opt linalg input

// -----// IR Dump After AssignTargetDevicesPass (iree-hal-assign-target-devices) //----- //
#executable_target_embedded_elf_x86_64_ = #hal.executable.target<"llvm-cpu", "embedded-elf-x86_64", {cpu = "generic", cpu_features = "", data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", native_vector_size = 16 : index, target_triple = "x86_64-unknown-unknown-eabi-elf", ukernels = "default"}>
#map = affine_map<(d0, d1, d2) -> (d0, d1)>
#map1 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map2 = affine_map<() -> ()>
#map3 = affine_map<(d0, d1) -> ()>
#map4 = affine_map<(d0, d1) -> (d0, d1)>
#map5 = affine_map<(d0) -> (d0)>
#map6 = affine_map<(d0, d1) -> (d1)>
#map7 = affine_map<(d0, d1) -> (d0, 0)>
#map8 = affine_map<(d0, d1, d2, d3) -> (0, 0, d2, d3)>
#map9 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map10 = affine_map<(d0, d1, d2, d3) -> (0, 0, 0, d3)>
#map11 = affine_map<(d0, d1, d2, d3) -> ()>
#map12 = affine_map<(d0, d1) -> (0, d1)>
#map13 = affine_map<(d0, d1, d2) -> (0, d1, d2)>
#map14 = affine_map<(d0, d1, d2) -> (d0, d1, 0)>
#map15 = affine_map<(d0, d1, d2) -> (0, d1, 0)>
#map16 = affine_map<(d0, d1, d2) -> (d2)>
#map17 = affine_map<(d0, d1) -> (d1, d0)>
#map18 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map19 = affine_map<(d0, d1, d2, d3) -> (0, d1, d2, d3)>
#map20 = affine_map<(d0, d1, d2) -> (d0, d2, d1)>
#device_target_llvm_cpu = #hal.device.target<"llvm-cpu", {executable_targets = [#executable_target_embedded_elf_x86_64_]}>
module attributes {hal.device.targets = [#device_target_llvm_cpu], torch.debug_module_name = "_lambda"} {
  ml_program.global private mutable @global_seed(dense<0> : tensor<i64>) : tensor<i64>
  func.func @forward(%arg0: tensor<1x8xi64>, %arg1: tensor<1x8xi64>) -> tensor<1x8x50272xf32> {
    %cst = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_0 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_1 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_2 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_3 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_4 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_5 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_6 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_7 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_8 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_9 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_10 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_11 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_12 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_13 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_14 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_15 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_16 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_17 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_18 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_19 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_20 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_21 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_22 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_23 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_24 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_25 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_26 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_27 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_28 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_29 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_30 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_31 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_32 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_33 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_34 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_35 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_36 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_37 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_38 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_39 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_40 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_41 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_42 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_43 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_44 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_45 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_46 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_47 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_48 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_49 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_50 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_51 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_52 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_53 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_54 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_55 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_56 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_57 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_58 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_59 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_60 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_61 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_62 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_63 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_64 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_65 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_66 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_67 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_68 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_69 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_70 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_71 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_72 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_73 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_74 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_75 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_76 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_77 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_78 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_79 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_80 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_81 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_82 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_83 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_84 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_85 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_86 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_87 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_88 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_89 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_90 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_91 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_92 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_93 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_94 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_95 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_96 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_97 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_98 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_99 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_100 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_101 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_102 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_103 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_104 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_105 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_106 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_107 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_108 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_109 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_110 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_111 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_112 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_113 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_114 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_115 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_116 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_117 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_118 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_119 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_120 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_121 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_122 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_123 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_124 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_125 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_126 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_127 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_128 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_129 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_130 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_131 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_132 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_133 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_134 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_135 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_136 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_137 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_138 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_139 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_140 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_141 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_142 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_143 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_144 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_145 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_146 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_147 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_148 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_149 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_150 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_151 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_152 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_153 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_154 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_155 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_156 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_157 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_158 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_159 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_160 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_161 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_162 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_163 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_164 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_165 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_166 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_167 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_168 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_169 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_170 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_171 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_172 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_173 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_174 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_175 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_176 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_177 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_178 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_179 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_180 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_181 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_182 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_183 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_184 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_185 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_186 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_187 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_188 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_189 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_190 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_191 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_192 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_193 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_194 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_195 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_196 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_197 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_198 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_199 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_200 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_201 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_202 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_203 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_204 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_205 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_206 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_207 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_208 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_209 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_210 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_211 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_212 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_213 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_214 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_215 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_216 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_217 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_218 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_219 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_220 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_221 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_222 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_223 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_224 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_225 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_226 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_227 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_228 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_229 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_230 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_231 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_232 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_233 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_234 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_235 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_236 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_237 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_238 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_239 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_240 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_241 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_242 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_243 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_244 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_245 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_246 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_247 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_248 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_249 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_250 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_251 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_252 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_253 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_254 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_255 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_256 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_257 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_258 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_259 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_260 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_261 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_262 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_263 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_264 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_265 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_266 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_267 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_268 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_269 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_270 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_271 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_272 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_273 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_274 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_275 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_276 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_277 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_278 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_279 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_280 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_281 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_282 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_283 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_284 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_285 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_286 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_287 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_288 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_289 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_290 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_291 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_292 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_293 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_294 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_295 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_296 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_297 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_298 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_299 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_300 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_301 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_302 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_303 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_304 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_305 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_306 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_307 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_308 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_309 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_310 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_311 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_312 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_313 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_314 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_315 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_316 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_317 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_318 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_319 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_320 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_321 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_322 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_323 = arith.constant dense_resource<__elided__> : tensor<2048x8192xf32>
    %cst_324 = arith.constant dense_resource<__elided__> : tensor<8192xf32>
    %cst_325 = arith.constant dense_resource<__elided__> : tensor<8192x2048xf32>
    %cst_326 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_327 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_328 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_329 = arith.constant dense<-3.40282347E+38> : tensor<f32>
    %cst_330 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_331 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_332 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_333 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_334 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_335 = arith.constant dense_resource<__elided__> : tensor<2048x2048xf32>
    %cst_336 = arith.constant dense_resource<__elided__> : tensor<2048xf32>
    %cst_337 = arith.constant dense<1.000000e+00> : tensor<2048xf32>
    %cst_338 = arith.constant dense_resource<__elided__> : tensor<2050x2048xf32>
    %cst_339 = arith.constant dense_resource<__elided__> : tensor<50272x2048xf32>
    %c50272 = arith.constant 50272 : index
    %cst_340 = arith.constant 0.000000e+00 : f32
    %c2050 = arith.constant 2050 : index
    %cst_341 = arith.constant 0xFF800000 : f32
    %c0_i64 = arith.constant 0 : i64
    %cst_342 = arith.constant 0.000000e+00 : f64
    %cst_343 = arith.constant -3.4028234663852886E+38 : f64
    %c2_i64 = arith.constant 2 : i64
    %cst_344 = arith.constant 1.000000e-05 : f64
    %cst_345 = arith.constant 1.000000e+00 : f32
    %c1_i64 = arith.constant 1 : i64
    %cst_346 = arith.constant 2.048000e+03 : f64
    %cst_347 = arith.constant 2.048000e+03 : f32
    %cst_348 = arith.constant 1.250000e-01 : f32
    %0 = tensor.empty() : tensor<1x8x2048xf32>
    %1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg0 : tensor<1x8xi64>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: i64, %out: f32):
      %1818 = arith.index_cast %in : i64 to index
      %1819 = linalg.index 2 : index
      %1820 = arith.cmpi slt, %1818, %c50272 : index
      cf.assert %1820, "index must be smaller than dim size"
      %1821 = arith.cmpi sge, %in, %c0_i64 : i64
      cf.assert %1821, "index must be larger or equal to 0"
      %extracted = tensor.extract %cst_339[%1818, %1819] : tensor<50272x2048xf32>
      linalg.yield %extracted : f32
    } -> tensor<1x8x2048xf32>
    %2 = tensor.empty() : tensor<f64>
    %3 = linalg.fill ins(%cst_343 : f64) outs(%2 : tensor<f64>) -> tensor<f64>
    %4 = tensor.empty() : tensor<f32>
    %5 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%3 : tensor<f64>) outs(%4 : tensor<f32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<f32>
    %6 = tensor.empty() : tensor<8x8xf32>
    %7 = linalg.generic {indexing_maps = [#map3, #map4], iterator_types = ["parallel", "parallel"]} ins(%5 : tensor<f32>) outs(%6 : tensor<8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8x8xf32>
    %8 = tensor.empty() : tensor<8xi64>
    %9 = linalg.generic {indexing_maps = [#map5], iterator_types = ["parallel"]} outs(%8 : tensor<8xi64>) {
    ^bb0(%out: i64):
      %1818 = linalg.index 0 : index
      %1819 = arith.index_cast %1818 : index to i64
      linalg.yield %1819 : i64
    } -> tensor<8xi64>
    %10 = linalg.generic {indexing_maps = [#map5, #map5], iterator_types = ["parallel"]} ins(%9 : tensor<8xi64>) outs(%8 : tensor<8xi64>) {
    ^bb0(%in: i64, %out: i64):
      %1818 = arith.addi %in, %c1_i64 : i64
      linalg.yield %1818 : i64
    } -> tensor<8xi64>
    %expanded = tensor.expand_shape %10 [[0, 1]] : tensor<8xi64> into tensor<8x1xi64>
    %11 = tensor.empty() : tensor<8x8xi1>
    %12 = linalg.generic {indexing_maps = [#map6, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%9, %expanded : tensor<8xi64>, tensor<8x1xi64>) outs(%11 : tensor<8x8xi1>) {
    ^bb0(%in: i64, %in_712: i64, %out: i1):
      %1818 = arith.cmpi slt, %in, %in_712 : i64
      linalg.yield %1818 : i1
    } -> tensor<8x8xi1>
    %13 = tensor.empty() : tensor<i64>
    %14 = linalg.fill ins(%c0_i64 : i64) outs(%13 : tensor<i64>) -> tensor<i64>
    %15 = linalg.generic {indexing_maps = [#map2, #map2], iterator_types = []} ins(%14 : tensor<i64>) outs(%4 : tensor<f32>) {
    ^bb0(%in: i64, %out: f32):
      %1818 = arith.sitofp %in : i64 to f32
      linalg.yield %1818 : f32
    } -> tensor<f32>
    %16 = linalg.generic {indexing_maps = [#map4, #map3, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%12, %15, %7 : tensor<8x8xi1>, tensor<f32>, tensor<8x8xf32>) outs(%6 : tensor<8x8xf32>) {
    ^bb0(%in: i1, %in_712: f32, %in_713: f32, %out: f32):
      %1818 = arith.select %in, %in_712, %in_713 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8xf32>
    %expanded_349 = tensor.expand_shape %16 [[0, 1, 2], [3]] : tensor<8x8xf32> into tensor<1x1x8x8xf32>
    %17 = tensor.empty() : tensor<1x1x8x8xf32>
    %18 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_349 : tensor<1x1x8x8xf32>) outs(%17 : tensor<1x1x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x1x8x8xf32>
    %expanded_350 = tensor.expand_shape %arg1 [[0], [1, 2, 3]] : tensor<1x8xi64> into tensor<1x1x1x8xi64>
    %19 = tensor.empty() : tensor<1x1x8x8xi64>
    %20 = linalg.generic {indexing_maps = [#map10, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_350 : tensor<1x1x1x8xi64>) outs(%19 : tensor<1x1x8x8xi64>) {
    ^bb0(%in: i64, %out: i64):
      linalg.yield %in : i64
    } -> tensor<1x1x8x8xi64>
    %21 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%20 : tensor<1x1x8x8xi64>) outs(%17 : tensor<1x1x8x8xf32>) {
    ^bb0(%in: i64, %out: f32):
      %1818 = arith.sitofp %in : i64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x1x8x8xf32>
    %22 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%21 : tensor<1x1x8x8xf32>) outs(%17 : tensor<1x1x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x1x8x8xf32>
    %23 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%22 : tensor<1x1x8x8xf32>) outs(%17 : tensor<1x1x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.subf %cst_345, %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x1x8x8xf32>
    %24 = tensor.empty() : tensor<1x1x8x8xi1>
    %25 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%23 : tensor<1x1x8x8xf32>) outs(%24 : tensor<1x1x8x8xi1>) {
    ^bb0(%in: f32, %out: i1):
      %1818 = arith.cmpf une, %in, %cst_340 : f32
      linalg.yield %1818 : i1
    } -> tensor<1x1x8x8xi1>
    %26 = linalg.generic {indexing_maps = [#map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%25 : tensor<1x1x8x8xi1>) outs(%24 : tensor<1x1x8x8xi1>) {
    ^bb0(%in: i1, %out: i1):
      linalg.yield %in : i1
    } -> tensor<1x1x8x8xi1>
    %27 = linalg.generic {indexing_maps = [#map8, #map11, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%26, %5, %23 : tensor<1x1x8x8xi1>, tensor<f32>, tensor<1x1x8x8xf32>) outs(%17 : tensor<1x1x8x8xf32>) {
    ^bb0(%in: i1, %in_712: f32, %in_713: f32, %out: f32):
      %1818 = arith.select %in, %in_712, %in_713 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x1x8x8xf32>
    %28 = linalg.generic {indexing_maps = [#map8, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%27, %18 : tensor<1x1x8x8xf32>, tensor<1x1x8x8xf32>) outs(%17 : tensor<1x1x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x1x8x8xf32>
    %29 = tensor.empty() : tensor<1x8xi64>
    %30 = linalg.fill ins(%c0_i64 : i64) outs(%29 : tensor<1x8xi64>) -> tensor<1x8xi64>
    %31 = tensor.empty() : tensor<1xi64>
    %32 = linalg.fill ins(%c0_i64 : i64) outs(%31 : tensor<1xi64>) -> tensor<1xi64>
    %33:2 = tm_tensor.scan dimension(1) inclusive(true) ins(%arg1 : tensor<1x8xi64>) outs(%30, %32 : tensor<1x8xi64>, tensor<1xi64>) {
    ^bb0(%arg2: i64, %arg3: i64):
      %1818 = arith.addi %arg2, %arg3 : i64
      tm_tensor.yield %1818 : i64
    } -> tensor<1x8xi64>, tensor<1xi64>
    %34 = linalg.generic {indexing_maps = [#map12, #map12, #map4], iterator_types = ["parallel", "parallel"]} ins(%33#0, %arg1 : tensor<1x8xi64>, tensor<1x8xi64>) outs(%29 : tensor<1x8xi64>) {
    ^bb0(%in: i64, %in_712: i64, %out: i64):
      %1818 = arith.muli %in, %in_712 : i64
      linalg.yield %1818 : i64
    } -> tensor<1x8xi64>
    %35 = linalg.generic {indexing_maps = [#map12, #map4], iterator_types = ["parallel", "parallel"]} ins(%34 : tensor<1x8xi64>) outs(%29 : tensor<1x8xi64>) {
    ^bb0(%in: i64, %out: i64):
      %1818 = arith.subi %in, %c1_i64 : i64
      linalg.yield %1818 : i64
    } -> tensor<1x8xi64>
    %36 = linalg.generic {indexing_maps = [#map12, #map4], iterator_types = ["parallel", "parallel"]} ins(%35 : tensor<1x8xi64>) outs(%29 : tensor<1x8xi64>) {
    ^bb0(%in: i64, %out: i64):
      %1818 = arith.addi %in, %c2_i64 : i64
      linalg.yield %1818 : i64
    } -> tensor<1x8xi64>
    %37 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%36 : tensor<1x8xi64>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: i64, %out: f32):
      %1818 = arith.index_cast %in : i64 to index
      %1819 = linalg.index 2 : index
      %1820 = arith.cmpi slt, %1818, %c2050 : index
      cf.assert %1820, "index must be smaller than dim size"
      %1821 = arith.cmpi sge, %in, %c0_i64 : i64
      cf.assert %1821, "index must be larger or equal to 0"
      %extracted = tensor.extract %cst_338[%1818, %1819] : tensor<2050x2048xf32>
      linalg.yield %extracted : f32
    } -> tensor<1x8x2048xf32>
    %38 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1, %37 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %39 = tensor.empty() : tensor<1x8x2048xf64>
    %40 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%38 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %41 = tensor.empty() : tensor<1x8x1xf64>
    %42 = linalg.fill ins(%cst_342 : f64) outs(%41 : tensor<1x8x1xf64>) -> tensor<1x8x1xf64>
    %43 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%40 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %44 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%43 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %45 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%40, %44 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %46 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%45, %45 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %47 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%46 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %48 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%47 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %49 = tensor.empty() : tensor<1x8x1xf32>
    %50 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%48 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %51 = linalg.fill ins(%cst_340 : f32) outs(%49 : tensor<1x8x1xf32>) -> tensor<1x8x1xf32>
    %52 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%38 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %53 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%52 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %54 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%50 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %55 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%54 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %56 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%38, %53 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %57 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%56, %55 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %58 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%57, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %59 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%58, %cst_336 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %60 = tensor.empty() : tensor<2048x2048xf32>
    %61 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_335 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed = tensor.collapse_shape %59 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %62 = tensor.empty() : tensor<8x2048xf32>
    %63 = linalg.fill ins(%cst_340 : f32) outs(%62 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %64 = linalg.matmul ins(%collapsed, %61 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %65 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_334, %64 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_351 = tensor.expand_shape %65 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %66 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_351 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %67 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_333 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %68 = linalg.matmul ins(%collapsed, %67 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %69 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_332, %68 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_352 = tensor.expand_shape %69 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %70 = tensor.empty() : tensor<1x32x8x64xf32>
