 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%580 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_282 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%581 = linalg.batch_matmul ins(%576, %580 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%582 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_932, %581 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%583 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_281 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%584 = linalg.batch_matmul ins(%565, %583 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%585 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_931, %584 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1198 = tensor.expand_shape %579 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%586 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1198 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1199 = tensor.expand_shape %582 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1200 = tensor.expand_shape %585 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%587 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1200 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%588 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1199 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%589 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%588 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%590 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%586 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%591 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%589 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1201 = tensor.collapse_shape %590 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1202 = tensor.collapse_shape %591 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%592 = linalg.batch_matmul ins(%collapsed_1201, %collapsed_1202 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1203 = tensor.expand_shape %592 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%593 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1203, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%594 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%593, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%595:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%594 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1204 = tensor.expand_shape %595#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%596 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%594, %expanded_1204 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%597 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%596 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%598 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%597 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%599 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%597, %598 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%600 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%599 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%601 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%587 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1205 = tensor.collapse_shape %600 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1206 = tensor.collapse_shape %601 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%602 = linalg.batch_matmul ins(%collapsed_1205, %collapsed_1206 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1207 = tensor.expand_shape %602 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%603 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1207 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1208 = tensor.collapse_shape %603 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%604 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1208 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%605 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_280 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%606 = linalg.batch_matmul ins(%604, %605 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%607 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_930, %606 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%608 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%607, %570 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%609 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%608, %cst_928 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%610 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%609, %cst_929 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%611 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%610 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%612 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_279 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%613 = linalg.batch_matmul ins(%611, %612 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%614 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_915, %613 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%615 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%614 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%616 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%615 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%617 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_278 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%618 = linalg.batch_matmul ins(%616, %617 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%619 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_914, %618 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%620 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%619, %610 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%621 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%620, %cst_912 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%622 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%621, %cst_913 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%623 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%622 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%624 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_277 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%625 = linalg.batch_matmul ins(%623, %624 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%626 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_911, %625 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%627 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%626 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%628 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%627 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%629 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_276 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%630 = linalg.batch_matmul ins(%628, %629 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%631 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_910, %630 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%632 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%631, %622 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%633 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%632, %cst_908 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%634 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%633, %cst_909 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%635 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%634 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%636 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_275 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%637 = linalg.batch_matmul ins(%635, %636 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%638 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_907, %637 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%639 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%638 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%640 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%639 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%641 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_274 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%642 = linalg.batch_matmul ins(%640, %641 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%643 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_906, %642 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%644 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%643, %634 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%645 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%644, %cst_904 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%646 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%645, %cst_905 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%647 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%646 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%648 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_273 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%649 = linalg.batch_matmul ins(%647, %648 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%650 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_927, %649 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%651 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%650 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%652 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%651 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%653 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_272 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%654 = linalg.batch_matmul ins(%652, %653 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%655 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_926, %654 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%656 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%655, %646 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%657 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%656, %cst_924 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%658 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%657, %cst_925 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%659 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%658 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%660 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_271 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%661 = linalg.batch_matmul ins(%659, %660 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%662 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_923, %661 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%663 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%662, %564 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%664 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%663, %cst_921 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%665 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%664, %cst_922 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%666 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%665 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%667 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_270 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%668 = linalg.batch_matmul ins(%666, %667 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%669 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_890, %668 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%670 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%669, %cst_889 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%671 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%670, %cst_900 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%672 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_269 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%673 = linalg.batch_matmul ins(%666, %672 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%674 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_888, %673 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%675 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%674, %cst_886 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%676 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%675, %cst_887 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%677 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%676 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%678 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_268 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%679 = linalg.batch_matmul ins(%677, %678 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%680 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_903, %679 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%681 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_267 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%682 = linalg.batch_matmul ins(%677, %681 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%683 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_902, %682 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%684 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_266 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%685 = linalg.batch_matmul ins(%666, %684 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%686 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_901, %685 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1209 = tensor.expand_shape %680 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%687 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1209 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1210 = tensor.expand_shape %683 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1211 = tensor.expand_shape %686 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%688 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1211 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%689 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1210 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%690 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%689 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%691 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%687 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%692 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%690 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1212 = tensor.collapse_shape %691 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1213 = tensor.collapse_shape %692 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%693 = linalg.batch_matmul ins(%collapsed_1212, %collapsed_1213 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1214 = tensor.expand_shape %693 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%694 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1214, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%695 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%694, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%696:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%695 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1215 = tensor.expand_shape %696#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%697 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%695, %expanded_1215 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%698 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%697 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%699 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%698 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%700 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%698, %699 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%701 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%700 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%702 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%688 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1216 = tensor.collapse_shape %701 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1217 = tensor.collapse_shape %702 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%703 = linalg.batch_matmul ins(%collapsed_1216, %collapsed_1217 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1218 = tensor.expand_shape %703 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%704 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1218 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1219 = tensor.collapse_shape %704 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%705 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1219 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%706 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_265 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%707 = linalg.batch_matmul ins(%705, %706 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%708 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_900, %707 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%709 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%708, %671 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%710 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%709, %cst_898 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%711 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%710, %cst_899 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%712 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%711 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%713 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_264 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%714 = linalg.batch_matmul ins(%712, %713 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%715 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_885, %714 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%716 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%715 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%717 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%716 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%718 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_263 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%719 = linalg.batch_matmul ins(%717, %718 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%720 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_884, %719 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%721 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%720, %711 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%722 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%721, %cst_882 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%723 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%722, %cst_883 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%724 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%723 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%725 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_262 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%726 = linalg.batch_matmul ins(%724, %725 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%727 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_881, %726 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%728 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%727 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%729 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%728 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%730 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_261 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%731 = linalg.batch_matmul ins(%729, %730 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%732 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_880, %731 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%733 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%732, %723 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%734 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%733, %cst_878 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%735 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%734, %cst_879 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%736 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%735 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%737 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_260 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%738 = linalg.batch_matmul ins(%736, %737 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%739 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_877, %738 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%740 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%739 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%741 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%740 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%742 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_259 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%743 = linalg.batch_matmul ins(%741, %742 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%744 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_876, %743 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%745 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%744, %735 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%746 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%745, %cst_874 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%747 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%746, %cst_875 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%748 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%747 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%749 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_258 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%750 = linalg.batch_matmul ins(%748, %749 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%751 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_897, %750 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%752 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%751 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%753 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%752 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%754 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_257 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%755 = linalg.batch_matmul ins(%753, %754 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%756 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_896, %755 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%757 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%756, %747 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%758 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%757, %cst_894 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%759 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%758, %cst_895 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%760 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%759 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%761 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_256 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%762 = linalg.batch_matmul ins(%760, %761 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%763 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_893, %762 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%764 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%763, %665 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%765 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%764, %cst_891 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%766 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%765, %cst_892 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%767 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%766 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%768 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_255 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%769 = linalg.batch_matmul ins(%767, %768 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%770 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_860, %769 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%771 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%770, %cst_859 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%772 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%771, %cst_870 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%773 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_254 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%774 = linalg.batch_matmul ins(%767, %773 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%775 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_858, %774 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%776 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%775, %cst_856 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%777 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%776, %cst_857 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%778 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%777 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%779 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_253 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%780 = linalg.batch_matmul ins(%778, %779 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%781 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_873, %780 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%782 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_252 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%783 = linalg.batch_matmul ins(%778, %782 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%784 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_872, %783 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%785 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_251 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%786 = linalg.batch_matmul ins(%767, %785 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%787 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_871, %786 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1220 = tensor.expand_shape %781 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%788 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1220 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1221 = tensor.expand_shape %784 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1222 = tensor.expand_shape %787 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%789 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1222 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%790 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1221 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%791 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%790 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%792 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%788 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%793 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%791 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1223 = tensor.collapse_shape %792 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1224 = tensor.collapse_shape %793 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%794 = linalg.batch_matmul ins(%collapsed_1223, %collapsed_1224 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1225 = tensor.expand_shape %794 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%795 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1225, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%796 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%795, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%797:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%796 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1226 = tensor.expand_shape %797#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%798 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%796, %expanded_1226 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%799 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%798 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%800 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%799 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%801 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%799, %800 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%802 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%801 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%803 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%789 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1227 = tensor.collapse_shape %802 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1228 = tensor.collapse_shape %803 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%804 = linalg.batch_matmul ins(%collapsed_1227, %collapsed_1228 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1229 = tensor.expand_shape %804 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%805 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1229 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1230 = tensor.collapse_shape %805 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%806 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1230 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%807 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_250 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%808 = linalg.batch_matmul ins(%806, %807 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%809 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_870, %808 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%810 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%809, %772 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%811 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%810, %cst_868 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%812 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%811, %cst_869 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%813 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%812 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%814 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_249 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%815 = linalg.batch_matmul ins(%813, %814 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%816 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_855, %815 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%817 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%816 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%818 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%817 