v_diff.mlir

#map0 = affine_map<(d0, d1) -> (d0, d1)>
#map1 = affine_map<(d0, d1) -> (d0, 0)>
#map2 = affine_map<(d0, d1) -> ()>
#map3 = affine_map<(d0, d1) -> (d1, d0)>
#map4 = affine_map<(d0, d1) -> (d1)>
#map5 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (0, d1, 0, d3, 0, 0, 0, 0)>
#map6 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5, d6, d7)>
#map7 = affine_map<(d0, d1, d2, d3) -> (d1)>
#map8 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map9 = affine_map<(d0, d1, d2, d3) -> (d0, d1, 0, 0)>
#map10 = affine_map<(d0, d1, d2, d3) -> (d0, 0, 0, 0)>
#map11 = affine_map<(d0, d1, d2, d3) -> (d0, 0, d2, d3)>
#map12 = affine_map<(d0, d1, d2, d3) -> ()>
#map13 = affine_map<(d0, d1, d2, d3) -> (d1, 0, 0)>
#map14 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d3, d2)>
#map15 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>
#map16 = affine_map<(d0, d1, d2, d3) -> (d2, 0)>
#map17 = affine_map<(d0, d1, d2, d3) -> (d3)>
module attributes {torch.debug_module_name = "model_inference"} {
  func.func @forward(%arg0: tensor<2x3x256x256xf32>, %arg1: tensor<2xf32>, %arg2: tensor<2x512xf32>) -> tensor<2x3x256x256xf32> {
    %cst = arith.constant 1.000000e+00 : f64
    %cst_0 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64x1xf32>
    %cst_1 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x640xf32>
    %cst_2 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_3 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_4 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_5 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x640xf32>
    %cst_6 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_7 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_8 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_9 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_10 = arith.constant dense<[[0.0985220074], [1.14068091], [0.969601631], [1.7801584], [0.453004777], [-0.784689903], [1.25971174], [1.0398941]]> : tensor<8x1xf32>
    %cst_11 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x19x3x3xf32>
    %cst_12 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_13 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_14 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_15 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_16 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_17 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x19x1x1xf32>
    %cst_18 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_19 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_20 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_21 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_22 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_23 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_24 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_25 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_26 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_27 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_28 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_29 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_30 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_31 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_32 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_33 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_34 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_35 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_36 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x128x3x3xf32>
    %cst_37 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_38 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_39 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_40 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_41 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_42 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x128x1x1xf32>
    %cst_43 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_44 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_45 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_46 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_47 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_48 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_49 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_50 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_51 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_52 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_53 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_54 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_55 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_56 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_57 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_58 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_59 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_60 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_61 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_62 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_63 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_64 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_65 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_66 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_67 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_68 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_69 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_70 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_71 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_72 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_73 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_74 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_75 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_76 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_77 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_78 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_79 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_80 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_81 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_82 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_83 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_84 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_85 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x256x3x3xf32>
    %cst_86 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_87 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_88 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_89 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_90 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_91 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x256x1x1xf32>
    %cst_92 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_93 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_94 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_95 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_96 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_97 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_98 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_99 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_100 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_101 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_102 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_103 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_104 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_105 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_106 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_107 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_108 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_109 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_110 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_111 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_112 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_113 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_114 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_115 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_116 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_117 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_118 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_119 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_120 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_121 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_122 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_123 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_124 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_125 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_126 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_127 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_128 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_129 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_130 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_131 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_132 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_133 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_134 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_135 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_136 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_137 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_138 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_139 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_140 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_141 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_142 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_143 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_144 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_145 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_146 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_147 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_148 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_149 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_150 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x512x3x3xf32>
    %cst_151 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_152 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_153 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_154 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_155 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_156 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x512x1x1xf32>
    %cst_157 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_158 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_159 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_160 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_161 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_162 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_163 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_164 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_165 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_166 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_167 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_168 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_169 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_170 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_171 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_172 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_173 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_174 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_175 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_176 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_177 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_178 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_179 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_180 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_181 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_182 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_183 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_184 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_185 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_186 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_187 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_188 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_189 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_190 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_191 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_192 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_193 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_194 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_195 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_196 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_197 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_198 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_199 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_200 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_201 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_202 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_203 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_204 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_205 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_206 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_207 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_208 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_209 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_210 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_211 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_212 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_213 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_214 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_215 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_216 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_217 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_218 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_219 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_220 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_221 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_222 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_223 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_224 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_225 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_226 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_227 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_228 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_229 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_230 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_231 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_232 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_233 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_234 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_235 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_236 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_237 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_238 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_239 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_240 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_241 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_242 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_243 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_244 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_245 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_246 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_247 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_248 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_249 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_250 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_251 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_252 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_253 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_254 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_255 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_256 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_257 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_258 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_259 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_260 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_261 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_262 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_263 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_264 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_265 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_266 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_267 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_268 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_269 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_270 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_271 = arith.constant dense<[[0], [0], [0], [1], [1], [2], [2], [3]]> : tensor<8x1xi64>
    %cst_272 = arith.constant dense<[0, 0, 0, 1, 1, 2, 2, 3]> : tensor<8xi64>
    %cst_273 = arith.constant dense<[[0], [1], [1], [2], [2], [3], [3], [3]]> : tensor<8x1xi64>
    %cst_274 = arith.constant dense<[0, 1, 1, 2, 2, 3, 3, 3]> : tensor<8xi64>
    %cst_275 = arith.constant dense<[[1.000000e+00], [7.500000e-01], [2.500000e-01], [7.500000e-01], [2.500000e-01], [7.500000e-01], [2.500000e-01], [7.500000e-01]]> : tensor<8x1xf32>
    %cst_276 = arith.constant dense<[[0.000000e+00], [2.500000e-01], [7.500000e-01], [2.500000e-01], [7.500000e-01], [2.500000e-01], [7.500000e-01], [2.500000e-01]]> : tensor<8x1xf32>
    %cst_277 = arith.constant dense<[1.000000e+00, 7.500000e-01, 2.500000e-01, 7.500000e-01, 2.500000e-01, 7.500000e-01, 2.500000e-01, 7.500000e-01]> : tensor<8xf32>
    %cst_278 = arith.constant dense<[0.000000e+00, 2.500000e-01, 7.500000e-01, 2.500000e-01, 7.500000e-01, 2.500000e-01, 7.500000e-01, 2.500000e-01]> : tensor<8xf32>
    %cst_279 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x2048x3x3xf32>
    %cst_280 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_281 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_282 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_283 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_284 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_285 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x2048x1x1xf32>
    %cst_286 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_287 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_288 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_289 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_290 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_291 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_292 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_293 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_294 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_295 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_296 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_297 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_298 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_299 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_300 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_301 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_302 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_303 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_304 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_305 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_306 = arith.constant dense<0.707106769> : tensor<1024x1x1xf32>
    %cst_307 = arith.constant dense<0.000000e+00> : tensor<1024x1x1xf32>
    %cst_308 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072x1024x1x1xf32>
    %cst_309 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3072xf32>
    %cst_310 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x1x1xf32>
    %cst_311 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_312 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024x3x3xf32>
    %cst_313 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024xf32>
    %cst_314 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<2048x1024xf32>
    %cst_315 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024x3x3xf32>
    %cst_316 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_317 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_318 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024x1x1xf32>
    %cst_319 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_320 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_321 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_322 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_323 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16x1xi64>
    %cst_324 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16xi64>
    %cst_325 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16x1xi64>
    %cst_326 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16xi64>
    %cst_327 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16x1xf32>
    %cst_328 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16x1xf32>
    %cst_329 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16xf32>
    %cst_330 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<16xf32>
    %cst_331 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024x3x3xf32>
    %cst_332 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_333 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_334 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_335 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_336 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_337 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024x1x1xf32>
    %cst_338 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_339 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_340 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_341 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_342 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_343 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_344 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_345 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_346 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_347 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_348 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_349 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_350 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_351 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_352 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_353 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_354 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_355 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_356 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_357 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_358 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_359 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_360 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_361 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_362 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_363 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_364 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_365 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_366 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_367 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_368 = arith.constant dense<0.707106769> : tensor<512x1x1xf32>
    %cst_369 = arith.constant dense<0.000000e+00> : tensor<512x1x1xf32>
    %cst_370 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536x512x1x1xf32>
    %cst_371 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1536xf32>
    %cst_372 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x1x1xf32>
    %cst_373 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_374 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32x1xi64>
    %cst_375 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32xi64>
    %cst_376 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32x1xi64>
    %cst_377 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32xi64>
    %cst_378 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32x1xf32>
    %cst_379 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32x1xf32>
    %cst_380 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32xf32>
    %cst_381 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<32xf32>
    %cst_382 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024x3x3xf32>
    %cst_383 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_384 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_385 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_386 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_387 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_388 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024x1x1xf32>
    %cst_389 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_390 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_391 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_392 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_393 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_394 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_395 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_396 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_397 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_398 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_399 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_400 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_401 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x512x3x3xf32>
    %cst_402 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512xf32>
    %cst_403 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<1024x1024xf32>
    %cst_404 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x512x3x3xf32>
    %cst_405 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_406 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_407 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x512x1x1xf32>
    %cst_408 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64x1xi64>
    %cst_409 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64xi64>
    %cst_410 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64x1xi64>
    %cst_411 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64xi64>
    %cst_412 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64x1xf32>
    %cst_413 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64x1xf32>
    %cst_414 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64xf32>
    %cst_415 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<64xf32>
    %cst_416 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x512x3x3xf32>
    %cst_417 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_418 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_419 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_420 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_421 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_422 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x512x1x1xf32>
    %cst_423 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_424 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_425 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_426 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_427 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_428 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_429 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_430 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_431 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_432 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_433 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_434 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_435 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_436 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_437 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_438 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_439 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_440 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_441 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x1xi64>
    %cst_442 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xi64>
    %cst_443 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x1xi64>
    %cst_444 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xi64>
    %cst_445 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x1xf32>
    %cst_446 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x1xf32>
    %cst_447 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_448 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_449 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x512x3x3xf32>
    %cst_450 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_451 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_452 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_453 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_454 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_455 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x512x1x1xf32>
    %cst_456 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_457 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_458 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_459 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_460 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_461 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_462 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_463 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_464 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_465 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_466 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_467 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_468 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x256x3x3xf32>
    %cst_469 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_470 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<512x1024xf32>
    %cst_471 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x256x3x3xf32>
    %cst_472 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_473 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_474 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x256x1x1xf32>
    %cst_475 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1xi64>
    %cst_476 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xi64>
    %cst_477 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1xi64>
    %cst_478 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xi64>
    %cst_479 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1xf32>
    %cst_480 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1xf32>
    %cst_481 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_482 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256xf32>
    %cst_483 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x256x3x3xf32>
    %cst_484 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_485 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_486 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_487 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_488 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_489 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x256x1x1xf32>
    %cst_490 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_491 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_492 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_493 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_494 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_495 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_496 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_497 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_498 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_499 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_500 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_501 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_502 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128x128x3x3xf32>
    %cst_503 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<128xf32>
    %cst_504 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<256x1024xf32>
    %cst_505 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3x128x3x3xf32>
    %cst_506 = arith.constant dense<[-0.0193842854, 0.0141116232, -0.00911529362]> : tensor<3xf32>
    %cst_507 = arith.constant opaque<"elided_large_const", "0xDEADBEEF"> : tensor<3x128x1x1xf32>
    %cst_508 = arith.constant dense<22.627416997969522> : tensor<f64>
    %cst_509 = arith.constant dense<6.2831853071795862> : tensor<f64>
    %cst_510 = arith.constant dense<1.000000e-05> : tensor<f64>
    %cst_511 = arith.constant dense<1> : tensor<i64>
    %cst_512 = arith.constant dense<0.35355339059327379> : tensor<f64>
    %cst_513 = arith.constant 9.9999999999999998E-13 : f64
    %cst_514 = arith.constant 0.000000e+00 : f32
    %cst_515 = arith.constant -3.40282347E+38 : f32
    %cst_516 = arith.constant 0x4B000000 : f32
    %cst_517 = arith.constant 0x4A800000 : f32
    %cst_518 = arith.constant 0x49800000 : f32
    %cst_519 = arith.constant 5.242880e+05 : f32
    %cst_520 = arith.constant 1.310720e+05 : f32
    %cst_521 = arith.constant 6.553600e+04 : f32
    %cst_522 = arith.constant 4.000000e+00 : f32
    %cst_523 = arith.constant 1.638400e+04 : f32
    %cst_524 = arith.constant 3.276800e+04 : f32
    %cst_525 = arith.constant 2.621440e+05 : f32
    %cst_526 = arith.constant 0x4A000000 : f32
    %c0_i64 = arith.constant 0 : i64
    %0 = linalg.init_tensor [2, 512] : tensor<2x512xf32>
    %1 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%arg2 : tensor<2x512xf32>) outs(%0 : tensor<2x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512xf32>
    %2 = linalg.init_tensor [2, 1] : tensor<2x1xf32>
    %3 = linalg.fill ins(%cst_514 : f32) outs(%2 : tensor<2x1xf32>) -> tensor<2x1xf32>
    %4 = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel", "reduction"]} ins(%1 : tensor<2x512xf32>) outs(%3 : tensor<2x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1xf32>
    %5 = linalg.generic {indexing_maps = [#map1, #map0], iterator_types = ["parallel", "parallel"]} ins(%4 : tensor<2x1xf32>) outs(%2 : tensor<2x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.sqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1xf32>
    %6 = linalg.generic {indexing_maps = [#map1, #map0], iterator_types = ["parallel", "parallel"]} ins(%5 : tensor<2x1xf32>) outs(%2 : tensor<2x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.truncf %cst_513 : f64 to f32
      %5341 = arith.cmpf ult, %arg3, %5340 : f32
      %5342 = arith.select %5341, %5340, %arg3 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1xf32>
    %7 = linalg.generic {indexing_maps = [#map1, #map0], iterator_types = ["parallel", "parallel"]} ins(%6 : tensor<2x1xf32>) outs(%0 : tensor<2x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512xf32>
    %8 = linalg.generic {indexing_maps = [#map0, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%arg2, %7 : tensor<2x512xf32>, tensor<2x512xf32>) outs(%0 : tensor<2x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512xf32>
    %9 = linalg.generic {indexing_maps = [#map0, #map2, #map0], iterator_types = ["parallel", "parallel"]} ins(%8, %cst_508 : tensor<2x512xf32>, tensor<f64>) outs(%0 : tensor<2x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512xf32>
    %10 = tensor.expand_shape %arg1 [[0, 1]] : tensor<2xf32> into tensor<2x1xf32>
    %11 = linalg.generic {indexing_maps = [#map1, #map2, #map0], iterator_types = ["parallel", "parallel"]} ins(%10, %cst_509 : tensor<2x1xf32>, tensor<f64>) outs(%2 : tensor<2x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1xf32>
    %12 = linalg.init_tensor [1, 64] : tensor<1x64xf32>
    %13 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_0 : tensor<64x1xf32>) outs(%12 : tensor<1x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1x64xf32>
    %14 = linalg.init_tensor [2, 64] : tensor<2x64xf32>
    %15 = linalg.fill ins(%cst_514 : f32) outs(%14 : tensor<2x64xf32>) -> tensor<2x64xf32>
    %16 = linalg.matmul ins(%11, %13 : tensor<2x1xf32>, tensor<1x64xf32>) outs(%15 : tensor<2x64xf32>) -> tensor<2x64xf32>
    %17 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%16 : tensor<2x64xf32>) outs(%14 : tensor<2x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.cos %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x64xf32>
    %18 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%16 : tensor<2x64xf32>) outs(%14 : tensor<2x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.sin %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x64xf32>
    %19 = linalg.init_tensor [2, 128] : tensor<2x128xf32>
    %20 = tensor.insert_slice %17 into %19[0, 0] [2, 64] [1, 1] : tensor<2x64xf32> into tensor<2x128xf32>
    %21 = tensor.insert_slice %18 into %20[0, 64] [2, 64] [1, 1] : tensor<2x64xf32> into tensor<2x128xf32>
    %22 = linalg.init_tensor [2, 640] : tensor<2x640xf32>
    %23 = tensor.insert_slice %9 into %22[0, 0] [2, 512] [1, 1] : tensor<2x512xf32> into tensor<2x640xf32>
    %24 = tensor.insert_slice %21 into %23[0, 512] [2, 128] [1, 1] : tensor<2x128xf32> into tensor<2x640xf32>
    %25 = linalg.init_tensor [640, 1024] : tensor<640x1024xf32>
    %26 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_1 : tensor<1024x640xf32>) outs(%25 : tensor<640x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<640x1024xf32>
    %27 = linalg.init_tensor [2, 1024] : tensor<2x1024xf32>
    %28 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %29 = linalg.matmul ins(%24, %26 : tensor<2x640xf32>, tensor<640x1024xf32>) outs(%28 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %30 = linalg.generic {indexing_maps = [#map4, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%cst_2, %29 : tensor<1024xf32>, tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024xf32>
    %31 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%30 : tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024xf32>
    %32 = linalg.init_tensor [1024, 1024] : tensor<1024x1024xf32>
    %33 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_3 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %34 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %35 = linalg.matmul ins(%31, %33 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%34 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %36 = linalg.generic {indexing_maps = [#map4, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%cst_4, %35 : tensor<1024xf32>, tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024xf32>
    %37 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%36 : tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024xf32>
    %38 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_5 : tensor<1024x640xf32>) outs(%25 : tensor<640x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<640x1024xf32>
    %39 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %40 = linalg.matmul ins(%24, %38 : tensor<2x640xf32>, tensor<640x1024xf32>) outs(%39 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %41 = linalg.generic {indexing_maps = [#map0, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%37, %40 : tensor<2x1024xf32>, tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024xf32>
    %42 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_6 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %43 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %44 = linalg.matmul ins(%41, %42 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%43 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %45 = linalg.generic {indexing_maps = [#map4, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%cst_7, %44 : tensor<1024xf32>, tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024xf32>
    %46 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%45 : tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024xf32>
    %47 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_8 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %48 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %49 = linalg.matmul ins(%46, %47 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%48 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %50 = linalg.generic {indexing_maps = [#map4, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%cst_9, %49 : tensor<1024xf32>, tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024xf32>
    %51 = linalg.generic {indexing_maps = [#map0, #map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%50, %41 : tensor<2x1024xf32>, tensor<2x1024xf32>) outs(%27 : tensor<2x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024xf32>
    %52 = linalg.generic {indexing_maps = [#map1, #map2, #map0], iterator_types = ["parallel", "parallel"]} ins(%10, %cst_509 : tensor<2x1xf32>, tensor<f64>) outs(%2 : tensor<2x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1xf32>
    %53 = linalg.init_tensor [1, 8] : tensor<1x8xf32>
    %54 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_10 : tensor<8x1xf32>) outs(%53 : tensor<1x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1x8xf32>
    %55 = linalg.init_tensor [2, 8] : tensor<2x8xf32>
    %56 = linalg.fill ins(%cst_514 : f32) outs(%55 : tensor<2x8xf32>) -> tensor<2x8xf32>
    %57 = linalg.matmul ins(%52, %54 : tensor<2x1xf32>, tensor<1x8xf32>) outs(%56 : tensor<2x8xf32>) -> tensor<2x8xf32>
    %58 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%57 : tensor<2x8xf32>) outs(%55 : tensor<2x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.cos %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8xf32>
    %59 = linalg.generic {indexing_maps = [#map0, #map0], iterator_types = ["parallel", "parallel"]} ins(%57 : tensor<2x8xf32>) outs(%55 : tensor<2x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.sin %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8xf32>
    %60 = linalg.init_tensor [2, 16] : tensor<2x16xf32>
    %61 = tensor.insert_slice %58 into %60[0, 0] [2, 8] [1, 1] : tensor<2x8xf32> into tensor<2x16xf32>
    %62 = tensor.insert_slice %59 into %61[0, 8] [2, 8] [1, 1] : tensor<2x8xf32> into tensor<2x16xf32>
    %63 = tensor.expand_shape %62 [[0, 1], [2, 3, 4, 5, 6, 7]] : tensor<2x16xf32> into tensor<1x2x1x16x1x1x1x1xf32>
    %64 = linalg.init_tensor [1, 2, 1, 16, 256, 1, 256, 1] : tensor<1x2x1x16x256x1x256x1xf32>
    %65 = linalg.generic {indexing_maps = [#map5, #map6], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%63 : tensor<1x2x1x16x1x1x1x1xf32>) outs(%64 : tensor<1x2x1x16x256x1x256x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1x2x1x16x256x1x256x1xf32>
    %66 = tensor.collapse_shape %65 [[0, 1], [2, 3], [4], [5, 6, 7]] : tensor<1x2x1x16x256x1x256x1xf32> into tensor<2x16x256x256xf32>
    %67 = linalg.init_tensor [2, 19, 256, 256] : tensor<2x19x256x256xf32>
    %68 = tensor.insert_slice %arg0 into %67[0, 0, 0, 0] [2, 3, 256, 256] [1, 1, 1, 1] : tensor<2x3x256x256xf32> into tensor<2x19x256x256xf32>
    %69 = tensor.insert_slice %66 into %68[0, 3, 0, 0] [2, 16, 256, 256] [1, 1, 1, 1] : tensor<2x16x256x256xf32> into tensor<2x19x256x256xf32>
    %70 = tensor.pad %69 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x19x256x256xf32> to tensor<2x19x258x258xf32>
    %71 = linalg.init_tensor [2, 128, 256, 256] : tensor<2x128x256x256xf32>
    %72 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_12 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %73 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%70, %cst_11 : tensor<2x19x258x258xf32>, tensor<128x19x3x3xf32>) outs(%72 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %74 = tensor.collapse_shape %73 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %75 = tensor.expand_shape %74 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %76 = linalg.init_tensor [2, 1, 1, 1] : tensor<2x1x1x1xf32>
    %77 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %78 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%75 : tensor<2x1x128x65536xf32>) outs(%77 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %79 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%78 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %80 = linalg.init_tensor [2, 1, 128, 65536] : tensor<2x1x128x65536xf32>
    %81 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%75, %79 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %82 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%81 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %83 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %84 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%82 : tensor<2x1x128x65536xf32>) outs(%83 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %85 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%84 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %86 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %87 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%75 : tensor<2x1x128x65536xf32>) outs(%86 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %88 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%87 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %89 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%85, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %90 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%89 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %91 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%75, %88 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %92 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%91, %90 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %93 = tensor.collapse_shape %92 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %94 = tensor.expand_shape %93 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %95 = linalg.init_tensor [1024, 256] : tensor<1024x256xf32>
    %96 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_13 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %97 = linalg.init_tensor [2, 256] : tensor<2x256xf32>
    %98 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %99 = linalg.matmul ins(%51, %96 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%98 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %100 = tensor.extract_slice %99[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %101 = tensor.extract_slice %99[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %102 = tensor.expand_shape %101 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %103 = tensor.expand_shape %100 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %104 = linalg.init_tensor [2, 128, 1, 1] : tensor<2x128x1x1xf32>