    %71 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_352 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %72 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%71 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %73 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_331 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %74 = linalg.matmul ins(%collapsed, %73 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %75 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_330, %74 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_353 = tensor.expand_shape %75 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %76 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_353 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %77 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%76 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_354 = tensor.expand_shape %66 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %78 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_354 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %79 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%78 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_355 = tensor.collapse_shape %79 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_356 = tensor.collapse_shape %72 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_357 = tensor.collapse_shape %77 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %80 = tensor.empty() : tensor<32x64x8xf32>
    %81 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_356 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %82 = tensor.empty() : tensor<32x8x8xf32>
    %83 = linalg.fill ins(%cst_340 : f32) outs(%82 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %84 = linalg.batch_matmul ins(%collapsed_355, %81 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_358 = tensor.expand_shape %84 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %85 = tensor.empty() : tensor<1x32x8x8xf32>
    %86 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_358, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %87 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%86, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_359 = tensor.collapse_shape %87 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %88 = tensor.empty() : tensor<32x8x1xi64>
    %89 = linalg.fill ins(%c0_i64 : i64) outs(%88 : tensor<32x8x1xi64>) -> tensor<32x8x1xi64>
    %90 = tensor.empty() : tensor<32x8x1xf32>
    %91 = linalg.fill ins(%cst_341 : f32) outs(%90 : tensor<32x8x1xf32>) -> tensor<32x8x1xf32>
    %92:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_359 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %93 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_359, %92#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %94 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%93 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %95 = linalg.fill ins(%cst_340 : f32) outs(%90 : tensor<32x8x1xf32>) -> tensor<32x8x1xf32>
    %96 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%94 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %97 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%94, %96 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %98 = tensor.empty() : tensor<32x8x64xf32>
    %99 = linalg.fill ins(%cst_340 : f32) outs(%98 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %100 = linalg.batch_matmul ins(%97, %collapsed_357 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_360 = tensor.expand_shape %100 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %101 = tensor.empty() : tensor<1x8x32x64xf32>
    %102 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_360 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %103 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%102 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %104 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_328 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_361 = tensor.collapse_shape %103 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %105 = linalg.matmul ins(%collapsed_361, %104 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %106 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_327, %105 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_362 = tensor.expand_shape %106 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %107 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%38, %expanded_362 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_363 = tensor.collapse_shape %107 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %108 = tensor.empty() : tensor<8x2048xf64>
    %109 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_363 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %110 = tensor.empty() : tensor<8x1xf64>
    %111 = linalg.fill ins(%cst_342 : f64) outs(%110 : tensor<8x1xf64>) -> tensor<8x1xf64>
    %112 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%109 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %113 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%112 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %114 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%109, %113 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %115 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%114, %114 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %116 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%115 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %117 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%116 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %118 = tensor.empty() : tensor<8x1xf32>
    %119 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%117 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %120 = linalg.fill ins(%cst_340 : f32) outs(%118 : tensor<8x1xf32>) -> tensor<8x1xf32>
    %121 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_363 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %122 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%121 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %123 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%119 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %124 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%123 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %125 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_363, %122 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %126 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%125, %124 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %127 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%126, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %128 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%127, %cst_326 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %129 = tensor.empty() : tensor<2048x8192xf32>
    %130 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_325 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %131 = tensor.empty() : tensor<8x8192xf32>
    %132 = linalg.fill ins(%cst_340 : f32) outs(%131 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %133 = linalg.matmul ins(%128, %130 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %134 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_324, %133 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %135 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%134 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %136 = tensor.empty() : tensor<8192x2048xf32>
    %137 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_323 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %138 = linalg.matmul ins(%135, %137 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %139 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_322, %138 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %140 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_363, %139 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_364 = tensor.expand_shape %140 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %141 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_364 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %142 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%141 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %143 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%142 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %144 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%141, %143 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %145 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%144, %144 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %146 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%145 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %147 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%146 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %148 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%147 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %149 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_364 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %150 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%149 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %151 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%148 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %152 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%151 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %153 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_364, %150 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %154 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%153, %152 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %155 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%154, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %156 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%155, %cst_321 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %157 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_320 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_365 = tensor.collapse_shape %156 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %158 = linalg.matmul ins(%collapsed_365, %157 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %159 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_319, %158 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_366 = tensor.expand_shape %159 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %160 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_366 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %161 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_318 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %162 = linalg.matmul ins(%collapsed_365, %161 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %163 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_317, %162 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_367 = tensor.expand_shape %163 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %164 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_367 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %165 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%164 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %166 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_316 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %167 = linalg.matmul ins(%collapsed_365, %166 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %168 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_315, %167 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_368 = tensor.expand_shape %168 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %169 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_368 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %170 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%169 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_369 = tensor.expand_shape %160 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %171 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_369 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %172 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%171 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_370 = tensor.collapse_shape %172 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_371 = tensor.collapse_shape %165 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_372 = tensor.collapse_shape %170 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %173 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_371 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %174 = linalg.batch_matmul ins(%collapsed_370, %173 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_373 = tensor.expand_shape %174 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %175 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_373, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %176 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%175, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_374 = tensor.collapse_shape %176 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %177:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_374 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %178 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_374, %177#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %179 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%178 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %180 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%179 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %181 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%179, %180 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %182 = linalg.batch_matmul ins(%181, %collapsed_372 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_375 = tensor.expand_shape %182 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %183 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_375 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %184 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%183 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %185 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_314 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_376 = tensor.collapse_shape %184 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %186 = linalg.matmul ins(%collapsed_376, %185 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %187 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_313, %186 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_377 = tensor.expand_shape %187 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %188 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_364, %expanded_377 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_378 = tensor.collapse_shape %188 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %189 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_378 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %190 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%189 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %191 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%190 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %192 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%189, %191 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %193 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%192, %192 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %194 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%193 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %195 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%194 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %196 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%195 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %197 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_378 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %198 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%197 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %199 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%196 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %200 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%199 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %201 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_378, %198 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %202 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%201, %200 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %203 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%202, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %204 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%203, %cst_312 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %205 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_311 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %206 = linalg.matmul ins(%204, %205 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %207 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_310, %206 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %208 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%207 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %209 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_309 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %210 = linalg.matmul ins(%208, %209 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %211 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_308, %210 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %212 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_378, %211 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_379 = tensor.expand_shape %212 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %213 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_379 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %214 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%213 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %215 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%214 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %216 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%213, %215 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %217 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%216, %216 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %218 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%217 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %219 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%218 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %220 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%219 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %221 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_379 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %222 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%221 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %223 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%220 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %224 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%223 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %225 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_379, %222 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %226 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%225, %224 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %227 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%226, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %228 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%227, %cst_307 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %229 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_306 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_380 = tensor.collapse_shape %228 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %230 = linalg.matmul ins(%collapsed_380, %229 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %231 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_305, %230 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_381 = tensor.expand_shape %231 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %232 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_381 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %233 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_304 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %234 = linalg.matmul ins(%collapsed_380, %233 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %235 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_303, %234 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_382 = tensor.expand_shape %235 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %236 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_382 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %237 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%236 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %238 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_302 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %239 = linalg.matmul ins(%collapsed_380, %238 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %240 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_301, %239 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_383 = tensor.expand_shape %240 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %241 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_383 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %242 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%241 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_384 = tensor.expand_shape %232 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %243 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_384 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %244 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%243 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_385 = tensor.collapse_shape %244 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_386 = tensor.collapse_shape %237 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_387 = tensor.collapse_shape %242 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %245 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_386 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %246 = linalg.batch_matmul ins(%collapsed_385, %245 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_388 = tensor.expand_shape %246 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %247 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_388, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %248 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%247, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_389 = tensor.collapse_shape %248 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %249:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_389 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %250 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_389, %249#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %251 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%250 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %252 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%251 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %253 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%251, %252 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %254 = linalg.batch_matmul ins(%253, %collapsed_387 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_390 = tensor.expand_shape %254 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %255 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_390 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %256 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%255 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %257 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_300 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_391 = tensor.collapse_shape %256 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %258 = linalg.matmul ins(%collapsed_391, %257 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %259 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_299, %258 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_392 = tensor.expand_shape %259 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %260 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_379, %expanded_392 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_393 = tensor.collapse_shape %260 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %261 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_393 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %262 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%261 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %263 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%262 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %264 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%261, %263 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %265 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%264, %264 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %266 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%265 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %267 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%266 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %268 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%267 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %269 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_393 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %270 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%269 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %271 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%268 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %272 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%271 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %273 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_393, %270 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %274 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%273, %272 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %275 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%274, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %276 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%275, %cst_298 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %277 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_297 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %278 = linalg.matmul ins(%276, %277 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %279 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_296, %278 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %280 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%279 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %281 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_295 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %282 = linalg.matmul ins(%280, %281 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %283 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_294, %282 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %284 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_393, %283 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_394 = tensor.expand_shape %284 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %285 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_394 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %286 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%285 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %287 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%286 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %288 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%285, %287 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %289 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%288, %288 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %290 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%289 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %291 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%290 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %292 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%291 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %293 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_394 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %294 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%293 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %295 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%292 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %296 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%295 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %297 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_394, %294 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %298 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%297, %296 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %299 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%298, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %300 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%299, %cst_293 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %301 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_292 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_395 = tensor.collapse_shape %300 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %302 = linalg.matmul ins(%collapsed_395, %301 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %303 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_291, %302 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_396 = tensor.expand_shape %303 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %304 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_396 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %305 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_290 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %306 = linalg.matmul ins(%collapsed_395, %305 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %307 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_289, %306 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_397 = tensor.expand_shape %307 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %308 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_397 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %309 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%308 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %310 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_288 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %311 = linalg.matmul ins(%collapsed_395, %310 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %312 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_287, %311 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_398 = tensor.expand_shape %312 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %313 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_398 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %314 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%313 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_399 = tensor.expand_shape %304 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %315 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_399 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %316 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%315 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_400 = tensor.collapse_shape %316 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_401 = tensor.collapse_shape %309 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_402 = tensor.collapse_shape %314 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %317 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_401 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %318 = linalg.batch_matmul ins(%collapsed_400, %317 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_403 = tensor.expand_shape %318 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %319 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_403, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %320 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%319, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_404 = tensor.collapse_shape %320 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %321:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_404 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %322 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_404, %321#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %323 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%322 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %324 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%323 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %325 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%323, %324 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %326 = linalg.batch_matmul ins(%325, %collapsed_402 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_405 = tensor.expand_shape %326 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %327 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_405 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %328 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%327 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %329 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_286 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_406 = tensor.collapse_shape %328 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %330 = linalg.matmul ins(%collapsed_406, %329 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %331 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_285, %330 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_407 = tensor.expand_shape %331 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %332 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_394, %expanded_407 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_408 = tensor.collapse_shape %332 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %333 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_408 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %334 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%333 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %335 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%334 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %336 