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%819 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_248 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%820 = linalg.batch_matmul ins(%818, %819 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%821 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_854, %820 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%822 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%821, %812 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%823 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%822, %cst_852 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%824 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%823, %cst_853 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%825 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%824 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%826 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_247 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%827 = linalg.batch_matmul ins(%825, %826 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%828 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_851, %827 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%829 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%828 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%830 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%829 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%831 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_246 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%832 = linalg.batch_matmul ins(%830, %831 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%833 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_850, %832 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%834 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%833, %824 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%835 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%834, %cst_848 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%836 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%835, %cst_849 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%837 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%836 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%838 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_245 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%839 = linalg.batch_matmul ins(%837, %838 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%840 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_847, %839 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%841 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%840 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%842 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%841 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%843 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_244 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%844 = linalg.batch_matmul ins(%842, %843 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%845 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_846, %844 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%846 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%845, %836 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%847 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%846, %cst_844 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%848 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%847, %cst_845 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%849 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%848 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%850 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_243 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%851 = linalg.batch_matmul ins(%849, %850 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%852 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_867, %851 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%853 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%852 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%854 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%853 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%855 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_242 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%856 = linalg.batch_matmul ins(%854, %855 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%857 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_866, %856 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%858 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%857, %848 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%859 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%858, %cst_864 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%860 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%859, %cst_865 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%861 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%860 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%862 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_241 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%863 = linalg.batch_matmul ins(%861, %862 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%864 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_863, %863 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%865 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%864, %766 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%866 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%865, %cst_861 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%867 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%866, %cst_862 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%868 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%867 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%869 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_240 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%870 = linalg.batch_matmul ins(%868, %869 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%871 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_830, %870 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%872 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%871, %cst_829 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%873 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%872, %cst_840 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%874 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_239 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%875 = linalg.batch_matmul ins(%868, %874 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%876 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_828, %875 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%877 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%876, %cst_826 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%878 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%877, %cst_827 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%879 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%878 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%880 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_238 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%881 = linalg.batch_matmul ins(%879, %880 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%882 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_843, %881 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%883 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_237 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%884 = linalg.batch_matmul ins(%879, %883 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%885 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_842, %884 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%886 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_236 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%887 = linalg.batch_matmul ins(%868, %886 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%888 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_841, %887 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1231 = tensor.expand_shape %882 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%889 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1231 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1232 = tensor.expand_shape %885 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1233 = tensor.expand_shape %888 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%890 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1233 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%891 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1232 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%892 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%891 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%893 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%889 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%894 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%892 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1234 = tensor.collapse_shape %893 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1235 = tensor.collapse_shape %894 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%895 = linalg.batch_matmul ins(%collapsed_1234, %collapsed_1235 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1236 = tensor.expand_shape %895 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%896 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1236, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%897 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%896, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%898:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%897 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1237 = tensor.expand_shape %898#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%899 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%897, %expanded_1237 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%900 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%899 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%901 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%900 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%902 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%900, %901 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%903 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%902 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%904 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%890 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1238 = tensor.collapse_shape %903 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1239 = tensor.collapse_shape %904 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%905 = linalg.batch_matmul ins(%collapsed_1238, %collapsed_1239 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1240 = tensor.expand_shape %905 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%906 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1240 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1241 = tensor.collapse_shape %906 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%907 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1241 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%908 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_235 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%909 = linalg.batch_matmul ins(%907, %908 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%910 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_840, %909 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%911 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%910, %873 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%912 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%911, %cst_838 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%913 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%912, %cst_839 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%914 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%913 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%915 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_234 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%916 = linalg.batch_matmul ins(%914, %915 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%917 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_825, %916 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%918 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%917 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%919 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%918 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%920 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_233 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%921 = linalg.batch_matmul ins(%919, %920 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%922 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_824, %921 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%923 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%922, %913 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%924 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%923, %cst_822 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%925 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%924, %cst_823 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%926 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%925 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%927 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_232 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%928 = linalg.batch_matmul ins(%926, %927 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%929 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_821, %928 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%930 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%929 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%931 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%930 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%932 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_231 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%933 = linalg.batch_matmul ins(%931, %932 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%934 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_820, %933 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%935 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%934, %925 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%936 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%935, %cst_818 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%937 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%936, %cst_819 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%938 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%937 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%939 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_230 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%940 = linalg.batch_matmul ins(%938, %939 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%941 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_817, %940 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%942 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%941 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%943 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%942 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%944 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_229 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%945 = linalg.batch_matmul ins(%943, %944 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%946 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_816, %945 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%947 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%946, %937 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%948 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%947, %cst_814 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%949 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%948, %cst_815 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%950 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%949 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%951 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_228 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%952 = linalg.batch_matmul ins(%950, %951 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%953 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_837, %952 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%954 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%953 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%955 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%954 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%956 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_227 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%957 = linalg.batch_matmul ins(%955, %956 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%958 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_836, %957 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%959 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%958, %949 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%960 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%959, %cst_834 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%961 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%960, %cst_835 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%962 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%961 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%963 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_226 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%964 = linalg.batch_matmul ins(%962, %963 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%965 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_833, %964 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%966 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%965, %867 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%967 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%966, %cst_831 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%968 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%967, %cst_832 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%969 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%968 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%970 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_225 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%971 = linalg.batch_matmul ins(%969, %970 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%972 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_800, %971 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%973 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%972, %cst_799 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%974 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%973, %cst_810 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%975 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_224 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%976 = linalg.batch_matmul ins(%969, %975 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%977 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_798, %976 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%978 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%977, %cst_796 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%979 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%978, %cst_797 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%980 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%979 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%981 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_223 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%982 = linalg.batch_matmul ins(%980, %981 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%983 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_813, %982 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%984 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_222 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%985 = linalg.batch_matmul ins(%980, %984 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%986 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_812, %985 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%987 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_221 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%988 = linalg.batch_matmul ins(%969, %987 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%989 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_811, %988 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1242 = tensor.expand_shape %983 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%990 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1242 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1243 = tensor.expand_shape %986 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1244 = tensor.expand_shape %989 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%991 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1244 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%992 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1243 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%993 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%992 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%994 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%990 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%995 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%993 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1245 = tensor.collapse_shape %994 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1246 = tensor.collapse_shape %995 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%996 = linalg.batch_matmul ins(%collapsed_1245, %collapsed_1246 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1247 = tensor.expand_shape %996 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%997 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1247, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%998 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%997, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%999:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%998 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1248 = tensor.expand_shape %999#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1000 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%998, %expanded_1248 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1001 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1000 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1002 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1001 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1003 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1001, %1002 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1004 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1003 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1005 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%991 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1249 = tensor.collapse_shape %1004 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1250 = tensor.collapse_shape %1005 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1006 = linalg.batch_matmul ins(%collapsed_1249, %collapsed_1250 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1251 = tensor.expand_shape %1006 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1007 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1251 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1252 = tensor.collapse_shape %1007 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1008 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1252 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1009 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_220 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1010 = linalg.batch_matmul ins(%1008, %1009 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1011 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_810, %1010 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1012 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1011, %974 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1013 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1012, %cst_808 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1014 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1013, %cst_809 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1015 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1014 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1016 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_219 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1017 = linalg.batch_matmul ins(%1015, %1016 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1018 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_795, %1017 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1019 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1018 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1020 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1019 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1021 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_218 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1022 = linalg.batch_matmul ins(%1020, %1021 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1023 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_794, %1022 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1024 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1023, %1014 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1025 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1024, %cst_792 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1026 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1025, %cst_793 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1027 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1026 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1028 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_217 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1029 = linalg.batch_matmul ins(%1027, %1028 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1030 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_791, %1029 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1031 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1030 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1032 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1031 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1033 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_216 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1034 = linalg.batch_matmul ins(%1032, %1033 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1035 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_790, %1034 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1036 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1035, %1026 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1037 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1036, %cst_788 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1038 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1037, %cst_789 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1039 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1038 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1040 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_215 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1041 = linalg.batch_matmul ins(%1039, %1040 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1042 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_787, %1041 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1043 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1042 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1044 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1043 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1045 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_214 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1046 = linalg.batch_matmul ins(%1044, %1045 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1047 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_786, %1046 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1048 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1047, %1038 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1049 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1048, %cst_784 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1050 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1049, %cst_785 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1051 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1050 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1052 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_213 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1053 = linalg.batch_matmul ins(%1051, %1052 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1054 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_807, %1053 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1055 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1054 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1056 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1055 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1057 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_212 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1058 = linalg.batch_matmul ins(%1056, %1057 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1059 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_806, %1058 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1060 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1059, %1050 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1061 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1060, %cst_804 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1062 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1061, %cst_805 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1063 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1062 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1064 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_211 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1065 = linalg.batch_matmul ins(%1063, %1064 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1066 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_803, %1065 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1067 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1066, %968 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1068 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1067, %cst_801 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1069 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1068, %cst_802 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1070 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1069 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1071 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_210 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1072 = linalg.batch_matmul ins(%1070, %1071 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1073 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_770, %1072 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1074 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1073, %cst_769 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1075 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1074, %cst_780 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1076 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_209 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1077 = linalg.batch_matmul ins(%1070, %1076 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1078 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_768, %1077 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1079 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1078, %cst_766 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1080 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1079, %cst_767 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1081 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1080 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1082 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_208 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1083 = linalg.batch_matmul ins(%1081, %1082 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1084 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_783, %1083 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1085 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_207 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1086 = linalg.batch_matmul ins(%1081, %1085 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1087 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_782, %1086 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1088 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_206 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1089 = linalg.batch_matmul ins(%1070, %1088 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1090 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_781, %1089 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1253 = tensor.expand_shape %1084 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1091 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1253 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1254 = tensor.expand_shape %1087 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1255 = tensor.expand_shape %1090 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1092 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1255 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1093 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1254 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1094 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1093 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1095 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1091 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1096 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1094 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1256 = tensor.collapse_shape %1095 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1257 = tensor.collapse_shape %1096 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1097 = linalg.batch_matmul ins(%collapsed_1256, %collapsed_1257 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1258 = tensor.expand_shape %1097 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1098 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1258, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1099 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1098, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1100:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1099 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1259 = tensor.expand_shape %1100#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1101 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1099, %expanded_1259 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1102 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1101 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1103 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1102 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1104 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1102, %1103 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1105 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1104 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1106 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1092 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1260 = tensor.collapse_shape %1105 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1261 = tensor.collapse_shape %1106 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1107 = linalg.batch_matmul ins(%collapsed_1260, %collapsed_1261 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1262 = tensor.expand_shape %1107 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1108 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1262 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1263 = tensor.collapse_shape %1108 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1109 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1263 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1110 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_205 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1111 = linalg.batch_matmul ins(%1109, %1110 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1112 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_780, %1111 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1113 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1112, %1075 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1114 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1113, %cst_778 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1115 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1114, %cst_779 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1116 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1115 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1117 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_204 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1118 = linalg.batch_matmul ins(%1116, %1117 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1119 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_765, %1118 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1120 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1119 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1121 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1120 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1122 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_203 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1123 = linalg.batch_matmul ins(%1121, %1122 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1124 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_764, %1123 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1125 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1124, %1115 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1126 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1125, %cst_762 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1127 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1126, %cst_763 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1128 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1127 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1129 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_202 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1130 = linalg.batch_matmul ins(%1128, %1129 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1131 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_761, %1130 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1132 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1131 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1133 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1132 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1134 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_201 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1135 = linalg.batch_matmul ins(%1133, %1134 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1136 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_760, %1135 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1137 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1136, %1127 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1138 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1137, %cst_758 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1139 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1138, %cst_759 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1140 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1139 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1141 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_200 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1142 = linalg.batch_matmul ins(%1140, %1141 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1143 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_757, %1142 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1144 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1143 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1145 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1144 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1146 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_199 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1147 = linalg.batch_matmul ins(%1145, %1146 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1148 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_756, %1147 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1149 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1148, %1139 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1150 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1149, %cst_754 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1151 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1150, %cst_755 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1152 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1151 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1153 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_198 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1154 = linalg.batch_matmul ins(%1152, %1153 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1155 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_777, %1154 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1156 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1155 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1157 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1156 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1158 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_197 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1159 = linalg.batch_matmul ins(%1157, %1158 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1160 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_776, %1159 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1161 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1160, %1151 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1162 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1161, %cst_774 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1163 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1162, %cst_775 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1164 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1163 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1165 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_196 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1166 = linalg.batch_matmul ins(%1164, %1165 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1167 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_773, %1166 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1168 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1167, %1069 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1169 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1168, %cst_771 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1170 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1169, %cst_772 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1171 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1170 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1172 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_195 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1173 = linalg.batch_matmul ins(%1171, %1172 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1174 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_740, %1173 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1175 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1174, %cst_739 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1176 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1175, %cst_750 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1177 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_194 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1178 = linalg.batch_matmul ins(%1171, %1177 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1179 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_738, %1178 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1180 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1179, %cst_736 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1181 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1180, %cst_737 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1182 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1181 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1183 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_193 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1184 = linalg.batch_matmul ins(%1182, %1183 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1185 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_753, %1184 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1186 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_192 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1187 = linalg.batch_matmul ins(%1182, %1186 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1188 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_752, %1187 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1189 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_191 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1190 = linalg.batch_matmul ins(%1171, %1189 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1191 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_751, %1190 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1264 = tensor.expand_shape %1185 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1192 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1264 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1265 = tensor.expand_shape %1188 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1266 = tensor.expand_shape %1191 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1193 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1266 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1194 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1265 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1195 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1194 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1196 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1192 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1197 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1195 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1267 = tensor.collapse_shape %1196 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1268 = tensor.collapse_shape %1197 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1198 = linalg.batch_matmul ins(%collapsed_1267, %collapsed_1268 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1269 = tensor.expand_shape %1198 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1199 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1269, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1200 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1199, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1201:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1200 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1270 = tensor.expand_shape %1201#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1202 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1200, %expanded_1270 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1203 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1202 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1204 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1203 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1205 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1203, %1204 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1206 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1205 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1207 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1193 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1271 = tensor.collapse_shape %1206 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1272 = tensor.collapse_shape %1207 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1208 = linalg.batch_matmul ins(%collapsed_1271, %collapsed_1272 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1273 = tensor.expand_shape %1208 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1209 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1273 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1274 = tensor.collapse_shape %1209 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1210 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1274 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1211 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_190 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1212 = linalg.batch_matmul ins(%1210, %1211 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1213 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_750, %1212 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1214 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1213, %1176 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1215 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1214, %cst_748 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1216 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1215, %cst_749 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1217 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1216 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1218 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_189 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1219 = linalg.batch_matmul ins(%1217, %1218 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1220 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_735, %1219 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1221 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1220 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1222 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1221 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1223 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_188 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1224 = linalg.batch_matmul ins(%1222, %1223 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1225 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_734, %1224 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1226 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1225, %1216 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1227 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1226, %cst_732 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1228 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1227, %cst_733 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1229 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1228 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1230 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_187 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1231 = linalg.batch_matmul ins(%1229, %1230 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1232 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_731, %1231 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1233 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1232 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1234 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1233 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1235 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_186 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1236 = linalg.batch_matmul ins(%1234, %1235 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1237 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_730, %1236 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1238 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1237, %1228 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1239 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1238, %cst_728 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1240 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1239, %cst_729 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1241 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1240 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1242 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_185 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1243 = linalg.batch_matmul ins(%1241, %1242 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1244 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_727, %1243 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1245 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1244 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1246 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1245 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1247 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_184 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1248 = linalg.batch_matmul ins(%1246, %1247 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1249 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_726, %1248 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1250 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1249, %1240 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1251 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1250, %cst_724 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1252 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1251, %cst_725 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1253 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1252 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1254 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_183 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1255 = linalg.batch_matmul ins(%1253, %1254 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1256 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_747, %1255 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1257 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1256 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1258 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1257 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1259 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_182 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1260 = linalg.batch_matmul ins(%1258, %1259 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1261 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_746, %1260 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1262 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1261, %1252 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1263 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1262, %cst_744 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1264 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1263, %cst_745 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1265 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1264 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1266 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_181 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1267 = linalg.batch_matmul ins(%1265, %1266 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1268 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_743, %1267 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1269 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1268, %1170 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1270 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1269, %cst_741 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1271 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1270, %cst_742 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1272 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1271 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1273 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_180 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1274 = linalg.batch_matmul ins(%1272, %1273 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1275 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_710, %1274 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1276 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1275, %cst_709 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1277 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1276, %cst_720 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1278 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_179 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1279 = linalg.batch_matmul ins(%1272, %1278 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1280 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_708, %1279 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1281 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1280, %cst_706 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1282 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1281, %cst_707 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1283 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1282 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1284 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_178 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1285 = linalg.batch_matmul ins(%1283, %1284 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1286 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_723, %1285 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1287 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_177 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1288 = linalg.batch_matmul ins(%1283, %1287 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1289 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_722, %1288 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1290 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_176 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1291 = linalg.batch_matmul ins(%1272, %1290 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1292 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_721, %1291 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1275 = tensor.expand_shape %1286 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1293 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1275 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1276 = tensor.expand_shape %1289 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1277 = tensor.expand_shape %1292 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1294 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1277 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1295 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1276 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1296 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1295 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1297 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1293 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1298 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1296 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1278 = tensor.collapse_shape %1297 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1279 = tensor.collapse_shape %1298 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1299 = linalg.batch_matmul ins(%collapsed_1278, %collapsed_1279 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1280 = tensor.expand_shape %1299 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1300 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1280, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1301 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1300, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1302:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1301 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1281 = tensor.expand_shape %1302#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1303 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1301, %expanded_1281 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1304 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1303 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1305 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1304 