    %105 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%103, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %106 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%94, %105 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %107 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%102, %106 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %108 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%107 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %109 = tensor.pad %108 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %110 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_15 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %111 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%109, %cst_14 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%110 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %112 = tensor.collapse_shape %111 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %113 = tensor.expand_shape %112 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %114 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %115 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%113 : tensor<2x1x128x65536xf32>) outs(%114 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %116 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%115 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %117 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%113, %116 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %118 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%117 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %119 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %120 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%118 : tensor<2x1x128x65536xf32>) outs(%119 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %121 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%120 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %122 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %123 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%113 : tensor<2x1x128x65536xf32>) outs(%122 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %124 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%123 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %125 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%121, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %126 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%125 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %127 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%113, %124 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %128 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%127, %126 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %129 = tensor.collapse_shape %128 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %130 = tensor.expand_shape %129 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %131 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_16 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %132 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %133 = linalg.matmul ins(%51, %131 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%132 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %134 = tensor.extract_slice %133[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %135 = tensor.extract_slice %133[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %136 = tensor.expand_shape %135 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %137 = tensor.expand_shape %134 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %138 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%137, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %139 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%130, %138 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %140 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%136, %139 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %141 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%140 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %142 = linalg.fill ins(%cst_514 : f32) outs(%71 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %143 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%69, %cst_17 : tensor<2x19x256x256xf32>, tensor<128x19x1x1xf32>) outs(%142 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %144 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%141, %143 : tensor<2x128x256x256xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %145 = tensor.pad %144 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %146 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_19 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %147 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%145, %cst_18 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%146 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %148 = tensor.collapse_shape %147 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %149 = tensor.expand_shape %148 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %150 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %151 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%149 : tensor<2x1x128x65536xf32>) outs(%150 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %152 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%151 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %153 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%149, %152 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %154 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%153 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %155 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %156 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%154 : tensor<2x1x128x65536xf32>) outs(%155 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %157 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%156 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %158 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %159 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%149 : tensor<2x1x128x65536xf32>) outs(%158 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %160 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%159 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %161 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%157, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %162 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%161 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %163 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%149, %160 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %164 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%163, %162 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %165 = tensor.collapse_shape %164 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %166 = tensor.expand_shape %165 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %167 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_20 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %168 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %169 = linalg.matmul ins(%51, %167 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%168 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %170 = tensor.extract_slice %169[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %171 = tensor.extract_slice %169[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %172 = tensor.expand_shape %171 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %173 = tensor.expand_shape %170 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %174 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%173, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %175 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%166, %174 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %176 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%172, %175 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %177 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%176 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %178 = tensor.pad %177 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %179 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_22 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %180 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%178, %cst_21 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%179 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %181 = tensor.collapse_shape %180 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %182 = tensor.expand_shape %181 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %183 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %184 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%182 : tensor<2x1x128x65536xf32>) outs(%183 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %185 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%184 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %186 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%182, %185 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %187 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%186 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %188 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %189 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%187 : tensor<2x1x128x65536xf32>) outs(%188 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %190 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%189 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %191 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %192 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%182 : tensor<2x1x128x65536xf32>) outs(%191 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %193 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%192 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %194 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%190, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %195 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%194 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %196 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%182, %193 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %197 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%196, %195 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %198 = tensor.collapse_shape %197 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %199 = tensor.expand_shape %198 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %200 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_23 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %201 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %202 = linalg.matmul ins(%51, %200 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%201 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %203 = tensor.extract_slice %202[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %204 = tensor.extract_slice %202[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %205 = tensor.expand_shape %204 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %206 = tensor.expand_shape %203 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %207 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%206, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %208 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%199, %207 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %209 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%205, %208 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %210 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%209 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %211 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%210, %144 : tensor<2x128x256x256xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %212 = tensor.pad %211 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %213 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_25 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %214 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%212, %cst_24 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%213 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %215 = tensor.collapse_shape %214 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %216 = tensor.expand_shape %215 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %217 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %218 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%216 : tensor<2x1x128x65536xf32>) outs(%217 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %219 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%218 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %220 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%216, %219 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %221 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%220 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %222 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %223 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%221 : tensor<2x1x128x65536xf32>) outs(%222 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %224 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%223 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %225 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %226 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%216 : tensor<2x1x128x65536xf32>) outs(%225 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %227 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%226 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %228 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%224, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %229 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%228 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %230 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%216, %227 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %231 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%230, %229 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %232 = tensor.collapse_shape %231 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %233 = tensor.expand_shape %232 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %234 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_26 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %235 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %236 = linalg.matmul ins(%51, %234 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%235 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %237 = tensor.extract_slice %236[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %238 = tensor.extract_slice %236[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %239 = tensor.expand_shape %238 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %240 = tensor.expand_shape %237 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %241 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%240, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %242 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%233, %241 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %243 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%239, %242 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %244 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%243 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %245 = tensor.pad %244 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %246 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_28 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %247 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%245, %cst_27 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%246 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %248 = tensor.collapse_shape %247 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %249 = tensor.expand_shape %248 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %250 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %251 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%249 : tensor<2x1x128x65536xf32>) outs(%250 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %252 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%251 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %253 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%249, %252 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %254 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%253 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %255 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %256 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%254 : tensor<2x1x128x65536xf32>) outs(%255 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %257 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%256 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %258 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %259 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%249 : tensor<2x1x128x65536xf32>) outs(%258 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %260 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%259 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %261 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%257, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %262 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%261 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %263 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%249, %260 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %264 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%263, %262 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %265 = tensor.collapse_shape %264 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %266 = tensor.expand_shape %265 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %267 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_29 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %268 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %269 = linalg.matmul ins(%51, %267 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%268 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %270 = tensor.extract_slice %269[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %271 = tensor.extract_slice %269[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %272 = tensor.expand_shape %271 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %273 = tensor.expand_shape %270 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %274 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%273, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %275 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%266, %274 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %276 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%272, %275 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %277 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%276 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %278 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%277, %211 : tensor<2x128x256x256xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %279 = tensor.pad %278 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %280 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_31 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %281 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%279, %cst_30 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%280 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %282 = tensor.collapse_shape %281 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %283 = tensor.expand_shape %282 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %284 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %285 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%283 : tensor<2x1x128x65536xf32>) outs(%284 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %286 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%285 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %287 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%283, %286 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %288 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%287 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %289 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %290 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%288 : tensor<2x1x128x65536xf32>) outs(%289 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %291 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%290 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %292 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %293 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%283 : tensor<2x1x128x65536xf32>) outs(%292 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %294 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%293 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %295 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%291, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %296 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%295 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %297 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%283, %294 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %298 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%297, %296 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %299 = tensor.collapse_shape %298 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %300 = tensor.expand_shape %299 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %301 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_32 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %302 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %303 = linalg.matmul ins(%51, %301 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%302 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %304 = tensor.extract_slice %303[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %305 = tensor.extract_slice %303[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %306 = tensor.expand_shape %305 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %307 = tensor.expand_shape %304 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %308 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%307, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %309 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%300, %308 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %310 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%306, %309 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %311 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%310 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %312 = tensor.pad %311 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x256x256xf32> to tensor<2x128x258x258xf32>
    %313 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_34 : tensor<128xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x128x256x256xf32>
    %314 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%312, %cst_33 : tensor<2x128x258x258xf32>, tensor<128x128x3x3xf32>) outs(%313 : tensor<2x128x256x256xf32>) -> tensor<2x128x256x256xf32>
    %315 = tensor.collapse_shape %314 [[0], [1], [2, 3]] : tensor<2x128x256x256xf32> into tensor<2x128x65536xf32>
    %316 = tensor.expand_shape %315 [[0], [1, 2], [3]] : tensor<2x128x65536xf32> into tensor<2x1x128x65536xf32>
    %317 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %318 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%316 : tensor<2x1x128x65536xf32>) outs(%317 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %319 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%318 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %320 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%316, %319 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x128x65536xf32>
    %321 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%320 : tensor<2x1x128x65536xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %322 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %323 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%321 : tensor<2x1x128x65536xf32>) outs(%322 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %324 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%323 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %325 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %326 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%316 : tensor<2x1x128x65536xf32>) outs(%325 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %327 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%326 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_516 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %328 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%324, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %329 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%328 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %330 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%316, %327 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %331 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%330, %329 : tensor<2x1x128x65536xf32>, tensor<2x1x1x1xf32>) outs(%80 : tensor<2x1x128x65536xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x128x65536xf32>
    %332 = tensor.collapse_shape %331 [[0], [1, 2], [3]] : tensor<2x1x128x65536xf32> into tensor<2x128x65536xf32>
    %333 = tensor.expand_shape %332 [[0], [1], [2, 3]] : tensor<2x128x65536xf32> into tensor<2x128x256x256xf32>
    %334 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_35 : tensor<256x1024xf32>) outs(%95 : tensor<1024x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x256xf32>
    %335 = linalg.fill ins(%cst_514 : f32) outs(%97 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %336 = linalg.matmul ins(%51, %334 : tensor<2x1024xf32>, tensor<1024x256xf32>) outs(%335 : tensor<2x256xf32>) -> tensor<2x256xf32>
    %337 = tensor.extract_slice %336[0, 0] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %338 = tensor.extract_slice %336[0, 128] [2, 128] [1, 1] : tensor<2x256xf32> to tensor<2x128xf32>
    %339 = tensor.expand_shape %338 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %340 = tensor.expand_shape %337 [[0], [1, 2, 3]] : tensor<2x128xf32> into tensor<2x128x1x1xf32>
    %341 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%340, %cst_511 : tensor<2x128x1x1xf32>, tensor<i64>) outs(%104 : tensor<2x128x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x1x1xf32>
    %342 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%333, %341 : tensor<2x128x256x256xf32>, tensor<2x128x1x1xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %343 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%339, %342 : tensor<2x128x1x1xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %344 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%343 : tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x128x256x256xf32>
    %345 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%344, %278 : tensor<2x128x256x256xf32>, tensor<2x128x256x256xf32>) outs(%71 : tensor<2x128x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x256x256xf32>
    %346 = linalg.init_tensor [2, 128, 128, 128] : tensor<2x128x128x128xf32>
    %347 = linalg.fill ins(%cst_514 : f32) outs(%346 : tensor<2x128x128x128xf32>) -> tensor<2x128x128x128xf32>
    %348 = linalg.init_tensor [2, 2] : tensor<2x2xf32>
    %349 = linalg.pooling_nchw_sum {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%345, %348 : tensor<2x128x256x256xf32>, tensor<2x2xf32>) outs(%347 : tensor<2x128x128x128xf32>) -> tensor<2x128x128x128xf32>
    %350 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%349 : tensor<2x128x128x128xf32>) outs(%346 : tensor<2x128x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_522 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x128x128x128xf32>
    %351 = tensor.pad %350 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x128x128x128xf32> to tensor<2x128x130x130xf32>
    %352 = linalg.init_tensor [2, 256, 128, 128] : tensor<2x256x128x128xf32>
    %353 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_37 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %354 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%351, %cst_36 : tensor<2x128x130x130xf32>, tensor<256x128x3x3xf32>) outs(%353 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %355 = tensor.collapse_shape %354 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %356 = tensor.expand_shape %355 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %357 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %358 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%356 : tensor<2x1x256x16384xf32>) outs(%357 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %359 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%358 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %360 = linalg.init_tensor [2, 1, 256, 16384] : tensor<2x1x256x16384xf32>
    %361 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%356, %359 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %362 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%361 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %363 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %364 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%362 : tensor<2x1x256x16384xf32>) outs(%363 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %365 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%364 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %366 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %367 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%356 : tensor<2x1x256x16384xf32>) outs(%366 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %368 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%367 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %369 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%365, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %370 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%369 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %371 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%356, %368 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %372 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%371, %370 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %373 = tensor.collapse_shape %372 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %374 = tensor.expand_shape %373 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %375 = linalg.init_tensor [1024, 512] : tensor<1024x512xf32>
    %376 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_38 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %377 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %378 = linalg.matmul ins(%51, %376 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%377 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %379 = tensor.extract_slice %378[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %380 = tensor.extract_slice %378[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %381 = tensor.expand_shape %380 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %382 = tensor.expand_shape %379 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %383 = linalg.init_tensor [2, 256, 1, 1] : tensor<2x256x1x1xf32>
    %384 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%382, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %385 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%374, %384 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %386 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%381, %385 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %387 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%386 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %388 = tensor.pad %387 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %389 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_40 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %390 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%388, %cst_39 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%389 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %391 = tensor.collapse_shape %390 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %392 = tensor.expand_shape %391 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %393 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %394 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%392 : tensor<2x1x256x16384xf32>) outs(%393 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %395 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%394 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %396 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%392, %395 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %397 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%396 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %398 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %399 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%397 : tensor<2x1x256x16384xf32>) outs(%398 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %400 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%399 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %401 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %402 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%392 : tensor<2x1x256x16384xf32>) outs(%401 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %403 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%402 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %404 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%400, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %405 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%404 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %406 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%392, %403 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %407 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%406, %405 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %408 = tensor.collapse_shape %407 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %409 = tensor.expand_shape %408 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %410 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_41 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %411 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %412 = linalg.matmul ins(%51, %410 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%411 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %413 = tensor.extract_slice %412[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %414 = tensor.extract_slice %412[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %415 = tensor.expand_shape %414 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %416 = tensor.expand_shape %413 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %417 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%416, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %418 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%409, %417 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %419 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%415, %418 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %420 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%419 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %421 = linalg.fill ins(%cst_514 : f32) outs(%352 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %422 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%350, %cst_42 : tensor<2x128x128x128xf32>, tensor<256x128x1x1xf32>) outs(%421 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %423 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%420, %422 : tensor<2x256x128x128xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %424 = tensor.pad %423 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %425 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_44 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %426 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%424, %cst_43 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%425 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %427 = tensor.collapse_shape %426 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %428 = tensor.expand_shape %427 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %429 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %430 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%428 : tensor<2x1x256x16384xf32>) outs(%429 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %431 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%430 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %432 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%428, %431 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %433 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%432 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %434 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %435 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%433 : tensor<2x1x256x16384xf32>) outs(%434 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %436 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%435 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %437 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %438 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%428 : tensor<2x1x256x16384xf32>) outs(%437 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %439 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%438 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %440 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%436, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %441 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%440 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %442 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%428, %439 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %443 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%442, %441 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %444 = tensor.collapse_shape %443 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %445 = tensor.expand_shape %444 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %446 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_45 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %447 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %448 = linalg.matmul ins(%51, %446 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%447 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %449 = tensor.extract_slice %448[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %450 = tensor.extract_slice %448[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %451 = tensor.expand_shape %450 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %452 = tensor.expand_shape %449 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %453 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%452, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %454 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%445, %453 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %455 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%451, %454 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %456 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%455 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %457 = tensor.pad %456 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %458 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_47 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %459 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%457, %cst_46 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%458 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %460 = tensor.collapse_shape %459 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %461 = tensor.expand_shape %460 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %462 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %463 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%461 : tensor<2x1x256x16384xf32>) outs(%462 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %464 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%463 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %465 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%461, %464 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %466 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%465 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %467 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %468 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%466 : tensor<2x1x256x16384xf32>) outs(%467 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %469 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%468 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %470 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %471 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%461 : tensor<2x1x256x16384xf32>) outs(%470 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %472 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%471 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %473 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%469, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %474 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%473 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %475 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%461, %472 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %476 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%475, %474 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %477 = tensor.collapse_shape %476 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %478 = tensor.expand_shape %477 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %479 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_48 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %480 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %481 = linalg.matmul ins(%51, %479 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%480 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %482 = tensor.extract_slice %481[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %483 = tensor.extract_slice %481[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %484 = tensor.expand_shape %483 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %485 = tensor.expand_shape %482 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %486 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%485, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %487 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%478, %486 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %488 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%484, %487 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %489 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%488 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %490 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%489, %423 : tensor<2x256x128x128xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %491 = tensor.pad %490 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %492 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_50 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %493 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%491, %cst_49 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%492 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %494 = tensor.collapse_shape %493 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %495 = tensor.expand_shape %494 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %496 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %497 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%495 : tensor<2x1x256x16384xf32>) outs(%496 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %498 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%497 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %499 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%495, %498 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %500 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%499 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %501 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %502 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%500 : tensor<2x1x256x16384xf32>) outs(%501 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %503 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%502 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %504 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %505 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%495 : tensor<2x1x256x16384xf32>) outs(%504 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %506 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%505 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %507 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%503, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %508 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%507 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %509 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%495, %506 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %510 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%509, %508 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %511 = tensor.collapse_shape %510 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %512 = tensor.expand_shape %511 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %513 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_51 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %514 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %515 = linalg.matmul ins(%51, %513 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%514 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %516 = tensor.extract_slice %515[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %517 = tensor.extract_slice %515[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %518 = tensor.expand_shape %517 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %519 = tensor.expand_shape %516 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %520 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%519, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %521 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%512, %520 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %522 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%518, %521 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %523 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%522 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %524 = tensor.pad %523 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %525 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_53 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %526 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%524, %cst_52 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%525 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %527 = tensor.collapse_shape %526 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %528 = tensor.expand_shape %527 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %529 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %530 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%528 : tensor<2x1x256x16384xf32>) outs(%529 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %531 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%530 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %532 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%528, %531 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %533 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%532 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %534 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %535 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%533 : tensor<2x1x256x16384xf32>) outs(%534 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %536 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%535 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %537 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %538 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%528 : tensor<2x1x256x16384xf32>) outs(%537 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %539 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%538 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %540 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%536, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %541 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%540 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %542 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%528, %539 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %543 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%542, %541 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %544 = tensor.collapse_shape %543 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %545 = tensor.expand_shape %544 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %546 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_54 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %547 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %548 = linalg.matmul ins(%51, %546 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%547 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %549 = tensor.extract_slice %548[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %550 = tensor.extract_slice %548[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %551 = tensor.expand_shape %550 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %552 = tensor.expand_shape %549 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %553 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%552, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %554 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%545, %553 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %555 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%551, %554 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %556 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%555 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %557 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%556, %490 : tensor<2x256x128x128xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %558 = tensor.pad %557 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %559 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_56 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %560 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%558, %cst_55 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%559 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %561 = tensor.collapse_shape %560 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %562 = tensor.expand_shape %561 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %563 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %564 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%562 : tensor<2x1x256x16384xf32>) outs(%563 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %565 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%564 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %566 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%562, %565 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %567 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%566 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %568 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %569 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%567 : tensor<2x1x256x16384xf32>) outs(%568 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %570 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%569 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %571 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %572 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%562 : tensor<2x1x256x16384xf32>) outs(%571 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %573 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%572 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %574 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%570, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %575 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%574 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %576 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%562, %573 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %577 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%576, %575 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %578 = tensor.collapse_shape %577 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %579 = tensor.expand_shape %578 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %580 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_57 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %581 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %582 = linalg.matmul ins(%51, %580 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%581 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %583 = tensor.extract_slice %582[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %584 = tensor.extract_slice %582[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %585 = tensor.expand_shape %584 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %586 = tensor.expand_shape %583 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %587 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%586, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %588 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%579, %587 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %589 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%585, %588 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %590 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%589 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %591 = tensor.pad %590 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x128x128xf32> to tensor<2x256x130x130xf32>
    %592 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_59 : tensor<256xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x128x128xf32>
    %593 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%591, %cst_58 : tensor<2x256x130x130xf32>, tensor<256x256x3x3xf32>) outs(%592 : tensor<2x256x128x128xf32>) -> tensor<2x256x128x128xf32>
    %594 = tensor.collapse_shape %593 [[0], [1], [2, 3]] : tensor<2x256x128x128xf32> into tensor<2x256x16384xf32>
    %595 = tensor.expand_shape %594 [[0], [1, 2], [3]] : tensor<2x256x16384xf32> into tensor<2x1x256x16384xf32>
    %596 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %597 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%595 : tensor<2x1x256x16384xf32>) outs(%596 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %598 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%597 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %599 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%595, %598 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x16384xf32>
    %600 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%599 : tensor<2x1x256x16384xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %601 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %602 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%600 : tensor<2x1x256x16384xf32>) outs(%601 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %603 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%602 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %604 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %605 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%595 : tensor<2x1x256x16384xf32>) outs(%604 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %606 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%605 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_517 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %607 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%603, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %608 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%607 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %609 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%595, %606 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %610 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%609, %608 : tensor<2x1x256x16384xf32>, tensor<2x1x1x1xf32>) outs(%360 : tensor<2x1x256x16384xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x16384xf32>
    %611 = tensor.collapse_shape %610 [[0], [1, 2], [3]] : tensor<2x1x256x16384xf32> into tensor<2x256x16384xf32>
    %612 = tensor.expand_shape %611 [[0], [1], [2, 3]] : tensor<2x256x16384xf32> into tensor<2x256x128x128xf32>
    %613 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_60 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %614 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %615 = linalg.matmul ins(%51, %613 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%614 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %616 = tensor.extract_slice %615[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %617 = tensor.extract_slice %615[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %618 = tensor.expand_shape %617 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %619 = tensor.expand_shape %616 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %620 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%619, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %621 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%612, %620 : tensor<2x256x128x128xf32>, tensor<2x256x1x1xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %622 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%618, %621 : tensor<2x256x1x1xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %623 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%622 : tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x128x128xf32>
    %624 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%623, %557 : tensor<2x256x128x128xf32>, tensor<2x256x128x128xf32>) outs(%352 : tensor<2x256x128x128xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x128x128xf32>
    %625 = linalg.init_tensor [2, 256, 64, 64] : tensor<2x256x64x64xf32>
    %626 = linalg.fill ins(%cst_514 : f32) outs(%625 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %627 = linalg.pooling_nchw_sum {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%624, %348 : tensor<2x256x128x128xf32>, tensor<2x2xf32>) outs(%626 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %628 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%627 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_522 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %629 = tensor.pad %628 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %630 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_62 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %631 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%629, %cst_61 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%630 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %632 = tensor.collapse_shape %631 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %633 = tensor.expand_shape %632 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %634 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %635 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%633 : tensor<2x1x256x4096xf32>) outs(%634 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %636 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%635 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %637 = linalg.init_tensor [2, 1, 256, 4096] : tensor<2x1x256x4096xf32>