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%333, %335 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %337 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%336, %336 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %338 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%337 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %339 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%338 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %340 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%339 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %341 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_408 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %342 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%341 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %343 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%340 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %344 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%343 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %345 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_408, %342 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %346 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%345, %344 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %347 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%346, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %348 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%347, %cst_284 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %349 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_283 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %350 = linalg.matmul ins(%348, %349 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %351 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_282, %350 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %352 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%351 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %353 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_281 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %354 = linalg.matmul ins(%352, %353 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %355 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_280, %354 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %356 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_408, %355 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_409 = tensor.expand_shape %356 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %357 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_409 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %358 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%357 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %359 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%358 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %360 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%357, %359 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %361 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%360, %360 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %362 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%361 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %363 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%362 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %364 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%363 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %365 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_409 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %366 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%365 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %367 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%364 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %368 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%367 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %369 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_409, %366 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %370 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%369, %368 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %371 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%370, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %372 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%371, %cst_279 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %373 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_278 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_410 = tensor.collapse_shape %372 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %374 = linalg.matmul ins(%collapsed_410, %373 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %375 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_277, %374 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_411 = tensor.expand_shape %375 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %376 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_411 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %377 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_276 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %378 = linalg.matmul ins(%collapsed_410, %377 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %379 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_275, %378 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_412 = tensor.expand_shape %379 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %380 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_412 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %381 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%380 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %382 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_274 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %383 = linalg.matmul ins(%collapsed_410, %382 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %384 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_273, %383 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_413 = tensor.expand_shape %384 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %385 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_413 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %386 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%385 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_414 = tensor.expand_shape %376 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %387 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_414 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %388 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%387 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_415 = tensor.collapse_shape %388 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_416 = tensor.collapse_shape %381 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_417 = tensor.collapse_shape %386 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %389 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_416 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %390 = linalg.batch_matmul ins(%collapsed_415, %389 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_418 = tensor.expand_shape %390 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %391 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_418, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %392 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%391, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_419 = tensor.collapse_shape %392 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %393:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_419 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %394 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_419, %393#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %395 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%394 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %396 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%395 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %397 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%395, %396 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %398 = linalg.batch_matmul ins(%397, %collapsed_417 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_420 = tensor.expand_shape %398 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %399 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_420 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %400 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%399 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %401 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_272 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_421 = tensor.collapse_shape %400 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %402 = linalg.matmul ins(%collapsed_421, %401 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %403 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_271, %402 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_422 = tensor.expand_shape %403 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %404 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_409, %expanded_422 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_423 = tensor.collapse_shape %404 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %405 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_423 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %406 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%405 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %407 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%406 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %408 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%405, %407 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %409 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%408, %408 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %410 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%409 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %411 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%410 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %412 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%411 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %413 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_423 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %414 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%413 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %415 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%412 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %416 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%415 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %417 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_423, %414 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %418 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%417, %416 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %419 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%418, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %420 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%419, %cst_270 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %421 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_269 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %422 = linalg.matmul ins(%420, %421 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %423 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_268, %422 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %424 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%423 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %425 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_267 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %426 = linalg.matmul ins(%424, %425 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %427 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_266, %426 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %428 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_423, %427 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_424 = tensor.expand_shape %428 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %429 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_424 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %430 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%429 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %431 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%430 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %432 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%429, %431 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %433 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%432, %432 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %434 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%433 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %435 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%434 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %436 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%435 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %437 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_424 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %438 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%437 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %439 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%436 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %440 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%439 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %441 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_424, %438 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %442 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%441, %440 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %443 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%442, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %444 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%443, %cst_265 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %445 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_264 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_425 = tensor.collapse_shape %444 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %446 = linalg.matmul ins(%collapsed_425, %445 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %447 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_263, %446 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_426 = tensor.expand_shape %447 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %448 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_426 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %449 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_262 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %450 = linalg.matmul ins(%collapsed_425, %449 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %451 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_261, %450 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_427 = tensor.expand_shape %451 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %452 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_427 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %453 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%452 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %454 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_260 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %455 = linalg.matmul ins(%collapsed_425, %454 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %456 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_259, %455 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_428 = tensor.expand_shape %456 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %457 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_428 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %458 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%457 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_429 = tensor.expand_shape %448 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %459 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_429 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %460 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%459 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_430 = tensor.collapse_shape %460 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_431 = tensor.collapse_shape %453 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_432 = tensor.collapse_shape %458 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %461 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_431 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %462 = linalg.batch_matmul ins(%collapsed_430, %461 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_433 = tensor.expand_shape %462 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %463 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_433, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %464 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%463, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_434 = tensor.collapse_shape %464 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %465:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_434 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %466 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_434, %465#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %467 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%466 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %468 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%467 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %469 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%467, %468 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %470 = linalg.batch_matmul ins(%469, %collapsed_432 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_435 = tensor.expand_shape %470 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %471 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_435 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %472 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%471 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %473 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_258 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_436 = tensor.collapse_shape %472 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %474 = linalg.matmul ins(%collapsed_436, %473 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %475 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_257, %474 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_437 = tensor.expand_shape %475 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %476 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_424, %expanded_437 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_438 = tensor.collapse_shape %476 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %477 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_438 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %478 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%477 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %479 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%478 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %480 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%477, %479 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %481 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%480, %480 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %482 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%481 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %483 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%482 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %484 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%483 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %485 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_438 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %486 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%485 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %487 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%484 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %488 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%487 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %489 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_438, %486 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %490 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%489, %488 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %491 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%490, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %492 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%491, %cst_256 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %493 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_255 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %494 = linalg.matmul ins(%492, %493 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %495 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_254, %494 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %496 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%495 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %497 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_253 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %498 = linalg.matmul ins(%496, %497 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %499 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_252, %498 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %500 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_438, %499 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_439 = tensor.expand_shape %500 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %501 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_439 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %502 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%501 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %503 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%502 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %504 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%501, %503 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %505 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%504, %504 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %506 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%505 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %507 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%506 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %508 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%507 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %509 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_439 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %510 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%509 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %511 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%508 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %512 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%511 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %513 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_439, %510 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %514 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%513, %512 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %515 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%514, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %516 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%515, %cst_251 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %517 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_250 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_440 = tensor.collapse_shape %516 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %518 = linalg.matmul ins(%collapsed_440, %517 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %519 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_249, %518 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_441 = tensor.expand_shape %519 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %520 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_441 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %521 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_248 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %522 = linalg.matmul ins(%collapsed_440, %521 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %523 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_247, %522 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_442 = tensor.expand_shape %523 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %524 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_442 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %525 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%524 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %526 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_246 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %527 = linalg.matmul ins(%collapsed_440, %526 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %528 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_245, %527 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_443 = tensor.expand_shape %528 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %529 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_443 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %530 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%529 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_444 = tensor.expand_shape %520 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %531 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_444 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %532 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%531 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_445 = tensor.collapse_shape %532 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_446 = tensor.collapse_shape %525 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_447 = tensor.collapse_shape %530 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %533 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_446 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %534 = linalg.batch_matmul ins(%collapsed_445, %533 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_448 = tensor.expand_shape %534 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %535 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_448, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %536 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%535, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_449 = tensor.collapse_shape %536 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %537:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_449 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %538 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_449, %537#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %539 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%538 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %540 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%539 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %541 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%539, %540 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %542 = linalg.batch_matmul ins(%541, %collapsed_447 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_450 = tensor.expand_shape %542 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %543 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_450 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %544 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%543 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %545 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_244 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_451 = tensor.collapse_shape %544 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %546 = linalg.matmul ins(%collapsed_451, %545 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %547 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_243, %546 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_452 = tensor.expand_shape %547 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %548 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_439, %expanded_452 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_453 = tensor.collapse_shape %548 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %549 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_453 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %550 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%549 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %551 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%550 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %552 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%549, %551 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %553 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%552, %552 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %554 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%553 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %555 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%554 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %556 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%555 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %557 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_453 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %558 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%557 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %559 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%556 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %560 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%559 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %561 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_453, %558 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %562 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%561, %560 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %563 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%562, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %564 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%563, %cst_242 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %565 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_241 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %566 = linalg.matmul ins(%564, %565 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %567 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_240, %566 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %568 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%567 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %569 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_239 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %570 = linalg.matmul ins(%568, %569 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %571 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_238, %570 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %572 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_453, %571 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_454 = tensor.expand_shape %572 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %573 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_454 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %574 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%573 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %575 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%574 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %576 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%573, %575 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %577 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%576, %576 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %578 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%577 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %579 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%578 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %580 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%579 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %581 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_454 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %582 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%581 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %583 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%580 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %584 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%583 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %585 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_454, %582 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %586 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%585, %584 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %587 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%586, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %588 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%587, %cst_237 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %589 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_236 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_455 = tensor.collapse_shape %588 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %590 = linalg.matmul ins(%collapsed_455, %589 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %591 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_235, %590 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_456 = tensor.expand_shape %591 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %592 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_456 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %593 