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1306 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1304, %1305 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1307 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1306 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1308 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1294 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1282 = tensor.collapse_shape %1307 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1283 = tensor.collapse_shape %1308 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1309 = linalg.batch_matmul ins(%collapsed_1282, %collapsed_1283 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1284 = tensor.expand_shape %1309 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1310 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1284 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1285 = tensor.collapse_shape %1310 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1311 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1285 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1312 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_175 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1313 = linalg.batch_matmul ins(%1311, %1312 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1314 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_720, %1313 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1315 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1314, %1277 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1316 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1315, %cst_718 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1317 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1316, %cst_719 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1318 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1317 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1319 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_174 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1320 = linalg.batch_matmul ins(%1318, %1319 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1321 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_705, %1320 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1322 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1321 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1323 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1322 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1324 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_173 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1325 = linalg.batch_matmul ins(%1323, %1324 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1326 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_704, %1325 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1327 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1326, %1317 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1328 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1327, %cst_702 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1329 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1328, %cst_703 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1330 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1329 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1331 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_172 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1332 = linalg.batch_matmul ins(%1330, %1331 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1333 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_701, %1332 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1334 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1333 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1335 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1334 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1336 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_171 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1337 = linalg.batch_matmul ins(%1335, %1336 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1338 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_700, %1337 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1339 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1338, %1329 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1340 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1339, %cst_698 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1341 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1340, %cst_699 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1342 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1341 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1343 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_170 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1344 = linalg.batch_matmul ins(%1342, %1343 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1345 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_697, %1344 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1346 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1345 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1347 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1346 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1348 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_169 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1349 = linalg.batch_matmul ins(%1347, %1348 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1350 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_696, %1349 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1351 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1350, %1341 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1352 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1351, %cst_694 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1353 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1352, %cst_695 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1354 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1353 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1355 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_168 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1356 = linalg.batch_matmul ins(%1354, %1355 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1357 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_717, %1356 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1358 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1357 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1359 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1358 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1360 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_167 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1361 = linalg.batch_matmul ins(%1359, %1360 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1362 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_716, %1361 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1363 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1362, %1353 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1364 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1363, %cst_714 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1365 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1364, %cst_715 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1366 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1365 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1367 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_166 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1368 = linalg.batch_matmul ins(%1366, %1367 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1369 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_713, %1368 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1370 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1369, %1271 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1371 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1370, %cst_711 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1372 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1371, %cst_712 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1373 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1372 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1374 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_165 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1375 = linalg.batch_matmul ins(%1373, %1374 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1376 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_680, %1375 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1377 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1376, %cst_679 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1378 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1377, %cst_690 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1379 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_164 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1380 = linalg.batch_matmul ins(%1373, %1379 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1381 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_678, %1380 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1382 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1381, %cst_676 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1383 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1382, %cst_677 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1384 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1383 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1385 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_163 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1386 = linalg.batch_matmul ins(%1384, %1385 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1387 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_693, %1386 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1388 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_162 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1389 = linalg.batch_matmul ins(%1384, %1388 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1390 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_692, %1389 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1391 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_161 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1392 = linalg.batch_matmul ins(%1373, %1391 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1393 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_691, %1392 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1286 = tensor.expand_shape %1387 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1394 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1286 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1287 = tensor.expand_shape %1390 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1288 = tensor.expand_shape %1393 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1395 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1288 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1396 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1287 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1397 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1396 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1398 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1394 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1399 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1397 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1289 = tensor.collapse_shape %1398 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1290 = tensor.collapse_shape %1399 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1400 = linalg.batch_matmul ins(%collapsed_1289, %collapsed_1290 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1291 = tensor.expand_shape %1400 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1401 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1291, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1402 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1401, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1403:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1402 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1292 = tensor.expand_shape %1403#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1404 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1402, %expanded_1292 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1405 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1404 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1406 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1405 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1407 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1405, %1406 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1408 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1407 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1409 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1395 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1293 = tensor.collapse_shape %1408 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1294 = tensor.collapse_shape %1409 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1410 = linalg.batch_matmul ins(%collapsed_1293, %collapsed_1294 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1295 = tensor.expand_shape %1410 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1411 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1295 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1296 = tensor.collapse_shape %1411 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1412 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1296 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1413 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_160 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1414 = linalg.batch_matmul ins(%1412, %1413 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1415 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_690, %1414 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1416 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1415, %1378 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1417 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1416, %cst_688 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1418 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1417, %cst_689 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1419 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1418 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1420 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_159 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1421 = linalg.batch_matmul ins(%1419, %1420 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1422 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_675, %1421 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1423 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1422 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1424 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1423 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1425 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_158 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1426 = linalg.batch_matmul ins(%1424, %1425 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1427 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_674, %1426 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1428 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1427, %1418 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1429 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1428, %cst_672 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1430 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1429, %cst_673 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1431 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1430 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1432 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_157 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1433 = linalg.batch_matmul ins(%1431, %1432 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1434 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_671, %1433 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1435 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1434 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1436 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1435 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1437 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_156 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1438 = linalg.batch_matmul ins(%1436, %1437 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1439 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_670, %1438 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1440 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1439, %1430 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1441 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1440, %cst_668 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1442 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1441, %cst_669 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1443 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1442 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1444 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_155 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1445 = linalg.batch_matmul ins(%1443, %1444 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1446 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_667, %1445 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1447 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1446 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1448 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1447 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1449 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_154 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1450 = linalg.batch_matmul ins(%1448, %1449 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1451 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_666, %1450 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1452 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1451, %1442 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1453 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1452, %cst_664 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1454 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1453, %cst_665 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1455 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1454 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1456 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_153 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1457 = linalg.batch_matmul ins(%1455, %1456 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1458 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_687, %1457 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1459 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1458 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1460 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1459 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1461 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_152 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1462 = linalg.batch_matmul ins(%1460, %1461 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1463 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_686, %1462 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1464 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1463, %1454 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1465 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1464, %cst_684 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1466 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1465, %cst_685 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1467 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1466 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1468 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_151 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1469 = linalg.batch_matmul ins(%1467, %1468 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1470 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_683, %1469 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1471 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1470, %1372 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1472 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1471, %cst_681 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1473 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1472, %cst_682 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1474 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1473 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1475 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_150 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1476 = linalg.batch_matmul ins(%1474, %1475 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1477 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_650, %1476 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1478 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1477, %cst_649 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1479 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1478, %cst_660 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1480 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_149 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1481 = linalg.batch_matmul ins(%1474, %1480 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1482 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_648, %1481 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1483 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1482, %cst_646 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1484 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1483, %cst_647 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1485 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1484 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1486 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_148 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1487 = linalg.batch_matmul ins(%1485, %1486 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1488 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_663, %1487 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1489 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_147 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1490 = linalg.batch_matmul ins(%1485, %1489 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1491 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_662, %1490 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1492 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_146 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1493 = linalg.batch_matmul ins(%1474, %1492 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1494 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_661, %1493 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1297 = tensor.expand_shape %1488 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1495 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1297 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1298 = tensor.expand_shape %1491 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1299 = tensor.expand_shape %1494 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1496 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1299 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1497 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1298 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1498 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1497 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1499 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1495 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1500 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1498 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1300 = tensor.collapse_shape %1499 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1301 = tensor.collapse_shape %1500 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1501 = linalg.batch_matmul ins(%collapsed_1300, %collapsed_1301 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1302 = tensor.expand_shape %1501 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1502 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1302, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1503 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1502, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1504:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1503 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1303 = tensor.expand_shape %1504#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1505 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1503, %expanded_1303 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1506 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1505 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1507 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1506 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1508 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1506, %1507 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1509 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1508 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1510 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1496 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1304 = tensor.collapse_shape %1509 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1305 = tensor.collapse_shape %1510 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1511 = linalg.batch_matmul ins(%collapsed_1304, %collapsed_1305 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1306 = tensor.expand_shape %1511 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1512 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1306 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1307 = tensor.collapse_shape %1512 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1513 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1307 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1514 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_145 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1515 = linalg.batch_matmul ins(%1513, %1514 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1516 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_660, %1515 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1517 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1516, %1479 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1518 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1517, %cst_658 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1519 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1518, %cst_659 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1520 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1519 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1521 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_144 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1522 = linalg.batch_matmul ins(%1520, %1521 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1523 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_645, %1522 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1524 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1523 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1525 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1524 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1526 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_143 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1527 = linalg.batch_matmul ins(%1525, %1526 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1528 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_644, %1527 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1529 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1528, %1519 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1530 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1529, %cst_642 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1531 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1530, %cst_643 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1532 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1531 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1533 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_142 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1534 = linalg.batch_matmul ins(%1532, %1533 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1535 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_641, %1534 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1536 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1535 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1537 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1536 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1538 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_141 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1539 = linalg.batch_matmul ins(%1537, %1538 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1540 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_640, %1539 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1541 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1540, %1531 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1542 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1541, %cst_638 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1543 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1542, %cst_639 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1544 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1543 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1545 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_140 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1546 = linalg.batch_matmul ins(%1544, %1545 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1547 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_637, %1546 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1548 