    %638 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%633, %636 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %639 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%638 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %640 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %641 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%639 : tensor<2x1x256x4096xf32>) outs(%640 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %642 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%641 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %643 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %644 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%633 : tensor<2x1x256x4096xf32>) outs(%643 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %645 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%644 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %646 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%642, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %647 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%646 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %648 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%633, %645 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %649 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%648, %647 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %650 = tensor.collapse_shape %649 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %651 = tensor.expand_shape %650 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %652 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_63 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %653 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %654 = linalg.matmul ins(%51, %652 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%653 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %655 = tensor.extract_slice %654[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %656 = tensor.extract_slice %654[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %657 = tensor.expand_shape %656 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %658 = tensor.expand_shape %655 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %659 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%658, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %660 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%651, %659 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %661 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%657, %660 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %662 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%661 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %663 = tensor.pad %662 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %664 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_65 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %665 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%663, %cst_64 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%664 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %666 = tensor.collapse_shape %665 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %667 = tensor.expand_shape %666 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %668 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %669 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%667 : tensor<2x1x256x4096xf32>) outs(%668 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %670 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%669 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %671 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%667, %670 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %672 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%671 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %673 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %674 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%672 : tensor<2x1x256x4096xf32>) outs(%673 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %675 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%674 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %676 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %677 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%667 : tensor<2x1x256x4096xf32>) outs(%676 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %678 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%677 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %679 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%675, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %680 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%679 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %681 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%667, %678 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %682 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%681, %680 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %683 = tensor.collapse_shape %682 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %684 = tensor.expand_shape %683 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %685 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_66 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %686 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %687 = linalg.matmul ins(%51, %685 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%686 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %688 = tensor.extract_slice %687[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %689 = tensor.extract_slice %687[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %690 = tensor.expand_shape %689 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %691 = tensor.expand_shape %688 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %692 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%691, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %693 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%684, %692 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %694 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%690, %693 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %695 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%694 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %696 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%695, %628 : tensor<2x256x64x64xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %697 = tensor.pad %696 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %698 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_68 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %699 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%697, %cst_67 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%698 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %700 = tensor.collapse_shape %699 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %701 = tensor.expand_shape %700 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %702 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %703 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%701 : tensor<2x1x256x4096xf32>) outs(%702 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %704 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%703 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %705 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%701, %704 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %706 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%705 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %707 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %708 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%706 : tensor<2x1x256x4096xf32>) outs(%707 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %709 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%708 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %710 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %711 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%701 : tensor<2x1x256x4096xf32>) outs(%710 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %712 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%711 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %713 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%709, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %714 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%713 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %715 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%701, %712 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %716 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%715, %714 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %717 = tensor.collapse_shape %716 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %718 = tensor.expand_shape %717 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %719 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_69 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %720 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %721 = linalg.matmul ins(%51, %719 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%720 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %722 = tensor.extract_slice %721[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %723 = tensor.extract_slice %721[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %724 = tensor.expand_shape %723 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %725 = tensor.expand_shape %722 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %726 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%725, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %727 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%718, %726 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %728 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%724, %727 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %729 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%728 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %730 = tensor.pad %729 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %731 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_71 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %732 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%730, %cst_70 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%731 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %733 = tensor.collapse_shape %732 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %734 = tensor.expand_shape %733 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %735 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %736 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%734 : tensor<2x1x256x4096xf32>) outs(%735 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %737 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%736 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %738 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%734, %737 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %739 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%738 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %740 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %741 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%739 : tensor<2x1x256x4096xf32>) outs(%740 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %742 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%741 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %743 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %744 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%734 : tensor<2x1x256x4096xf32>) outs(%743 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %745 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%744 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %746 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%742, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %747 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%746 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %748 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%734, %745 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %749 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%748, %747 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %750 = tensor.collapse_shape %749 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %751 = tensor.expand_shape %750 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %752 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_72 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %753 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %754 = linalg.matmul ins(%51, %752 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%753 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %755 = tensor.extract_slice %754[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %756 = tensor.extract_slice %754[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %757 = tensor.expand_shape %756 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %758 = tensor.expand_shape %755 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %759 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%758, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %760 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%751, %759 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %761 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%757, %760 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %762 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%761 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %763 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%762, %696 : tensor<2x256x64x64xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %764 = tensor.pad %763 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %765 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_74 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %766 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%764, %cst_73 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%765 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %767 = tensor.collapse_shape %766 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %768 = tensor.expand_shape %767 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %769 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %770 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%768 : tensor<2x1x256x4096xf32>) outs(%769 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %771 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%770 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %772 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%768, %771 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %773 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%772 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %774 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %775 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%773 : tensor<2x1x256x4096xf32>) outs(%774 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %776 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%775 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %777 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %778 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%768 : tensor<2x1x256x4096xf32>) outs(%777 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %779 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%778 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %780 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%776, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %781 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%780 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %782 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%768, %779 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %783 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%782, %781 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %784 = tensor.collapse_shape %783 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %785 = tensor.expand_shape %784 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %786 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_75 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %787 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %788 = linalg.matmul ins(%51, %786 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%787 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %789 = tensor.extract_slice %788[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %790 = tensor.extract_slice %788[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %791 = tensor.expand_shape %790 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %792 = tensor.expand_shape %789 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %793 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%792, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %794 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%785, %793 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %795 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%791, %794 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %796 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%795 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %797 = tensor.pad %796 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %798 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_77 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %799 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%797, %cst_76 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%798 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %800 = tensor.collapse_shape %799 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %801 = tensor.expand_shape %800 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %802 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %803 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%801 : tensor<2x1x256x4096xf32>) outs(%802 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %804 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%803 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %805 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%801, %804 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %806 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%805 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %807 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %808 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%806 : tensor<2x1x256x4096xf32>) outs(%807 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %809 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%808 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %810 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %811 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%801 : tensor<2x1x256x4096xf32>) outs(%810 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %812 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%811 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %813 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%809, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %814 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%813 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %815 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%801, %812 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %816 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%815, %814 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %817 = tensor.collapse_shape %816 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %818 = tensor.expand_shape %817 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %819 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_78 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %820 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %821 = linalg.matmul ins(%51, %819 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%820 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %822 = tensor.extract_slice %821[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %823 = tensor.extract_slice %821[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %824 = tensor.expand_shape %823 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %825 = tensor.expand_shape %822 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %826 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%825, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %827 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%818, %826 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %828 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%824, %827 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %829 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%828 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %830 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%829, %763 : tensor<2x256x64x64xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %831 = tensor.pad %830 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %832 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_80 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %833 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%831, %cst_79 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%832 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %834 = tensor.collapse_shape %833 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %835 = tensor.expand_shape %834 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %836 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %837 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%835 : tensor<2x1x256x4096xf32>) outs(%836 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %838 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%837 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %839 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%835, %838 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %840 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%839 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %841 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %842 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%840 : tensor<2x1x256x4096xf32>) outs(%841 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %843 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%842 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %844 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %845 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%835 : tensor<2x1x256x4096xf32>) outs(%844 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %846 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%845 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %847 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%843, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %848 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%847 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %849 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%835, %846 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %850 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%849, %848 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %851 = tensor.collapse_shape %850 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %852 = tensor.expand_shape %851 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %853 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_81 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %854 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %855 = linalg.matmul ins(%51, %853 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%854 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %856 = tensor.extract_slice %855[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %857 = tensor.extract_slice %855[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %858 = tensor.expand_shape %857 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %859 = tensor.expand_shape %856 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %860 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%859, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %861 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%852, %860 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %862 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%858, %861 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %863 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%862 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %864 = tensor.pad %863 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x64x64xf32> to tensor<2x256x66x66xf32>
    %865 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_83 : tensor<256xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x256x64x64xf32>
    %866 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%864, %cst_82 : tensor<2x256x66x66xf32>, tensor<256x256x3x3xf32>) outs(%865 : tensor<2x256x64x64xf32>) -> tensor<2x256x64x64xf32>
    %867 = tensor.collapse_shape %866 [[0], [1], [2, 3]] : tensor<2x256x64x64xf32> into tensor<2x256x4096xf32>
    %868 = tensor.expand_shape %867 [[0], [1, 2], [3]] : tensor<2x256x4096xf32> into tensor<2x1x256x4096xf32>
    %869 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %870 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%868 : tensor<2x1x256x4096xf32>) outs(%869 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %871 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%870 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %872 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%868, %871 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x256x4096xf32>
    %873 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%872 : tensor<2x1x256x4096xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %874 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %875 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%873 : tensor<2x1x256x4096xf32>) outs(%874 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %876 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%875 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %877 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %878 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%868 : tensor<2x1x256x4096xf32>) outs(%877 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %879 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%878 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_518 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %880 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%876, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %881 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%880 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %882 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%868, %879 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %883 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%882, %881 : tensor<2x1x256x4096xf32>, tensor<2x1x1x1xf32>) outs(%637 : tensor<2x1x256x4096xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x256x4096xf32>
    %884 = tensor.collapse_shape %883 [[0], [1, 2], [3]] : tensor<2x1x256x4096xf32> into tensor<2x256x4096xf32>
    %885 = tensor.expand_shape %884 [[0], [1], [2, 3]] : tensor<2x256x4096xf32> into tensor<2x256x64x64xf32>
    %886 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_84 : tensor<512x1024xf32>) outs(%375 : tensor<1024x512xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x512xf32>
    %887 = linalg.fill ins(%cst_514 : f32) outs(%0 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %888 = linalg.matmul ins(%51, %886 : tensor<2x1024xf32>, tensor<1024x512xf32>) outs(%887 : tensor<2x512xf32>) -> tensor<2x512xf32>
    %889 = tensor.extract_slice %888[0, 0] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %890 = tensor.extract_slice %888[0, 256] [2, 256] [1, 1] : tensor<2x512xf32> to tensor<2x256xf32>
    %891 = tensor.expand_shape %890 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %892 = tensor.expand_shape %889 [[0], [1, 2, 3]] : tensor<2x256xf32> into tensor<2x256x1x1xf32>
    %893 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%892, %cst_511 : tensor<2x256x1x1xf32>, tensor<i64>) outs(%383 : tensor<2x256x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x1x1xf32>
    %894 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%885, %893 : tensor<2x256x64x64xf32>, tensor<2x256x1x1xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %895 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%891, %894 : tensor<2x256x1x1xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %896 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%895 : tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x256x64x64xf32>
    %897 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%896, %830 : tensor<2x256x64x64xf32>, tensor<2x256x64x64xf32>) outs(%625 : tensor<2x256x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x64x64xf32>
    %898 = linalg.init_tensor [2, 256, 32, 32] : tensor<2x256x32x32xf32>
    %899 = linalg.fill ins(%cst_514 : f32) outs(%898 : tensor<2x256x32x32xf32>) -> tensor<2x256x32x32xf32>
    %900 = linalg.pooling_nchw_sum {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%897, %348 : tensor<2x256x64x64xf32>, tensor<2x2xf32>) outs(%899 : tensor<2x256x32x32xf32>) -> tensor<2x256x32x32xf32>
    %901 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%900 : tensor<2x256x32x32xf32>) outs(%898 : tensor<2x256x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_522 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x256x32x32xf32>
    %902 = tensor.pad %901 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x256x32x32xf32> to tensor<2x256x34x34xf32>
    %903 = linalg.init_tensor [2, 512, 32, 32] : tensor<2x512x32x32xf32>
    %904 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_86 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %905 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%902, %cst_85 : tensor<2x256x34x34xf32>, tensor<512x256x3x3xf32>) outs(%904 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %906 = tensor.collapse_shape %905 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %907 = tensor.expand_shape %906 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %908 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %909 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%907 : tensor<2x1x512x1024xf32>) outs(%908 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %910 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%909 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %911 = linalg.init_tensor [2, 1, 512, 1024] : tensor<2x1x512x1024xf32>
    %912 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%907, %910 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %913 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%912 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %914 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %915 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%913 : tensor<2x1x512x1024xf32>) outs(%914 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %916 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%915 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %917 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %918 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%907 : tensor<2x1x512x1024xf32>) outs(%917 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %919 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%918 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %920 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%916, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %921 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%920 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %922 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%907, %919 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %923 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%922, %921 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %924 = tensor.collapse_shape %923 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %925 = tensor.expand_shape %924 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %926 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_87 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %927 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %928 = linalg.matmul ins(%51, %926 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%927 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %929 = tensor.extract_slice %928[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %930 = tensor.extract_slice %928[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %931 = tensor.expand_shape %930 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %932 = tensor.expand_shape %929 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %933 = linalg.init_tensor [2, 512, 1, 1] : tensor<2x512x1x1xf32>
    %934 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%932, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %935 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%925, %934 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %936 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%931, %935 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %937 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%936 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %938 = tensor.pad %937 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %939 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_89 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %940 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%938, %cst_88 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%939 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %941 = tensor.collapse_shape %940 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %942 = tensor.expand_shape %941 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %943 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %944 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%942 : tensor<2x1x512x1024xf32>) outs(%943 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %945 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%944 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %946 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%942, %945 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %947 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%946 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %948 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %949 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%947 : tensor<2x1x512x1024xf32>) outs(%948 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %950 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%949 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %951 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %952 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%942 : tensor<2x1x512x1024xf32>) outs(%951 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %953 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%952 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %954 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%950, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %955 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%954 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %956 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%942, %953 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %957 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%956, %955 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %958 = tensor.collapse_shape %957 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %959 = tensor.expand_shape %958 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %960 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_90 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %961 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %962 = linalg.matmul ins(%51, %960 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%961 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %963 = tensor.extract_slice %962[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %964 = tensor.extract_slice %962[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %965 = tensor.expand_shape %964 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %966 = tensor.expand_shape %963 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %967 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%966, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %968 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%959, %967 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %969 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%965, %968 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %970 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%969 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %971 = linalg.fill ins(%cst_514 : f32) outs(%903 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %972 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%901, %cst_91 : tensor<2x256x32x32xf32>, tensor<512x256x1x1xf32>) outs(%971 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %973 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%970, %972 : tensor<2x512x32x32xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %974 = tensor.pad %973 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %975 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_93 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %976 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%974, %cst_92 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%975 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %977 = tensor.collapse_shape %976 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %978 = tensor.expand_shape %977 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %979 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %980 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%978 : tensor<2x1x512x1024xf32>) outs(%979 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %981 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%980 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %982 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%978, %981 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %983 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%982 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %984 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %985 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%983 : tensor<2x1x512x1024xf32>) outs(%984 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %986 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%985 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %987 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %988 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%978 : tensor<2x1x512x1024xf32>) outs(%987 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %989 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%988 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %990 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%986, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %991 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%990 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %992 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%978, %989 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %993 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%992, %991 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %994 = tensor.collapse_shape %993 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %995 = tensor.expand_shape %994 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %996 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_94 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %997 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %998 = linalg.matmul ins(%51, %996 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%997 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %999 = tensor.extract_slice %998[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1000 = tensor.extract_slice %998[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1001 = tensor.expand_shape %1000 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1002 = tensor.expand_shape %999 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1003 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1002, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1004 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%995, %1003 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1005 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1001, %1004 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1006 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1005 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %1007 = tensor.pad %1006 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %1008 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_96 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %1009 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1007, %cst_95 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%1008 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %1010 = tensor.collapse_shape %1009 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %1011 = tensor.expand_shape %1010 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %1012 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1013 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1011 : tensor<2x1x512x1024xf32>) outs(%1012 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1014 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1013 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1015 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1011, %1014 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %1016 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1015 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1017 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1018 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1016 : tensor<2x1x512x1024xf32>) outs(%1017 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1019 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1018 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1020 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1021 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1011 : tensor<2x1x512x1024xf32>) outs(%1020 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1022 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1021 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1023 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1019, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1024 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1023 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1025 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1011, %1022 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1026 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1025, %1024 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1027 = tensor.collapse_shape %1026 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %1028 = tensor.expand_shape %1027 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %1029 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_97 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1030 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1031 = linalg.matmul ins(%51, %1029 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1030 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1032 = tensor.extract_slice %1031[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1033 = tensor.extract_slice %1031[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1034 = tensor.expand_shape %1033 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1035 = tensor.expand_shape %1032 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1036 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1035, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1037 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1028, %1036 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1038 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1034, %1037 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1039 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1038 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %1040 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1039, %973 : tensor<2x512x32x32xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1041 = tensor.pad %1040 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %1042 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_99 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %1043 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1041, %cst_98 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%1042 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %1044 = tensor.collapse_shape %1043 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %1045 = tensor.expand_shape %1044 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %1046 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1047 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1045 : tensor<2x1x512x1024xf32>) outs(%1046 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1048 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1047 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1049 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1045, %1048 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %1050 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1049 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1051 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1052 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1050 : tensor<2x1x512x1024xf32>) outs(%1051 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1053 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1052 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1054 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1055 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1045 : tensor<2x1x512x1024xf32>) outs(%1054 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1056 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1055 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1057 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1053, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1058 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1057 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1059 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1045, %1056 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1060 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1059, %1058 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1061 = tensor.collapse_shape %1060 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %1062 = tensor.expand_shape %1061 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %1063 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_100 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1064 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1065 = linalg.matmul ins(%51, %1063 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1064 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1066 = tensor.extract_slice %1065[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1067 = tensor.extract_slice %1065[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1068 = tensor.expand_shape %1067 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1069 = tensor.expand_shape %1066 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1070 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1069, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1071 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1062, %1070 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1072 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1068, %1071 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1073 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1072 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %1074 = tensor.pad %1073 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %1075 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_102 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %1076 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1074, %cst_101 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%1075 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %1077 = tensor.collapse_shape %1076 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %1078 = tensor.expand_shape %1077 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %1079 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1080 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1078 : tensor<2x1x512x1024xf32>) outs(%1079 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1081 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1080 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1082 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1078, %1081 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %1083 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1082 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1084 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1085 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1083 : tensor<2x1x512x1024xf32>) outs(%1084 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1086 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1085 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1087 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1088 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1078 : tensor<2x1x512x1024xf32>) outs(%1087 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1089 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1088 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1090 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1086, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1091 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1090 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1092 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1078, %1089 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1093 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1092, %1091 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1094 = tensor.collapse_shape %1093 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %1095 = tensor.expand_shape %1094 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %1096 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_103 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1097 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1098 = linalg.matmul ins(%51, %1096 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1097 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1099 = tensor.extract_slice %1098[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1100 = tensor.extract_slice %1098[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1101 = tensor.expand_shape %1100 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1102 = tensor.expand_shape %1099 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1103 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1102, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1104 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1095, %1103 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1105 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1101, %1104 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1106 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1105 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %1107 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1106, %1040 : tensor<2x512x32x32xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1108 = tensor.pad %1107 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %1109 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_105 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %1110 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1108, %cst_104 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%1109 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %1111 = tensor.collapse_shape %1110 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %1112 = tensor.expand_shape %1111 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %1113 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1114 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1112 : tensor<2x1x512x1024xf32>) outs(%1113 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1115 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1114 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1116 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1112, %1115 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %1117 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1116 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1118 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1119 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1117 : tensor<2x1x512x1024xf32>) outs(%1118 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1120 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1119 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1121 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1122 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1112 : tensor<2x1x512x1024xf32>) outs(%1121 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1123 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1122 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1124 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1120, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1125 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1124 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1126 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1112, %1123 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1127 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1126, %1125 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1128 = tensor.collapse_shape %1127 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %1129 = tensor.expand_shape %1128 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %1130 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_106 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1131 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1132 = linalg.matmul ins(%51, %1130 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1131 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1133 = tensor.extract_slice %1132[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1134 = tensor.extract_slice %1132[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1135 = tensor.expand_shape %1134 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1136 = tensor.expand_shape %1133 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1137 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1136, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1138 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1129, %1137 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1139 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1135, %1138 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1140 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1139 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %1141 = tensor.pad %1140 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x32x32xf32> to tensor<2x512x34x34xf32>