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_234 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %594 = linalg.matmul ins(%collapsed_455, %593 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %595 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_233, %594 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_457 = tensor.expand_shape %595 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %596 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_457 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %597 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%596 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %598 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_232 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %599 = linalg.matmul ins(%collapsed_455, %598 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %600 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_231, %599 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_458 = tensor.expand_shape %600 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %601 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_458 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %602 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%601 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_459 = tensor.expand_shape %592 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %603 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_459 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %604 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%603 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_460 = tensor.collapse_shape %604 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_461 = tensor.collapse_shape %597 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_462 = tensor.collapse_shape %602 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %605 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_461 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %606 = linalg.batch_matmul ins(%collapsed_460, %605 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_463 = tensor.expand_shape %606 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %607 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_463, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %608 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%607, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_464 = tensor.collapse_shape %608 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %609:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_464 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %610 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_464, %609#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %611 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%610 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %612 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%611 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %613 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%611, %612 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %614 = linalg.batch_matmul ins(%613, %collapsed_462 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_465 = tensor.expand_shape %614 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %615 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_465 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %616 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%615 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %617 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_230 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_466 = tensor.collapse_shape %616 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %618 = linalg.matmul ins(%collapsed_466, %617 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %619 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_229, %618 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_467 = tensor.expand_shape %619 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %620 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_454, %expanded_467 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_468 = tensor.collapse_shape %620 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %621 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_468 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %622 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%621 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %623 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%622 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %624 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%621, %623 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %625 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%624, %624 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %626 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%625 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %627 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%626 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %628 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%627 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %629 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_468 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %630 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%629 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %631 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%628 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %632 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%631 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %633 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_468, %630 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %634 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%633, %632 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %635 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%634, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %636 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%635, %cst_228 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %637 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_227 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %638 = linalg.matmul ins(%636, %637 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %639 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_226, %638 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %640 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%639 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %641 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_225 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %642 = linalg.matmul ins(%640, %641 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %643 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_224, %642 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %644 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_468, %643 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_469 = tensor.expand_shape %644 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %645 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_469 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %646 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%645 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %647 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%646 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %648 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%645, %647 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %649 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%648, %648 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %650 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%649 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %651 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%650 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %652 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%651 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %653 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_469 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %654 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%653 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %655 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%652 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %656 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%655 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %657 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_469, %654 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %658 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%657, %656 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %659 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%658, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %660 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%659, %cst_223 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %661 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_222 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_470 = tensor.collapse_shape %660 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %662 = linalg.matmul ins(%collapsed_470, %661 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %663 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_221, %662 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_471 = tensor.expand_shape %663 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %664 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_471 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %665 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_220 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %666 = linalg.matmul ins(%collapsed_470, %665 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %667 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_219, %666 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_472 = tensor.expand_shape %667 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %668 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_472 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %669 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%668 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %670 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_218 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %671 = linalg.matmul ins(%collapsed_470, %670 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %672 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_217, %671 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_473 = tensor.expand_shape %672 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %673 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_473 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %674 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%673 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_474 = tensor.expand_shape %664 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %675 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_474 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %676 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%675 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_475 = tensor.collapse_shape %676 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_476 = tensor.collapse_shape %669 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_477 = tensor.collapse_shape %674 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %677 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_476 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %678 = linalg.batch_matmul ins(%collapsed_475, %677 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_478 = tensor.expand_shape %678 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %679 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_478, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %680 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%679, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_479 = tensor.collapse_shape %680 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %681:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_479 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %682 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_479, %681#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %683 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%682 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %684 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%683 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %685 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%683, %684 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %686 = linalg.batch_matmul ins(%685, %collapsed_477 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_480 = tensor.expand_shape %686 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %687 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_480 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %688 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%687 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %689 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_216 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_481 = tensor.collapse_shape %688 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %690 = linalg.matmul ins(%collapsed_481, %689 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %691 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_215, %690 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_482 = tensor.expand_shape %691 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %692 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_469, %expanded_482 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_483 = tensor.collapse_shape %692 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %693 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_483 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %694 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%693 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %695 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%694 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %696 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%693, %695 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %697 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%696, %696 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %698 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%697 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %699 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%698 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %700 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%699 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %701 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_483 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %702 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%701 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %703 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%700 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %704 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%703 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %705 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_483, %702 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %706 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%705, %704 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %707 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%706, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %708 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%707, %cst_214 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %709 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_213 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %710 = linalg.matmul ins(%708, %709 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %711 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_212, %710 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %712 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%711 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %713 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_211 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %714 = linalg.matmul ins(%712, %713 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %715 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_210, %714 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %716 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_483, %715 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_484 = tensor.expand_shape %716 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %717 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_484 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %718 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%717 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %719 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%718 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %720 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%717, %719 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %721 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%720, %720 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %722 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%721 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %723 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%722 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %724 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%723 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %725 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_484 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %726 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%725 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %727 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%724 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %728 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%727 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %729 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_484, %726 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %730 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%729, %728 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %731 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%730, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %732 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%731, %cst_209 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %733 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_208 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_485 = tensor.collapse_shape %732 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %734 = linalg.matmul ins(%collapsed_485, %733 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %735 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_207, %734 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_486 = tensor.expand_shape %735 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %736 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_486 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %737 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_206 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %738 = linalg.matmul ins(%collapsed_485, %737 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %739 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_205, %738 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_487 = tensor.expand_shape %739 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %740 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_487 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %741 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%740 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %742 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_204 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %743 = linalg.matmul ins(%collapsed_485, %742 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %744 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_203, %743 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_488 = tensor.expand_shape %744 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %745 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_488 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %746 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%745 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_489 = tensor.expand_shape %736 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %747 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_489 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %748 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%747 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_490 = tensor.collapse_shape %748 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_491 = tensor.collapse_shape %741 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_492 = tensor.collapse_shape %746 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %749 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_491 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %750 = linalg.batch_matmul ins(%collapsed_490, %749 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_493 = tensor.expand_shape %750 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %751 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_493, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %752 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%751, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_494 = tensor.collapse_shape %752 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %753:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_494 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %754 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_494, %753#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %755 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%754 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %756 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%755 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %757 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%755, %756 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %758 = linalg.batch_matmul ins(%757, %collapsed_492 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_495 = tensor.expand_shape %758 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %759 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_495 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %760 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%759 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %761 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_202 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_496 = tensor.collapse_shape %760 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %762 = linalg.matmul ins(%collapsed_496, %761 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %763 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_201, %762 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_497 = tensor.expand_shape %763 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %764 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_484, %expanded_497 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_498 = tensor.collapse_shape %764 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %765 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_498 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %766 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%765 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %767 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%766 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %768 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%765, %767 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %769 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%768, %768 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %770 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%769 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %771 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%770 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %772 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%771 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %773 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_498 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %774 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%773 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %775 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%772 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %776 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%775 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %777 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_498, %774 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %778 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%777, %776 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %779 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%778, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %780 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%779, %cst_200 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %781 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_199 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %782 = linalg.matmul ins(%780, %781 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %783 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_198, %782 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %784 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%783 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %785 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_197 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %786 = linalg.matmul ins(%784, %785 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %787 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_196, %786 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %788 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_498, %787 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_499 = tensor.expand_shape %788 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %789 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_499 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %790 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%789 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %791 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%790 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %792 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%789, %791 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %793 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%792, %792 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %794 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%793 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %795 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%794 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %796 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%795 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %797 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_499 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %798 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%797 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %799 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%796 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %800 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%799 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %801 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_499, %798 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %802 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%801, %800 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %803 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%802, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %804 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%803, %cst_195 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %805 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_194 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_500 = tensor.collapse_shape %804 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %806 = linalg.matmul ins(%collapsed_500, %805 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %807 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_193, %806 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_501 = tensor.expand_shape %807 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %808 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_501 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %809 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_192 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %810 = linalg.matmul ins(%collapsed_500, %809 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %811 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_191, %810 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_502 = tensor.expand_shape %811 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %812 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_502 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %813 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%812 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %814 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_190 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %815 = linalg.matmul ins(%collapsed_500, %814 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %816 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_189, %815 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_503 = tensor.expand_shape %816 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %817 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_503 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %818 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%817 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_504 = tensor.expand_shape %808 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %819 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_504 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %820 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%819 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_505 = tensor.collapse_shape %820 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_506 = tensor.collapse_shape %813 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_507 = tensor.collapse_shape %818 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %821 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_506 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %822 = linalg.batch_matmul ins(%collapsed_505, %821 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_508 = tensor.expand_shape %822 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %823 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_508, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %824 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%823, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_509 = tensor.collapse_shape %824 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %825:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_509 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %826 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_509, %825#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %827 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%826 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %828 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%827 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %829 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%827, %828 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %830 = linalg.batch_matmul ins(%829, %collapsed_507 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_510 = tensor.expand_shape %830 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %831 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_510 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %832 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%831 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %833 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_188 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_511 = tensor.collapse_shape %832 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %834 = linalg.matmul ins(%collapsed_511, %833 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %835 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_187, %834 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_512 = tensor.expand_shape %835 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %836 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_499, %expanded_512 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_513 = tensor.collapse_shape %836 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %837 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_513 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %838 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%837 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %839 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%838 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %840 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%837, %839 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %841 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%840, %840 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %842 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%841 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %843 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%842 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %844 