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1547 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1549 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1548 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1550 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_139 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1551 = linalg.batch_matmul ins(%1549, %1550 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1552 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_636, %1551 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1553 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1552, %1543 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1554 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1553, %cst_634 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1555 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1554, %cst_635 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1556 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1555 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1557 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_138 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1558 = linalg.batch_matmul ins(%1556, %1557 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1559 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_657, %1558 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1560 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1559 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1561 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1560 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1562 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_137 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1563 = linalg.batch_matmul ins(%1561, %1562 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1564 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_656, %1563 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1565 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1564, %1555 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1566 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1565, %cst_654 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1567 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1566, %cst_655 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1568 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1567 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1569 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_136 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1570 = linalg.batch_matmul ins(%1568, %1569 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1571 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_653, %1570 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1572 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1571, %1473 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1573 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1572, %cst_651 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1574 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1573, %cst_652 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1575 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1574 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1576 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_135 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1577 = linalg.batch_matmul ins(%1575, %1576 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1578 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_620, %1577 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1579 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1578, %cst_619 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1580 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1579, %cst_630 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1581 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_134 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1582 = linalg.batch_matmul ins(%1575, %1581 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1583 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_618, %1582 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1584 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1583, %cst_616 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1585 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1584, %cst_617 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1586 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1585 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1587 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_133 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1588 = linalg.batch_matmul ins(%1586, %1587 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1589 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_633, %1588 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1590 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_132 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1591 = linalg.batch_matmul ins(%1586, %1590 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1592 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_632, %1591 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1593 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_131 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1594 = linalg.batch_matmul ins(%1575, %1593 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1595 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_631, %1594 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1308 = tensor.expand_shape %1589 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1596 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1308 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1309 = tensor.expand_shape %1592 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1310 = tensor.expand_shape %1595 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1597 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1310 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1598 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1309 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1599 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1598 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1600 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1596 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1601 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1599 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1311 = tensor.collapse_shape %1600 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1312 = tensor.collapse_shape %1601 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1602 = linalg.batch_matmul ins(%collapsed_1311, %collapsed_1312 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1313 = tensor.expand_shape %1602 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1603 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1313, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1604 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1603, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1605:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1604 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1314 = tensor.expand_shape %1605#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1606 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1604, %expanded_1314 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1607 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1606 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1608 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1607 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1609 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1607, %1608 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1610 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1609 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1611 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1597 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1315 = tensor.collapse_shape %1610 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1316 = tensor.collapse_shape %1611 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1612 = linalg.batch_matmul ins(%collapsed_1315, %collapsed_1316 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1317 = tensor.expand_shape %1612 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1613 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1317 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1318 = tensor.collapse_shape %1613 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1614 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1318 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1615 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_130 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1616 = linalg.batch_matmul ins(%1614, %1615 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1617 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_630, %1616 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1618 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1617, %1580 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1619 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1618, %cst_628 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1620 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1619, %cst_629 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1621 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1620 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1622 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_129 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1623 = linalg.batch_matmul ins(%1621, %1622 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1624 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_615, %1623 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1625 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1624 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1626 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1625 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1627 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_128 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1628 = linalg.batch_matmul ins(%1626, %1627 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1629 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_614, %1628 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1630 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1629, %1620 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1631 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1630, %cst_612 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1632 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1631, %cst_613 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1633 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1632 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1634 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_127 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1635 = linalg.batch_matmul ins(%1633, %1634 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1636 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_611, %1635 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1637 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1636 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1638 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1637 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1639 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_126 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1640 = linalg.batch_matmul ins(%1638, %1639 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1641 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_610, %1640 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1642 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1641, %1632 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1643 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1642, %cst_608 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1644 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1643, %cst_609 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1645 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1644 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1646 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_125 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1647 = linalg.batch_matmul ins(%1645, %1646 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1648 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_607, %1647 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1649 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1648 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1650 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1649 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1651 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_124 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1652 = linalg.batch_matmul ins(%1650, %1651 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1653 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_606, %1652 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1654 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1653, %1644 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1655 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1654, %cst_604 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1656 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1655, %cst_605 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1657 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1656 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1658 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_123 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1659 = linalg.batch_matmul ins(%1657, %1658 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1660 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_627, %1659 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1661 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1660 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1662 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1661 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1663 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_122 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1664 = linalg.batch_matmul ins(%1662, %1663 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1665 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_626, %1664 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1666 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1665, %1656 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1667 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1666, %cst_624 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1668 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1667, %cst_625 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1669 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1668 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1670 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_121 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1671 = linalg.batch_matmul ins(%1669, %1670 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1672 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_623, %1671 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1673 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1672, %1574 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1674 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1673, %cst_621 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1675 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1674, %cst_622 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1676 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1675 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1677 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_120 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1678 = linalg.batch_matmul ins(%1676, %1677 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1679 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_590, %1678 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1680 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1679, %cst_589 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1681 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1680, %cst_600 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1682 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_119 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1683 = linalg.batch_matmul ins(%1676, %1682 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1684 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_588, %1683 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1685 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1684, %cst_586 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1686 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1685, %cst_587 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1687 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1686 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1688 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_118 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1689 = linalg.batch_matmul ins(%1687, %1688 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1690 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_603, %1689 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1691 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_117 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1692 = linalg.batch_matmul ins(%1687, %1691 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1693 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_602, %1692 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1694 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_116 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1695 = linalg.batch_matmul ins(%1676, %1694 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1696 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_601, %1695 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1319 = tensor.expand_shape %1690 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1697 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1319 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1320 = tensor.expand_shape %1693 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1321 = tensor.expand_shape %1696 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1698 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1321 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1699 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1320 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1700 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1699 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1701 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1697 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1702 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1700 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1322 = tensor.collapse_shape %1701 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1323 = tensor.collapse_shape %1702 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1703 = linalg.batch_matmul ins(%collapsed_1322, %collapsed_1323 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1324 = tensor.expand_shape %1703 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1704 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1324, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1705 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1704, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1706:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1705 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1325 = tensor.expand_shape %1706#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1707 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1705, %expanded_1325 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1708 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1707 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1709 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1708 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1710 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1708, %1709 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1711 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1710 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1712 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1698 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1326 = tensor.collapse_shape %1711 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1327 = tensor.collapse_shape %1712 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1713 = linalg.batch_matmul ins(%collapsed_1326, %collapsed_1327 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1328 = tensor.expand_shape %1713 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1714 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1328 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1329 = tensor.collapse_shape %1714 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1715 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1329 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1716 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_115 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1717 = linalg.batch_matmul ins(%1715, %1716 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1718 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_600, %1717 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1719 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1718, %1681 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1720 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1719, %cst_598 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1721 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1720, %cst_599 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1722 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1721 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1723 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_114 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1724 = linalg.batch_matmul ins(%1722, %1723 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1725 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_585, %1724 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1726 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1725 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1727 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1726 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1728 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_113 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1729 = linalg.batch_matmul ins(%1727, %1728 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1730 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_584, %1729 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1731 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1730, %1721 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1732 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1731, %cst_582 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1733 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1732, %cst_583 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1734 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1733 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1735 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_112 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1736 = linalg.batch_matmul ins(%1734, %1735 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1737 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_581, %1736 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1738 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1737 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1739 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1738 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1740 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_111 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1741 = linalg.batch_matmul ins(%1739, %1740 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1742 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_580, %1741 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1743 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1742, %1733 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1744 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1743, %cst_578 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1745 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1744, %cst_579 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1746 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1745 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1747 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_110 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1748 = linalg.batch_matmul ins(%1746, %1747 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1749 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_577, %1748 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1750 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1749 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1751 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1750 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1752 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_109 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1753 = linalg.batch_matmul ins(%1751, %1752 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1754 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_576, %1753 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1755 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1754, %1745 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1756 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1755, %cst_574 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1757 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1756, %cst_575 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1758 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1757 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1759 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_108 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1760 = linalg.batch_matmul ins(%1758, %1759 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1761 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_597, %1760 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1762 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1761 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1763 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1762 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1764 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_107 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1765 = linalg.batch_matmul ins(%1763, %1764 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1766 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_596, %1765 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1767 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1766, %1757 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1768 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1767, %cst_594 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1769 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1768, %cst_595 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1770 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1769 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1771 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_106 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1772 = linalg.batch_matmul ins(%1770, %1771 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1773 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_593, %1772 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1774 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1773, %1675 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1775 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1774, %cst_591 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1776 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1775, %cst_592 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1777 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1776 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1778 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_105 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1779 = linalg.batch_matmul ins(%1777, %1778 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1780 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_560, %1779 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1781 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1780, %cst_559 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1782 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1781, %cst_570 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1783 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_104 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1784 = linalg.batch_matmul ins(%1777, %1783 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1785 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_558, %1784 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1786 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1785, %cst_556 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1787 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1786, %cst_557 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1788 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1787 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1789 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_103 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1790 = linalg.batch_matmul ins(%1788, %1789 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1791 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_573, %1790 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1792 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_102 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1793 = linalg.batch_matmul ins(%1788, %1792 