    %1142 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_108 : tensor<512xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x32x32xf32>
    %1143 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1141, %cst_107 : tensor<2x512x34x34xf32>, tensor<512x512x3x3xf32>) outs(%1142 : tensor<2x512x32x32xf32>) -> tensor<2x512x32x32xf32>
    %1144 = tensor.collapse_shape %1143 [[0], [1], [2, 3]] : tensor<2x512x32x32xf32> into tensor<2x512x1024xf32>
    %1145 = tensor.expand_shape %1144 [[0], [1, 2], [3]] : tensor<2x512x1024xf32> into tensor<2x1x512x1024xf32>
    %1146 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1147 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1145 : tensor<2x1x512x1024xf32>) outs(%1146 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1148 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1147 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1149 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1145, %1148 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x1024xf32>
    %1150 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1149 : tensor<2x1x512x1024xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1151 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1152 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1150 : tensor<2x1x512x1024xf32>) outs(%1151 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1153 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1152 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1154 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1155 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1145 : tensor<2x1x512x1024xf32>) outs(%1154 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1156 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1155 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_519 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1157 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1153, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1158 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1157 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1159 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1145, %1156 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1160 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1159, %1158 : tensor<2x1x512x1024xf32>, tensor<2x1x1x1xf32>) outs(%911 : tensor<2x1x512x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x1024xf32>
    %1161 = tensor.collapse_shape %1160 [[0], [1, 2], [3]] : tensor<2x1x512x1024xf32> into tensor<2x512x1024xf32>
    %1162 = tensor.expand_shape %1161 [[0], [1], [2, 3]] : tensor<2x512x1024xf32> into tensor<2x512x32x32xf32>
    %1163 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_109 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1164 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1165 = linalg.matmul ins(%51, %1163 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1164 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1166 = tensor.extract_slice %1165[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1167 = tensor.extract_slice %1165[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1168 = tensor.expand_shape %1167 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1169 = tensor.expand_shape %1166 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1170 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1169, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1171 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1162, %1170 : tensor<2x512x32x32xf32>, tensor<2x512x1x1xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1172 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1168, %1171 : tensor<2x512x1x1xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1173 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1172 : tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x32x32xf32>
    %1174 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1173, %1107 : tensor<2x512x32x32xf32>, tensor<2x512x32x32xf32>) outs(%903 : tensor<2x512x32x32xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x32x32xf32>
    %1175 = linalg.init_tensor [2, 512, 16, 16] : tensor<2x512x16x16xf32>
    %1176 = linalg.fill ins(%cst_514 : f32) outs(%1175 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1177 = linalg.pooling_nchw_sum {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%1174, %348 : tensor<2x512x32x32xf32>, tensor<2x2xf32>) outs(%1176 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1178 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1177 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_522 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1179 = tensor.pad %1178 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1180 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_111 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1181 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1179, %cst_110 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1180 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1182 = tensor.collapse_shape %1181 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1183 = tensor.expand_shape %1182 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1184 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1185 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1183 : tensor<2x1x512x256xf32>) outs(%1184 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1186 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1185 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1187 = linalg.init_tensor [2, 1, 512, 256] : tensor<2x1x512x256xf32>
    %1188 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1183, %1186 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1189 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1188 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1190 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1191 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1189 : tensor<2x1x512x256xf32>) outs(%1190 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1192 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1191 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1193 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1194 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1183 : tensor<2x1x512x256xf32>) outs(%1193 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1195 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1194 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1196 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1192, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1197 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1196 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1198 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1183, %1195 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1199 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1198, %1197 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1200 = tensor.collapse_shape %1199 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1201 = tensor.expand_shape %1200 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1202 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_112 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1203 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1204 = linalg.matmul ins(%51, %1202 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1203 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1205 = tensor.extract_slice %1204[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1206 = tensor.extract_slice %1204[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1207 = tensor.expand_shape %1206 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1208 = tensor.expand_shape %1205 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1209 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1208, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1210 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1201, %1209 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1211 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1207, %1210 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1212 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1211 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1213 = tensor.pad %1212 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1214 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_114 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1215 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1213, %cst_113 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1214 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1216 = tensor.collapse_shape %1215 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1217 = tensor.expand_shape %1216 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1218 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1219 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1217 : tensor<2x1x512x256xf32>) outs(%1218 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1220 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1219 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1221 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1217, %1220 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1222 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1221 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1223 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1224 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1222 : tensor<2x1x512x256xf32>) outs(%1223 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1225 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1224 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1226 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1227 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1217 : tensor<2x1x512x256xf32>) outs(%1226 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1228 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1227 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1229 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1225, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1230 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1229 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1231 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1217, %1228 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1232 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1231, %1230 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1233 = tensor.collapse_shape %1232 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1234 = tensor.expand_shape %1233 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1235 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_115 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1236 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1237 = linalg.matmul ins(%51, %1235 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1236 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1238 = tensor.extract_slice %1237[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1239 = tensor.extract_slice %1237[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1240 = tensor.expand_shape %1239 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1241 = tensor.expand_shape %1238 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1242 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1241, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1243 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1234, %1242 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1244 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1240, %1243 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1245 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1244 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1246 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1245, %1178 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1247 = tensor.collapse_shape %1246 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1248 = tensor.expand_shape %1247 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1249 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1250 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1248 : tensor<2x1x512x256xf32>) outs(%1249 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1251 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1250 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1252 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1248, %1251 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1253 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1252 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1254 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1255 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1253 : tensor<2x1x512x256xf32>) outs(%1254 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1256 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1255 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1257 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1258 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1248 : tensor<2x1x512x256xf32>) outs(%1257 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1259 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1258 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1260 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1256, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1261 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1260 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1262 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1248, %1259 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1263 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1262, %1261 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1264 = tensor.collapse_shape %1263 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1265 = tensor.expand_shape %1264 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1266 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1265, %cst_368 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1267 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1266, %cst_369 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1268 = linalg.init_tensor [2, 1536, 16, 16] : tensor<2x1536x16x16xf32>
    %1269 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_117 : tensor<1536xf32>) outs(%1268 : tensor<2x1536x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1536x16x16xf32>
    %1270 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1267, %cst_116 : tensor<2x512x16x16xf32>, tensor<1536x512x1x1xf32>) outs(%1269 : tensor<2x1536x16x16xf32>) -> tensor<2x1536x16x16xf32>
    %1271 = tensor.collapse_shape %1270 [[0], [1], [2, 3]] : tensor<2x1536x16x16xf32> into tensor<2x1536x256xf32>
    %1272 = tensor.expand_shape %1271 [[0], [1, 2], [3]] : tensor<2x1536x256xf32> into tensor<2x24x64x256xf32>
    %1273 = linalg.init_tensor [2, 24, 256, 64] : tensor<2x24x256x64xf32>
    %1274 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1272 : tensor<2x24x64x256xf32>) outs(%1273 : tensor<2x24x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x24x256x64xf32>
    %1275 = tensor.extract_slice %1274[0, 0, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1276 = tensor.extract_slice %1274[0, 8, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1277 = tensor.extract_slice %1274[0, 16, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1278 = linalg.init_tensor [2, 8, 256, 64] : tensor<2x8x256x64xf32>
    %1279 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1275, %cst_512 : tensor<2x8x256x64xf32>, tensor<f64>) outs(%1278 : tensor<2x8x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x256x64xf32>
    %1280 = linalg.init_tensor [2, 8, 64, 256] : tensor<2x8x64x256xf32>
    %1281 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1276 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1282 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1281, %cst_512 : tensor<2x8x64x256xf32>, tensor<f64>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x64x256xf32>
    %1283 = tensor.collapse_shape %1279 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1284 = tensor.collapse_shape %1282 [[0, 1], [2], [3]] : tensor<2x8x64x256xf32> into tensor<16x64x256xf32>
    %1285 = linalg.init_tensor [16, 256, 256] : tensor<16x256x256xf32>
    %1286 = linalg.fill ins(%cst_514 : f32) outs(%1285 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1287 = linalg.batch_matmul ins(%1283, %1284 : tensor<16x256x64xf32>, tensor<16x64x256xf32>) outs(%1286 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1288 = tensor.expand_shape %1287 [[0, 1], [2], [3]] : tensor<16x256x256xf32> into tensor<2x8x256x256xf32>
    %1289 = linalg.init_tensor [2, 8, 256, 1] : tensor<2x8x256x1xi64>
    %1290 = linalg.fill ins(%c0_i64 : i64) outs(%1289 : tensor<2x8x256x1xi64>) -> tensor<2x8x256x1xi64>
    %1291 = linalg.init_tensor [2, 8, 256, 1] : tensor<2x8x256x1xf32>
    %1292 = linalg.fill ins(%cst_515 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1293:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1288 : tensor<2x8x256x256xf32>) outs(%1292, %1290 : tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>)
    %1294 = linalg.init_tensor [2, 8, 256, 256] : tensor<2x8x256x256xf32>
    %1295 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1288, %1293#0 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x8x256x256xf32>
    %1296 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1295 : tensor<2x8x256x256xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1297 = linalg.fill ins(%cst_514 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1298 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1296 : tensor<2x8x256x256xf32>) outs(%1297 : tensor<2x8x256x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x1xf32>
    %1299 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1296, %1298 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1300 = tensor.collapse_shape %1299 [[0, 1], [2], [3]] : tensor<2x8x256x256xf32> into tensor<16x256x256xf32>
    %1301 = tensor.collapse_shape %1277 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1302 = linalg.init_tensor [16, 256, 64] : tensor<16x256x64xf32>
    %1303 = linalg.fill ins(%cst_514 : f32) outs(%1302 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1304 = linalg.batch_matmul ins(%1300, %1301 : tensor<16x256x256xf32>, tensor<16x256x64xf32>) outs(%1303 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1305 = tensor.expand_shape %1304 [[0, 1], [2], [3]] : tensor<16x256x64xf32> into tensor<2x8x256x64xf32>
    %1306 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1305 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1307 = tensor.collapse_shape %1306 [[0], [1, 2], [3]] : tensor<2x8x64x256xf32> into tensor<2x512x256xf32>
    %1308 = tensor.expand_shape %1307 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1309 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_119 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1310 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1308, %cst_118 : tensor<2x512x16x16xf32>, tensor<512x512x1x1xf32>) outs(%1309 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1311 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1246, %1310 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1312 = tensor.pad %1311 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1313 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_121 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1314 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1312, %cst_120 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1313 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1315 = tensor.collapse_shape %1314 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1316 = tensor.expand_shape %1315 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1317 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1318 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1316 : tensor<2x1x512x256xf32>) outs(%1317 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1319 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1318 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1320 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1316, %1319 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1321 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1320 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1322 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1323 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1321 : tensor<2x1x512x256xf32>) outs(%1322 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1324 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1323 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1325 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1326 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1316 : tensor<2x1x512x256xf32>) outs(%1325 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1327 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1326 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1328 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1324, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1329 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1328 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1330 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1316, %1327 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1331 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1330, %1329 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1332 = tensor.collapse_shape %1331 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1333 = tensor.expand_shape %1332 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1334 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_122 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1335 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1336 = linalg.matmul ins(%51, %1334 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1335 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1337 = tensor.extract_slice %1336[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1338 = tensor.extract_slice %1336[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1339 = tensor.expand_shape %1338 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1340 = tensor.expand_shape %1337 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1341 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1340, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1342 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1333, %1341 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1343 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1339, %1342 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1344 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1343 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1345 = tensor.pad %1344 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1346 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_124 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1347 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1345, %cst_123 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1346 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1348 = tensor.collapse_shape %1347 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1349 = tensor.expand_shape %1348 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1350 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1351 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1349 : tensor<2x1x512x256xf32>) outs(%1350 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1352 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1351 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1353 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1349, %1352 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1354 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1353 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1355 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1356 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1354 : tensor<2x1x512x256xf32>) outs(%1355 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1357 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1356 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1358 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1359 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1349 : tensor<2x1x512x256xf32>) outs(%1358 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1360 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1359 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1361 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1357, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1362 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1361 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1363 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1349, %1360 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1364 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1363, %1362 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1365 = tensor.collapse_shape %1364 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1366 = tensor.expand_shape %1365 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1367 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_125 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1368 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1369 = linalg.matmul ins(%51, %1367 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1368 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1370 = tensor.extract_slice %1369[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1371 = tensor.extract_slice %1369[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1372 = tensor.expand_shape %1371 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1373 = tensor.expand_shape %1370 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1374 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1373, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1375 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1366, %1374 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1376 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1372, %1375 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1377 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1376 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1378 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1377, %1311 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1379 = tensor.collapse_shape %1378 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1380 = tensor.expand_shape %1379 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1381 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1382 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1380 : tensor<2x1x512x256xf32>) outs(%1381 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1383 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1382 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1384 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1380, %1383 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1385 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1384 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1386 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1387 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1385 : tensor<2x1x512x256xf32>) outs(%1386 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1388 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1387 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1389 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1390 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1380 : tensor<2x1x512x256xf32>) outs(%1389 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1391 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1390 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1392 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1388, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1393 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1392 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1394 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1380, %1391 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1395 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1394, %1393 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1396 = tensor.collapse_shape %1395 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1397 = tensor.expand_shape %1396 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1398 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1397, %cst_368 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1399 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1398, %cst_369 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1400 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_127 : tensor<1536xf32>) outs(%1268 : tensor<2x1536x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1536x16x16xf32>
    %1401 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1399, %cst_126 : tensor<2x512x16x16xf32>, tensor<1536x512x1x1xf32>) outs(%1400 : tensor<2x1536x16x16xf32>) -> tensor<2x1536x16x16xf32>
    %1402 = tensor.collapse_shape %1401 [[0], [1], [2, 3]] : tensor<2x1536x16x16xf32> into tensor<2x1536x256xf32>
    %1403 = tensor.expand_shape %1402 [[0], [1, 2], [3]] : tensor<2x1536x256xf32> into tensor<2x24x64x256xf32>
    %1404 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1403 : tensor<2x24x64x256xf32>) outs(%1273 : tensor<2x24x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x24x256x64xf32>
    %1405 = tensor.extract_slice %1404[0, 0, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1406 = tensor.extract_slice %1404[0, 8, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1407 = tensor.extract_slice %1404[0, 16, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1408 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1405, %cst_512 : tensor<2x8x256x64xf32>, tensor<f64>) outs(%1278 : tensor<2x8x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x256x64xf32>
    %1409 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1406 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1410 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1409, %cst_512 : tensor<2x8x64x256xf32>, tensor<f64>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x64x256xf32>
    %1411 = tensor.collapse_shape %1408 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1412 = tensor.collapse_shape %1410 [[0, 1], [2], [3]] : tensor<2x8x64x256xf32> into tensor<16x64x256xf32>
    %1413 = linalg.fill ins(%cst_514 : f32) outs(%1285 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1414 = linalg.batch_matmul ins(%1411, %1412 : tensor<16x256x64xf32>, tensor<16x64x256xf32>) outs(%1413 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1415 = tensor.expand_shape %1414 [[0, 1], [2], [3]] : tensor<16x256x256xf32> into tensor<2x8x256x256xf32>
    %1416 = linalg.fill ins(%c0_i64 : i64) outs(%1289 : tensor<2x8x256x1xi64>) -> tensor<2x8x256x1xi64>
    %1417 = linalg.fill ins(%cst_515 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1418:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1415 : tensor<2x8x256x256xf32>) outs(%1417, %1416 : tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>)
    %1419 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1415, %1418#0 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x8x256x256xf32>
    %1420 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1419 : tensor<2x8x256x256xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1421 = linalg.fill ins(%cst_514 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1422 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1420 : tensor<2x8x256x256xf32>) outs(%1421 : tensor<2x8x256x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x1xf32>
    %1423 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1420, %1422 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1424 = tensor.collapse_shape %1423 [[0, 1], [2], [3]] : tensor<2x8x256x256xf32> into tensor<16x256x256xf32>
    %1425 = tensor.collapse_shape %1407 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1426 = linalg.fill ins(%cst_514 : f32) outs(%1302 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1427 = linalg.batch_matmul ins(%1424, %1425 : tensor<16x256x256xf32>, tensor<16x256x64xf32>) outs(%1426 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1428 = tensor.expand_shape %1427 [[0, 1], [2], [3]] : tensor<16x256x64xf32> into tensor<2x8x256x64xf32>
    %1429 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1428 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1430 = tensor.collapse_shape %1429 [[0], [1, 2], [3]] : tensor<2x8x64x256xf32> into tensor<2x512x256xf32>
    %1431 = tensor.expand_shape %1430 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1432 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_129 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1433 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1431, %cst_128 : tensor<2x512x16x16xf32>, tensor<512x512x1x1xf32>) outs(%1432 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1434 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1378, %1433 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1435 = tensor.pad %1434 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1436 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_131 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1437 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1435, %cst_130 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1436 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1438 = tensor.collapse_shape %1437 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1439 = tensor.expand_shape %1438 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1440 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1441 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1439 : tensor<2x1x512x256xf32>) outs(%1440 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1442 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1441 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1443 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1439, %1442 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1444 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1443 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1445 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1446 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1444 : tensor<2x1x512x256xf32>) outs(%1445 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1447 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1446 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1448 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1449 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1439 : tensor<2x1x512x256xf32>) outs(%1448 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1450 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1449 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1451 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1447, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1452 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1451 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1453 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1439, %1450 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1454 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1453, %1452 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1455 = tensor.collapse_shape %1454 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1456 = tensor.expand_shape %1455 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1457 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_132 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1458 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1459 = linalg.matmul ins(%51, %1457 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1458 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1460 = tensor.extract_slice %1459[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1461 = tensor.extract_slice %1459[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1462 = tensor.expand_shape %1461 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1463 = tensor.expand_shape %1460 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1464 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1463, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1465 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1456, %1464 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1466 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1462, %1465 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1467 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1466 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1468 = tensor.pad %1467 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1469 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_134 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1470 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1468, %cst_133 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1469 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1471 = tensor.collapse_shape %1470 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1472 = tensor.expand_shape %1471 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1473 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1474 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1472 : tensor<2x1x512x256xf32>) outs(%1473 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1475 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1474 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1476 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1472, %1475 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1477 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1476 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1478 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1479 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1477 : tensor<2x1x512x256xf32>) outs(%1478 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1480 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1479 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1481 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1482 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1472 : tensor<2x1x512x256xf32>) outs(%1481 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1483 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1482 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1484 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1480, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1485 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1484 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1486 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1472, %1483 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1487 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1486, %1485 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1488 = tensor.collapse_shape %1487 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1489 = tensor.expand_shape %1488 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1490 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_135 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1491 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1492 = linalg.matmul ins(%51, %1490 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1491 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1493 = tensor.extract_slice %1492[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1494 = tensor.extract_slice %1492[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1495 = tensor.expand_shape %1494 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1496 = tensor.expand_shape %1493 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1497 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1496, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1498 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1489, %1497 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1499 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1495, %1498 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1500 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1499 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1501 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1500, %1434 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1502 = tensor.collapse_shape %1501 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1503 = tensor.expand_shape %1502 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1504 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1505 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1503 : tensor<2x1x512x256xf32>) outs(%1504 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1506 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1505 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1507 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1503, %1506 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1508 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1507 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1509 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1510 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1508 : tensor<2x1x512x256xf32>) outs(%1509 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1511 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1510 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1512 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1513 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1503 : tensor<2x1x512x256xf32>) outs(%1512 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1514 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1513 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1515 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1511, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1516 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1515 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1517 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1503, %1514 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1518 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1517, %1516 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1519 = tensor.collapse_shape %1518 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1520 = tensor.expand_shape %1519 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1521 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1520, %cst_368 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1522 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1521, %cst_369 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1523 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_137 : tensor<1536xf32>) outs(%1268 : tensor<2x1536x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1536x16x16xf32>
    %1524 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1522, %cst_136 : tensor<2x512x16x16xf32>, tensor<1536x512x1x1xf32>) outs(%1523 : tensor<2x1536x16x16xf32>) -> tensor<2x1536x16x16xf32>
    %1525 = tensor.collapse_shape %1524 [[0], [1], [2, 3]] : tensor<2x1536x16x16xf32> into tensor<2x1536x256xf32>
    %1526 = tensor.expand_shape %1525 [[0], [1, 2], [3]] : tensor<2x1536x256xf32> into tensor<2x24x64x256xf32>
    %1527 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1526 : tensor<2x24x64x256xf32>) outs(%1273 : tensor<2x24x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x24x256x64xf32>
    %1528 = tensor.extract_slice %1527[0, 0, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1529 = tensor.extract_slice %1527[0, 8, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1530 = tensor.extract_slice %1527[0, 16, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1531 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1528, %cst_512 : tensor<2x8x256x64xf32>, tensor<f64>) outs(%1278 : tensor<2x8x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x256x64xf32>
    %1532 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1529 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1533 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1532, %cst_512 : tensor<2x8x64x256xf32>, tensor<f64>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x64x256xf32>
    %1534 = tensor.collapse_shape %1531 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1535 = tensor.collapse_shape %1533 [[0, 1], [2], [3]] : tensor<2x8x64x256xf32> into tensor<16x64x256xf32>
    %1536 = linalg.fill ins(%cst_514 : f32) outs(%1285 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1537 = linalg.batch_matmul ins(%1534, %1535 : tensor<16x256x64xf32>, tensor<16x64x256xf32>) outs(%1536 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1538 = tensor.expand_shape %1537 [[0, 1], [2], [3]] : tensor<16x256x256xf32> into tensor<2x8x256x256xf32>
    %1539 = linalg.fill ins(%c0_i64 : i64) outs(%1289 : tensor<2x8x256x1xi64>) -> tensor<2x8x256x1xi64>
    %1540 = linalg.fill ins(%cst_515 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1541:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1538 : tensor<2x8x256x256xf32>) outs(%1540, %1539 : tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>)
    %1542 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1538, %1541#0 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x8x256x256xf32>
    %1543 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1542 : tensor<2x8x256x256xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1544 = linalg.fill ins(%cst_514 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1545 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1543 : tensor<2x8x256x256xf32>) outs(%1544 : tensor<2x8x256x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x1xf32>
    %1546 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1543, %1545 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1547 = tensor.collapse_shape %1546 [[0, 1], [2], [3]] : tensor<2x8x256x256xf32> into tensor<16x256x256xf32>
    %1548 = tensor.collapse_shape %1530 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1549 = linalg.fill ins(%cst_514 : f32) outs(%1302 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1550 = linalg.batch_matmul ins(%1547, %1548 : tensor<16x256x256xf32>, tensor<16x256x64xf32>) outs(%1549 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1551 = tensor.expand_shape %1550 [[0, 1], [2], [3]] : tensor<16x256x64xf32> into tensor<2x8x256x64xf32>
    %1552 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1551 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1553 = tensor.collapse_shape %1552 [[0], [1, 2], [3]] : tensor<2x8x64x256xf32> into tensor<2x512x256xf32>
    %1554 = tensor.expand_shape %1553 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1555 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_139 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1556 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1554, %cst_138 : tensor<2x512x16x16xf32>, tensor<512x512x1x1xf32>) outs(%1555 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1557 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1501, %1556 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1558 = tensor.pad %1557 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1559 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_141 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1560 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1558, %cst_140 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1559 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1561 = tensor.collapse_shape %1560 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1562 = tensor.expand_shape %1561 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1563 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1564 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1562 : tensor<2x1x512x256xf32>) outs(%1563 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1565 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1564 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1566 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1562, %1565 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1567 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1566 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1568 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1569 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1567 : tensor<2x1x512x256xf32>) outs(%1568 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1570 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1569 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1571 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1572 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1562 : tensor<2x1x512x256xf32>) outs(%1571 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1573 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1572 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1574 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1570, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1575 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1574 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1576 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1562, %1573 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1577 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1576, %1575 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1578 = tensor.collapse_shape %1577 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1579 = tensor.expand_shape %1578 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1580 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_142 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1581 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1582 = linalg.matmul ins(%51, %1580 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1581 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1583 = tensor.extract_slice %1582[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1584 = tensor.extract_slice %1582[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1585 = tensor.expand_shape %1584 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1586 = tensor.expand_shape %1583 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1587 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1586, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1588 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1579, %1587 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1589 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1585, %1588 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1590 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1589 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1591 = tensor.pad %1590 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x16x16xf32> to tensor<2x512x18x18xf32>
    %1592 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_144 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1593 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1591, %cst_143 : tensor<2x512x18x18xf32>, tensor<512x512x3x3xf32>) outs(%1592 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1594 = tensor.collapse_shape %1593 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1595 = tensor.expand_shape %1594 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1596 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1597 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1595 : tensor<2x1x512x256xf32>) outs(%1596 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1598 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1597 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1599 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1595, %1598 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1600 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1599 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1601 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1602 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1600 : tensor<2x1x512x256xf32>) outs(%1601 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1603 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1602 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1604 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1605 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1595 : tensor<2x1x512x256xf32>) outs(%1604 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1606 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1605 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1607 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1603, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1608 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1607 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1609 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1595, %1606 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1610 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1609, %1608 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1611 = tensor.collapse_shape %1610 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1612 = tensor.expand_shape %1611 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1613 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_145 : tensor<1024x1024xf32>) outs(%32 : tensor<1024x1024xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x1024xf32>