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%843 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %845 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_513 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %846 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%845 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %847 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%844 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %848 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%847 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %849 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_513, %846 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %850 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%849, %848 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %851 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%850, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %852 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%851, %cst_186 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %853 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_185 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %854 = linalg.matmul ins(%852, %853 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %855 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_184, %854 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %856 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%855 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %857 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_183 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %858 = linalg.matmul ins(%856, %857 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %859 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_182, %858 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %860 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_513, %859 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_514 = tensor.expand_shape %860 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %861 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_514 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %862 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%861 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %863 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%862 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %864 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%861, %863 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %865 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%864, %864 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %866 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%865 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %867 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%866 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %868 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%867 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %869 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_514 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %870 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%869 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %871 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%868 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %872 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%871 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %873 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_514, %870 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %874 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%873, %872 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %875 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%874, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %876 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%875, %cst_181 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %877 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_180 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_515 = tensor.collapse_shape %876 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %878 = linalg.matmul ins(%collapsed_515, %877 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %879 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_179, %878 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_516 = tensor.expand_shape %879 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %880 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_516 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %881 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_178 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %882 = linalg.matmul ins(%collapsed_515, %881 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %883 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_177, %882 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_517 = tensor.expand_shape %883 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %884 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_517 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %885 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%884 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %886 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_176 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %887 = linalg.matmul ins(%collapsed_515, %886 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %888 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_175, %887 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_518 = tensor.expand_shape %888 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %889 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_518 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %890 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%889 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_519 = tensor.expand_shape %880 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %891 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_519 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %892 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%891 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_520 = tensor.collapse_shape %892 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_521 = tensor.collapse_shape %885 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_522 = tensor.collapse_shape %890 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %893 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_521 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %894 = linalg.batch_matmul ins(%collapsed_520, %893 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_523 = tensor.expand_shape %894 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %895 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_523, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %896 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%895, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_524 = tensor.collapse_shape %896 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %897:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_524 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %898 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_524, %897#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %899 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%898 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %900 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%899 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %901 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%899, %900 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %902 = linalg.batch_matmul ins(%901, %collapsed_522 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_525 = tensor.expand_shape %902 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %903 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_525 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %904 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%903 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %905 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_174 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_526 = tensor.collapse_shape %904 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %906 = linalg.matmul ins(%collapsed_526, %905 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %907 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_173, %906 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_527 = tensor.expand_shape %907 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %908 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_514, %expanded_527 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_528 = tensor.collapse_shape %908 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %909 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_528 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %910 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%909 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %911 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%910 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %912 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%909, %911 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %913 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%912, %912 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %914 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%913 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %915 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%914 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %916 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%915 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %917 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_528 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %918 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%917 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %919 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%916 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %920 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%919 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %921 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_528, %918 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %922 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%921, %920 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %923 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%922, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %924 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%923, %cst_172 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %925 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_171 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %926 = linalg.matmul ins(%924, %925 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %927 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_170, %926 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %928 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%927 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %929 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_169 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %930 = linalg.matmul ins(%928, %929 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %931 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_168, %930 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %932 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_528, %931 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_529 = tensor.expand_shape %932 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %933 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_529 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %934 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%933 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %935 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%934 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %936 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%933, %935 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %937 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%936, %936 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %938 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%937 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %939 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%938 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %940 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%939 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %941 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_529 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %942 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%941 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %943 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%940 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %944 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%943 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %945 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_529, %942 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %946 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%945, %944 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %947 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%946, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %948 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%947, %cst_167 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %949 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_166 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_530 = tensor.collapse_shape %948 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %950 = linalg.matmul ins(%collapsed_530, %949 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %951 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_165, %950 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_531 = tensor.expand_shape %951 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %952 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_531 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %953 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_164 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %954 = linalg.matmul ins(%collapsed_530, %953 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %955 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_163, %954 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_532 = tensor.expand_shape %955 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %956 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_532 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %957 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%956 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %958 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_162 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %959 = linalg.matmul ins(%collapsed_530, %958 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %960 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_161, %959 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_533 = tensor.expand_shape %960 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %961 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_533 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %962 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%961 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_534 = tensor.expand_shape %952 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %963 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_534 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %964 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%963 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_535 = tensor.collapse_shape %964 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_536 = tensor.collapse_shape %957 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_537 = tensor.collapse_shape %962 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %965 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_536 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %966 = linalg.batch_matmul ins(%collapsed_535, %965 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_538 = tensor.expand_shape %966 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %967 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_538, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %968 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%967, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_539 = tensor.collapse_shape %968 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %969:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_539 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %970 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_539, %969#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %971 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%970 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %972 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%971 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %973 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%971, %972 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %974 = linalg.batch_matmul ins(%973, %collapsed_537 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_540 = tensor.expand_shape %974 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %975 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_540 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %976 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%975 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %977 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_160 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_541 = tensor.collapse_shape %976 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %978 = linalg.matmul ins(%collapsed_541, %977 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %979 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_159, %978 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_542 = tensor.expand_shape %979 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %980 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_529, %expanded_542 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_543 = tensor.collapse_shape %980 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %981 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_543 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %982 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%981 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %983 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%982 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %984 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%981, %983 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %985 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%984, %984 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %986 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%985 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %987 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%986 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %988 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%987 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %989 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_543 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %990 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%989 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %991 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%988 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %992 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%991 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %993 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_543, %990 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %994 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%993, %992 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %995 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%994, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %996 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%995, %cst_158 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %997 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_157 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %998 = linalg.matmul ins(%996, %997 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %999 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_156, %998 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1000 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%999 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1001 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_155 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1002 = linalg.matmul ins(%1000, %1001 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1003 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_154, %1002 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1004 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_543, %1003 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_544 = tensor.expand_shape %1004 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1005 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_544 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1006 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1005 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1007 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1006 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1008 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1005, %1007 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1009 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1008, %1008 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1010 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1009 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1011 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1010 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1012 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1011 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1013 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_544 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1014 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1013 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1015 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1012 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1016 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1015 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1017 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_544, %1014 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1018 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1017, %1016 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1019 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1018, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1020 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1019, %cst_153 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1021 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_152 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_545 = tensor.collapse_shape %1020 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1022 = linalg.matmul ins(%collapsed_545, %1021 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1023 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_151, %1022 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_546 = tensor.expand_shape %1023 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1024 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_546 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1025 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_150 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1026 = linalg.matmul ins(%collapsed_545, %1025 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1027 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_149, %1026 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_547 = tensor.expand_shape %1027 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1028 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_547 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1029 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1028 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1030 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_148 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1031 = linalg.matmul ins(%collapsed_545, %1030 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1032 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_147, %1031 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_548 = tensor.expand_shape %1032 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1033 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_548 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1034 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1033 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_549 = tensor.expand_shape %1024 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1035 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_549 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1036 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1035 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_550 = tensor.collapse_shape %1036 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_551 = tensor.collapse_shape %1029 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_552 = tensor.collapse_shape %1034 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1037 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_551 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1038 = linalg.batch_matmul ins(%collapsed_550, %1037 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_553 = tensor.expand_shape %1038 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1039 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_553, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1040 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1039, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_554 = tensor.collapse_shape %1040 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1041:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_554 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1042 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_554, %1041#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1043 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1042 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1044 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1043 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1045 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1043, %1044 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1046 = linalg.batch_matmul ins(%1045, %collapsed_552 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_555 = tensor.expand_shape %1046 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1047 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_555 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1048 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1047 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1049 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_146 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_556 = tensor.collapse_shape %1048 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1050 = linalg.matmul ins(%collapsed_556, %1049 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1051 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_145, %1050 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_557 = tensor.expand_shape %1051 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1052 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_544, %expanded_557 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_558 = tensor.collapse_shape %1052 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1053 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_558 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1054 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1053 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1055 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1054 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1056 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1053, %1055 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1057 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1056, %1056 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1058 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1057 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1059 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1058 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1060 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1059 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1061 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_558 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1062 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1061 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1063 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1060 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1064 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1063 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1065 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_558, %1062 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1066 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1065, %1064 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1067 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1066, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1068 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1067, %cst_144 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1069 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_143 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1070 = linalg.matmul ins(%1068, %1069 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1071 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_142, %1070 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1072 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1071 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1073 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_141 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1074 = linalg.matmul ins(%1072, %1073 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1075 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_140, %1074 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1076 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_558, %1075 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_559 = tensor.expand_shape %1076 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1077 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_559 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1078 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1077 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1079 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1078 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1080 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1077, %1079 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1081 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1080, %1080 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1082 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1081 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1083 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1082 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1084 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1083 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1085 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_559 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1086 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1085 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1087 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1084 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1088 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1087 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1089 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_559, %1086 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1090 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1089, %1088 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1091 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1090, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1092 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1091, %cst_139 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1093 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_138 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_560 = tensor.collapse_shape %1092 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1094 = linalg.matmul ins(%collapsed_560, %1093 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1095 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_137, %1094 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_561 = tensor.expand_shape %1095 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1096 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_561 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1097 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_136 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1098 = linalg.matmul ins(%collapsed_560, %1097 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1099 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_135, %1098 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_562 = tensor.expand_shape %1099 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1100 