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1794 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_572, %1793 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1795 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_101 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1796 = linalg.batch_matmul ins(%1777, %1795 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1797 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_571, %1796 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1330 = tensor.expand_shape %1791 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1798 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1330 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1331 = tensor.expand_shape %1794 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1332 = tensor.expand_shape %1797 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1799 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1332 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1800 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1331 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1801 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1800 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1802 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1798 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1803 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1801 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1333 = tensor.collapse_shape %1802 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1334 = tensor.collapse_shape %1803 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1804 = linalg.batch_matmul ins(%collapsed_1333, %collapsed_1334 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1335 = tensor.expand_shape %1804 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1805 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1335, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1806 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1805, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1807:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1806 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1336 = tensor.expand_shape %1807#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1808 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1806, %expanded_1336 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1809 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1808 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1810 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1809 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1811 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1809, %1810 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1812 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1811 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1813 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1799 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1337 = tensor.collapse_shape %1812 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1338 = tensor.collapse_shape %1813 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1814 = linalg.batch_matmul ins(%collapsed_1337, %collapsed_1338 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1339 = tensor.expand_shape %1814 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1815 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1339 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1340 = tensor.collapse_shape %1815 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1816 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1340 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1817 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_100 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1818 = linalg.batch_matmul ins(%1816, %1817 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1819 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_570, %1818 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1820 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1819, %1782 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1821 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1820, %cst_568 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1822 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1821, %cst_569 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1823 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1822 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1824 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_99 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1825 = linalg.batch_matmul ins(%1823, %1824 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1826 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_555, %1825 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1827 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1826 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1828 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1827 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1829 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_98 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1830 = linalg.batch_matmul ins(%1828, %1829 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1831 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_554, %1830 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1832 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1831, %1822 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1833 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1832, %cst_552 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1834 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1833, %cst_553 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1835 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1834 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1836 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_97 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1837 = linalg.batch_matmul ins(%1835, %1836 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1838 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_551, %1837 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1839 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1838 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1840 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1839 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1841 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_96 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1842 = linalg.batch_matmul ins(%1840, %1841 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1843 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_550, %1842 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1844 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1843, %1834 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1845 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1844, %cst_548 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1846 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1845, %cst_549 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1847 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1846 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1848 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_95 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1849 = linalg.batch_matmul ins(%1847, %1848 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1850 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_547, %1849 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1851 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1850 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1852 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1851 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1853 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_94 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1854 = linalg.batch_matmul ins(%1852, %1853 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1855 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_546, %1854 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1856 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1855, %1846 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1857 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1856, %cst_544 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1858 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1857, %cst_545 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1859 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1858 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1860 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_93 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1861 = linalg.batch_matmul ins(%1859, %1860 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1862 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_567, %1861 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1863 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1862 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1864 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1863 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1865 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_92 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1866 = linalg.batch_matmul ins(%1864, %1865 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1867 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_566, %1866 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1868 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1867, %1858 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1869 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1868, %cst_564 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1870 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1869, %cst_565 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1871 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1870 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1872 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_91 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1873 = linalg.batch_matmul ins(%1871, %1872 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1874 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_563, %1873 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1875 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1874, %1776 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1876 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1875, %cst_561 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1877 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1876, %cst_562 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1878 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1877 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1879 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_90 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1880 = linalg.batch_matmul ins(%1878, %1879 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1881 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_530, %1880 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1882 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1881, %cst_529 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1883 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1882, %cst_540 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1884 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_89 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1885 = linalg.batch_matmul ins(%1878, %1884 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1886 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_528, %1885 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1887 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1886, %cst_526 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1888 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1887, %cst_527 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1889 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1888 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1890 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_88 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1891 = linalg.batch_matmul ins(%1889, %1890 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1892 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_543, %1891 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1893 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_87 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1894 = linalg.batch_matmul ins(%1889, %1893 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1895 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_542, %1894 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1896 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_86 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1897 = linalg.batch_matmul ins(%1878, %1896 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1898 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_541, %1897 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1341 = tensor.expand_shape %1892 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1899 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1341 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1342 = tensor.expand_shape %1895 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1343 = tensor.expand_shape %1898 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%1900 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1343 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1901 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1342 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1902 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1901 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%1903 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1899 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%1904 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1902 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1344 = tensor.collapse_shape %1903 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1345 = tensor.collapse_shape %1904 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%1905 = linalg.batch_matmul ins(%collapsed_1344, %collapsed_1345 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1346 = tensor.expand_shape %1905 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%1906 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1346, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1907 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1906, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1908:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1907 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1347 = tensor.expand_shape %1908#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%1909 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1907, %expanded_1347 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1910 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1909 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1911 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1910 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%1912 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1910, %1911 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%1913 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1912 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%1914 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1900 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1348 = tensor.collapse_shape %1913 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1349 = tensor.collapse_shape %1914 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%1915 = linalg.batch_matmul ins(%collapsed_1348, %collapsed_1349 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1350 = tensor.expand_shape %1915 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%1916 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1350 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1351 = tensor.collapse_shape %1916 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%1917 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1351 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1918 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_85 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1919 = linalg.batch_matmul ins(%1917, %1918 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1920 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_540, %1919 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1921 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1920, %1883 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1922 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1921, %cst_538 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1923 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1922, %cst_539 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1924 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1923 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1925 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_84 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1926 = linalg.batch_matmul ins(%1924, %1925 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1927 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_525, %1926 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1928 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1927 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1929 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1928 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1930 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_83 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1931 = linalg.batch_matmul ins(%1929, %1930 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1932 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_524, %1931 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1933 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1932, %1923 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1934 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1933, %cst_522 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1935 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1934, %cst_523 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1936 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1935 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1937 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_82 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1938 = linalg.batch_matmul ins(%1936, %1937 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1939 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_521, %1938 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1940 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1939 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1941 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1940 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1942 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_81 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1943 = linalg.batch_matmul ins(%1941, %1942 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1944 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_520, %1943 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1945 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1944, %1935 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1946 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1945, %cst_518 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1947 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1946, %cst_519 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1948 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1947 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1949 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_80 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1950 = linalg.batch_matmul ins(%1948, %1949 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1951 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_517, %1950 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1952 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1951 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1953 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1952 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1954 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_79 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1955 = linalg.batch_matmul ins(%1953, %1954 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1956 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_516, %1955 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1957 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1956, %1947 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1958 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1957, %cst_514 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1959 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1958, %cst_515 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1960 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1959 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1961 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_78 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1962 = linalg.batch_matmul ins(%1960, %1961 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1963 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_537, %1962 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1964 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1963 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%1965 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1964 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1966 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_77 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1967 = linalg.batch_matmul ins(%1965, %1966 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1968 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_536, %1967 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1969 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1968, %1959 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1970 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1969, %cst_534 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1971 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1970, %cst_535 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1972 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1971 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1973 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_76 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1974 = linalg.batch_matmul ins(%1972, %1973 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%1975 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_533, %1974 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1976 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1975, %1877 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1977 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1976, %cst_531 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1978 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1977, %cst_532 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%1979 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1978 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%1980 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_75 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1981 = linalg.batch_matmul ins(%1979, %1980 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1982 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_500, %1981 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1983 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1982, %cst_499 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1984 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1983, %cst_510 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1985 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_74 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1986 = linalg.batch_matmul ins(%1979, %1985 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1987 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_498, %1986 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1988 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1987, %cst_496 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1989 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1988, %cst_497 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1990 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1989 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1991 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_73 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1992 = linalg.batch_matmul ins(%1990, %1991 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1993 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_513, %1992 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1994 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_72 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%1995 = linalg.batch_matmul ins(%1990, %1994 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1996 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_512, %1995 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%1997 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_71 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%1998 = linalg.batch_matmul ins(%1979, %1997 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%1999 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_511, %1998 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1352 = tensor.expand_shape %1993 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2000 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1352 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1353 = tensor.expand_shape %1996 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1354 = tensor.expand_shape %1999 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2001 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1354 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2002 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1353 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2003 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2002 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%2004 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2000 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2005 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2003 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1355 = tensor.collapse_shape %2004 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1356 = tensor.collapse_shape %2005 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%2006 = linalg.batch_matmul ins(%collapsed_1355, %collapsed_1356 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1357 = tensor.expand_shape %2006 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%2007 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1357, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2008 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2007, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2009:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2008 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1358 = tensor.expand_shape %2009#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%2010 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2008, %expanded_1358 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2011 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2010 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2012 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2011 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%2013 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2011, %2012 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2014 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2013 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%2015 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2001 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1359 = tensor.collapse_shape %2014 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1360 = tensor.collapse_shape %2015 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%2016 = linalg.batch_matmul ins(%collapsed_1359, %collapsed_1360 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1361 = tensor.expand_shape %2016 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%2017 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1361 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1362 = tensor.collapse_shape %2017 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%2018 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1362 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2019 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_70 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2020 = linalg.batch_matmul ins(%2018, %2019 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2021 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_510, %2020 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2022 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2021, %1984 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2023 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2022, %cst_508 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2024 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2023, %cst_509 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2025 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2024 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2026 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_69 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2027 = linalg.batch_matmul ins(%2025, %2026 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2028 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_495, %2027 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2029 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2028 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2030 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2029 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2031 