    %1614 = linalg.fill ins(%cst_514 : f32) outs(%27 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1615 = linalg.matmul ins(%51, %1613 : tensor<2x1024xf32>, tensor<1024x1024xf32>) outs(%1614 : tensor<2x1024xf32>) -> tensor<2x1024xf32>
    %1616 = tensor.extract_slice %1615[0, 0] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1617 = tensor.extract_slice %1615[0, 512] [2, 512] [1, 1] : tensor<2x1024xf32> to tensor<2x512xf32>
    %1618 = tensor.expand_shape %1617 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1619 = tensor.expand_shape %1616 [[0], [1, 2, 3]] : tensor<2x512xf32> into tensor<2x512x1x1xf32>
    %1620 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1619, %cst_511 : tensor<2x512x1x1xf32>, tensor<i64>) outs(%933 : tensor<2x512x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x1x1xf32>
    %1621 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1612, %1620 : tensor<2x512x16x16xf32>, tensor<2x512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1622 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1618, %1621 : tensor<2x512x1x1xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1623 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1622 : tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x512x16x16xf32>
    %1624 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1623, %1557 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1625 = tensor.collapse_shape %1624 [[0], [1], [2, 3]] : tensor<2x512x16x16xf32> into tensor<2x512x256xf32>
    %1626 = tensor.expand_shape %1625 [[0], [1, 2], [3]] : tensor<2x512x256xf32> into tensor<2x1x512x256xf32>
    %1627 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1628 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1626 : tensor<2x1x512x256xf32>) outs(%1627 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1629 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1628 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1630 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1626, %1629 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x512x256xf32>
    %1631 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1630 : tensor<2x1x512x256xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1632 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1633 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1631 : tensor<2x1x512x256xf32>) outs(%1632 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1634 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1633 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1635 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1636 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1626 : tensor<2x1x512x256xf32>) outs(%1635 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1637 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1636 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_520 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1638 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1634, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1639 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1638 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1640 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1626, %1637 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1641 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1640, %1639 : tensor<2x1x512x256xf32>, tensor<2x1x1x1xf32>) outs(%1187 : tensor<2x1x512x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x512x256xf32>
    %1642 = tensor.collapse_shape %1641 [[0], [1, 2], [3]] : tensor<2x1x512x256xf32> into tensor<2x512x256xf32>
    %1643 = tensor.expand_shape %1642 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1644 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1643, %cst_368 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1645 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1644, %cst_369 : tensor<2x512x16x16xf32>, tensor<512x1x1xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1646 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_147 : tensor<1536xf32>) outs(%1268 : tensor<2x1536x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1536x16x16xf32>
    %1647 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1645, %cst_146 : tensor<2x512x16x16xf32>, tensor<1536x512x1x1xf32>) outs(%1646 : tensor<2x1536x16x16xf32>) -> tensor<2x1536x16x16xf32>
    %1648 = tensor.collapse_shape %1647 [[0], [1], [2, 3]] : tensor<2x1536x16x16xf32> into tensor<2x1536x256xf32>
    %1649 = tensor.expand_shape %1648 [[0], [1, 2], [3]] : tensor<2x1536x256xf32> into tensor<2x24x64x256xf32>
    %1650 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1649 : tensor<2x24x64x256xf32>) outs(%1273 : tensor<2x24x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x24x256x64xf32>
    %1651 = tensor.extract_slice %1650[0, 0, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1652 = tensor.extract_slice %1650[0, 8, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1653 = tensor.extract_slice %1650[0, 16, 0, 0] [2, 8, 256, 64] [1, 1, 1, 1] : tensor<2x24x256x64xf32> to tensor<2x8x256x64xf32>
    %1654 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1651, %cst_512 : tensor<2x8x256x64xf32>, tensor<f64>) outs(%1278 : tensor<2x8x256x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x256x64xf32>
    %1655 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1652 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1656 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1655, %cst_512 : tensor<2x8x64x256xf32>, tensor<f64>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x8x64x256xf32>
    %1657 = tensor.collapse_shape %1654 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1658 = tensor.collapse_shape %1656 [[0, 1], [2], [3]] : tensor<2x8x64x256xf32> into tensor<16x64x256xf32>
    %1659 = linalg.fill ins(%cst_514 : f32) outs(%1285 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1660 = linalg.batch_matmul ins(%1657, %1658 : tensor<16x256x64xf32>, tensor<16x64x256xf32>) outs(%1659 : tensor<16x256x256xf32>) -> tensor<16x256x256xf32>
    %1661 = tensor.expand_shape %1660 [[0, 1], [2], [3]] : tensor<16x256x256xf32> into tensor<2x8x256x256xf32>
    %1662 = linalg.fill ins(%c0_i64 : i64) outs(%1289 : tensor<2x8x256x1xi64>) -> tensor<2x8x256x1xi64>
    %1663 = linalg.fill ins(%cst_515 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1664:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1661 : tensor<2x8x256x256xf32>) outs(%1663, %1662 : tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x8x256x1xf32>, tensor<2x8x256x1xi64>)
    %1665 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1661, %1664#0 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x8x256x256xf32>
    %1666 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1665 : tensor<2x8x256x256xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1667 = linalg.fill ins(%cst_514 : f32) outs(%1291 : tensor<2x8x256x1xf32>) -> tensor<2x8x256x1xf32>
    %1668 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1666 : tensor<2x8x256x256xf32>) outs(%1667 : tensor<2x8x256x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x1xf32>
    %1669 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1666, %1668 : tensor<2x8x256x256xf32>, tensor<2x8x256x1xf32>) outs(%1294 : tensor<2x8x256x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x8x256x256xf32>
    %1670 = tensor.collapse_shape %1669 [[0, 1], [2], [3]] : tensor<2x8x256x256xf32> into tensor<16x256x256xf32>
    %1671 = tensor.collapse_shape %1653 [[0, 1], [2], [3]] : tensor<2x8x256x64xf32> into tensor<16x256x64xf32>
    %1672 = linalg.fill ins(%cst_514 : f32) outs(%1302 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1673 = linalg.batch_matmul ins(%1670, %1671 : tensor<16x256x256xf32>, tensor<16x256x64xf32>) outs(%1672 : tensor<16x256x64xf32>) -> tensor<16x256x64xf32>
    %1674 = tensor.expand_shape %1673 [[0, 1], [2], [3]] : tensor<16x256x64xf32> into tensor<2x8x256x64xf32>
    %1675 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1674 : tensor<2x8x256x64xf32>) outs(%1280 : tensor<2x8x64x256xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x8x64x256xf32>
    %1676 = tensor.collapse_shape %1675 [[0], [1, 2], [3]] : tensor<2x8x64x256xf32> into tensor<2x512x256xf32>
    %1677 = tensor.expand_shape %1676 [[0], [1], [2, 3]] : tensor<2x512x256xf32> into tensor<2x512x16x16xf32>
    %1678 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_149 : tensor<512xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x512x16x16xf32>
    %1679 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1677, %cst_148 : tensor<2x512x16x16xf32>, tensor<512x512x1x1xf32>) outs(%1678 : tensor<2x512x16x16xf32>) -> tensor<2x512x16x16xf32>
    %1680 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1624, %1679 : tensor<2x512x16x16xf32>, tensor<2x512x16x16xf32>) outs(%1175 : tensor<2x512x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x16x16xf32>
    %1681 = linalg.init_tensor [2, 512, 8, 8] : tensor<2x512x8x8xf32>
    %1682 = linalg.fill ins(%cst_514 : f32) outs(%1681 : tensor<2x512x8x8xf32>) -> tensor<2x512x8x8xf32>
    %1683 = linalg.pooling_nchw_sum {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%1680, %348 : tensor<2x512x16x16xf32>, tensor<2x2xf32>) outs(%1682 : tensor<2x512x8x8xf32>) -> tensor<2x512x8x8xf32>
    %1684 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1683 : tensor<2x512x8x8xf32>) outs(%1681 : tensor<2x512x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_522 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x512x8x8xf32>
    %1685 = tensor.pad %1684 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x512x8x8xf32> to tensor<2x512x10x10xf32>
    %1686 = linalg.init_tensor [2, 1024, 8, 8] : tensor<2x1024x8x8xf32>
    %1687 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_151 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1688 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1685, %cst_150 : tensor<2x512x10x10xf32>, tensor<1024x512x3x3xf32>) outs(%1687 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1689 = tensor.collapse_shape %1688 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1690 = tensor.expand_shape %1689 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1691 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1692 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1690 : tensor<2x1x1024x64xf32>) outs(%1691 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1693 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1692 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1694 = linalg.init_tensor [2, 1, 1024, 64] : tensor<2x1x1024x64xf32>
    %1695 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1690, %1693 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1696 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1695 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1697 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1698 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1696 : tensor<2x1x1024x64xf32>) outs(%1697 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1699 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1698 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1700 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1701 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1690 : tensor<2x1x1024x64xf32>) outs(%1700 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1702 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1701 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1703 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1699, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1704 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1703 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1705 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1690, %1702 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1706 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1705, %1704 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1707 = tensor.collapse_shape %1706 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1708 = tensor.expand_shape %1707 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1709 = linalg.init_tensor [1024, 2048] : tensor<1024x2048xf32>
    %1710 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_152 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %1711 = linalg.init_tensor [2, 2048] : tensor<2x2048xf32>
    %1712 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1713 = linalg.matmul ins(%51, %1710 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%1712 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1714 = tensor.extract_slice %1713[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1715 = tensor.extract_slice %1713[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1716 = tensor.expand_shape %1715 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1717 = tensor.expand_shape %1714 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1718 = linalg.init_tensor [2, 1024, 1, 1] : tensor<2x1024x1x1xf32>
    %1719 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1717, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %1720 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1708, %1719 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1721 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1716, %1720 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1722 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1721 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %1723 = tensor.pad %1722 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %1724 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_154 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1725 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1723, %cst_153 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%1724 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1726 = tensor.collapse_shape %1725 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1727 = tensor.expand_shape %1726 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1728 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1729 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1727 : tensor<2x1x1024x64xf32>) outs(%1728 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1730 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1729 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1731 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1727, %1730 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1732 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1731 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1733 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1734 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1732 : tensor<2x1x1024x64xf32>) outs(%1733 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1735 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1734 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1736 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1737 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1727 : tensor<2x1x1024x64xf32>) outs(%1736 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1738 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1737 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1739 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1735, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1740 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1739 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1741 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1727, %1738 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1742 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1741, %1740 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1743 = tensor.collapse_shape %1742 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1744 = tensor.expand_shape %1743 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1745 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_155 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %1746 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1747 = linalg.matmul ins(%51, %1745 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%1746 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1748 = tensor.extract_slice %1747[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1749 = tensor.extract_slice %1747[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1750 = tensor.expand_shape %1749 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1751 = tensor.expand_shape %1748 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1752 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1751, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %1753 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1744, %1752 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1754 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1750, %1753 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1755 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1754 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %1756 = linalg.fill ins(%cst_514 : f32) outs(%1686 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1757 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1684, %cst_156 : tensor<2x512x8x8xf32>, tensor<1024x512x1x1xf32>) outs(%1756 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1758 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1755, %1757 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1759 = tensor.collapse_shape %1758 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1760 = tensor.expand_shape %1759 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1761 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1762 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1760 : tensor<2x1x1024x64xf32>) outs(%1761 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1763 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1762 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1764 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1760, %1763 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1765 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1764 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1766 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1767 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1765 : tensor<2x1x1024x64xf32>) outs(%1766 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1768 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1767 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1769 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1770 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1760 : tensor<2x1x1024x64xf32>) outs(%1769 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1771 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1770 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1772 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1768, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1773 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1772 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1774 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1760, %1771 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1775 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1774, %1773 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1776 = tensor.collapse_shape %1775 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1777 = tensor.expand_shape %1776 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1778 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1777, %cst_306 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1779 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1778, %cst_307 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1780 = linalg.init_tensor [2, 3072, 8, 8] : tensor<2x3072x8x8xf32>
    %1781 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_158 : tensor<3072xf32>) outs(%1780 : tensor<2x3072x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x8x8xf32>
    %1782 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1779, %cst_157 : tensor<2x1024x8x8xf32>, tensor<3072x1024x1x1xf32>) outs(%1781 : tensor<2x3072x8x8xf32>) -> tensor<2x3072x8x8xf32>
    %1783 = tensor.collapse_shape %1782 [[0], [1], [2, 3]] : tensor<2x3072x8x8xf32> into tensor<2x3072x64xf32>
    %1784 = tensor.expand_shape %1783 [[0], [1, 2], [3]] : tensor<2x3072x64xf32> into tensor<2x48x64x64xf32>
    %1785 = linalg.init_tensor [2, 48, 64, 64] : tensor<2x48x64x64xf32>
    %1786 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1784 : tensor<2x48x64x64xf32>) outs(%1785 : tensor<2x48x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x64x64xf32>
    %1787 = tensor.extract_slice %1786[0, 0, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %1788 = tensor.extract_slice %1786[0, 16, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %1789 = tensor.extract_slice %1786[0, 32, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %1790 = linalg.init_tensor [2, 16, 64, 64] : tensor<2x16x64x64xf32>
    %1791 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1787, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %1792 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1788 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %1793 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1792, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %1794 = tensor.collapse_shape %1791 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1795 = tensor.collapse_shape %1793 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1796 = linalg.init_tensor [32, 64, 64] : tensor<32x64x64xf32>
    %1797 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1798 = linalg.batch_matmul ins(%1794, %1795 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%1797 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1799 = tensor.expand_shape %1798 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %1800 = linalg.init_tensor [2, 16, 64, 1] : tensor<2x16x64x1xi64>
    %1801 = linalg.fill ins(%c0_i64 : i64) outs(%1800 : tensor<2x16x64x1xi64>) -> tensor<2x16x64x1xi64>
    %1802 = linalg.init_tensor [2, 16, 64, 1] : tensor<2x16x64x1xf32>
    %1803 = linalg.fill ins(%cst_515 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %1804:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1799 : tensor<2x16x64x64xf32>) outs(%1803, %1801 : tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>)
    %1805 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1799, %1804#0 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x64x64xf32>
    %1806 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1805 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %1807 = linalg.fill ins(%cst_514 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %1808 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1806 : tensor<2x16x64x64xf32>) outs(%1807 : tensor<2x16x64x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x1xf32>
    %1809 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1806, %1808 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %1810 = tensor.collapse_shape %1809 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1811 = tensor.collapse_shape %1789 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1812 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1813 = linalg.batch_matmul ins(%1810, %1811 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%1812 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1814 = tensor.expand_shape %1813 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %1815 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1814 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %1816 = tensor.collapse_shape %1815 [[0], [1, 2], [3]] : tensor<2x16x64x64xf32> into tensor<2x1024x64xf32>
    %1817 = tensor.expand_shape %1816 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1818 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_160 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1819 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1817, %cst_159 : tensor<2x1024x8x8xf32>, tensor<1024x1024x1x1xf32>) outs(%1818 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1820 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1758, %1819 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1821 = tensor.pad %1820 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %1822 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_162 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1823 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1821, %cst_161 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%1822 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1824 = tensor.collapse_shape %1823 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1825 = tensor.expand_shape %1824 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1826 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1827 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1825 : tensor<2x1x1024x64xf32>) outs(%1826 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1828 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1827 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1829 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1825, %1828 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1830 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1829 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1831 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1832 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1830 : tensor<2x1x1024x64xf32>) outs(%1831 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1833 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1832 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1834 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1835 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1825 : tensor<2x1x1024x64xf32>) outs(%1834 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1836 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1835 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1837 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1833, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1838 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1837 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1839 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1825, %1836 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1840 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1839, %1838 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1841 = tensor.collapse_shape %1840 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1842 = tensor.expand_shape %1841 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1843 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_163 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %1844 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1845 = linalg.matmul ins(%51, %1843 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%1844 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1846 = tensor.extract_slice %1845[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1847 = tensor.extract_slice %1845[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1848 = tensor.expand_shape %1847 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1849 = tensor.expand_shape %1846 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1850 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1849, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %1851 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1842, %1850 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1852 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1848, %1851 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1853 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1852 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %1854 = tensor.pad %1853 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %1855 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_165 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1856 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1854, %cst_164 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%1855 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1857 = tensor.collapse_shape %1856 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1858 = tensor.expand_shape %1857 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1859 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1860 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1858 : tensor<2x1x1024x64xf32>) outs(%1859 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1861 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1860 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1862 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1858, %1861 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1863 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1862 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1864 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1865 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1863 : tensor<2x1x1024x64xf32>) outs(%1864 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1866 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1865 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1867 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1868 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1858 : tensor<2x1x1024x64xf32>) outs(%1867 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1869 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1868 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1870 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1866, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1871 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1870 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1872 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1858, %1869 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1873 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1872, %1871 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1874 = tensor.collapse_shape %1873 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1875 = tensor.expand_shape %1874 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1876 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_166 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %1877 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1878 = linalg.matmul ins(%51, %1876 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%1877 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1879 = tensor.extract_slice %1878[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1880 = tensor.extract_slice %1878[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1881 = tensor.expand_shape %1880 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1882 = tensor.expand_shape %1879 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1883 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1882, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %1884 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1875, %1883 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1885 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1881, %1884 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1886 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1885 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %1887 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1886, %1820 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1888 = tensor.collapse_shape %1887 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1889 = tensor.expand_shape %1888 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1890 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1891 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1889 : tensor<2x1x1024x64xf32>) outs(%1890 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1892 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1891 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1893 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1889, %1892 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1894 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1893 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1895 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1896 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1894 : tensor<2x1x1024x64xf32>) outs(%1895 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1897 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1896 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1898 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1899 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1889 : tensor<2x1x1024x64xf32>) outs(%1898 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1900 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1899 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1901 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1897, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1902 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1901 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1903 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1889, %1900 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1904 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1903, %1902 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1905 = tensor.collapse_shape %1904 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1906 = tensor.expand_shape %1905 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1907 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1906, %cst_306 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1908 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1907, %cst_307 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1909 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_168 : tensor<3072xf32>) outs(%1780 : tensor<2x3072x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x8x8xf32>
    %1910 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1908, %cst_167 : tensor<2x1024x8x8xf32>, tensor<3072x1024x1x1xf32>) outs(%1909 : tensor<2x3072x8x8xf32>) -> tensor<2x3072x8x8xf32>
    %1911 = tensor.collapse_shape %1910 [[0], [1], [2, 3]] : tensor<2x3072x8x8xf32> into tensor<2x3072x64xf32>
    %1912 = tensor.expand_shape %1911 [[0], [1, 2], [3]] : tensor<2x3072x64xf32> into tensor<2x48x64x64xf32>
    %1913 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1912 : tensor<2x48x64x64xf32>) outs(%1785 : tensor<2x48x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x64x64xf32>
    %1914 = tensor.extract_slice %1913[0, 0, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %1915 = tensor.extract_slice %1913[0, 16, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %1916 = tensor.extract_slice %1913[0, 32, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %1917 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1914, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %1918 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1915 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %1919 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1918, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %1920 = tensor.collapse_shape %1917 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1921 = tensor.collapse_shape %1919 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1922 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1923 = linalg.batch_matmul ins(%1920, %1921 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%1922 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1924 = tensor.expand_shape %1923 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %1925 = linalg.fill ins(%c0_i64 : i64) outs(%1800 : tensor<2x16x64x1xi64>) -> tensor<2x16x64x1xi64>
    %1926 = linalg.fill ins(%cst_515 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %1927:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1924 : tensor<2x16x64x64xf32>) outs(%1926, %1925 : tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>)
    %1928 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1924, %1927#0 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x64x64xf32>
    %1929 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1928 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %1930 = linalg.fill ins(%cst_514 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %1931 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%1929 : tensor<2x16x64x64xf32>) outs(%1930 : tensor<2x16x64x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x1xf32>
    %1932 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1929, %1931 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %1933 = tensor.collapse_shape %1932 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1934 = tensor.collapse_shape %1916 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %1935 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1936 = linalg.batch_matmul ins(%1933, %1934 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%1935 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %1937 = tensor.expand_shape %1936 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %1938 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1937 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %1939 = tensor.collapse_shape %1938 [[0], [1, 2], [3]] : tensor<2x16x64x64xf32> into tensor<2x1024x64xf32>
    %1940 = tensor.expand_shape %1939 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1941 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_170 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1942 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1940, %cst_169 : tensor<2x1024x8x8xf32>, tensor<1024x1024x1x1xf32>) outs(%1941 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1943 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1887, %1942 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1944 = tensor.pad %1943 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %1945 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_172 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1946 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1944, %cst_171 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%1945 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1947 = tensor.collapse_shape %1946 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1948 = tensor.expand_shape %1947 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1949 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1950 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1948 : tensor<2x1x1024x64xf32>) outs(%1949 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1951 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1950 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1952 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1948, %1951 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1953 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1952 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1954 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1955 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1953 : tensor<2x1x1024x64xf32>) outs(%1954 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1956 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1955 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1957 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1958 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1948 : tensor<2x1x1024x64xf32>) outs(%1957 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1959 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1958 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1960 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1956, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1961 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1960 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1962 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1948, %1959 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1963 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1962, %1961 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1964 = tensor.collapse_shape %1963 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1965 = tensor.expand_shape %1964 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1966 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_173 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %1967 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1968 = linalg.matmul ins(%51, %1966 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%1967 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %1969 = tensor.extract_slice %1968[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1970 = tensor.extract_slice %1968[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %1971 = tensor.expand_shape %1970 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1972 = tensor.expand_shape %1969 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %1973 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1972, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %1974 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1965, %1973 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1975 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1971, %1974 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %1976 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1975 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %1977 = tensor.pad %1976 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %1978 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_175 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %1979 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%1977, %cst_174 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%1978 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %1980 = tensor.collapse_shape %1979 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %1981 = tensor.expand_shape %1980 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %1982 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1983 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1981 : tensor<2x1x1024x64xf32>) outs(%1982 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1984 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1983 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1985 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1981, %1984 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %1986 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1985 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1987 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1988 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1986 : tensor<2x1x1024x64xf32>) outs(%1987 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1989 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1988 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1990 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %1991 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%1981 : tensor<2x1x1024x64xf32>) outs(%1990 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1992 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1991 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1993 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1989, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %1994 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1993 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %1995 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1981, %1992 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1996 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1995, %1994 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %1997 = tensor.collapse_shape %1996 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %1998 = tensor.expand_shape %1997 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %1999 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_176 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2000 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2001 = linalg.matmul ins(%51, %1999 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2000 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2002 = tensor.extract_slice %2001[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2003 = tensor.extract_slice %2001[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2004 = tensor.expand_shape %2003 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2005 = tensor.expand_shape %2002 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2006 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2005, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2007 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1998, %2006 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2008 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2004, %2007 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2009 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2008 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %2010 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2009, %1943 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2011 = tensor.collapse_shape %2010 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %2012 = tensor.expand_shape %2011 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %2013 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2014 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2012 : tensor<2x1x1024x64xf32>) outs(%2013 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2015 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2014 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2016 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2012, %2015 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %2017 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2016 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2018 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2019 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2017 : tensor<2x1x1024x64xf32>) outs(%2018 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2020 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2019 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2021 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2022 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2012 : tensor<2x1x1024x64xf32>) outs(%2021 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2023 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2022 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2024 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2020, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2025 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2024 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2026 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2012, %2023 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2027 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2026, %2025 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2028 = tensor.collapse_shape %2027 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %2029 = tensor.expand_shape %2028 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %2030 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2029, %cst_306 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2031 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2030, %cst_307 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2032 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_178 : tensor<3072xf32>) outs(%1780 : tensor<2x3072x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x8x8xf32>
    %2033 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2031, %cst_177 : tensor<2x1024x8x8xf32>, tensor<3072x1024x1x1xf32>) outs(%2032 : tensor<2x3072x8x8xf32>) -> tensor<2x3072x8x8xf32>
    %2034 = tensor.collapse_shape %2033 [[0], [1], [2, 3]] : tensor<2x3072x8x8xf32> into tensor<2x3072x64xf32>
    %2035 = tensor.expand_shape %2034 [[0], [1, 2], [3]] : tensor<2x3072x64xf32> into tensor<2x48x64x64xf32>
    %2036 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2035 : tensor<2x48x64x64xf32>) outs(%1785 : tensor<2x48x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x64x64xf32>
    %2037 = tensor.extract_slice %2036[0, 0, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %2038 = tensor.extract_slice %2036[0, 16, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %2039 = tensor.extract_slice %2036[0, 32, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %2040 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2037, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %2041 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2038 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %2042 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2041, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %2043 = tensor.collapse_shape %2040 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2044 = tensor.collapse_shape %2042 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2045 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2046 = linalg.batch_matmul ins(%2043, %2044 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%2045 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2047 = tensor.expand_shape %2046 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %2048 = linalg.fill ins(%c0_i64 : i64) outs(%1800 : tensor<2x16x64x1xi64>) -> tensor<2x16x64x1xi64>
    %2049 = linalg.fill ins(%cst_515 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %2050:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2047 : tensor<2x16x64x64xf32>) outs(%2049, %2048 : tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>)
    %2051 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2047, %2050#0 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x64x64xf32>
    %2052 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2051 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %2053 = linalg.fill ins(%cst_514 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %2054 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2052 : tensor<2x16x64x64xf32>) outs(%2053 : tensor<2x16x64x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x1xf32>
    %2055 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2052, %2054 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %2056 = tensor.collapse_shape %2055 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2057 = tensor.collapse_shape %2039 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2058 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2059 = linalg.batch_matmul ins(%2056, %2057 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%2058 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2060 = tensor.expand_shape %2059 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %2061 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2060 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %2062 = tensor.collapse_shape %2061 [[0], [1, 2], [3]] : tensor<2x16x64x64xf32> into tensor<2x1024x64xf32>
    %2063 = tensor.expand_shape %2062 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %2064 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_180 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %2065 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2063, %cst_179 : tensor<2x1024x8x8xf32>, tensor<1024x1024x1x1xf32>) outs(%2064 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %2066 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2010, %2065 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2067 = tensor.pad %2066 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %2068 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_182 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %2069 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2067, %cst_181 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%2068 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %2070 = tensor.collapse_shape %2069 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %2071 = tensor.expand_shape %2070 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %2072 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2073 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2071 : tensor<2x1x1024x64xf32>) outs(%2072 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2074 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2073 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2075 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2071, %2074 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %2076 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2075 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2077 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2078 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2076 : tensor<2x1x1024x64xf32>) outs(%2077 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2079 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2078 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2080 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2081 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2071 : tensor<2x1x1024x64xf32>) outs(%2080 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2082 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2081 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2083 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2079, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2084 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2083 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2085 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2071, %2082 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2086 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2085, %2084 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2087 = tensor.collapse_shape %2086 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %2088 = tensor.expand_shape %2087 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %2089 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_183 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2090 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2091 = linalg.matmul ins(%51, %2089 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2090 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2092 = tensor.extract_slice %2091[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2093 = tensor.extract_slice %2091[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2094 = tensor.expand_shape %2093 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2095 = tensor.expand_shape %2092 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2096 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2095, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2097 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2088, %2096 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2098 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2094, %2097 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2099 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2098 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %2100 = tensor.pad %2099 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %2101 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_185 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %2102 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2100, %cst_184 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%2101 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %2103 = tensor.collapse_shape %2102 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %2104 = tensor.expand_shape %2103 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %2105 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2106 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2104 : tensor<2x1x1024x64xf32>) outs(%2105 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2107 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2106 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2108 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2104, %2107 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %2109 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2108 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2110 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2111 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2109 : tensor<2x1x1024x64xf32>) outs(%2110 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2112 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2111 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2113 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2114 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2104 : tensor<2x1x1024x64xf32>) outs(%2113 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2115 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2114 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2116 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2112, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2117 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2116 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2118 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2104, %2115 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2119 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2118, %2117 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2120 = tensor.collapse_shape %2119 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %2121 = tensor.expand_shape %2120 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %2122 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_186 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2123 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2124 = linalg.matmul ins(%51, %2122 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2123 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2125 = tensor.extract_slice %2124[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2126 = tensor.extract_slice %2124[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2127 = tensor.expand_shape %2126 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2128 = tensor.expand_shape %2125 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2129 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2128, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2130 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2121, %2129 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2131 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2127, %2130 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2132 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2131 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %2133 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2132, %2066 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2134 = tensor.collapse_shape %2133 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %2135 = tensor.expand_shape %2134 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %2136 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2137 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2135 : tensor<2x1x1024x64xf32>) outs(%2136 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2138 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2137 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2139 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2135, %2138 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %2140 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2139 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2141 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2142 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2140 : tensor<2x1x1024x64xf32>) outs(%2141 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2143 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2142 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2144 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2145 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2135 : tensor<2x1x1024x64xf32>) outs(%2144 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2146 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2145 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2147 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2143, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2148 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2147 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2149 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2135, %2146 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2150 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2149, %2148 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %2151 = tensor.collapse_shape %2150 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %2152 = tensor.expand_shape %2151 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %2153 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2152, %cst_306 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2154 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2153, %cst_307 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2155 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_188 : tensor<3072xf32>) outs(%1780 : tensor<2x3072x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x8x8xf32>
    %2156 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2154, %cst_187 : tensor<2x1024x8x8xf32>, tensor<3072x1024x1x1xf32>) outs(%2155 : tensor<2x3072x8x8xf32>) -> tensor<2x3072x8x8xf32>
    %2157 = tensor.collapse_shape %2156 [[0], [1], [2, 3]] : tensor<2x3072x8x8xf32> into tensor<2x3072x64xf32>
    %2158 = tensor.expand_shape %2157 [[0], [1, 2], [3]] : tensor<2x3072x64xf32> into tensor<2x48x64x64xf32>
    %2159 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2158 : tensor<2x48x64x64xf32>) outs(%1785 : tensor<2x48x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x64x64xf32>
    %2160 = tensor.extract_slice %2159[0, 0, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %2161 = tensor.extract_slice %2159[0, 16, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %2162 = tensor.extract_slice %2159[0, 32, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %2163 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2160, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %2164 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2161 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %2165 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2164, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %2166 = tensor.collapse_shape %2163 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2167 = tensor.collapse_shape %2165 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2168 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2169 = linalg.batch_matmul ins(%2166, %2167 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%2168 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2170 = tensor.expand_shape %2169 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %2171 = linalg.fill ins(%c0_i64 : i64) outs(%1800 : tensor<2x16x64x1xi64>) -> tensor<2x16x64x1xi64>
    %2172 = linalg.fill ins(%cst_515 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %2173:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2170 : tensor<2x16x64x64xf32>) outs(%2172, %2171 : tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>)
    %2174 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2170, %2173#0 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x64x64xf32>
    %2175 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2174 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %2176 = linalg.fill ins(%cst_514 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %2177 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2175 : tensor<2x16x64x64xf32>) outs(%2176 : tensor<2x16x64x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x1xf32>
    %2178 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2175, %2177 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %2179 = tensor.collapse_shape %2178 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2180 = tensor.collapse_shape %2162 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %2181 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2182 = linalg.batch_matmul ins(%2179, %2180 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%2181 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %2183 = tensor.expand_shape %2182 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %2184 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2183 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %2185 = tensor.collapse_shape %2184 [[0], [1, 2], [3]] : tensor<2x16x64x64xf32> into tensor<2x1024x64xf32>
    %2186 = tensor.expand_shape %2185 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %2187 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_190 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %2188 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2186, %cst_189 : tensor<2x1024x8x8xf32>, tensor<1024x1024x1x1xf32>) outs(%2187 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %2189 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2133, %2188 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %2190 = linalg.init_tensor [2, 1024, 4, 4] : tensor<2x1024x4x4xf32>
    %2191 = linalg.fill ins(%cst_514 : f32) outs(%2190 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2192 = linalg.pooling_nchw_sum {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%2189, %348 : tensor<2x1024x8x8xf32>, tensor<2x2xf32>) outs(%2191 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2193 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2192 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_522 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2194 = tensor.pad %2193 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2195 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_192 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2196 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2194, %cst_191 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2195 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2197 = tensor.collapse_shape %2196 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2198 = tensor.expand_shape %2197 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2199 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2200 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2198 : tensor<2x1x1024x16xf32>) outs(%2199 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2201 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2200 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2202 = linalg.init_tensor [2, 1, 1024, 16] : tensor<2x1x1024x16xf32>
    %2203 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2198, %2201 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2204 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2203 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2205 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2206 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2204 : tensor<2x1x1024x16xf32>) outs(%2205 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2207 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2206 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2208 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2209 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2198 : tensor<2x1x1024x16xf32>) outs(%2208 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2210 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2209 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2211 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2207, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2212 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2211 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2213 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2198, %2210 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2214 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2213, %2212 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2215 = tensor.collapse_shape %2214 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2216 = tensor.expand_shape %2215 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2217 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_193 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2218 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2219 = linalg.matmul ins(%51, %2217 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2218 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2220 = tensor.extract_slice %2219[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2221 = tensor.extract_slice %2219[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2222 = tensor.expand_shape %2221 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2223 = tensor.expand_shape %2220 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2224 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2223, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2225 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2216, %2224 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2226 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2222, %2225 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2227 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2226 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2228 = tensor.pad %2227 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2229 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_195 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2230 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2228, %cst_194 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2229 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2231 = tensor.collapse_shape %2230 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2232 = tensor.expand_shape %2231 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2233 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2234 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2232 : tensor<2x1x1024x16xf32>) outs(%2233 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2235 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2234 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2236 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2232, %2235 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2237 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2236 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2238 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2239 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2237 : tensor<2x1x1024x16xf32>) outs(%2238 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2240 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2239 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2241 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2242 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2232 : tensor<2x1x1024x16xf32>) outs(%2241 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2243 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2242 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2244 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2240, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2245 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2244 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2246 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2232, %2243 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2247 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2246, %2245 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2248 = tensor.collapse_shape %2247 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2249 = tensor.expand_shape %2248 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2250 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_196 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2251 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2252 = linalg.matmul ins(%51, %2250 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2251 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2253 = tensor.extract_slice %2252[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2254 = tensor.extract_slice %2252[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2255 = tensor.expand_shape %2254 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2256 = tensor.expand_shape %2253 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2257 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2256, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2258 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2249, %2257 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2259 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2255, %2258 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2260 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2259 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2261 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2260, %2193 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2262 = tensor.collapse_shape %2261 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2263 = tensor.expand_shape %2262 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2264 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2265 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2263 : tensor<2x1x1024x16xf32>) outs(%2264 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2266 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2265 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2267 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2263, %2266 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2268 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2267 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2269 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2270 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2268 : tensor<2x1x1024x16xf32>) outs(%2269 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2271 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2270 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2272 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2273 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2263 : tensor<2x1x1024x16xf32>) outs(%2272 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2274 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2273 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2275 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2271, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2276 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2275 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2277 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2263, %2274 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2278 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2277, %2276 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2279 = tensor.collapse_shape %2278 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2280 = tensor.expand_shape %2279 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2281 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2280, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2282 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2281, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2283 = linalg.init_tensor [2, 3072, 4, 4] : tensor<2x3072x4x4xf32>
    %2284 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_198 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %2285 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2282, %cst_197 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%2284 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %2286 = tensor.collapse_shape %2285 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %2287 = tensor.expand_shape %2286 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %2288 = linalg.init_tensor [2, 48, 16, 64] : tensor<2x48x16x64xf32>
    %2289 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2287 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %2290 = tensor.extract_slice %2289[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2291 = tensor.extract_slice %2289[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2292 = tensor.extract_slice %2289[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2293 = linalg.init_tensor [2, 16, 16, 64] : tensor<2x16x16x64xf32>
    %2294 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2290, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %2295 = linalg.init_tensor [2, 16, 64, 16] : tensor<2x16x64x16xf32>
    %2296 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2291 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2297 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2296, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %2298 = tensor.collapse_shape %2294 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2299 = tensor.collapse_shape %2297 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %2300 = linalg.init_tensor [32, 16, 16] : tensor<32x16x16xf32>
    %2301 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2302 = linalg.batch_matmul ins(%2298, %2299 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%2301 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2303 = tensor.expand_shape %2302 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %2304 = linalg.init_tensor [2, 16, 16, 1] : tensor<2x16x16x1xi64>
    %2305 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %2306 = linalg.init_tensor [2, 16, 16, 1] : tensor<2x16x16x1xf32>
    %2307 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2308:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2303 : tensor<2x16x16x16xf32>) outs(%2307, %2305 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %2309 = linalg.init_tensor [2, 16, 16, 16] : tensor<2x16x16x16xf32>
    %2310 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2303, %2308#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %2311 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2310 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2312 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2313 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2311 : tensor<2x16x16x16xf32>) outs(%2312 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %2314 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2311, %2313 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2315 = tensor.collapse_shape %2314 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %2316 = tensor.collapse_shape %2292 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2317 = linalg.init_tensor [32, 16, 64] : tensor<32x16x64xf32>
    %2318 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2319 = linalg.batch_matmul ins(%2315, %2316 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%2318 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2320 = tensor.expand_shape %2319 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %2321 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2320 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2322 = tensor.collapse_shape %2321 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %2323 = tensor.expand_shape %2322 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2324 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_200 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2325 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2323, %cst_199 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%2324 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2326 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2261, %2325 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2327 = tensor.pad %2326 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2328 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_202 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2329 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2327, %cst_201 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2328 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2330 = tensor.collapse_shape %2329 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2331 = tensor.expand_shape %2330 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2332 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2333 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2331 : tensor<2x1x1024x16xf32>) outs(%2332 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2334 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2333 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2335 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2331, %2334 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2336 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2335 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2337 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2338 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2336 : tensor<2x1x1024x16xf32>) outs(%2337 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2339 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2338 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2340 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2341 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2331 : tensor<2x1x1024x16xf32>) outs(%2340 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2342 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2341 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2343 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2339, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2344 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2343 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2345 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2331, %2342 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2346 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2345, %2344 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2347 = tensor.collapse_shape %2346 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2348 = tensor.expand_shape %2347 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2349 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_203 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2350 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2351 = linalg.matmul ins(%51, %2349 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2350 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2352 = tensor.extract_slice %2351[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2353 = tensor.extract_slice %2351[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2354 = tensor.expand_shape %2353 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2355 = tensor.expand_shape %2352 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2356 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2355, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2357 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2348, %2356 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2358 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2354, %2357 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2359 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2358 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2360 = tensor.pad %2359 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2361 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_205 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2362 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2360, %cst_204 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2361 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2363 = tensor.collapse_shape %2362 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2364 = tensor.expand_shape %2363 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2365 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2366 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2364 : tensor<2x1x1024x16xf32>) outs(%2365 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2367 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2366 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2368 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2364, %2367 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2369 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2368 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2370 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2371 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2369 : tensor<2x1x1024x16xf32>) outs(%2370 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2372 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2371 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2373 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2374 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2364 : tensor<2x1x1024x16xf32>) outs(%2373 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2375 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2374 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2376 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2372, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2377 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2376 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2378 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2364, %2375 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2379 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2378, %2377 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2380 = tensor.collapse_shape %2379 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2381 = tensor.expand_shape %2380 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2382 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_206 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2383 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2384 = linalg.matmul ins(%51, %2382 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2383 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2385 = tensor.extract_slice %2384[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2386 = tensor.extract_slice %2384[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2387 = tensor.expand_shape %2386 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2388 = tensor.expand_shape %2385 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2389 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2388, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2390 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2381, %2389 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2391 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2387, %2390 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2392 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2391 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2393 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2392, %2326 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2394 = tensor.collapse_shape %2393 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2395 = tensor.expand_shape %2394 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2396 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2397 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2395 : tensor<2x1x1024x16xf32>) outs(%2396 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2398 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2397 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2399 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2395, %2398 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2400 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2399 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2401 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2402 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2400 : tensor<2x1x1024x16xf32>) outs(%2401 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2403 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2402 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2404 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2405 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2395 : tensor<2x1x1024x16xf32>) outs(%2404 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2406 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2405 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2407 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2403, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2408 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2407 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2409 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2395, %2406 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2410 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2409, %2408 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2411 = tensor.collapse_shape %2410 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2412 = tensor.expand_shape %2411 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2413 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2412, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2414 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2413, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2415 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_208 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %2416 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2414, %cst_207 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%2415 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %2417 = tensor.collapse_shape %2416 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %2418 = tensor.expand_shape %2417 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %2419 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2418 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %2420 = tensor.extract_slice %2419[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2421 = tensor.extract_slice %2419[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2422 = tensor.extract_slice %2419[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2423 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2420, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %2424 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2421 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2425 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2424, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %2426 = tensor.collapse_shape %2423 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2427 = tensor.collapse_shape %2425 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %2428 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2429 = linalg.batch_matmul ins(%2426, %2427 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%2428 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2430 = tensor.expand_shape %2429 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %2431 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %2432 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2433:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2430 : tensor<2x16x16x16xf32>) outs(%2432, %2431 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %2434 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2430, %2433#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %2435 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2434 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2436 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2437 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2435 : tensor<2x16x16x16xf32>) outs(%2436 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %2438 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2435, %2437 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2439 = tensor.collapse_shape %2438 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %2440 = tensor.collapse_shape %2422 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2441 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2442 = linalg.batch_matmul ins(%2439, %2440 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%2441 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2443 = tensor.expand_shape %2442 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %2444 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2443 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2445 = tensor.collapse_shape %2444 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %2446 = tensor.expand_shape %2445 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2447 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_210 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2448 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2446, %cst_209 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%2447 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2449 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2393, %2448 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2450 = tensor.pad %2449 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2451 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_212 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2452 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2450, %cst_211 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2451 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2453 = tensor.collapse_shape %2452 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2454 = tensor.expand_shape %2453 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2455 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2456 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2454 : tensor<2x1x1024x16xf32>) outs(%2455 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2457 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2456 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2458 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2454, %2457 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2459 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2458 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2460 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2461 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2459 : tensor<2x1x1024x16xf32>) outs(%2460 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2462 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2461 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2463 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2464 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2454 : tensor<2x1x1024x16xf32>) outs(%2463 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2465 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2464 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2466 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2462, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2467 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2466 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2468 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2454, %2465 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2469 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2468, %2467 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2470 = tensor.collapse_shape %2469 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2471 = tensor.expand_shape %2470 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2472 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_213 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2473 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2474 = linalg.matmul ins(%51, %2472 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2473 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2475 = tensor.extract_slice %2474[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2476 = tensor.extract_slice %2474[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2477 = tensor.expand_shape %2476 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2478 = tensor.expand_shape %2475 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2479 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2478, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2480 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2471, %2479 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2481 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2477, %2480 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2482 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2481 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2483 = tensor.pad %2482 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2484 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_215 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2485 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2483, %cst_214 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2484 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2486 = tensor.collapse_shape %2485 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2487 = tensor.expand_shape %2486 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2488 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2489 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2487 : tensor<2x1x1024x16xf32>) outs(%2488 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2490 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2489 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2491 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2487, %2490 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2492 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2491 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2493 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2494 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2492 : tensor<2x1x1024x16xf32>) outs(%2493 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2495 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2494 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2496 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2497 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2487 : tensor<2x1x1024x16xf32>) outs(%2496 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2498 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2497 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2499 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2495, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2500 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2499 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2501 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2487, %2498 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2502 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2501, %2500 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2503 = tensor.collapse_shape %2502 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2504 = tensor.expand_shape %2503 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2505 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_216 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2506 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2507 = linalg.matmul ins(%51, %2505 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2506 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2508 = tensor.extract_slice %2507[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2509 = tensor.extract_slice %2507[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2510 = tensor.expand_shape %2509 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2511 = tensor.expand_shape %2508 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2512 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2511, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2513 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2504, %2512 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2514 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2510, %2513 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2515 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2514 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2516 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2515, %2449 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2517 = tensor.collapse_shape %2516 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2518 = tensor.expand_shape %2517 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2519 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2520 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2518 : tensor<2x1x1024x16xf32>) outs(%2519 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2521 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2520 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2522 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2518, %2521 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2523 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2522 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2524 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2525 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2523 : tensor<2x1x1024x16xf32>) outs(%2524 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2526 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2525 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2527 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2528 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2518 : tensor<2x1x1024x16xf32>) outs(%2527 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2529 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2528 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2530 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2526, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2531 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2530 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2532 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2518, %2529 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2533 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2532, %2531 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2534 = tensor.collapse_shape %2533 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2535 = tensor.expand_shape %2534 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2536 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2535, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2537 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2536, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2538 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_218 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %2539 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2537, %cst_217 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%2538 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %2540 = tensor.collapse_shape %2539 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %2541 = tensor.expand_shape %2540 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %2542 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2541 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %2543 = tensor.extract_slice %2542[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2544 = tensor.extract_slice %2542[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2545 = tensor.extract_slice %2542[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2546 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2543, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %2547 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2544 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2548 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2547, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %2549 = tensor.collapse_shape %2546 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2550 = tensor.collapse_shape %2548 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %2551 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2552 = linalg.batch_matmul ins(%2549, %2550 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%2551 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2553 = tensor.expand_shape %2552 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %2554 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %2555 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2556:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2553 : tensor<2x16x16x16xf32>) outs(%2555, %2554 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %2557 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2553, %2556#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %2558 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2557 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2559 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2560 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2558 : tensor<2x16x16x16xf32>) outs(%2559 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %2561 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2558, %2560 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2562 = tensor.collapse_shape %2561 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %2563 = tensor.collapse_shape %2545 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2564 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2565 = linalg.batch_matmul ins(%2562, %2563 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%2564 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2566 = tensor.expand_shape %2565 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %2567 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2566 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2568 = tensor.collapse_shape %2567 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %2569 = tensor.expand_shape %2568 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2570 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_220 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2571 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2569, %cst_219 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%2570 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2572 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2516, %2571 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2573 = tensor.pad %2572 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2574 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_222 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2575 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2573, %cst_221 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2574 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2576 = tensor.collapse_shape %2575 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2577 = tensor.expand_shape %2576 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2578 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2579 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2577 : tensor<2x1x1024x16xf32>) outs(%2578 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2580 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2579 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2581 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2577, %2580 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2582 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2581 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2583 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2584 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2582 : tensor<2x1x1024x16xf32>) outs(%2583 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2585 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2584 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2586 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2587 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2577 : tensor<2x1x1024x16xf32>) outs(%2586 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2588 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2587 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2589 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2585, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2590 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2589 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2591 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2577, %2588 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2592 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2591, %2590 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2593 = tensor.collapse_shape %2592 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2594 = tensor.expand_shape %2593 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2595 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_223 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2596 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2597 = linalg.matmul ins(%51, %2595 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2596 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2598 = tensor.extract_slice %2597[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2599 = tensor.extract_slice %2597[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2600 = tensor.expand_shape %2599 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2601 = tensor.expand_shape %2598 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2602 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2601, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2603 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2594, %2602 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2604 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2600, %2603 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2605 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2604 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2606 = tensor.pad %2605 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2607 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_225 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2608 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2606, %cst_224 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2607 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2609 = tensor.collapse_shape %2608 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2610 = tensor.expand_shape %2609 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2611 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2612 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2610 : tensor<2x1x1024x16xf32>) outs(%2611 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2613 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2612 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2614 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2610, %2613 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2615 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2614 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2616 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2617 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2615 : tensor<2x1x1024x16xf32>) outs(%2616 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2618 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2617 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2619 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2620 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2610 : tensor<2x1x1024x16xf32>) outs(%2619 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2621 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2620 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2622 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2618, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2623 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2622 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2624 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2610, %2621 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2625 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2624, %2623 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2626 = tensor.collapse_shape %2625 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2627 = tensor.expand_shape %2626 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2628 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_226 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2629 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2630 = linalg.matmul ins(%51, %2628 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2629 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2631 = tensor.extract_slice %2630[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2632 = tensor.extract_slice %2630[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2633 = tensor.expand_shape %2632 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2634 = tensor.expand_shape %2631 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2635 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2634, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2636 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2627, %2635 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2637 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2633, %2636 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2638 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2637 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2639 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2638, %2572 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2640 = tensor.collapse_shape %2639 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2641 = tensor.expand_shape %2640 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2642 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2643 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2641 : tensor<2x1x1024x16xf32>) outs(%2642 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2644 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2643 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2645 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2641, %2644 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2646 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2645 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2647 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2648 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2646 : tensor<2x1x1024x16xf32>) outs(%2647 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2649 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2648 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2650 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2651 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2641 : tensor<2x1x1024x16xf32>) outs(%2650 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2652 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2651 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2653 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2649, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2654 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2653 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2655 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2641, %2652 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2656 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2655, %2654 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2657 = tensor.collapse_shape %2656 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2658 = tensor.expand_shape %2657 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2659 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2658, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2660 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2659, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2661 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_228 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %2662 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2660, %cst_227 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%2661 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %2663 = tensor.collapse_shape %2662 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %2664 = tensor.expand_shape %2663 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %2665 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2664 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %2666 = tensor.extract_slice %2665[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2667 = tensor.extract_slice %2665[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2668 = tensor.extract_slice %2665[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2669 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2666, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %2670 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2667 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2671 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2670, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %2672 = tensor.collapse_shape %2669 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2673 = tensor.collapse_shape %2671 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %2674 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2675 = linalg.batch_matmul ins(%2672, %2673 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%2674 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2676 = tensor.expand_shape %2675 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %2677 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %2678 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2679:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2676 : tensor<2x16x16x16xf32>) outs(%2678, %2677 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %2680 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2676, %2679#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %2681 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2680 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2682 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2683 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2681 : tensor<2x16x16x16xf32>) outs(%2682 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %2684 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2681, %2683 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2685 = tensor.collapse_shape %2684 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %2686 = tensor.collapse_shape %2668 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2687 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2688 = linalg.batch_matmul ins(%2685, %2686 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%2687 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2689 = tensor.expand_shape %2688 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %2690 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2689 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2691 = tensor.collapse_shape %2690 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %2692 = tensor.expand_shape %2691 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2693 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_230 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2694 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2692, %cst_229 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%2693 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2695 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2639, %2694 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2696 = tensor.pad %2695 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2697 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_232 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2698 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2696, %cst_231 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2697 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2699 = tensor.collapse_shape %2698 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2700 = tensor.expand_shape %2699 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2701 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2702 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2700 : tensor<2x1x1024x16xf32>) outs(%2701 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2703 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2702 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2704 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2700, %2703 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2705 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2704 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2706 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2707 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2705 : tensor<2x1x1024x16xf32>) outs(%2706 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2708 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2707 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2709 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2710 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2700 : tensor<2x1x1024x16xf32>) outs(%2709 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2711 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2710 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2712 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2708, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2713 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2712 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2714 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2700, %2711 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2715 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2714, %2713 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2716 = tensor.collapse_shape %2715 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2717 = tensor.expand_shape %2716 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2718 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_233 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2719 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2720 = linalg.matmul ins(%51, %2718 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2719 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2721 = tensor.extract_slice %2720[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2722 = tensor.extract_slice %2720[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2723 = tensor.expand_shape %2722 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2724 = tensor.expand_shape %2721 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2725 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2724, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2726 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2717, %2725 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2727 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2723, %2726 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2728 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2727 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2729 = tensor.pad %2728 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2730 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_235 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2731 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2729, %cst_234 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2730 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2732 = tensor.collapse_shape %2731 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2733 = tensor.expand_shape %2732 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2734 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2735 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2733 : tensor<2x1x1024x16xf32>) outs(%2734 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2736 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2735 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2737 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2733, %2736 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2738 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2737 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2739 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2740 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2738 : tensor<2x1x1024x16xf32>) outs(%2739 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2741 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2740 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2742 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2743 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2733 : tensor<2x1x1024x16xf32>) outs(%2742 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2744 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2743 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2745 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2741, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2746 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2745 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2747 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2733, %2744 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2748 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2747, %2746 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2749 = tensor.collapse_shape %2748 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2750 = tensor.expand_shape %2749 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2751 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_236 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2752 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2753 = linalg.matmul ins(%51, %2751 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2752 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2754 = tensor.extract_slice %2753[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2755 = tensor.extract_slice %2753[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2756 = tensor.expand_shape %2755 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2757 = tensor.expand_shape %2754 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2758 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2757, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2759 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2750, %2758 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2760 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2756, %2759 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2761 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2760 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2762 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2761, %2695 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2763 = tensor.collapse_shape %2762 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2764 = tensor.expand_shape %2763 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2765 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2766 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2764 : tensor<2x1x1024x16xf32>) outs(%2765 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2767 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2766 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2768 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2764, %2767 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2769 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2768 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2770 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2771 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2769 : tensor<2x1x1024x16xf32>) outs(%2770 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2772 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2771 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2773 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2774 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2764 : tensor<2x1x1024x16xf32>) outs(%2773 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2775 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2774 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2776 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2772, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2777 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2776 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2778 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2764, %2775 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2779 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2778, %2777 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2780 = tensor.collapse_shape %2779 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2781 = tensor.expand_shape %2780 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2782 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2781, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2783 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2782, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2784 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_238 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %2785 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2783, %cst_237 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%2784 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %2786 = tensor.collapse_shape %2785 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %2787 = tensor.expand_shape %2786 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %2788 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2787 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %2789 = tensor.extract_slice %2788[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2790 = tensor.extract_slice %2788[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2791 = tensor.extract_slice %2788[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2792 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2789, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %2793 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2790 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2794 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2793, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %2795 = tensor.collapse_shape %2792 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2796 = tensor.collapse_shape %2794 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %2797 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2798 = linalg.batch_matmul ins(%2795, %2796 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%2797 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2799 = tensor.expand_shape %2798 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %2800 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %2801 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2802:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2799 : tensor<2x16x16x16xf32>) outs(%2801, %2800 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %2803 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2799, %2802#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %2804 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2803 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2805 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2806 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2804 : tensor<2x16x16x16xf32>) outs(%2805 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %2807 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2804, %2806 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2808 = tensor.collapse_shape %2807 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %2809 = tensor.collapse_shape %2791 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2810 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2811 = linalg.batch_matmul ins(%2808, %2809 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%2810 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2812 = tensor.expand_shape %2811 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %2813 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2812 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2814 = tensor.collapse_shape %2813 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %2815 = tensor.expand_shape %2814 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2816 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_240 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2817 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2815, %cst_239 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%2816 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2818 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2762, %2817 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2819 = tensor.pad %2818 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2820 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_242 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2821 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2819, %cst_241 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2820 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2822 = tensor.collapse_shape %2821 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2823 = tensor.expand_shape %2822 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2824 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2825 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2823 : tensor<2x1x1024x16xf32>) outs(%2824 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2826 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2825 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2827 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2823, %2826 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2828 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2827 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2829 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2830 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2828 : tensor<2x1x1024x16xf32>) outs(%2829 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2831 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2830 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2832 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2833 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2823 : tensor<2x1x1024x16xf32>) outs(%2832 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2834 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2833 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2835 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2831, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2836 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2835 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2837 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2823, %2834 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2838 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2837, %2836 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2839 = tensor.collapse_shape %2838 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2840 = tensor.expand_shape %2839 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2841 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_243 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2842 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2843 = linalg.matmul ins(%51, %2841 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2842 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2844 = tensor.extract_slice %2843[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2845 = tensor.extract_slice %2843[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2846 = tensor.expand_shape %2845 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2847 = tensor.expand_shape %2844 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2848 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2847, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2849 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2840, %2848 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2850 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2846, %2849 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2851 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2850 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2852 = tensor.pad %2851 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2853 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_245 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2854 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2852, %cst_244 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2853 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2855 = tensor.collapse_shape %2854 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2856 = tensor.expand_shape %2855 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2857 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2858 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2856 : tensor<2x1x1024x16xf32>) outs(%2857 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2859 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2858 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2860 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2856, %2859 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2861 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2860 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2862 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2863 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2861 : tensor<2x1x1024x16xf32>) outs(%2862 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2864 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2863 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2865 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2866 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2856 : tensor<2x1x1024x16xf32>) outs(%2865 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2867 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2866 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2868 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2864, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2869 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2868 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2870 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2856, %2867 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2871 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2870, %2869 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2872 = tensor.collapse_shape %2871 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2873 = tensor.expand_shape %2872 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2874 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_246 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2875 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2876 = linalg.matmul ins(%51, %2874 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2875 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2877 = tensor.extract_slice %2876[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2878 = tensor.extract_slice %2876[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2879 = tensor.expand_shape %2878 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2880 = tensor.expand_shape %2877 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2881 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2880, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2882 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2873, %2881 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2883 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2879, %2882 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2884 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2883 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2885 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2884, %2818 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2886 = tensor.collapse_shape %2885 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2887 = tensor.expand_shape %2886 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2888 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2889 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2887 : tensor<2x1x1024x16xf32>) outs(%2888 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2890 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2889 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2891 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2887, %2890 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2892 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2891 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2893 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2894 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2892 : tensor<2x1x1024x16xf32>) outs(%2893 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2895 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2894 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2896 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2897 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2887 : tensor<2x1x1024x16xf32>) outs(%2896 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2898 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2897 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2899 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2895, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2900 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2899 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2901 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2887, %2898 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2902 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2901, %2900 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2903 = tensor.collapse_shape %2902 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2904 = tensor.expand_shape %2903 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2905 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2904, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2906 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2905, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2907 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_248 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %2908 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2906, %cst_247 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%2907 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %2909 = tensor.collapse_shape %2908 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %2910 = tensor.expand_shape %2909 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %2911 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2910 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %2912 = tensor.extract_slice %2911[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2913 = tensor.extract_slice %2911[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2914 = tensor.extract_slice %2911[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %2915 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2912, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %2916 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2913 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2917 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2916, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %2918 = tensor.collapse_shape %2915 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2919 = tensor.collapse_shape %2917 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %2920 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2921 = linalg.batch_matmul ins(%2918, %2919 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%2920 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %2922 = tensor.expand_shape %2921 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %2923 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %2924 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2925:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2922 : tensor<2x16x16x16xf32>) outs(%2924, %2923 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %2926 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2922, %2925#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %2927 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2926 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2928 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %2929 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%2927 : tensor<2x16x16x16xf32>) outs(%2928 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %2930 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2927, %2929 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %2931 = tensor.collapse_shape %2930 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %2932 = tensor.collapse_shape %2914 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %2933 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2934 = linalg.batch_matmul ins(%2931, %2932 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%2933 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %2935 = tensor.expand_shape %2934 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %2936 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2935 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %2937 = tensor.collapse_shape %2936 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %2938 = tensor.expand_shape %2937 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2939 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_250 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2940 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2938, %cst_249 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%2939 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2941 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2885, %2940 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2942 = tensor.pad %2941 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2943 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_252 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2944 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2942, %cst_251 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2943 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2945 = tensor.collapse_shape %2944 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2946 = tensor.expand_shape %2945 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2947 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2948 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2946 : tensor<2x1x1024x16xf32>) outs(%2947 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2949 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2948 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2950 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2946, %2949 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2951 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2950 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2952 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2953 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2951 : tensor<2x1x1024x16xf32>) outs(%2952 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2954 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2953 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2955 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2956 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2946 : tensor<2x1x1024x16xf32>) outs(%2955 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2957 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2956 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2958 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2954, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2959 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2958 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2960 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2946, %2957 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2961 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2960, %2959 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2962 = tensor.collapse_shape %2961 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2963 = tensor.expand_shape %2962 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2964 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_253 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2965 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2966 = linalg.matmul ins(%51, %2964 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2965 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2967 = tensor.extract_slice %2966[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2968 = tensor.extract_slice %2966[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %2969 = tensor.expand_shape %2968 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2970 = tensor.expand_shape %2967 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %2971 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2970, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %2972 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2963, %2971 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2973 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2969, %2972 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %2974 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2973 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %2975 = tensor.pad %2974 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %2976 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_255 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %2977 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%2975, %cst_254 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%2976 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %2978 = tensor.collapse_shape %2977 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %2979 = tensor.expand_shape %2978 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %2980 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2981 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2979 : tensor<2x1x1024x16xf32>) outs(%2980 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2982 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2981 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2983 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2979, %2982 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %2984 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2983 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2985 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2986 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2984 : tensor<2x1x1024x16xf32>) outs(%2985 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2987 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2986 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2988 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %2989 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%2979 : tensor<2x1x1024x16xf32>) outs(%2988 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2990 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2989 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2991 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2987, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %2992 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2991 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %2993 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2979, %2990 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2994 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2993, %2992 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %2995 = tensor.collapse_shape %2994 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %2996 = tensor.expand_shape %2995 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %2997 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_256 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %2998 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %2999 = linalg.matmul ins(%51, %2997 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%2998 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3000 = tensor.extract_slice %2999[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3001 = tensor.extract_slice %2999[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3002 = tensor.expand_shape %3001 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3003 = tensor.expand_shape %3000 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3004 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3003, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3005 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%2996, %3004 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3006 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3002, %3005 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3007 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3006 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %3008 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3007, %2941 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3009 = tensor.collapse_shape %3008 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %3010 = tensor.expand_shape %3009 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %3011 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3012 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3010 : tensor<2x1x1024x16xf32>) outs(%3011 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3013 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3012 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3014 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3010, %3013 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %3015 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3014 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3016 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3017 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3015 : tensor<2x1x1024x16xf32>) outs(%3016 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3018 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3017 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3019 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3020 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3010 : tensor<2x1x1024x16xf32>) outs(%3019 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3021 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3020 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3022 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3018, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3023 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3022 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3024 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3010, %3021 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3025 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3024, %3023 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3026 = tensor.collapse_shape %3025 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %3027 = tensor.expand_shape %3026 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %3028 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3027, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3029 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3028, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3030 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_258 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %3031 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3029, %cst_257 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%3030 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %3032 = tensor.collapse_shape %3031 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %3033 = tensor.expand_shape %3032 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %3034 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3033 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %3035 = tensor.extract_slice %3034[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %3036 = tensor.extract_slice %3034[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %3037 = tensor.extract_slice %3034[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %3038 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3035, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %3039 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3036 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %3040 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3039, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %3041 = tensor.collapse_shape %3038 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %3042 = tensor.collapse_shape %3040 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %3043 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %3044 = linalg.batch_matmul ins(%3041, %3042 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%3043 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %3045 = tensor.expand_shape %3044 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %3046 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %3047 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %3048:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3045 : tensor<2x16x16x16xf32>) outs(%3047, %3046 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %3049 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3045, %3048#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %3050 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3049 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %3051 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %3052 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3050 : tensor<2x16x16x16xf32>) outs(%3051 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %3053 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3050, %3052 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %3054 = tensor.collapse_shape %3053 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %3055 = tensor.collapse_shape %3037 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %3056 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %3057 = linalg.batch_matmul ins(%3054, %3055 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%3056 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %3058 = tensor.expand_shape %3057 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %3059 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3058 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %3060 = tensor.collapse_shape %3059 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %3061 = tensor.expand_shape %3060 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %3062 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_260 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %3063 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3061, %cst_259 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%3062 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %3064 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3008, %3063 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3065 = tensor.pad %3064 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %3066 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_262 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %3067 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3065, %cst_261 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%3066 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %3068 = tensor.collapse_shape %3067 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %3069 = tensor.expand_shape %3068 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %3070 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3071 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3069 : tensor<2x1x1024x16xf32>) outs(%3070 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3072 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3071 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3073 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3069, %3072 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %3074 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3073 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3075 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3076 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3074 : tensor<2x1x1024x16xf32>) outs(%3075 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3077 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3076 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3078 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3079 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3069 : tensor<2x1x1024x16xf32>) outs(%3078 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3080 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3079 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3081 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3077, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3082 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3081 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3083 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3069, %3080 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3084 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3083, %3082 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3085 = tensor.collapse_shape %3084 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %3086 = tensor.expand_shape %3085 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %3087 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_263 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3088 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3089 = linalg.matmul ins(%51, %3087 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3088 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3090 = tensor.extract_slice %3089[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3091 = tensor.extract_slice %3089[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3092 = tensor.expand_shape %3091 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3093 = tensor.expand_shape %3090 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3094 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3093, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3095 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3086, %3094 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3096 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3092, %3095 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3097 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3096 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %3098 = tensor.pad %3097 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x4x4xf32> to tensor<2x1024x6x6xf32>
    %3099 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_265 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %3100 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3098, %cst_264 : tensor<2x1024x6x6xf32>, tensor<1024x1024x3x3xf32>) outs(%3099 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %3101 = tensor.collapse_shape %3100 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %3102 = tensor.expand_shape %3101 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %3103 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3104 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3102 : tensor<2x1x1024x16xf32>) outs(%3103 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3105 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3104 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3106 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3102, %3105 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %3107 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3106 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3108 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3109 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3107 : tensor<2x1x1024x16xf32>) outs(%3108 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3110 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3109 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3111 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3112 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3102 : tensor<2x1x1024x16xf32>) outs(%3111 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3113 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3112 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3114 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3110, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3115 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3114 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3116 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3102, %3113 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3117 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3116, %3115 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3118 = tensor.collapse_shape %3117 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %3119 = tensor.expand_shape %3118 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %3120 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_266 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3121 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3122 = linalg.matmul ins(%51, %3120 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3121 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3123 = tensor.extract_slice %3122[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3124 = tensor.extract_slice %3122[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3125 = tensor.expand_shape %3124 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3126 = tensor.expand_shape %3123 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3127 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3126, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3128 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3119, %3127 : tensor<2x1024x4x4xf32>, tensor<2x1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3129 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3125, %3128 : tensor<2x1024x1x1xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3130 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3129 : tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x4x4xf32>
    %3131 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3130, %3064 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3132 = tensor.collapse_shape %3131 [[0], [1], [2, 3]] : tensor<2x1024x4x4xf32> into tensor<2x1024x16xf32>
    %3133 = tensor.expand_shape %3132 [[0], [1, 2], [3]] : tensor<2x1024x16xf32> into tensor<2x1x1024x16xf32>
    %3134 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3135 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3133 : tensor<2x1x1024x16xf32>) outs(%3134 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3136 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3135 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3137 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3133, %3136 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x16xf32>
    %3138 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3137 : tensor<2x1x1024x16xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3139 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3140 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3138 : tensor<2x1x1024x16xf32>) outs(%3139 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3141 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3140 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3142 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3143 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3133 : tensor<2x1x1024x16xf32>) outs(%3142 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3144 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3143 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_523 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3145 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3141, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3146 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3145 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3147 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3133, %3144 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3148 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3147, %3146 : tensor<2x1x1024x16xf32>, tensor<2x1x1x1xf32>) outs(%2202 : tensor<2x1x1024x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x16xf32>
    %3149 = tensor.collapse_shape %3148 [[0], [1, 2], [3]] : tensor<2x1x1024x16xf32> into tensor<2x1024x16xf32>
    %3150 = tensor.expand_shape %3149 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %3151 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3150, %cst_306 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3152 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3151, %cst_307 : tensor<2x1024x4x4xf32>, tensor<1024x1x1xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3153 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_268 : tensor<3072xf32>) outs(%2283 : tensor<2x3072x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x4x4xf32>
    %3154 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3152, %cst_267 : tensor<2x1024x4x4xf32>, tensor<3072x1024x1x1xf32>) outs(%3153 : tensor<2x3072x4x4xf32>) -> tensor<2x3072x4x4xf32>
    %3155 = tensor.collapse_shape %3154 [[0], [1], [2, 3]] : tensor<2x3072x4x4xf32> into tensor<2x3072x16xf32>
    %3156 = tensor.expand_shape %3155 [[0], [1, 2], [3]] : tensor<2x3072x16xf32> into tensor<2x48x64x16xf32>
    %3157 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3156 : tensor<2x48x64x16xf32>) outs(%2288 : tensor<2x48x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x16x64xf32>
    %3158 = tensor.extract_slice %3157[0, 0, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %3159 = tensor.extract_slice %3157[0, 16, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %3160 = tensor.extract_slice %3157[0, 32, 0, 0] [2, 16, 16, 64] [1, 1, 1, 1] : tensor<2x48x16x64xf32> to tensor<2x16x16x64xf32>
    %3161 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3158, %cst_512 : tensor<2x16x16x64xf32>, tensor<f64>) outs(%2293 : tensor<2x16x16x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x16x64xf32>
    %3162 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3159 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %3163 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3162, %cst_512 : tensor<2x16x64x16xf32>, tensor<f64>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x16xf32>
    %3164 = tensor.collapse_shape %3161 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %3165 = tensor.collapse_shape %3163 [[0, 1], [2], [3]] : tensor<2x16x64x16xf32> into tensor<32x64x16xf32>
    %3166 = linalg.fill ins(%cst_514 : f32) outs(%2300 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %3167 = linalg.batch_matmul ins(%3164, %3165 : tensor<32x16x64xf32>, tensor<32x64x16xf32>) outs(%3166 : tensor<32x16x16xf32>) -> tensor<32x16x16xf32>
    %3168 = tensor.expand_shape %3167 [[0, 1], [2], [3]] : tensor<32x16x16xf32> into tensor<2x16x16x16xf32>
    %3169 = linalg.fill ins(%c0_i64 : i64) outs(%2304 : tensor<2x16x16x1xi64>) -> tensor<2x16x16x1xi64>
    %3170 = linalg.fill ins(%cst_515 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %3171:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3168 : tensor<2x16x16x16xf32>) outs(%3170, %3169 : tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x16x1xf32>, tensor<2x16x16x1xi64>)
    %3172 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3168, %3171#0 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x16x16xf32>
    %3173 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3172 : tensor<2x16x16x16xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %3174 = linalg.fill ins(%cst_514 : f32) outs(%2306 : tensor<2x16x16x1xf32>) -> tensor<2x16x16x1xf32>
    %3175 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3173 : tensor<2x16x16x16xf32>) outs(%3174 : tensor<2x16x16x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x1xf32>
    %3176 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3173, %3175 : tensor<2x16x16x16xf32>, tensor<2x16x16x1xf32>) outs(%2309 : tensor<2x16x16x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x16x16xf32>
    %3177 = tensor.collapse_shape %3176 [[0, 1], [2], [3]] : tensor<2x16x16x16xf32> into tensor<32x16x16xf32>
    %3178 = tensor.collapse_shape %3160 [[0, 1], [2], [3]] : tensor<2x16x16x64xf32> into tensor<32x16x64xf32>
    %3179 = linalg.fill ins(%cst_514 : f32) outs(%2317 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %3180 = linalg.batch_matmul ins(%3177, %3178 : tensor<32x16x16xf32>, tensor<32x16x64xf32>) outs(%3179 : tensor<32x16x64xf32>) -> tensor<32x16x64xf32>
    %3181 = tensor.expand_shape %3180 [[0, 1], [2], [3]] : tensor<32x16x64xf32> into tensor<2x16x16x64xf32>
    %3182 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3181 : tensor<2x16x16x64xf32>) outs(%2295 : tensor<2x16x64x16xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x16xf32>
    %3183 = tensor.collapse_shape %3182 [[0], [1, 2], [3]] : tensor<2x16x64x16xf32> into tensor<2x1024x16xf32>
    %3184 = tensor.expand_shape %3183 [[0], [1], [2, 3]] : tensor<2x1024x16xf32> into tensor<2x1024x4x4xf32>
    %3185 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_270 : tensor<1024xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x4x4xf32>
    %3186 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3184, %cst_269 : tensor<2x1024x4x4xf32>, tensor<1024x1024x1x1xf32>) outs(%3185 : tensor<2x1024x4x4xf32>) -> tensor<2x1024x4x4xf32>
    %3187 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3131, %3186 : tensor<2x1024x4x4xf32>, tensor<2x1024x4x4xf32>) outs(%2190 : tensor<2x1024x4x4xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x4x4xf32>
    %3188 = linalg.generic {indexing_maps = [#map16, #map17, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_271, %cst_272 : tensor<8x1xi64>, tensor<8xi64>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: i64, %arg4: i64, %arg5: f32):
      %5340 = linalg.index 0 : index
      %5341 = linalg.index 1 : index
      %5342 = arith.index_cast %arg3 : i64 to index
      %5343 = arith.index_cast %arg4 : i64 to index
      %5344 = tensor.extract %3187[%5340, %5341, %5342, %5343] : tensor<2x1024x4x4xf32>
      linalg.yield %5344 : f32
    } -> tensor<2x1024x8x8xf32>
    %3189 = linalg.generic {indexing_maps = [#map16, #map17, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_273, %cst_272 : tensor<8x1xi64>, tensor<8xi64>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: i64, %arg4: i64, %arg5: f32):
      %5340 = linalg.index 0 : index
      %5341 = linalg.index 1 : index
      %5342 = arith.index_cast %arg3 : i64 to index
      %5343 = arith.index_cast %arg4 : i64 to index
      %5344 = tensor.extract %3187[%5340, %5341, %5342, %5343] : tensor<2x1024x4x4xf32>
      linalg.yield %5344 : f32
    } -> tensor<2x1024x8x8xf32>
    %3190 = linalg.generic {indexing_maps = [#map16, #map17, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_271, %cst_274 : tensor<8x1xi64>, tensor<8xi64>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: i64, %arg4: i64, %arg5: f32):
      %5340 = linalg.index 0 : index
      %5341 = linalg.index 1 : index
      %5342 = arith.index_cast %arg3 : i64 to index
      %5343 = arith.index_cast %arg4 : i64 to index
      %5344 = tensor.extract %3187[%5340, %5341, %5342, %5343] : tensor<2x1024x4x4xf32>
      linalg.yield %5344 : f32
    } -> tensor<2x1024x8x8xf32>
    %3191 = linalg.generic {indexing_maps = [#map16, #map17, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_273, %cst_274 : tensor<8x1xi64>, tensor<8xi64>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: i64, %arg4: i64, %arg5: f32):
      %5340 = linalg.index 0 : index
      %5341 = linalg.index 1 : index
      %5342 = arith.index_cast %arg3 : i64 to index
      %5343 = arith.index_cast %arg4 : i64 to index
      %5344 = tensor.extract %3187[%5340, %5341, %5342, %5343] : tensor<2x1024x4x4xf32>
      linalg.yield %5344 : f32
    } -> tensor<2x1024x8x8xf32>
    %3192 = linalg.generic {indexing_maps = [#map8, #map16, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3188, %cst_275 : tensor<2x1024x8x8xf32>, tensor<8x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3193 = linalg.generic {indexing_maps = [#map8, #map16, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3189, %cst_276 : tensor<2x1024x8x8xf32>, tensor<8x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3194 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3192, %3193 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3195 = linalg.generic {indexing_maps = [#map8, #map16, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3190, %cst_275 : tensor<2x1024x8x8xf32>, tensor<8x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3196 = linalg.generic {indexing_maps = [#map8, #map16, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3191, %cst_276 : tensor<2x1024x8x8xf32>, tensor<8x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3197 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3195, %3196 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3198 = linalg.generic {indexing_maps = [#map8, #map17, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3194, %cst_277 : tensor<2x1024x8x8xf32>, tensor<8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3199 = linalg.generic {indexing_maps = [#map8, #map17, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3197, %cst_278 : tensor<2x1024x8x8xf32>, tensor<8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3200 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3198, %3199 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3201 = linalg.init_tensor [2, 2048, 8, 8] : tensor<2x2048x8x8xf32>
    %3202 = tensor.insert_slice %3200 into %3201[0, 0, 0, 0] [2, 1024, 8, 8] [1, 1, 1, 1] : tensor<2x1024x8x8xf32> into tensor<2x2048x8x8xf32>
    %3203 = tensor.insert_slice %2189 into %3202[0, 1024, 0, 0] [2, 1024, 8, 8] [1, 1, 1, 1] : tensor<2x1024x8x8xf32> into tensor<2x2048x8x8xf32>
    %3204 = tensor.pad %3203 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x2048x8x8xf32> to tensor<2x2048x10x10xf32>
    %3205 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_280 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3206 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3204, %cst_279 : tensor<2x2048x10x10xf32>, tensor<1024x2048x3x3xf32>) outs(%3205 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3207 = tensor.collapse_shape %3206 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3208 = tensor.expand_shape %3207 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3209 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3210 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3208 : tensor<2x1x1024x64xf32>) outs(%3209 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3211 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3210 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3212 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3208, %3211 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3213 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3212 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3214 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3215 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3213 : tensor<2x1x1024x64xf32>) outs(%3214 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3216 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3215 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3217 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3218 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3208 : tensor<2x1x1024x64xf32>) outs(%3217 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3219 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3218 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3220 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3216, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3221 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3220 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3222 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3208, %3219 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3223 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3222, %3221 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3224 = tensor.collapse_shape %3223 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3225 = tensor.expand_shape %3224 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3226 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_281 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3227 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3228 = linalg.matmul ins(%51, %3226 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3227 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3229 = tensor.extract_slice %3228[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3230 = tensor.extract_slice %3228[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3231 = tensor.expand_shape %3230 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3232 = tensor.expand_shape %3229 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3233 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3232, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3234 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3225, %3233 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3235 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3231, %3234 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3236 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3235 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %3237 = tensor.pad %3236 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %3238 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_283 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3239 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3237, %cst_282 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%3238 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3240 = tensor.collapse_shape %3239 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3241 = tensor.expand_shape %3240 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3242 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3243 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3241 : tensor<2x1x1024x64xf32>) outs(%3242 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3244 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3243 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3245 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3241, %3244 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3246 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3245 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3247 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3248 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3246 : tensor<2x1x1024x64xf32>) outs(%3247 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3249 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3248 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3250 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3251 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3241 : tensor<2x1x1024x64xf32>) outs(%3250 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3252 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3251 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3253 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3249, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3254 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3253 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3255 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3241, %3252 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3256 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3255, %3254 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3257 = tensor.collapse_shape %3256 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3258 = tensor.expand_shape %3257 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3259 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_284 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3260 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3261 = linalg.matmul ins(%51, %3259 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3260 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3262 = tensor.extract_slice %3261[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3263 = tensor.extract_slice %3261[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3264 = tensor.expand_shape %3263 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3265 = tensor.expand_shape %3262 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3266 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3265, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3267 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3258, %3266 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3268 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3264, %3267 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3269 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3268 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %3270 = linalg.fill ins(%cst_514 : f32) outs(%1686 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3271 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3203, %cst_285 : tensor<2x2048x8x8xf32>, tensor<1024x2048x1x1xf32>) outs(%3270 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3272 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3269, %3271 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3273 = tensor.collapse_shape %3272 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3274 = tensor.expand_shape %3273 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3275 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3276 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3274 : tensor<2x1x1024x64xf32>) outs(%3275 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3277 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3276 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3278 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3274, %3277 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3279 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3278 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3280 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3281 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3279 : tensor<2x1x1024x64xf32>) outs(%3280 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3282 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3281 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3283 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3284 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3274 : tensor<2x1x1024x64xf32>) outs(%3283 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3285 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3284 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3286 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3282, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3287 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3286 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3288 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3274, %3285 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3289 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3288, %3287 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3290 = tensor.collapse_shape %3289 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3291 = tensor.expand_shape %3290 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3292 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3291, %cst_306 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3293 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3292, %cst_307 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3294 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_287 : tensor<3072xf32>) outs(%1780 : tensor<2x3072x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x8x8xf32>
    %3295 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3293, %cst_286 : tensor<2x1024x8x8xf32>, tensor<3072x1024x1x1xf32>) outs(%3294 : tensor<2x3072x8x8xf32>) -> tensor<2x3072x8x8xf32>
    %3296 = tensor.collapse_shape %3295 [[0], [1], [2, 3]] : tensor<2x3072x8x8xf32> into tensor<2x3072x64xf32>
    %3297 = tensor.expand_shape %3296 [[0], [1, 2], [3]] : tensor<2x3072x64xf32> into tensor<2x48x64x64xf32>
    %3298 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3297 : tensor<2x48x64x64xf32>) outs(%1785 : tensor<2x48x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x64x64xf32>
    %3299 = tensor.extract_slice %3298[0, 0, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %3300 = tensor.extract_slice %3298[0, 16, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %3301 = tensor.extract_slice %3298[0, 32, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %3302 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3299, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %3303 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3300 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %3304 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3303, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %3305 = tensor.collapse_shape %3302 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3306 = tensor.collapse_shape %3304 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3307 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3308 = linalg.batch_matmul ins(%3305, %3306 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%3307 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3309 = tensor.expand_shape %3308 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %3310 = linalg.fill ins(%c0_i64 : i64) outs(%1800 : tensor<2x16x64x1xi64>) -> tensor<2x16x64x1xi64>
    %3311 = linalg.fill ins(%cst_515 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %3312:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3309 : tensor<2x16x64x64xf32>) outs(%3311, %3310 : tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>)
    %3313 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3309, %3312#0 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x64x64xf32>
    %3314 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3313 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %3315 = linalg.fill ins(%cst_514 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %3316 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3314 : tensor<2x16x64x64xf32>) outs(%3315 : tensor<2x16x64x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x1xf32>
    %3317 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3314, %3316 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %3318 = tensor.collapse_shape %3317 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3319 = tensor.collapse_shape %3301 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3320 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3321 = linalg.batch_matmul ins(%3318, %3319 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%3320 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3322 = tensor.expand_shape %3321 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %3323 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3322 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %3324 = tensor.collapse_shape %3323 [[0], [1, 2], [3]] : tensor<2x16x64x64xf32> into tensor<2x1024x64xf32>
    %3325 = tensor.expand_shape %3324 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3326 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_289 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3327 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3325, %cst_288 : tensor<2x1024x8x8xf32>, tensor<1024x1024x1x1xf32>) outs(%3326 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3328 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3272, %3327 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3329 = tensor.pad %3328 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %3330 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_291 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3331 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3329, %cst_290 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%3330 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3332 = tensor.collapse_shape %3331 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3333 = tensor.expand_shape %3332 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3334 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3335 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3333 : tensor<2x1x1024x64xf32>) outs(%3334 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3336 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3335 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3337 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3333, %3336 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3338 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3337 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3339 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3340 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3338 : tensor<2x1x1024x64xf32>) outs(%3339 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3341 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3340 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3342 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3343 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3333 : tensor<2x1x1024x64xf32>) outs(%3342 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3344 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3343 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3345 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3341, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3346 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3345 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3347 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3333, %3344 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3348 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3347, %3346 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3349 = tensor.collapse_shape %3348 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3350 = tensor.expand_shape %3349 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3351 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_292 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3352 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3353 = linalg.matmul ins(%51, %3351 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3352 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3354 = tensor.extract_slice %3353[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3355 = tensor.extract_slice %3353[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3356 = tensor.expand_shape %3355 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3357 = tensor.expand_shape %3354 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3358 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3357, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3359 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3350, %3358 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3360 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3356, %3359 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3361 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3360 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %3362 = tensor.pad %3361 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %3363 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_294 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3364 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3362, %cst_293 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%3363 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3365 = tensor.collapse_shape %3364 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3366 = tensor.expand_shape %3365 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3367 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3368 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3366 : tensor<2x1x1024x64xf32>) outs(%3367 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3369 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3368 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3370 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3366, %3369 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3371 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3370 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3372 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3373 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3371 : tensor<2x1x1024x64xf32>) outs(%3372 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3374 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3373 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3375 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3376 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3366 : tensor<2x1x1024x64xf32>) outs(%3375 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3377 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3376 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3378 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3374, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3379 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3378 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3380 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3366, %3377 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3381 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3380, %3379 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3382 = tensor.collapse_shape %3381 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3383 = tensor.expand_shape %3382 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3384 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_295 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3385 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3386 = linalg.matmul ins(%51, %3384 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3385 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3387 = tensor.extract_slice %3386[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3388 = tensor.extract_slice %3386[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3389 = tensor.expand_shape %3388 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3390 = tensor.expand_shape %3387 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3391 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3390, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf32>
    %3392 = linalg.generic {indexing_maps = [#map8, #map9, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3383, %3391 : tensor<2x1024x8x8xf32>, tensor<2x1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3393 = linalg.generic {indexing_maps = [#map9, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3389, %3392 : tensor<2x1024x1x1xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3394 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3393 : tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.cmpf ugt, %arg3, %cst_514 : f32
      %5341 = arith.select %5340, %arg3, %cst_514 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x8x8xf32>
    %3395 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3394, %3328 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3396 = tensor.collapse_shape %3395 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3397 = tensor.expand_shape %3396 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3398 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3399 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3397 : tensor<2x1x1024x64xf32>) outs(%3398 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3400 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3399 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3401 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3397, %3400 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3402 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3401 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3403 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3404 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3402 : tensor<2x1x1024x64xf32>) outs(%3403 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3405 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3404 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3406 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3407 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3397 : tensor<2x1x1024x64xf32>) outs(%3406 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3408 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3407 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3409 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3405, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3410 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3409 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3411 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3397, %3408 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3412 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3411, %3410 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3413 = tensor.collapse_shape %3412 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3414 = tensor.expand_shape %3413 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3415 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3414, %cst_306 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3416 = linalg.generic {indexing_maps = [#map8, #map13, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3415, %cst_307 : tensor<2x1024x8x8xf32>, tensor<1024x1x1xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3417 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_297 : tensor<3072xf32>) outs(%1780 : tensor<2x3072x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x3072x8x8xf32>
    %3418 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3416, %cst_296 : tensor<2x1024x8x8xf32>, tensor<3072x1024x1x1xf32>) outs(%3417 : tensor<2x3072x8x8xf32>) -> tensor<2x3072x8x8xf32>
    %3419 = tensor.collapse_shape %3418 [[0], [1], [2, 3]] : tensor<2x3072x8x8xf32> into tensor<2x3072x64xf32>
    %3420 = tensor.expand_shape %3419 [[0], [1, 2], [3]] : tensor<2x3072x64xf32> into tensor<2x48x64x64xf32>
    %3421 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3420 : tensor<2x48x64x64xf32>) outs(%1785 : tensor<2x48x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x48x64x64xf32>
    %3422 = tensor.extract_slice %3421[0, 0, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %3423 = tensor.extract_slice %3421[0, 16, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %3424 = tensor.extract_slice %3421[0, 32, 0, 0] [2, 16, 64, 64] [1, 1, 1, 1] : tensor<2x48x64x64xf32> to tensor<2x16x64x64xf32>
    %3425 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3422, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %3426 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3423 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %3427 = linalg.generic {indexing_maps = [#map8, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3426, %cst_512 : tensor<2x16x64x64xf32>, tensor<f64>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.mulf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x16x64x64xf32>
    %3428 = tensor.collapse_shape %3425 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3429 = tensor.collapse_shape %3427 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3430 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3431 = linalg.batch_matmul ins(%3428, %3429 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%3430 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3432 = tensor.expand_shape %3431 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %3433 = linalg.fill ins(%c0_i64 : i64) outs(%1800 : tensor<2x16x64x1xi64>) -> tensor<2x16x64x1xi64>
    %3434 = linalg.fill ins(%cst_515 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %3435:2 = linalg.generic {indexing_maps = [#map8, #map15, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3432 : tensor<2x16x64x64xf32>) outs(%3434, %3433 : tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: i64):
      %5340 = linalg.index 3 : index
      %5341 = arith.index_cast %5340 : index to i64
      %5342 = arith.cmpf ogt, %arg3, %arg4 : f32
      %5343 = arith.select %5342, %arg3, %arg4 : f32
      %5344 = arith.select %5342, %5341, %arg5 : i64
      linalg.yield %5343, %5344 : f32, i64
    } -> (tensor<2x16x64x1xf32>, tensor<2x16x64x1xi64>)
    %3436 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3432, %3435#0 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x16x64x64xf32>
    %3437 = linalg.generic {indexing_maps = [#map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3436 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.exp %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %3438 = linalg.fill ins(%cst_514 : f32) outs(%1802 : tensor<2x16x64x1xf32>) -> tensor<2x16x64x1xf32>
    %3439 = linalg.generic {indexing_maps = [#map8, #map15], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%3437 : tensor<2x16x64x64xf32>) outs(%3438 : tensor<2x16x64x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x1xf32>
    %3440 = linalg.generic {indexing_maps = [#map8, #map15, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3437, %3439 : tensor<2x16x64x64xf32>, tensor<2x16x64x1xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.divf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x16x64x64xf32>
    %3441 = tensor.collapse_shape %3440 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3442 = tensor.collapse_shape %3424 [[0, 1], [2], [3]] : tensor<2x16x64x64xf32> into tensor<32x64x64xf32>
    %3443 = linalg.fill ins(%cst_514 : f32) outs(%1796 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3444 = linalg.batch_matmul ins(%3441, %3442 : tensor<32x64x64xf32>, tensor<32x64x64xf32>) outs(%3443 : tensor<32x64x64xf32>) -> tensor<32x64x64xf32>
    %3445 = tensor.expand_shape %3444 [[0, 1], [2], [3]] : tensor<32x64x64xf32> into tensor<2x16x64x64xf32>
    %3446 = linalg.generic {indexing_maps = [#map8, #map14], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3445 : tensor<2x16x64x64xf32>) outs(%1790 : tensor<2x16x64x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x16x64x64xf32>
    %3447 = tensor.collapse_shape %3446 [[0], [1, 2], [3]] : tensor<2x16x64x64xf32> into tensor<2x1024x64xf32>
    %3448 = tensor.expand_shape %3447 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3449 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_299 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3450 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3448, %cst_298 : tensor<2x1024x8x8xf32>, tensor<1024x1024x1x1xf32>) outs(%3449 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3451 = linalg.generic {indexing_maps = [#map8, #map8, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3395, %3450 : tensor<2x1024x8x8xf32>, tensor<2x1024x8x8xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1024x8x8xf32>
    %3452 = tensor.pad %3451 low[0, 0, 1, 1] high[0, 0, 1, 1] {
    ^bb0(%arg3: index, %arg4: index, %arg5: index, %arg6: index):
      tensor.yield %cst_514 : f32
    } : tensor<2x1024x8x8xf32> to tensor<2x1024x10x10xf32>
    %3453 = linalg.generic {indexing_maps = [#map7, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%cst_301 : tensor<1024xf32>) outs(%1686 : tensor<2x1024x8x8xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<2x1024x8x8xf32>
    %3454 = linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%3452, %cst_300 : tensor<2x1024x10x10xf32>, tensor<1024x1024x3x3xf32>) outs(%3453 : tensor<2x1024x8x8xf32>) -> tensor<2x1024x8x8xf32>
    %3455 = tensor.collapse_shape %3454 [[0], [1], [2, 3]] : tensor<2x1024x8x8xf32> into tensor<2x1024x64xf32>
    %3456 = tensor.expand_shape %3455 [[0], [1, 2], [3]] : tensor<2x1024x64xf32> into tensor<2x1x1024x64xf32>
    %3457 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3458 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3456 : tensor<2x1x1024x64xf32>) outs(%3457 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3459 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3458 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3460 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3456, %3459 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.truncf %cst : f64 to f32
      %5341 = arith.mulf %arg4, %5340 : f32
      %5342 = arith.subf %arg3, %5341 : f32
      linalg.yield %5342 : f32
    } -> tensor<2x1x1024x64xf32>
    %3461 = linalg.generic {indexing_maps = [#map11, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3460 : tensor<2x1x1024x64xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.mulf %arg3, %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3462 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3463 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3461 : tensor<2x1x1024x64xf32>) outs(%3462 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3464 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3463 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3465 = linalg.fill ins(%cst_514 : f32) outs(%76 : tensor<2x1x1x1xf32>) -> tensor<2x1x1x1xf32>
    %3466 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%3456 : tensor<2x1x1024x64xf32>) outs(%3465 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.addf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3467 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3466 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = arith.divf %arg3, %cst_521 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3468 = linalg.generic {indexing_maps = [#map10, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3464, %cst_510 : tensor<2x1x1x1xf32>, tensor<f64>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f64, %arg5: f32):
      %5340 = arith.truncf %arg4 : f64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1x1x1xf32>
    %3469 = linalg.generic {indexing_maps = [#map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3468 : tensor<2x1x1x1xf32>) outs(%76 : tensor<2x1x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      %5340 = math.rsqrt %arg3 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1x1xf32>
    %3470 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3456, %3467 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.subf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3471 = linalg.generic {indexing_maps = [#map11, #map10, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3470, %3469 : tensor<2x1x1024x64xf32>, tensor<2x1x1x1xf32>) outs(%1694 : tensor<2x1x1024x64xf32>) {
    ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
      %5340 = arith.mulf %arg3, %arg4 : f32
      linalg.yield %5340 : f32
    } -> tensor<2x1x1024x64xf32>
    %3472 = tensor.collapse_shape %3471 [[0], [1, 2], [3]] : tensor<2x1x1024x64xf32> into tensor<2x1024x64xf32>
    %3473 = tensor.expand_shape %3472 [[0], [1], [2, 3]] : tensor<2x1024x64xf32> into tensor<2x1024x8x8xf32>
    %3474 = linalg.generic {indexing_maps = [#map0, #map3], iterator_types = ["parallel", "parallel"]} ins(%cst_302 : tensor<2048x1024xf32>) outs(%1709 : tensor<1024x2048xf32>) {
    ^bb0(%arg3: f32, %arg4: f32):
      linalg.yield %arg3 : f32
    } -> tensor<1024x2048xf32>
    %3475 = linalg.fill ins(%cst_514 : f32) outs(%1711 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3476 = linalg.matmul ins(%51, %3474 : tensor<2x1024xf32>, tensor<1024x2048xf32>) outs(%3475 : tensor<2x2048xf32>) -> tensor<2x2048xf32>
    %3477 = tensor.extract_slice %3476[0, 0] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3478 = tensor.extract_slice %3476[0, 1024] [2, 1024] [1, 1] : tensor<2x2048xf32> to tensor<2x1024xf32>
    %3479 = tensor.expand_shape %3478 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3480 = tensor.expand_shape %3477 [[0], [1, 2, 3]] : tensor<2x1024xf32> into tensor<2x1024x1x1xf32>
    %3481 = linalg.generic {indexing_maps = [#map9, #map12, #map8], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%3480, %cst_511 : tensor<2x1024x1x1xf32>, tensor<i64>) outs(%1718 : tensor<2x1024x1x1xf32>) {
    ^bb0(%arg3: f32, %arg4: i64, %arg5: f32):
      %5340 = arith.sitofp %arg4 : i64 to f32
      %5341 = arith.addf %arg3, %5340 : f32
      linalg.yield %5341 : f32
    } -> tensor<2x1024x1x1xf