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_562 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1101 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1100 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1102 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_134 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1103 = linalg.matmul ins(%collapsed_560, %1102 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1104 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_133, %1103 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_563 = tensor.expand_shape %1104 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1105 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_563 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1106 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1105 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_564 = tensor.expand_shape %1096 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1107 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_564 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1108 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1107 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_565 = tensor.collapse_shape %1108 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_566 = tensor.collapse_shape %1101 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_567 = tensor.collapse_shape %1106 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1109 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_566 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1110 = linalg.batch_matmul ins(%collapsed_565, %1109 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_568 = tensor.expand_shape %1110 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1111 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_568, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1112 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1111, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_569 = tensor.collapse_shape %1112 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1113:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_569 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1114 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_569, %1113#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1115 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1114 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1116 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1115 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1117 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1115, %1116 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1118 = linalg.batch_matmul ins(%1117, %collapsed_567 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_570 = tensor.expand_shape %1118 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1119 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_570 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1120 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1119 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1121 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_132 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_571 = tensor.collapse_shape %1120 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1122 = linalg.matmul ins(%collapsed_571, %1121 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1123 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_131, %1122 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_572 = tensor.expand_shape %1123 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1124 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_559, %expanded_572 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_573 = tensor.collapse_shape %1124 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1125 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_573 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1126 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1125 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1127 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1126 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1128 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1125, %1127 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1129 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1128, %1128 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1130 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1129 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1131 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1130 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1132 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1131 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1133 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_573 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1134 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1133 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1135 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1132 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1136 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1135 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1137 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_573, %1134 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1138 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1137, %1136 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1139 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1138, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1140 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1139, %cst_130 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1141 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_129 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1142 = linalg.matmul ins(%1140, %1141 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1143 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_128, %1142 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1144 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1143 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1145 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_127 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1146 = linalg.matmul ins(%1144, %1145 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1147 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_126, %1146 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1148 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_573, %1147 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_574 = tensor.expand_shape %1148 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1149 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_574 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1150 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1149 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1151 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1150 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1152 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1149, %1151 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1153 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1152, %1152 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1154 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1153 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1155 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1154 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1156 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1155 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1157 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_574 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1158 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1157 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1159 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1156 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1160 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1159 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1161 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_574, %1158 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1162 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1161, %1160 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1163 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1162, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1164 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1163, %cst_125 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1165 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_124 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_575 = tensor.collapse_shape %1164 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1166 = linalg.matmul ins(%collapsed_575, %1165 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1167 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_123, %1166 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_576 = tensor.expand_shape %1167 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1168 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_576 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1169 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_122 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1170 = linalg.matmul ins(%collapsed_575, %1169 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1171 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_121, %1170 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_577 = tensor.expand_shape %1171 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1172 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_577 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1173 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1172 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1174 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_120 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1175 = linalg.matmul ins(%collapsed_575, %1174 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1176 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_119, %1175 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_578 = tensor.expand_shape %1176 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1177 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_578 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1178 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1177 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_579 = tensor.expand_shape %1168 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1179 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_579 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1180 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1179 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_580 = tensor.collapse_shape %1180 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_581 = tensor.collapse_shape %1173 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_582 = tensor.collapse_shape %1178 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1181 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_581 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1182 = linalg.batch_matmul ins(%collapsed_580, %1181 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_583 = tensor.expand_shape %1182 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1183 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_583, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1184 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1183, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_584 = tensor.collapse_shape %1184 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1185:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_584 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1186 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_584, %1185#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1187 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1186 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1188 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1187 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1189 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1187, %1188 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1190 = linalg.batch_matmul ins(%1189, %collapsed_582 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_585 = tensor.expand_shape %1190 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1191 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_585 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1192 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1191 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1193 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_118 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_586 = tensor.collapse_shape %1192 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1194 = linalg.matmul ins(%collapsed_586, %1193 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1195 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_117, %1194 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_587 = tensor.expand_shape %1195 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1196 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_574, %expanded_587 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_588 = tensor.collapse_shape %1196 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1197 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_588 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1198 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1197 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1199 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1198 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1200 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1197, %1199 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1201 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1200, %1200 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1202 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1201 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1203 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1202 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1204 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1203 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1205 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_588 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1206 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1205 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1207 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1204 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1208 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1207 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1209 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_588, %1206 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1210 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1209, %1208 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1211 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1210, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1212 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1211, %cst_116 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1213 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_115 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1214 = linalg.matmul ins(%1212, %1213 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1215 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_114, %1214 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1216 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1215 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1217 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_113 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1218 = linalg.matmul ins(%1216, %1217 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1219 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_112, %1218 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1220 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_588, %1219 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_589 = tensor.expand_shape %1220 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1221 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_589 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1222 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1221 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1223 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1222 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1224 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1221, %1223 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1225 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1224, %1224 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1226 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1225 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1227 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1226 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1228 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1227 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1229 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_589 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1230 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1229 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1231 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1228 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1232 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1231 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1233 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_589, %1230 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1234 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1233, %1232 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1235 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1234, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1236 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1235, %cst_111 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1237 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_110 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_590 = tensor.collapse_shape %1236 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1238 = linalg.matmul ins(%collapsed_590, %1237 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1239 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_109, %1238 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_591 = tensor.expand_shape %1239 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1240 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_591 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1241 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_108 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1242 = linalg.matmul ins(%collapsed_590, %1241 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1243 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_107, %1242 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_592 = tensor.expand_shape %1243 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1244 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_592 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1245 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1244 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1246 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_106 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1247 = linalg.matmul ins(%collapsed_590, %1246 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1248 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_105, %1247 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_593 = tensor.expand_shape %1248 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1249 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_593 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1250 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1249 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_594 = tensor.expand_shape %1240 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1251 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_594 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1252 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1251 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_595 = tensor.collapse_shape %1252 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_596 = tensor.collapse_shape %1245 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_597 = tensor.collapse_shape %1250 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1253 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_596 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1254 = linalg.batch_matmul ins(%collapsed_595, %1253 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_598 = tensor.expand_shape %1254 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1255 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_598, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1256 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1255, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_599 = tensor.collapse_shape %1256 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1257:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_599 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1258 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_599, %1257#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1259 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1258 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1260 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1259 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1261 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1259, %1260 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1262 = linalg.batch_matmul ins(%1261, %collapsed_597 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_600 = tensor.expand_shape %1262 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1263 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_600 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1264 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1263 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1265 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_104 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_601 = tensor.collapse_shape %1264 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1266 = linalg.matmul ins(%collapsed_601, %1265 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1267 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_103, %1266 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_602 = tensor.expand_shape %1267 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1268 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_589, %expanded_602 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_603 = tensor.collapse_shape %1268 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1269 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_603 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1270 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1269 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1271 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1270 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1272 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1269, %1271 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1273 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1272, %1272 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1274 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1273 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1275 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1274 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1276 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1275 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1277 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_603 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1278 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1277 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1279 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1276 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1280 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1279 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1281 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_603, %1278 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1282 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1281, %1280 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1283 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1282, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1284 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1283, %cst_102 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1285 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_101 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1286 = linalg.matmul ins(%1284, %1285 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1287 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_100, %1286 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1288 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1287 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1289 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_99 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1290 = linalg.matmul ins(%1288, %1289 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1291 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_98, %1290 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1292 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_603, %1291 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_604 = tensor.expand_shape %1292 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1293 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_604 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1294 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1293 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1295 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1294 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1296 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1293, %1295 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1297 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1296, %1296 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1298 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1297 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1299 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1298 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1300 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1299 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1301 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_604 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1302 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1301 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1303 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1300 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1304 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1303 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1305 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_604, %1302 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1306 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1305, %1304 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1307 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1306, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1308 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1307, %cst_97 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1309 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_96 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_605 = tensor.collapse_shape %1308 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1310 = linalg.matmul ins(%collapsed_605, %1309 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1311 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_95, %1310 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_606 = tensor.expand_shape %1311 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1312 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_606 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1313 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_94 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1314 = linalg.matmul ins(%collapsed_605, %1313 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1315 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_93, %1314 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_607 = tensor.expand_shape %1315 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1316 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_607 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1317 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1316 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1318 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_92 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1319 = linalg.matmul ins(%collapsed_605, %1318 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1320 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_91, %1319 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_608 = tensor.expand_shape %1320 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1321 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_608 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1322 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1321 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_609 = tensor.expand_shape %1312 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1323 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_609 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1324 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1323 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_610 = tensor.collapse_shape %1324 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_611 = tensor.collapse_shape %1317 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_612 = tensor.collapse_shape %1322 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1325 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_611 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1326 = linalg.batch_matmul ins(%collapsed_610, %1325 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_613 = tensor.expand_shape %1326 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1327 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_613, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1328 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1327, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_614 = tensor.collapse_shape %1328 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1329:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_614 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1330 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_614, %1329#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1331 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1330 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1332 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1331 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1333 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1331, %1332 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1334 = linalg.batch_matmul ins(%1333, %collapsed_612 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_615 = tensor.expand_shape %1334 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1335 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_615 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1336 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1335 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1337 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_90 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_616 = tensor.collapse_shape %1336 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1338 = linalg.matmul ins(%collapsed_616, %1337 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1339 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_89, %1338 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_617 = tensor.expand_shape %1339 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1340 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_604, %expanded_617 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_618 = tensor.collapse_shape %1340 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1341 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_618 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1342 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1341 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1343 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1342 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1344 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1341, %1343 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1345 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1344, %1344 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1346 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1345 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1347 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1346 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1348 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1347 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1349 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_618 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1350 