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_68 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2032 = linalg.batch_matmul ins(%2030, %2031 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2033 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_494, %2032 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2034 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2033, %2024 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2035 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2034, %cst_492 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2036 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2035, %cst_493 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2037 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2036 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2038 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_67 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2039 = linalg.batch_matmul ins(%2037, %2038 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2040 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_491, %2039 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2041 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2040 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2042 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2041 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2043 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_66 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2044 = linalg.batch_matmul ins(%2042, %2043 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2045 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_490, %2044 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2046 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2045, %2036 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2047 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2046, %cst_488 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2048 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2047, %cst_489 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2049 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2048 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2050 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_65 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2051 = linalg.batch_matmul ins(%2049, %2050 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2052 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_487, %2051 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2053 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2052 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2054 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2053 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2055 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_64 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2056 = linalg.batch_matmul ins(%2054, %2055 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2057 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_486, %2056 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2058 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2057, %2048 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2059 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2058, %cst_484 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2060 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2059, %cst_485 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2061 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2060 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2062 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_63 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2063 = linalg.batch_matmul ins(%2061, %2062 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2064 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_507, %2063 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2065 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2064 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2066 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2065 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2067 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_62 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2068 = linalg.batch_matmul ins(%2066, %2067 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2069 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_506, %2068 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2070 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2069, %2060 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2071 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2070, %cst_504 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2072 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2071, %cst_505 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2073 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2072 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2074 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_61 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2075 = linalg.batch_matmul ins(%2073, %2074 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2076 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_503, %2075 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2077 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2076, %1978 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2078 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2077, %cst_501 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2079 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2078, %cst_502 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2080 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2079 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2081 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_60 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2082 = linalg.batch_matmul ins(%2080, %2081 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2083 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_470, %2082 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2084 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2083, %cst_469 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2085 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2084, %cst_480 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2086 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_59 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2087 = linalg.batch_matmul ins(%2080, %2086 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2088 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_468, %2087 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2089 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2088, %cst_466 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2090 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2089, %cst_467 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2091 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2090 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2092 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_58 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2093 = linalg.batch_matmul ins(%2091, %2092 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2094 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_483, %2093 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2095 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_57 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2096 = linalg.batch_matmul ins(%2091, %2095 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2097 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_482, %2096 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2098 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_56 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2099 = linalg.batch_matmul ins(%2080, %2098 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2100 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_481, %2099 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1363 = tensor.expand_shape %2094 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2101 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1363 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1364 = tensor.expand_shape %2097 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1365 = tensor.expand_shape %2100 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2102 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1365 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2103 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1364 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2104 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2103 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%2105 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2101 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2106 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2104 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1366 = tensor.collapse_shape %2105 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1367 = tensor.collapse_shape %2106 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%2107 = linalg.batch_matmul ins(%collapsed_1366, %collapsed_1367 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1368 = tensor.expand_shape %2107 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%2108 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1368, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2109 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2108, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2110:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2109 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1369 = tensor.expand_shape %2110#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%2111 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2109, %expanded_1369 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2112 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2111 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2113 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2112 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%2114 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2112, %2113 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2115 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2114 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%2116 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2102 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1370 = tensor.collapse_shape %2115 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1371 = tensor.collapse_shape %2116 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%2117 = linalg.batch_matmul ins(%collapsed_1370, %collapsed_1371 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1372 = tensor.expand_shape %2117 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%2118 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1372 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1373 = tensor.collapse_shape %2118 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%2119 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1373 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2120 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_55 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2121 = linalg.batch_matmul ins(%2119, %2120 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2122 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_480, %2121 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2123 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2122, %2085 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2124 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2123, %cst_478 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2125 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2124, %cst_479 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2126 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2125 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2127 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_54 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2128 = linalg.batch_matmul ins(%2126, %2127 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2129 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_465, %2128 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2130 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2129 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2131 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2130 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2132 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_53 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2133 = linalg.batch_matmul ins(%2131, %2132 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2134 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_464, %2133 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2135 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2134, %2125 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2136 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2135, %cst_462 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2137 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2136, %cst_463 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2138 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2137 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2139 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_52 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2140 = linalg.batch_matmul ins(%2138, %2139 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2141 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_461, %2140 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2142 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2141 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2143 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2142 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2144 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_51 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2145 = linalg.batch_matmul ins(%2143, %2144 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2146 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_460, %2145 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2147 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2146, %2137 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2148 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2147, %cst_458 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2149 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2148, %cst_459 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2150 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2149 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2151 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_50 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2152 = linalg.batch_matmul ins(%2150, %2151 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2153 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_457, %2152 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2154 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2153 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2155 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2154 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2156 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_49 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2157 = linalg.batch_matmul ins(%2155, %2156 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2158 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_456, %2157 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2159 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2158, %2149 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2160 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2159, %cst_454 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2161 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2160, %cst_455 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2162 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2161 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2163 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_48 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2164 = linalg.batch_matmul ins(%2162, %2163 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2165 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_477, %2164 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2166 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2165 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2167 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2166 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2168 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_47 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2169 = linalg.batch_matmul ins(%2167, %2168 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2170 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_476, %2169 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2171 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2170, %2161 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2172 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2171, %cst_474 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2173 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2172, %cst_475 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2174 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2173 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2175 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_46 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2176 = linalg.batch_matmul ins(%2174, %2175 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2177 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_473, %2176 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2178 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2177, %2079 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2179 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2178, %cst_471 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2180 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2179, %cst_472 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2181 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2180 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2182 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_45 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2183 = linalg.batch_matmul ins(%2181, %2182 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2184 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_440, %2183 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2185 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2184, %cst_439 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2186 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2185, %cst_450 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2187 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_44 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2188 = linalg.batch_matmul ins(%2181, %2187 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2189 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_438, %2188 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2190 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2189, %cst_436 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2191 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2190, %cst_437 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2192 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2191 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2193 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_43 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2194 = linalg.batch_matmul ins(%2192, %2193 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2195 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_453, %2194 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2196 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_42 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2197 = linalg.batch_matmul ins(%2192, %2196 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2198 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_452, %2197 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2199 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_41 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2200 = linalg.batch_matmul ins(%2181, %2199 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2201 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_451, %2200 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1374 = tensor.expand_shape %2195 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2202 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1374 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1375 = tensor.expand_shape %2198 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1376 = tensor.expand_shape %2201 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2203 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1376 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2204 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1375 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2205 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2204 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%2206 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2202 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2207 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2205 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1377 = tensor.collapse_shape %2206 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1378 = tensor.collapse_shape %2207 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%2208 = linalg.batch_matmul ins(%collapsed_1377, %collapsed_1378 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1379 = tensor.expand_shape %2208 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%2209 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1379, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2210 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2209, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2211:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2210 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1380 = tensor.expand_shape %2211#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%2212 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2210, %expanded_1380 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2213 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2212 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2214 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2213 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%2215 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2213, %2214 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2216 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2215 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%2217 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2203 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1381 = tensor.collapse_shape %2216 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1382 = tensor.collapse_shape %2217 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%2218 = linalg.batch_matmul ins(%collapsed_1381, %collapsed_1382 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1383 = tensor.expand_shape %2218 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%2219 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1383 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1384 = tensor.collapse_shape %2219 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%2220 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1384 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2221 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_40 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2222 = linalg.batch_matmul ins(%2220, %2221 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2223 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_450, %2222 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2224 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2223, %2186 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2225 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2224, %cst_448 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2226 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2225, %cst_449 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2227 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2226 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2228 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_39 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2229 = linalg.batch_matmul ins(%2227, %2228 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2230 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_435, %2229 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2231 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2230 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2232 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2231 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2233 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_38 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2234 = linalg.batch_matmul ins(%2232, %2233 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2235 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_434, %2234 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2236 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2235, %2226 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2237 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2236, %cst_432 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2238 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2237, %cst_433 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2239 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2238 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2240 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_37 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2241 = linalg.batch_matmul ins(%2239, %2240 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2242 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_431, %2241 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2243 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2242 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2244 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2243 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2245 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_36 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2246 = linalg.batch_matmul ins(%2244, %2245 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2247 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_430, %2246 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2248 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2247, %2238 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2249 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2248, %cst_428 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2250 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2249, %cst_429 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2251 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2250 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2252 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_35 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2253 = linalg.batch_matmul ins(%2251, %2252 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2254 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_427, %2253 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2255 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2254 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2256 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2255 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2257 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_34 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2258 = linalg.batch_matmul ins(%2256, %2257 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2259 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_426, %2258 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2260 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2259, %2250 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2261 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2260, %cst_424 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2262 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2261, %cst_425 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2263 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2262 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2264 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_33 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2265 = linalg.batch_matmul ins(%2263, %2264 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2266 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_447, %2265 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2267 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2266 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2268 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2267 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2269 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_32 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2270 = linalg.batch_matmul ins(%2268, %2269 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2271 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_446, %2270 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2272 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2271, %2262 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2273 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2272, %cst_444 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2274 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2273, %cst_445 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2275 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2274 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2276 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_31 