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1349 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1351 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1348 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1352 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1351 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1353 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_618, %1350 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1354 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1353, %1352 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1355 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1354, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1356 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1355, %cst_88 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1357 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_87 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1358 = linalg.matmul ins(%1356, %1357 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1359 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_86, %1358 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1360 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1359 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1361 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_85 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1362 = linalg.matmul ins(%1360, %1361 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1363 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_84, %1362 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1364 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_618, %1363 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_619 = tensor.expand_shape %1364 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1365 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_619 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1366 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1365 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1367 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1366 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1368 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1365, %1367 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1369 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1368, %1368 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1370 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1369 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1371 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1370 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1372 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1371 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1373 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_619 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1374 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1373 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1375 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1372 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1376 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1375 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1377 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_619, %1374 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1378 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1377, %1376 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1379 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1378, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1380 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1379, %cst_83 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1381 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_82 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_620 = tensor.collapse_shape %1380 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1382 = linalg.matmul ins(%collapsed_620, %1381 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1383 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_81, %1382 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_621 = tensor.expand_shape %1383 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1384 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_621 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1385 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_80 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1386 = linalg.matmul ins(%collapsed_620, %1385 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1387 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_79, %1386 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_622 = tensor.expand_shape %1387 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1388 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_622 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1389 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1388 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1390 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_78 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1391 = linalg.matmul ins(%collapsed_620, %1390 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1392 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_77, %1391 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_623 = tensor.expand_shape %1392 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1393 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_623 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1394 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1393 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_624 = tensor.expand_shape %1384 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1395 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_624 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1396 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1395 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_625 = tensor.collapse_shape %1396 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_626 = tensor.collapse_shape %1389 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_627 = tensor.collapse_shape %1394 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1397 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_626 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1398 = linalg.batch_matmul ins(%collapsed_625, %1397 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_628 = tensor.expand_shape %1398 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1399 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_628, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1400 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1399, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_629 = tensor.collapse_shape %1400 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1401:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_629 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1402 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_629, %1401#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1403 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1402 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1404 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1403 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1405 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1403, %1404 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1406 = linalg.batch_matmul ins(%1405, %collapsed_627 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_630 = tensor.expand_shape %1406 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1407 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_630 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1408 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1407 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1409 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_76 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_631 = tensor.collapse_shape %1408 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1410 = linalg.matmul ins(%collapsed_631, %1409 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1411 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_75, %1410 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_632 = tensor.expand_shape %1411 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1412 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_619, %expanded_632 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_633 = tensor.collapse_shape %1412 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1413 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_633 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1414 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1413 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1415 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1414 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1416 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1413, %1415 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1417 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1416, %1416 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1418 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1417 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1419 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1418 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1420 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1419 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1421 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_633 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1422 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1421 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1423 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1420 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1424 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1423 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1425 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_633, %1422 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1426 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1425, %1424 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1427 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1426, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1428 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1427, %cst_74 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1429 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_73 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1430 = linalg.matmul ins(%1428, %1429 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1431 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_72, %1430 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1432 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1431 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1433 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_71 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1434 = linalg.matmul ins(%1432, %1433 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1435 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_70, %1434 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1436 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_633, %1435 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_634 = tensor.expand_shape %1436 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1437 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_634 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1438 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1437 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1439 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1438 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1440 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1437, %1439 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1441 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1440, %1440 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1442 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1441 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1443 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1442 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1444 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1443 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1445 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_634 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1446 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1445 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1447 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1444 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1448 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1447 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1449 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_634, %1446 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1450 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1449, %1448 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1451 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1450, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1452 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1451, %cst_69 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1453 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_68 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_635 = tensor.collapse_shape %1452 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1454 = linalg.matmul ins(%collapsed_635, %1453 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1455 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_67, %1454 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_636 = tensor.expand_shape %1455 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1456 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_636 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1457 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_66 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1458 = linalg.matmul ins(%collapsed_635, %1457 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1459 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_65, %1458 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_637 = tensor.expand_shape %1459 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1460 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_637 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1461 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1460 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1462 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_64 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1463 = linalg.matmul ins(%collapsed_635, %1462 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1464 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_63, %1463 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_638 = tensor.expand_shape %1464 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1465 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_638 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1466 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1465 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_639 = tensor.expand_shape %1456 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1467 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_639 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1468 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1467 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_640 = tensor.collapse_shape %1468 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_641 = tensor.collapse_shape %1461 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_642 = tensor.collapse_shape %1466 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1469 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_641 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1470 = linalg.batch_matmul ins(%collapsed_640, %1469 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_643 = tensor.expand_shape %1470 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1471 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_643, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1472 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1471, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_644 = tensor.collapse_shape %1472 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1473:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_644 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1474 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_644, %1473#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1475 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1474 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1476 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1475 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1477 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1475, %1476 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1478 = linalg.batch_matmul ins(%1477, %collapsed_642 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_645 = tensor.expand_shape %1478 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1479 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_645 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1480 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1479 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1481 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_62 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_646 = tensor.collapse_shape %1480 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1482 = linalg.matmul ins(%collapsed_646, %1481 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1483 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_61, %1482 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_647 = tensor.expand_shape %1483 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1484 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_634, %expanded_647 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_648 = tensor.collapse_shape %1484 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1485 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_648 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1486 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1485 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1487 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1486 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1488 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1485, %1487 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1489 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1488, %1488 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1490 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1489 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1491 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1490 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1492 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1491 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1493 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_648 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1494 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1493 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1495 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1492 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1496 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1495 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1497 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_648, %1494 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1498 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1497, %1496 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1499 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1498, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1500 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1499, %cst_60 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1501 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_59 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1502 = linalg.matmul ins(%1500, %1501 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1503 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_58, %1502 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1504 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1503 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1505 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_57 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1506 = linalg.matmul ins(%1504, %1505 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1507 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_56, %1506 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1508 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_648, %1507 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_649 = tensor.expand_shape %1508 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1509 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_649 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1510 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1509 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1511 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1510 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1512 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1509, %1511 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1513 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1512, %1512 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1514 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1513 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1515 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1514 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1516 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1515 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1517 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_649 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1518 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1517 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1519 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1516 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1520 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1519 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1521 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_649, %1518 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1522 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1521, %1520 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1523 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1522, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1524 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1523, %cst_55 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1525 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_54 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_650 = tensor.collapse_shape %1524 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1526 = linalg.matmul ins(%collapsed_650, %1525 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1527 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_53, %1526 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_651 = tensor.expand_shape %1527 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1528 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_651 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1529 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_52 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1530 = linalg.matmul ins(%collapsed_650, %1529 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1531 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_51, %1530 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_652 = tensor.expand_shape %1531 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1532 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_652 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1533 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1532 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1534 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_50 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1535 = linalg.matmul ins(%collapsed_650, %1534 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1536 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_49, %1535 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_653 = tensor.expand_shape %1536 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1537 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_653 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1538 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1537 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_654 = tensor.expand_shape %1528 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1539 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_654 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1540 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1539 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_655 = tensor.collapse_shape %1540 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_656 = tensor.collapse_shape %1533 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_657 = tensor.collapse_shape %1538 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1541 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_656 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1542 = linalg.batch_matmul ins(%collapsed_655, %1541 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_658 = tensor.expand_shape %1542 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1543 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_658, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1544 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1543, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_659 = tensor.collapse_shape %1544 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1545:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_659 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1546 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_659, %1545#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1547 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1546 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1548 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1547 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1549 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1547, %1548 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1550 = linalg.batch_matmul ins(%1549, %collapsed_657 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_660 = tensor.expand_shape %1550 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1551 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_660 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1552 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1551 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1553 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_48 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_661 = tensor.collapse_shape %1552 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1554 = linalg.matmul ins(%collapsed_661, %1553 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1555 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_47, %1554 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_662 = tensor.expand_shape %1555 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1556 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_649, %expanded_662 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_663 = tensor.collapse_shape %1556 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1557 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_663 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1558 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1557 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1559 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1558 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1560 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1557, %1559 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1561 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1560, %1560 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1562 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1561 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1563 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1562 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1564 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1563 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1565 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_663 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1566 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1565 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1567 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1564 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1568 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1567 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1569 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_663, %1566 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1570 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1569, %1568 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1571 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1570, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1572 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1571, %cst_46 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1573 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_45 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1574 = linalg.matmul ins(%1572, %1573 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1575 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_44, %1574 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1576 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1575 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1577 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_43 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1578 = linalg.matmul ins(%1576, %1577 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1579 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_42, %1578 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1580 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_663, %1579 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_664 = tensor.expand_shape %1580 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1581 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_664 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1582 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1581 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1583 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1582 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1584 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1581, %1583 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1585 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1584, %1584 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1586 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1585 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1587 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1586 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1588 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1587 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1589 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_664 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1590 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1589 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1591 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1588 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1592 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1591 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1593 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_664, %1590 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1594 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1593, %1592 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1595 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1594, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1596 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1595, %cst_41 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1597 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_40 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_665 = tensor.collapse_shape %1596 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1598 = linalg.matmul ins(%collapsed_665, %1597 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1599 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_39, %1598 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_666 = tensor.expand_shape %1599 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1600 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_666 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1601 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_38 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1602 = linalg.matmul ins(%collapsed_665, %1601 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1603 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_37, %1602 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_667 = tensor.expand_shape %1603 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1604 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_667 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1605 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1604 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1606 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_36 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1607 = linalg.matmul ins(%collapsed_665, %1606 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1608 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_35, %1607 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_668 = tensor.expand_shape %1608 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1609 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_668 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1610 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1609 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_669 = tensor.expand_shape %1600 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1611 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_669 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1612 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1611 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_670 = tensor.collapse_shape %1612 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_671 = tensor.collapse_shape %1605 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_672 = tensor.collapse_shape %1610 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1613 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_671 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1614 = linalg.batch_matmul ins(%collapsed_670, %1613 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_673 = tensor.expand_shape %1614 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1615 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_673, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1616 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1615, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_674 = tensor.collapse_shape %1616 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1617:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_674 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1618 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_674, %1617#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1619 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1618 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1620 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1619 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1621 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1619, %1620 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1622 = linalg.batch_matmul ins(%1621, %collapsed_672 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_675 = tensor.expand_shape %1622 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1623 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_675 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1624 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1623 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1625 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_34 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_676 = tensor.collapse_shape %1624 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1626 = linalg.matmul ins(%collapsed_676, %1625 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1627 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_33, %1626 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_677 = tensor.expand_shape %1627 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1628 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_664, %expanded_677 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_678 = tensor.collapse_shape %1628 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1629 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_678 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1630 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1629 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1631 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1630 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1632 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1629, %1631 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1633 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1632, %1632 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1634 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1633 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1635 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1634 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1636 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1635 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1637 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_678 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1638 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1637 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1639 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1636 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1640 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1639 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1641 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_678, %1638 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1642 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1641, %1640 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1643 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1642, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1644 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1643, %cst_32 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1645 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_31 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1646 = linalg.matmul ins(%1644, %1645 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1647 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_30, %1646 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1648 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1647 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1649 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_29 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1650 = linalg.matmul ins(%1648, %1649 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1651 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_28, %1650 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1652 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_678, %1651 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_679 = tensor.expand_shape %1652 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1653 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_679 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1654 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1653 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1655 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1654 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1656 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1653, %1655 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1657 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1656, %1656 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1658 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1657 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1659 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1658 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1660 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1659 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1661 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_679 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1662 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1661 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1663 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1660 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1664 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1663 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1665 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_679, %1662 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1666 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1665, %1664 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1667 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1666, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1668 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1667, %cst_27 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1669 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_26 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_680 = tensor.collapse_shape %1668 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1670 = linalg.matmul ins(%collapsed_680, %1669 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1671 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_25, %1670 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_681 = tensor.expand_shape %1671 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1672 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_681 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1673 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_24 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1674 = linalg.matmul ins(%collapsed_680, %1673 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1675 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_23, %1674 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_682 = tensor.expand_shape %1675 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1676 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_682 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1677 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1676 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1678 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_22 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1679 = linalg.matmul ins(%collapsed_680, %1678 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1680 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_21, %1679 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_683 = tensor.expand_shape %1680 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1681 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_683 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1682 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1681 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_684 = tensor.expand_shape %1672 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1683 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_684 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1684 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1683 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_685 = tensor.collapse_shape %1684 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_686 = tensor.collapse_shape %1677 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_687 = tensor.collapse_shape %1682 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1685 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_686 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1686 = linalg.batch_matmul ins(%collapsed_685, %1685 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_688 = tensor.expand_shape %1686 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1687 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_688, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1688 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1687, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_689 = tensor.collapse_shape %1688 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1689:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_689 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1690 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_689, %1689#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1691 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1690 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1692 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1691 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1693 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1691, %1692 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1694 = linalg.batch_matmul ins(%1693, %collapsed_687 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_690 = tensor.expand_shape %1694 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1695 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_690 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1696 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1695 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1697 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_20 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_691 = tensor.collapse_shape %1696 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1698 = linalg.matmul ins(%collapsed_691, %1697 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1699 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_19, %1698 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_692 = tensor.expand_shape %1699 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1700 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_679, %expanded_692 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_693 = tensor.collapse_shape %1700 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1701 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_693 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1702 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1701 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1703 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1702 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1704 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1701, %1703 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1705 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1704, %1704 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1706 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1705 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1707 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1706 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1708 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1707 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1709 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_693 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1710 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1709 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1711 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1708 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1712 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1711 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1713 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_693, %1710 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1714 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1713, %1712 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1715 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1714, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1716 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1715, %cst_18 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1717 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_17 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1718 = linalg.matmul ins(%1716, %1717 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1719 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_16, %1718 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1720 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1719 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1721 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_15 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1722 = linalg.matmul ins(%1720, %1721 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1723 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_14, %1722 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1724 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_693, %1723 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_694 = tensor.expand_shape %1724 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1725 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_694 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1726 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1725 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1727 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1726 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1728 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1725, %1727 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1729 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1728, %1728 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1730 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1729 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1731 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1730 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1732 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1731 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1733 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_694 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1734 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1733 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1735 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1732 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1736 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1735 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1737 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_694, %1734 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1738 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1737, %1736 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1739 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1738, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1740 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1739, %cst_13 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1741 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_12 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_695 = tensor.collapse_shape %1740 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1742 = linalg.matmul ins(%collapsed_695, %1741 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1743 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_11, %1742 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_696 = tensor.expand_shape %1743 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1744 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_696 : tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.mulf %in, %cst_348 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1745 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_10 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1746 = linalg.matmul ins(%collapsed_695, %1745 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1747 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_9, %1746 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_697 = tensor.expand_shape %1747 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1748 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_697 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1749 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1748 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1750 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_8 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %1751 = linalg.matmul ins(%collapsed_695, %1750 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1752 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_7, %1751 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_698 = tensor.expand_shape %1752 [[0, 1], [2, 3]] : tensor<8x2048xf32> into tensor<1x8x32x64xf32>
    %1753 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_698 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1754 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1753 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %expanded_699 = tensor.expand_shape %1744 [[0], [1], [2, 3]] : tensor<1x8x2048xf32> into tensor<1x8x32x64xf32>
    %1755 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_699 : tensor<1x8x32x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %1756 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1755 : tensor<1x32x8x64xf32>) outs(%70 : tensor<1x32x8x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x32x8x64xf32>
    %collapsed_700 = tensor.collapse_shape %1756 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_701 = tensor.collapse_shape %1749 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %collapsed_702 = tensor.collapse_shape %1754 [[0, 1], [2], [3]] : tensor<1x32x8x64xf32> into tensor<32x8x64xf32>
    %1757 = linalg.generic {indexing_maps = [#map1, #map20], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_701 : tensor<32x8x64xf32>) outs(%80 : tensor<32x64x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<32x64x8xf32>
    %1758 = linalg.batch_matmul ins(%collapsed_700, %1757 : tensor<32x8x64xf32>, tensor<32x64x8xf32>) outs(%83 : tensor<32x8x8xf32>) -> tensor<32x8x8xf32>
    %expanded_703 = tensor.expand_shape %1758 [[0, 1], [2], [3]] : tensor<32x8x8xf32> into tensor<1x32x8x8xf32>
    %1759 = linalg.generic {indexing_maps = [#map19, #map8, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_703, %28 : tensor<1x32x8x8xf32>, tensor<1x1x8x8xf32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x32x8x8xf32>
    %1760 = linalg.generic {indexing_maps = [#map19, #map11, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1759, %cst_329 : tensor<1x32x8x8xf32>, tensor<f32>) outs(%85 : tensor<1x32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %in_712 : f32
      %1819 = arith.select %1818, %in, %in_712 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x32x8x8xf32>
    %collapsed_704 = tensor.collapse_shape %1760 [[0, 1], [2], [3]] : tensor<1x32x8x8xf32> into tensor<32x8x8xf32>
    %1761:2 = linalg.generic {indexing_maps = [#map1, #map14, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%collapsed_704 : tensor<32x8x8xf32>) outs(%91, %89 : tensor<32x8x1xf32>, tensor<32x8x1xi64>) {
    ^bb0(%in: f32, %out: f32, %out_712: i64):
      %1818 = linalg.index 2 : index
      %1819 = arith.index_cast %1818 : index to i64
      %1820 = arith.maximumf %in, %out : f32
      %1821 = arith.cmpf ogt, %in, %out : f32
      %1822 = arith.select %1821, %1819, %out_712 : i64
      linalg.yield %1820, %1822 : f32, i64
    } -> (tensor<32x8x1xf32>, tensor<32x8x1xi64>)
    %1762 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%collapsed_704, %1761#0 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1763 = linalg.generic {indexing_maps = [#map1, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1762 : tensor<32x8x8xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.exp %in : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1764 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1763 : tensor<32x8x8xf32>) outs(%95 : tensor<32x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x1xf32>
    %1765 = linalg.generic {indexing_maps = [#map1, #map14, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1763, %1764 : tensor<32x8x8xf32>, tensor<32x8x1xf32>) outs(%82 : tensor<32x8x8xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.divf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<32x8x8xf32>
    %1766 = linalg.batch_matmul ins(%1765, %collapsed_702 : tensor<32x8x8xf32>, tensor<32x8x64xf32>) outs(%99 : tensor<32x8x64xf32>) -> tensor<32x8x64xf32>
    %expanded_705 = tensor.expand_shape %1766 [[0, 1], [2], [3]] : tensor<32x8x64xf32> into tensor<1x32x8x64xf32>
    %1767 = linalg.generic {indexing_maps = [#map9, #map18], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%expanded_705 : tensor<1x32x8x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1768 = linalg.generic {indexing_maps = [#map19, #map9], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1767 : tensor<1x8x32x64xf32>) outs(%101 : tensor<1x8x32x64xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<1x8x32x64xf32>
    %1769 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_6 : tensor<2048x2048xf32>) outs(%60 : tensor<2048x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x2048xf32>
    %collapsed_706 = tensor.collapse_shape %1768 [[0, 1], [2, 3]] : tensor<1x8x32x64xf32> into tensor<8x2048xf32>
    %1770 = linalg.matmul ins(%collapsed_706, %1769 : tensor<8x2048xf32>, tensor<2048x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1771 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_5, %1770 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_707 = tensor.expand_shape %1771 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1772 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_694, %expanded_707 : tensor<1x8x2048xf32>, tensor<1x8x2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %collapsed_708 = tensor.collapse_shape %1772 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1773 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_708 : tensor<8x2048xf32>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1774 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1773 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1775 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1774 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1776 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1773, %1775 : tensor<8x2048xf64>, tensor<8x1xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1777 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1776, %1776 : tensor<8x2048xf64>, tensor<8x2048xf64>) outs(%108 : tensor<8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x2048xf64>
    %1778 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%1777 : tensor<8x2048xf64>) outs(%111 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1779 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1778 : tensor<8x1xf64>) outs(%110 : tensor<8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<8x1xf64>
    %1780 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1779 : tensor<8x1xf64>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1781 = linalg.generic {indexing_maps = [#map4, #map7], iterator_types = ["parallel", "reduction"]} ins(%collapsed_708 : tensor<8x2048xf32>) outs(%120 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1782 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1781 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1783 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1780 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x1xf32>
    %1784 = linalg.generic {indexing_maps = [#map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1783 : tensor<8x1xf32>) outs(%118 : tensor<8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<8x1xf32>
    %1785 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_708, %1782 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1786 = linalg.generic {indexing_maps = [#map4, #map7, #map4], iterator_types = ["parallel", "parallel"]} ins(%1785, %1784 : tensor<8x2048xf32>, tensor<8x1xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1787 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1786, %cst_337 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1788 = linalg.generic {indexing_maps = [#map4, #map6, #map4], iterator_types = ["parallel", "parallel"]} ins(%1787, %cst_4 : tensor<8x2048xf32>, tensor<2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1789 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_3 : tensor<8192x2048xf32>) outs(%129 : tensor<2048x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x8192xf32>
    %1790 = linalg.matmul ins(%1788, %1789 : tensor<8x2048xf32>, tensor<2048x8192xf32>) outs(%132 : tensor<8x8192xf32>) -> tensor<8x8192xf32>
    %1791 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_2, %1790 : tensor<8192xf32>, tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x8192xf32>
    %1792 = linalg.generic {indexing_maps = [#map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%1791 : tensor<8x8192xf32>) outs(%131 : tensor<8x8192xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.cmpf ugt, %in, %cst_340 : f32
      %1819 = arith.select %1818, %in, %cst_340 : f32
      linalg.yield %1819 : f32
    } -> tensor<8x8192xf32>
    %1793 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_1 : tensor<2048x8192xf32>) outs(%136 : tensor<8192x2048xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<8192x2048xf32>
    %1794 = linalg.matmul ins(%1792, %1793 : tensor<8x8192xf32>, tensor<8192x2048xf32>) outs(%63 : tensor<8x2048xf32>) -> tensor<8x2048xf32>
    %1795 = linalg.generic {indexing_maps = [#map6, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%cst_0, %1794 : tensor<2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %1796 = linalg.generic {indexing_maps = [#map4, #map4, #map4], iterator_types = ["parallel", "parallel"]} ins(%collapsed_708, %1795 : tensor<8x2048xf32>, tensor<8x2048xf32>) outs(%62 : tensor<8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<8x2048xf32>
    %expanded_709 = tensor.expand_shape %1796 [[0, 1], [2]] : tensor<8x2048xf32> into tensor<1x8x2048xf32>
    %1797 = linalg.generic {indexing_maps = [#map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_709 : tensor<1x8x2048xf32>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f32, %out: f64):
      %1818 = arith.extf %in : f32 to f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1798 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1797 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1799 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1798 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1800 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1797, %1799 : tensor<1x8x2048xf64>, tensor<1x8x1xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.subf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1801 = linalg.generic {indexing_maps = [#map13, #map13, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1800, %1800 : tensor<1x8x2048xf64>, tensor<1x8x2048xf64>) outs(%39 : tensor<1x8x2048xf64>) {
    ^bb0(%in: f64, %in_712: f64, %out: f64):
      %1818 = arith.mulf %in, %in_712 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x2048xf64>
    %1802 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1801 : tensor<1x8x2048xf64>) outs(%42 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.addf %in, %out : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1803 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1802 : tensor<1x8x1xf64>) outs(%41 : tensor<1x8x1xf64>) {
    ^bb0(%in: f64, %out: f64):
      %1818 = arith.divf %in, %cst_346 : f64
      linalg.yield %1818 : f64
    } -> tensor<1x8x1xf64>
    %1804 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1803 : tensor<1x8x1xf64>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f64, %out: f32):
      %1818 = arith.truncf %in : f64 to f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1805 = linalg.generic {indexing_maps = [#map1, #map14], iterator_types = ["parallel", "parallel", "reduction"]} ins(%expanded_709 : tensor<1x8x2048xf32>) outs(%51 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.addf %in, %out : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1806 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1805 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.divf %in, %cst_347 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1807 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1804 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = arith.truncf %cst_344 : f64 to f32
      %1819 = arith.addf %in, %1818 : f32
      linalg.yield %1819 : f32
    } -> tensor<1x8x1xf32>
    %1808 = linalg.generic {indexing_maps = [#map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1807 : tensor<1x8x1xf32>) outs(%49 : tensor<1x8x1xf32>) {
    ^bb0(%in: f32, %out: f32):
      %1818 = math.rsqrt %in : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x1xf32>
    %1809 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%expanded_709, %1806 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.subf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1810 = linalg.generic {indexing_maps = [#map13, #map15, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1809, %1808 : tensor<1x8x2048xf32>, tensor<1x8x1xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1811 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1810, %cst_337 : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.mulf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1812 = linalg.generic {indexing_maps = [#map13, #map16, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%1811, %cst : tensor<1x8x2048xf32>, tensor<2048xf32>) outs(%0 : tensor<1x8x2048xf32>) {
    ^bb0(%in: f32, %in_712: f32, %out: f32):
      %1818 = arith.addf %in, %in_712 : f32
      linalg.yield %1818 : f32
    } -> tensor<1x8x2048xf32>
    %1813 = tensor.empty() : tensor<2048x50272xf32>
    %1814 = linalg.generic {indexing_maps = [#map4, #map17], iterator_types = ["parallel", "parallel"]} ins(%cst_339 : tensor<50272x2048xf32>) outs(%1813 : tensor<2048x50272xf32>) {
    ^bb0(%in: f32, %out: f32):
      linalg.yield %in : f32
    } -> tensor<2048x50272xf32>
    %collapsed_710 = tensor.collapse_shape %1812 [[0, 1], [2]] : tensor<1x8x2048xf32> into tensor<8x2048xf32>
    %1815 = tensor.empty() : tensor<8x50272xf32>
    %1816 = linalg.fill ins(%cst_340 : f32) outs(%1815 : tensor<8x50272xf32>) -> tensor<8x50272xf32>
    %1817 = linalg.matmul ins(%collapsed_710, %1814 : tensor<8x2048xf32>, tensor<2048x50272xf32>) outs(%1816 : tensor<8x50272xf32>) -> tensor<8x50272xf32>
    %expanded_711 = tensor.expand_shape %1817 [[0, 1], [2]] : tensor<8x50272xf32> into tensor<1x8x50272xf32>
    return %expanded_711 : tensor<1x8x50272xf32>
  }
}