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2277 = linalg.batch_matmul ins(%2275, %2276 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2278 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_443, %2277 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2279 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2278, %2180 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2280 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2279, %cst_441 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2281 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2280, %cst_442 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2282 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2281 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2283 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_30 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2284 = linalg.batch_matmul ins(%2282, %2283 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2285 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_410, %2284 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2286 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2285, %cst_409 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2287 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2286, %cst_420 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2288 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_29 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2289 = linalg.batch_matmul ins(%2282, %2288 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2290 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_408, %2289 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2291 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2290, %cst_406 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2292 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2291, %cst_407 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2293 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2292 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2294 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_28 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2295 = linalg.batch_matmul ins(%2293, %2294 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2296 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_423, %2295 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2297 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_27 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2298 = linalg.batch_matmul ins(%2293, %2297 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2299 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_422, %2298 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2300 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_26 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2301 = linalg.batch_matmul ins(%2282, %2300 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2302 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_421, %2301 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1385 = tensor.expand_shape %2296 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2303 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1385 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1386 = tensor.expand_shape %2299 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1387 = tensor.expand_shape %2302 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2304 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1387 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2305 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1386 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2306 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2305 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%2307 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2303 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2308 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2306 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1388 = tensor.collapse_shape %2307 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1389 = tensor.collapse_shape %2308 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%2309 = linalg.batch_matmul ins(%collapsed_1388, %collapsed_1389 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1390 = tensor.expand_shape %2309 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%2310 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1390, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2311 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2310, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2312:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2311 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1391 = tensor.expand_shape %2312#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%2313 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2311, %expanded_1391 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2314 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2313 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2315 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2314 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%2316 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2314, %2315 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2317 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2316 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%2318 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2304 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1392 = tensor.collapse_shape %2317 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1393 = tensor.collapse_shape %2318 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%2319 = linalg.batch_matmul ins(%collapsed_1392, %collapsed_1393 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1394 = tensor.expand_shape %2319 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%2320 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1394 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1395 = tensor.collapse_shape %2320 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%2321 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1395 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2322 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_25 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2323 = linalg.batch_matmul ins(%2321, %2322 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2324 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_420, %2323 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2325 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2324, %2287 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2326 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2325, %cst_418 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2327 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2326, %cst_419 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2328 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2327 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2329 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_24 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2330 = linalg.batch_matmul ins(%2328, %2329 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2331 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_405, %2330 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2332 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2331 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2333 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2332 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2334 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_23 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2335 = linalg.batch_matmul ins(%2333, %2334 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2336 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_404, %2335 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2337 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2336, %2327 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2338 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2337, %cst_402 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2339 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2338, %cst_403 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2340 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2339 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2341 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_22 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2342 = linalg.batch_matmul ins(%2340, %2341 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2343 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_401, %2342 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2344 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2343 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2345 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2344 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2346 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_21 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2347 = linalg.batch_matmul ins(%2345, %2346 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2348 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_400, %2347 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2349 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2348, %2339 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2350 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2349, %cst_398 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2351 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2350, %cst_399 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2352 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2351 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2353 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_20 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2354 = linalg.batch_matmul ins(%2352, %2353 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2355 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_397, %2354 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2356 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2355 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2357 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2356 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2358 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_19 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2359 = linalg.batch_matmul ins(%2357, %2358 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2360 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_396, %2359 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2361 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2360, %2351 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2362 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2361, %cst_394 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2363 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2362, %cst_395 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2364 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2363 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2365 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_18 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2366 = linalg.batch_matmul ins(%2364, %2365 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2367 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_417, %2366 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2368 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2367 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2369 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2368 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2370 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_17 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2371 = linalg.batch_matmul ins(%2369, %2370 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2372 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_416, %2371 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2373 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2372, %2363 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2374 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2373, %cst_414 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2375 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2374, %cst_415 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2376 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2375 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2377 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_16 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2378 = linalg.batch_matmul ins(%2376, %2377 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2379 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_413, %2378 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2380 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2379, %2281 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2381 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2380, %cst_411 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2382 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2381, %cst_412 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2383 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2382 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2384 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_15 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2385 = linalg.batch_matmul ins(%2383, %2384 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2386 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_380, %2385 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2387 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2386, %cst_379 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2388 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2387, %cst_390 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2389 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_14 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2390 = linalg.batch_matmul ins(%2383, %2389 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2391 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_378, %2390 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2392 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2391, %cst_376 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2393 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2392, %cst_377 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2394 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2393 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2395 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_13 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2396 = linalg.batch_matmul ins(%2394, %2395 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2397 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_393, %2396 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2398 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_12 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2399 = linalg.batch_matmul ins(%2394, %2398 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2400 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_392, %2399 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2401 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_11 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2402 = linalg.batch_matmul ins(%2383, %2401 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2403 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_391, %2402 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%expanded_1396 = tensor.expand_shape %2397 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2404 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1396 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%expanded_1397 = tensor.expand_shape %2400 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%expanded_1398 = tensor.expand_shape %2403 [[0], [1], [2, 3]] output_shape [1, 128, 4, 32] : tensor<1x128x128xf32> into tensor<1x128x4x32xf32>
%2405 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1398 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2406 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1397 : tensor<1x128x4x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2407 = linalg.generic {indexing_maps = [#map9, #map11], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2406 : tensor<1x4x128x32xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%2408 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2404 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%2409 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2407 : tensor<1x4x32x128xf32>) outs(%71 : tensor<1x4x32x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x32x128xf32>
%collapsed_1399 = tensor.collapse_shape %2408 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%collapsed_1400 = tensor.collapse_shape %2409 [[0, 1], [2], [3]] : tensor<1x4x32x128xf32> into tensor<4x32x128xf32>
%2410 = linalg.batch_matmul ins(%collapsed_1399, %collapsed_1400 : tensor<4x128x32xf32>, tensor<4x32x128xf32>) outs(%76 : tensor<4x128x128xf32>) -> tensor<4x128x128xf32>
%expanded_1401 = tensor.expand_shape %2410 [[0, 1], [2], [3]] output_shape [1, 4, 128, 128] : tensor<4x128x128xf32> into tensor<1x4x128x128xf32>
%2411 = linalg.generic {indexing_maps = [#map12, #map13, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1401, %cst_1092 : tensor<1x4x128x128xf32>, tensor<f32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2412 = linalg.generic {indexing_maps = [#map12, #map14, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2411, %cst_1093 : tensor<1x4x128x128xf32>, tensor<1x1x1x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2413:2 = linalg.generic {indexing_maps = [#map9, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2412 : tensor<1x4x128x128xf32>) outs(%84, %82 : tensor<1x4x128xf32>, tensor<1x4x128xi64>) {
^bb0(%in: f32, %out: f32, %out_1408: i64):
%2492 = linalg.index 3 : index
%2493 = arith.index_cast %2492 : index to i64
%2494 = arith.maximumf %in, %out : f32
%2495 = arith.cmpf ogt, %in, %out : f32
%2496 = arith.select %2495, %2493, %out_1408 : i64
linalg.yield %2494, %2496 : f32, i64
} -> (tensor<1x4x128xf32>, tensor<1x4x128xi64>)
%expanded_1402 = tensor.expand_shape %2413#0 [[0], [1], [2, 3]] output_shape [1, 4, 128, 1] : tensor<1x4x128xf32> into tensor<1x4x128x1xf32>
%2414 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2412, %expanded_1402 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.subf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2415 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2414 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = math.exp %in : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2416 = linalg.generic {indexing_maps = [#map9, #map17], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2415 : tensor<1x4x128x128xf32>) outs(%89 : tensor<1x4x128x1xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.addf %in, %out : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x1xf32>
%2417 = linalg.generic {indexing_maps = [#map12, #map16, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2415, %2416 : tensor<1x4x128x128xf32>, tensor<1x4x128x1xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.divf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x4x128x128xf32>
%2418 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2417 : tensor<1x4x128x128xf32>) outs(%78 : tensor<1x4x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x128xf32>
%2419 = linalg.generic {indexing_maps = [#map12, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2405 : tensor<1x4x128x32xf32>) outs(%67 : tensor<1x4x128x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x4x128x32xf32>
%collapsed_1403 = tensor.collapse_shape %2418 [[0, 1], [2], [3]] : tensor<1x4x128x128xf32> into tensor<4x128x128xf32>
%collapsed_1404 = tensor.collapse_shape %2419 [[0, 1], [2], [3]] : tensor<1x4x128x32xf32> into tensor<4x128x32xf32>
%2420 = linalg.batch_matmul ins(%collapsed_1403, %collapsed_1404 : tensor<4x128x128xf32>, tensor<4x128x32xf32>) outs(%95 : tensor<4x128x32xf32>) -> tensor<4x128x32xf32>
%expanded_1405 = tensor.expand_shape %2420 [[0, 1], [2], [3]] output_shape [1, 4, 128, 32] : tensor<4x128x32xf32> into tensor<1x4x128x32xf32>
%2421 = linalg.generic {indexing_maps = [#map9, #map10], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_1405 : tensor<1x4x128x32xf32>) outs(%97 : tensor<1x128x4x32xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x4x32xf32>
%collapsed_1406 = tensor.collapse_shape %2421 [[0], [1], [2, 3]] : tensor<1x128x4x32xf32> into tensor<1x128x128xf32>
%2422 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_1406 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2423 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_10 : tensor<128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2424 = linalg.batch_matmul ins(%2422, %2423 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2425 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_390, %2424 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2426 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2425, %2388 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2427 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2426, %cst_388 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2428 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2427, %cst_389 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2429 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2428 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2430 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_9 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2431 = linalg.batch_matmul ins(%2429, %2430 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2432 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_375, %2431 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2433 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2432 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2434 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2433 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2435 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_8 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2436 = linalg.batch_matmul ins(%2434, %2435 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2437 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_374, %2436 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2438 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2437, %2428 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2439 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2438, %cst_372 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2440 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2439, %cst_373 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2441 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2440 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2442 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_7 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2443 = linalg.batch_matmul ins(%2441, %2442 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2444 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_371, %2443 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2445 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2444 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2446 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2445 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2447 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_6 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2448 = linalg.batch_matmul ins(%2446, %2447 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2449 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_370, %2448 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2450 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2449, %2440 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2451 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2450, %cst_368 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2452 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2451, %cst_369 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2453 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2452 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2454 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_5 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2455 = linalg.batch_matmul ins(%2453, %2454 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2456 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_367, %2455 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2457 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2456 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2458 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2457 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2459 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_4 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2460 = linalg.batch_matmul ins(%2458, %2459 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2461 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_366, %2460 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2462 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2461, %2452 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2463 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2462, %cst_364 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2464 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2463, %cst_365 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2465 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2464 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2466 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_3 : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2467 = linalg.batch_matmul ins(%2465, %2466 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2468 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_387, %2467 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2469 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2468 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
%2492 = arith.cmpf ugt, %in, %cst_0 : f32
%2493 = arith.select %2492, %in, %cst_0 : f32
linalg.yield %2493 : f32
} -> tensor<1x128x512xf32>
%2470 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2469 : tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2471 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_2 : tensor<512x128xf32>) outs(%44 : tensor<1x512x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x512x128xf32>
%2472 = linalg.batch_matmul ins(%2470, %2471 : tensor<1x128x512xf32>, tensor<1x512x128xf32>) outs(%47 : tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
%2473 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_386, %2472 : tensor<128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2474 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2473, %2464 : tensor<1x128x128xf32>, tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2475 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2474, %cst_384 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2476 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2475, %cst_385 : tensor<1x128x128xf32>, tensor<128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x128xf32>
%2477 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2476 : tensor<1x128x128xf32>) outs(%46 : tensor<1x128x128xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x128xf32>
%2478 = linalg.generic {indexing_maps = [#map7, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst : tensor<128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<1x128x512xf32>
%2479 = linalg.batch_matmul ins(%2477, %2478 : tensor<1x128x128xf32>, tensor<1x128x512xf32>) outs(%30 : tensor<1x128x512xf32>) -> tensor<1x128x512xf32>
%2480 = linalg.generic {indexing_maps = [#map8, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_383, %2479 : tensor<512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2481 = linalg.generic {indexing_maps = [#map5, #map5, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2480, %2382 : tensor<1x128x512xf32>, tensor<1x128x512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2482 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2481, %cst_381 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.mulf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2483 = linalg.generic {indexing_maps = [#map5, #map8, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%2482, %cst_382 : tensor<1x128x512xf32>, tensor<512xf32>) outs(%29 : tensor<1x128x512xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x128x512xf32>
%2484 = tensor.empty() : tensor<1x1x512xf32>
%2485 = linalg.generic {indexing_maps = [#map18, #map6], iterator_types = ["parallel", "parallel", "parallel"]} ins(%cst_1096 : tensor<1xi64>) outs(%2484 : tensor<1x1x512xf32>) {
^bb0(%in: i64, %out: f32):
%2492 = arith.index_cast %in : i64 to index
%2493 = linalg.index 0 : index
%2494 = linalg.index 2 : index
%extracted_1408 = tensor.extract %2483[%2493, %2492, %2494] : tensor<1x128x512xf32>
linalg.yield %extracted_1408 : f32
} -> tensor<1x1x512xf32>
%collapsed_1407 = tensor.collapse_shape %2485 [[0], [1, 2]] : tensor<1x1x512xf32> into tensor<1x512xf32>
%2486 = tensor.empty() : tensor<512x2xf32>
%2487 = linalg.generic {indexing_maps = [#map1, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_363 : tensor<2x512xf32>) outs(%2486 : tensor<512x2xf32>) {
^bb0(%in: f32, %out: f32):
linalg.yield %in : f32
} -> tensor<512x2xf32>
%2488 = tensor.empty() : tensor<1x2xf32>
%2489 = linalg.fill ins(%cst_0 : f32) outs(%2488 : tensor<1x2xf32>) -> tensor<1x2xf32>
%2490 = linalg.matmul ins(%collapsed_1407, %2487 : tensor<1x512xf32>, tensor<512x2xf32>) outs(%2489 : tensor<1x2xf32>) -> tensor<1x2xf32>
%2491 = linalg.generic {indexing_maps = [#map, #map19, #map1], iterator_types = ["parallel", "parallel"]} ins(%2490, %cst_362 : tensor<1x2xf32>, tensor<2xf32>) outs(%2488 : tensor<1x2xf32>) {
^bb0(%in: f32, %in_1408: f32, %out: f32):
%2492 = arith.addf %in, %in_1408 : f32
linalg.yield %2492 : f32
} -> tensor<1x2xf32>
return %2491 : tensor<1x2xf32>
}
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment