Skip to content

Instantly share code, notes, and snippets.

@AmosLewis

AmosLewis/gpt2_torch_tosa_elide_2023.mlir

Created January 2, 2023 05:15
Show Gist options
  • Save AmosLewis/fd8b242db2b5f328fa56e9afa270cbd6 to your computer and use it in GitHub Desktop.
Save AmosLewis/fd8b242db2b5f328fa56e9afa270cbd6 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
gpt2_torch_tosa_elide_2023.mlir
module attributes {torch.debug_module_name = "_lambda"} {
func.func @forward(%arg0: tensor<1x5xi64>) -> tensor<1x5x50257xf32> {
%0 = "tosa.const"() {value = dense<-3.40282347E+38> : tensor<f32>} : () -> tensor<f32>
%1 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x1024x1024xui8>} : () -> tensor<1x1x1024x1024xi8>
%2 = "tosa.const"() {value = dense<8.000000e+00> : tensor<f32>} : () -> tensor<f32>
%3 = "tosa.const"() {value = dense_resource<__elided__> : tensor<50257x768xf32>} : () -> tensor<50257x768xf32>
%4 = "tosa.const"() {value = dense<7.680000e+02> : tensor<1xf32>} : () -> tensor<1xf32>
%5 = "tosa.const"() {value = dense<[0, 2, 1, 3]> : tensor<4xi64>} : () -> tensor<4xi64>
%6 = "tosa.const"() {value = dense<[0, 1, 3, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
%7 = "tosa.const"() {value = dense<0> : tensor<1x1x5x5xi8>} : () -> tensor<1x1x5x5xi8>
%8 = "tosa.const"() {value = dense<[1, 0]> : tensor<2xi32>} : () -> tensor<2xi32>
%9 = "tosa.const"() {value = dense<0.797884583> : tensor<1x1x1xf32>} : () -> tensor<1x1x1xf32>
%10 = "tosa.const"() {value = dense<4.471500e-02> : tensor<1x1x1xf32>} : () -> tensor<1x1x1xf32>
%11 = "tosa.const"() {value = dense<3.000000e+00> : tensor<1x1x1xf32>} : () -> tensor<1x1x1xf32>
%12 = "tosa.const"() {value = dense<5.000000e-01> : tensor<1x1x1xf32>} : () -> tensor<1x1x1xf32>
%13 = "tosa.const"() {value = dense<9.99999974E-6> : tensor<1x1x1xf32>} : () -> tensor<1x1x1xf32>
%14 = "tosa.const"() {value = dense<1.000000e+00> : tensor<1x1x1xf32>} : () -> tensor<1x1x1xf32>
%15 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x5xi64>} : () -> tensor<1x5xi64>
%16 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1024x768xf32>} : () -> tensor<1x1024x768xf32>
%17 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%18 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%19 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%20 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%21 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%22 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%23 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%24 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%25 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%26 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%27 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%28 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%29 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%30 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%31 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%32 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%33 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%34 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%35 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%36 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%37 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%38 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%39 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%40 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%41 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%42 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%43 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%44 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%45 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%46 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%47 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%48 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%49 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%50 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%51 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%52 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%53 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%54 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%55 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%56 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%57 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%58 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%59 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%60 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%61 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%62 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%63 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%64 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%65 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%66 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%67 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%68 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%69 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%70 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%71 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%72 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%73 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%74 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%75 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%76 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%77 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%78 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%79 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%80 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%81 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%82 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%83 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%84 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%85 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%86 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%87 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%88 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%89 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%90 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%91 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%92 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%93 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%94 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%95 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%96 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%97 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%98 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%99 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%100 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%101 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%102 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%103 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%104 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%105 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%106 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%107 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%108 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%109 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%110 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%111 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%112 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%113 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%114 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%115 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%116 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%117 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%118 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%119 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%120 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%121 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%122 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%123 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%124 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%125 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%126 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%127 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%128 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%129 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%130 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%131 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%132 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%133 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%134 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%135 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%136 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%137 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%138 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%139 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%140 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%141 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%142 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%143 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%144 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%145 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%146 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%147 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%148 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%149 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%150 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%151 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x2304xf32>} : () -> tensor<1x768x2304xf32>
%152 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x2304xf32>} : () -> tensor<1x2304xf32>
%153 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x768xf32>} : () -> tensor<1x768x768xf32>
%154 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%155 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%156 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%157 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768x3072xf32>} : () -> tensor<1x768x3072xf32>
%158 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072xf32>} : () -> tensor<1x3072xf32>
%159 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x3072x768xf32>} : () -> tensor<1x3072x768xf32>
%160 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x768xf32>} : () -> tensor<1x768xf32>
%161 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%162 = "tosa.const"() {value = dense_resource<__elided__> : tensor<1x1x768xf32>} : () -> tensor<1x1x768xf32>
%163 = "tosa.reshape"(%3) {new_shape = [1, 50257, 768]} : (tensor<50257x768xf32>) -> tensor<1x50257x768xf32>
%164 = "tosa.cast"(%arg0) : (tensor<1x5xi64>) -> tensor<1x5xi32>
%165 = "tosa.gather"(%163, %164) : (tensor<1x50257x768xf32>, tensor<1x5xi32>) -> tensor<1x5x768xf32>
%166 = "tosa.cast"(%15) : (tensor<1x5xi64>) -> tensor<1x5xi32>
%167 = "tosa.gather"(%16, %166) : (tensor<1x1024x768xf32>, tensor<1x5xi32>) -> tensor<1x5x768xf32>
%168 = "tosa.add"(%165, %167) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%169 = "tosa.reciprocal"(%4) : (tensor<1xf32>) -> tensor<1xf32>
%170 = "tosa.reduce_sum"(%168) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%171 = "tosa.reshape"(%169) {new_shape = [1, 1, 1]} : (tensor<1xf32>) -> tensor<1x1x1xf32>
%172 = "tosa.mul"(%170, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%173 = "tosa.sub"(%168, %172) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%174 = "tosa.mul"(%173, %173) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%175 = "tosa.reduce_sum"(%174) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%176 = "tosa.mul"(%175, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%177 = "tosa.add"(%176, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%178 = "tosa.rsqrt"(%177) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%179 = "tosa.mul"(%173, %178) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%180 = "tosa.mul"(%179, %17) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%181 = "tosa.add"(%180, %18) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%182 = "tosa.matmul"(%181, %19) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%183 = "tosa.reshape"(%182) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%184 = "tosa.add"(%183, %20) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%185 = "tosa.reshape"(%184) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%186 = "tosa.slice"(%185) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%187 = "tosa.slice"(%185) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%188 = "tosa.slice"(%185) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%189 = "tosa.reshape"(%186) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%190 = "tosa.transpose"(%189, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%191 = "tosa.reshape"(%187) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%192 = "tosa.transpose"(%191, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%193 = "tosa.reshape"(%188) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%194 = "tosa.transpose"(%193, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%195 = "tosa.transpose"(%192, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%196 = "tosa.reshape"(%190) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%197 = "tosa.reshape"(%195) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%198 = "tosa.matmul"(%196, %197) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%199 = "tosa.reshape"(%198) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%200 = "tosa.reciprocal"(%2) : (tensor<f32>) -> tensor<f32>
%201 = "tosa.reshape"(%200) {new_shape = [1, 1, 1, 1]} : (tensor<f32>) -> tensor<1x1x1x1xf32>
%202 = "tosa.mul"(%199, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%203 = "tosa.slice"(%1) {size = [1, 1, 5, 1024], start = [0, 0, 0, 0]} : (tensor<1x1x1024x1024xi8>) -> tensor<1x1x5x1024xi8>
%204 = "tosa.slice"(%203) {size = [1, 1, 5, 5], start = [0, 0, 0, 0]} : (tensor<1x1x5x1024xi8>) -> tensor<1x1x5x5xi8>
%205 = "tosa.equal"(%204, %7) : (tensor<1x1x5x5xi8>, tensor<1x1x5x5xi8>) -> tensor<1x1x5x5xi1>
%206 = "tosa.select"(%205, %0, %202) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%207 = "tosa.reduce_max"(%206) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%208 = "tosa.sub"(%206, %207) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%209 = "tosa.exp"(%208) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%210 = "tosa.reduce_sum"(%209) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%211 = "tosa.reciprocal"(%210) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%212 = "tosa.mul"(%209, %211) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%213 = "tosa.reshape"(%212) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%214 = "tosa.reshape"(%194) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%215 = "tosa.matmul"(%213, %214) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%216 = "tosa.reshape"(%215) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%217 = "tosa.transpose"(%216, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%218 = "tosa.reshape"(%217) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%219 = "tosa.matmul"(%218, %21) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%220 = "tosa.reshape"(%219) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%221 = "tosa.add"(%220, %22) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%222 = "tosa.reshape"(%221) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%223 = "tosa.add"(%222, %168) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%224 = "tosa.reduce_sum"(%223) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%225 = "tosa.mul"(%224, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%226 = "tosa.sub"(%223, %225) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%227 = "tosa.mul"(%226, %226) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%228 = "tosa.reduce_sum"(%227) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%229 = "tosa.mul"(%228, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%230 = "tosa.add"(%229, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%231 = "tosa.rsqrt"(%230) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%232 = "tosa.mul"(%226, %231) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%233 = "tosa.mul"(%232, %23) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%234 = "tosa.add"(%233, %24) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%235 = "tosa.matmul"(%234, %25) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%236 = "tosa.reshape"(%235) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%237 = "tosa.add"(%236, %26) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%238 = "tosa.reshape"(%237) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%239 = "tosa.mul"(%238, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%240 = "tosa.pow"(%238, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%241 = "tosa.mul"(%240, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%242 = "tosa.add"(%238, %241) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%243 = "tosa.mul"(%242, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%244 = "tosa.tanh"(%243) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%245 = "tosa.add"(%244, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%246 = "tosa.mul"(%239, %245) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%247 = "tosa.matmul"(%246, %27) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%248 = "tosa.reshape"(%247) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%249 = "tosa.add"(%248, %28) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%250 = "tosa.reshape"(%249) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%251 = "tosa.add"(%223, %250) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%252 = "tosa.reduce_sum"(%251) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%253 = "tosa.mul"(%252, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%254 = "tosa.sub"(%251, %253) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%255 = "tosa.mul"(%254, %254) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%256 = "tosa.reduce_sum"(%255) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%257 = "tosa.mul"(%256, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%258 = "tosa.add"(%257, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%259 = "tosa.rsqrt"(%258) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%260 = "tosa.mul"(%254, %259) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%261 = "tosa.mul"(%260, %29) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%262 = "tosa.add"(%261, %30) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%263 = "tosa.matmul"(%262, %31) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%264 = "tosa.reshape"(%263) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%265 = "tosa.add"(%264, %32) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%266 = "tosa.reshape"(%265) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%267 = "tosa.slice"(%266) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%268 = "tosa.slice"(%266) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%269 = "tosa.slice"(%266) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%270 = "tosa.reshape"(%267) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%271 = "tosa.transpose"(%270, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%272 = "tosa.reshape"(%268) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%273 = "tosa.transpose"(%272, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%274 = "tosa.reshape"(%269) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%275 = "tosa.transpose"(%274, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%276 = "tosa.transpose"(%273, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%277 = "tosa.reshape"(%271) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%278 = "tosa.reshape"(%276) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%279 = "tosa.matmul"(%277, %278) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%280 = "tosa.reshape"(%279) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%281 = "tosa.mul"(%280, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%282 = "tosa.select"(%205, %0, %281) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%283 = "tosa.reduce_max"(%282) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%284 = "tosa.sub"(%282, %283) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%285 = "tosa.exp"(%284) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%286 = "tosa.reduce_sum"(%285) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%287 = "tosa.reciprocal"(%286) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%288 = "tosa.mul"(%285, %287) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%289 = "tosa.reshape"(%288) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%290 = "tosa.reshape"(%275) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%291 = "tosa.matmul"(%289, %290) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%292 = "tosa.reshape"(%291) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%293 = "tosa.transpose"(%292, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%294 = "tosa.reshape"(%293) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%295 = "tosa.matmul"(%294, %33) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%296 = "tosa.reshape"(%295) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%297 = "tosa.add"(%296, %34) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%298 = "tosa.reshape"(%297) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%299 = "tosa.add"(%298, %251) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%300 = "tosa.reduce_sum"(%299) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%301 = "tosa.mul"(%300, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%302 = "tosa.sub"(%299, %301) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%303 = "tosa.mul"(%302, %302) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%304 = "tosa.reduce_sum"(%303) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%305 = "tosa.mul"(%304, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%306 = "tosa.add"(%305, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%307 = "tosa.rsqrt"(%306) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%308 = "tosa.mul"(%302, %307) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%309 = "tosa.mul"(%308, %35) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%310 = "tosa.add"(%309, %36) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%311 = "tosa.matmul"(%310, %37) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%312 = "tosa.reshape"(%311) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%313 = "tosa.add"(%312, %38) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%314 = "tosa.reshape"(%313) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%315 = "tosa.mul"(%314, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%316 = "tosa.pow"(%314, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%317 = "tosa.mul"(%316, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%318 = "tosa.add"(%314, %317) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%319 = "tosa.mul"(%318, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%320 = "tosa.tanh"(%319) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%321 = "tosa.add"(%320, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%322 = "tosa.mul"(%315, %321) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%323 = "tosa.matmul"(%322, %39) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%324 = "tosa.reshape"(%323) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%325 = "tosa.add"(%324, %40) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%326 = "tosa.reshape"(%325) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%327 = "tosa.add"(%299, %326) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%328 = "tosa.reduce_sum"(%327) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%329 = "tosa.mul"(%328, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%330 = "tosa.sub"(%327, %329) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%331 = "tosa.mul"(%330, %330) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%332 = "tosa.reduce_sum"(%331) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%333 = "tosa.mul"(%332, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%334 = "tosa.add"(%333, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%335 = "tosa.rsqrt"(%334) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%336 = "tosa.mul"(%330, %335) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%337 = "tosa.mul"(%336, %41) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%338 = "tosa.add"(%337, %42) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%339 = "tosa.matmul"(%338, %43) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%340 = "tosa.reshape"(%339) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%341 = "tosa.add"(%340, %44) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%342 = "tosa.reshape"(%341) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%343 = "tosa.slice"(%342) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%344 = "tosa.slice"(%342) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%345 = "tosa.slice"(%342) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%346 = "tosa.reshape"(%343) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%347 = "tosa.transpose"(%346, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%348 = "tosa.reshape"(%344) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%349 = "tosa.transpose"(%348, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%350 = "tosa.reshape"(%345) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%351 = "tosa.transpose"(%350, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%352 = "tosa.transpose"(%349, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%353 = "tosa.reshape"(%347) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%354 = "tosa.reshape"(%352) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%355 = "tosa.matmul"(%353, %354) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%356 = "tosa.reshape"(%355) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%357 = "tosa.mul"(%356, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%358 = "tosa.select"(%205, %0, %357) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%359 = "tosa.reduce_max"(%358) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%360 = "tosa.sub"(%358, %359) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%361 = "tosa.exp"(%360) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%362 = "tosa.reduce_sum"(%361) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%363 = "tosa.reciprocal"(%362) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%364 = "tosa.mul"(%361, %363) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%365 = "tosa.reshape"(%364) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%366 = "tosa.reshape"(%351) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%367 = "tosa.matmul"(%365, %366) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%368 = "tosa.reshape"(%367) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%369 = "tosa.transpose"(%368, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%370 = "tosa.reshape"(%369) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%371 = "tosa.matmul"(%370, %45) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%372 = "tosa.reshape"(%371) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%373 = "tosa.add"(%372, %46) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%374 = "tosa.reshape"(%373) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%375 = "tosa.add"(%374, %327) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%376 = "tosa.reduce_sum"(%375) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%377 = "tosa.mul"(%376, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%378 = "tosa.sub"(%375, %377) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%379 = "tosa.mul"(%378, %378) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%380 = "tosa.reduce_sum"(%379) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%381 = "tosa.mul"(%380, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%382 = "tosa.add"(%381, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%383 = "tosa.rsqrt"(%382) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%384 = "tosa.mul"(%378, %383) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%385 = "tosa.mul"(%384, %47) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%386 = "tosa.add"(%385, %48) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%387 = "tosa.matmul"(%386, %49) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%388 = "tosa.reshape"(%387) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%389 = "tosa.add"(%388, %50) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%390 = "tosa.reshape"(%389) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%391 = "tosa.mul"(%390, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%392 = "tosa.pow"(%390, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%393 = "tosa.mul"(%392, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%394 = "tosa.add"(%390, %393) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%395 = "tosa.mul"(%394, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%396 = "tosa.tanh"(%395) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%397 = "tosa.add"(%396, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%398 = "tosa.mul"(%391, %397) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%399 = "tosa.matmul"(%398, %51) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%400 = "tosa.reshape"(%399) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%401 = "tosa.add"(%400, %52) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%402 = "tosa.reshape"(%401) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%403 = "tosa.add"(%375, %402) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%404 = "tosa.reduce_sum"(%403) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%405 = "tosa.mul"(%404, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%406 = "tosa.sub"(%403, %405) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%407 = "tosa.mul"(%406, %406) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%408 = "tosa.reduce_sum"(%407) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%409 = "tosa.mul"(%408, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%410 = "tosa.add"(%409, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%411 = "tosa.rsqrt"(%410) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%412 = "tosa.mul"(%406, %411) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%413 = "tosa.mul"(%412, %53) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%414 = "tosa.add"(%413, %54) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%415 = "tosa.matmul"(%414, %55) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%416 = "tosa.reshape"(%415) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%417 = "tosa.add"(%416, %56) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%418 = "tosa.reshape"(%417) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%419 = "tosa.slice"(%418) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%420 = "tosa.slice"(%418) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%421 = "tosa.slice"(%418) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%422 = "tosa.reshape"(%419) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%423 = "tosa.transpose"(%422, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%424 = "tosa.reshape"(%420) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%425 = "tosa.transpose"(%424, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%426 = "tosa.reshape"(%421) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%427 = "tosa.transpose"(%426, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%428 = "tosa.transpose"(%425, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%429 = "tosa.reshape"(%423) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%430 = "tosa.reshape"(%428) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%431 = "tosa.matmul"(%429, %430) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%432 = "tosa.reshape"(%431) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%433 = "tosa.mul"(%432, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%434 = "tosa.select"(%205, %0, %433) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%435 = "tosa.reduce_max"(%434) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%436 = "tosa.sub"(%434, %435) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%437 = "tosa.exp"(%436) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%438 = "tosa.reduce_sum"(%437) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%439 = "tosa.reciprocal"(%438) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%440 = "tosa.mul"(%437, %439) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%441 = "tosa.reshape"(%440) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%442 = "tosa.reshape"(%427) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%443 = "tosa.matmul"(%441, %442) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%444 = "tosa.reshape"(%443) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%445 = "tosa.transpose"(%444, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%446 = "tosa.reshape"(%445) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%447 = "tosa.matmul"(%446, %57) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%448 = "tosa.reshape"(%447) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%449 = "tosa.add"(%448, %58) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%450 = "tosa.reshape"(%449) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%451 = "tosa.add"(%450, %403) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%452 = "tosa.reduce_sum"(%451) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%453 = "tosa.mul"(%452, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%454 = "tosa.sub"(%451, %453) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%455 = "tosa.mul"(%454, %454) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%456 = "tosa.reduce_sum"(%455) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%457 = "tosa.mul"(%456, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%458 = "tosa.add"(%457, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%459 = "tosa.rsqrt"(%458) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%460 = "tosa.mul"(%454, %459) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%461 = "tosa.mul"(%460, %59) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%462 = "tosa.add"(%461, %60) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%463 = "tosa.matmul"(%462, %61) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%464 = "tosa.reshape"(%463) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%465 = "tosa.add"(%464, %62) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%466 = "tosa.reshape"(%465) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%467 = "tosa.mul"(%466, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%468 = "tosa.pow"(%466, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%469 = "tosa.mul"(%468, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%470 = "tosa.add"(%466, %469) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%471 = "tosa.mul"(%470, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%472 = "tosa.tanh"(%471) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%473 = "tosa.add"(%472, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%474 = "tosa.mul"(%467, %473) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%475 = "tosa.matmul"(%474, %63) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%476 = "tosa.reshape"(%475) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%477 = "tosa.add"(%476, %64) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%478 = "tosa.reshape"(%477) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%479 = "tosa.add"(%451, %478) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%480 = "tosa.reduce_sum"(%479) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%481 = "tosa.mul"(%480, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%482 = "tosa.sub"(%479, %481) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%483 = "tosa.mul"(%482, %482) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%484 = "tosa.reduce_sum"(%483) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%485 = "tosa.mul"(%484, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%486 = "tosa.add"(%485, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%487 = "tosa.rsqrt"(%486) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%488 = "tosa.mul"(%482, %487) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%489 = "tosa.mul"(%488, %65) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%490 = "tosa.add"(%489, %66) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%491 = "tosa.matmul"(%490, %67) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%492 = "tosa.reshape"(%491) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%493 = "tosa.add"(%492, %68) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%494 = "tosa.reshape"(%493) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%495 = "tosa.slice"(%494) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%496 = "tosa.slice"(%494) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%497 = "tosa.slice"(%494) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%498 = "tosa.reshape"(%495) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%499 = "tosa.transpose"(%498, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%500 = "tosa.reshape"(%496) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%501 = "tosa.transpose"(%500, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%502 = "tosa.reshape"(%497) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%503 = "tosa.transpose"(%502, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%504 = "tosa.transpose"(%501, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%505 = "tosa.reshape"(%499) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%506 = "tosa.reshape"(%504) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%507 = "tosa.matmul"(%505, %506) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%508 = "tosa.reshape"(%507) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%509 = "tosa.mul"(%508, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%510 = "tosa.select"(%205, %0, %509) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%511 = "tosa.reduce_max"(%510) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%512 = "tosa.sub"(%510, %511) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%513 = "tosa.exp"(%512) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%514 = "tosa.reduce_sum"(%513) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%515 = "tosa.reciprocal"(%514) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%516 = "tosa.mul"(%513, %515) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%517 = "tosa.reshape"(%516) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%518 = "tosa.reshape"(%503) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%519 = "tosa.matmul"(%517, %518) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%520 = "tosa.reshape"(%519) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%521 = "tosa.transpose"(%520, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%522 = "tosa.reshape"(%521) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%523 = "tosa.matmul"(%522, %69) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%524 = "tosa.reshape"(%523) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%525 = "tosa.add"(%524, %70) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%526 = "tosa.reshape"(%525) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%527 = "tosa.add"(%526, %479) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%528 = "tosa.reduce_sum"(%527) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%529 = "tosa.mul"(%528, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%530 = "tosa.sub"(%527, %529) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%531 = "tosa.mul"(%530, %530) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%532 = "tosa.reduce_sum"(%531) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%533 = "tosa.mul"(%532, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%534 = "tosa.add"(%533, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%535 = "tosa.rsqrt"(%534) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%536 = "tosa.mul"(%530, %535) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%537 = "tosa.mul"(%536, %71) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%538 = "tosa.add"(%537, %72) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%539 = "tosa.matmul"(%538, %73) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%540 = "tosa.reshape"(%539) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%541 = "tosa.add"(%540, %74) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%542 = "tosa.reshape"(%541) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%543 = "tosa.mul"(%542, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%544 = "tosa.pow"(%542, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%545 = "tosa.mul"(%544, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%546 = "tosa.add"(%542, %545) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%547 = "tosa.mul"(%546, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%548 = "tosa.tanh"(%547) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%549 = "tosa.add"(%548, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%550 = "tosa.mul"(%543, %549) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%551 = "tosa.matmul"(%550, %75) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%552 = "tosa.reshape"(%551) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%553 = "tosa.add"(%552, %76) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%554 = "tosa.reshape"(%553) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%555 = "tosa.add"(%527, %554) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%556 = "tosa.reduce_sum"(%555) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%557 = "tosa.mul"(%556, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%558 = "tosa.sub"(%555, %557) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%559 = "tosa.mul"(%558, %558) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%560 = "tosa.reduce_sum"(%559) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%561 = "tosa.mul"(%560, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%562 = "tosa.add"(%561, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%563 = "tosa.rsqrt"(%562) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%564 = "tosa.mul"(%558, %563) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%565 = "tosa.mul"(%564, %77) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%566 = "tosa.add"(%565, %78) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%567 = "tosa.matmul"(%566, %79) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%568 = "tosa.reshape"(%567) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%569 = "tosa.add"(%568, %80) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%570 = "tosa.reshape"(%569) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%571 = "tosa.slice"(%570) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%572 = "tosa.slice"(%570) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%573 = "tosa.slice"(%570) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%574 = "tosa.reshape"(%571) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%575 = "tosa.transpose"(%574, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%576 = "tosa.reshape"(%572) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%577 = "tosa.transpose"(%576, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%578 = "tosa.reshape"(%573) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%579 = "tosa.transpose"(%578, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%580 = "tosa.transpose"(%577, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%581 = "tosa.reshape"(%575) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%582 = "tosa.reshape"(%580) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%583 = "tosa.matmul"(%581, %582) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%584 = "tosa.reshape"(%583) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%585 = "tosa.mul"(%584, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%586 = "tosa.select"(%205, %0, %585) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%587 = "tosa.reduce_max"(%586) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%588 = "tosa.sub"(%586, %587) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%589 = "tosa.exp"(%588) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%590 = "tosa.reduce_sum"(%589) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%591 = "tosa.reciprocal"(%590) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%592 = "tosa.mul"(%589, %591) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%593 = "tosa.reshape"(%592) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%594 = "tosa.reshape"(%579) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%595 = "tosa.matmul"(%593, %594) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%596 = "tosa.reshape"(%595) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%597 = "tosa.transpose"(%596, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%598 = "tosa.reshape"(%597) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%599 = "tosa.matmul"(%598, %81) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%600 = "tosa.reshape"(%599) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%601 = "tosa.add"(%600, %82) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%602 = "tosa.reshape"(%601) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%603 = "tosa.add"(%602, %555) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%604 = "tosa.reduce_sum"(%603) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%605 = "tosa.mul"(%604, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%606 = "tosa.sub"(%603, %605) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%607 = "tosa.mul"(%606, %606) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%608 = "tosa.reduce_sum"(%607) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%609 = "tosa.mul"(%608, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%610 = "tosa.add"(%609, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%611 = "tosa.rsqrt"(%610) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%612 = "tosa.mul"(%606, %611) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%613 = "tosa.mul"(%612, %83) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%614 = "tosa.add"(%613, %84) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%615 = "tosa.matmul"(%614, %85) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%616 = "tosa.reshape"(%615) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%617 = "tosa.add"(%616, %86) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%618 = "tosa.reshape"(%617) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%619 = "tosa.mul"(%618, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%620 = "tosa.pow"(%618, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%621 = "tosa.mul"(%620, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%622 = "tosa.add"(%618, %621) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%623 = "tosa.mul"(%622, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%624 = "tosa.tanh"(%623) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%625 = "tosa.add"(%624, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%626 = "tosa.mul"(%619, %625) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%627 = "tosa.matmul"(%626, %87) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%628 = "tosa.reshape"(%627) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%629 = "tosa.add"(%628, %88) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%630 = "tosa.reshape"(%629) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%631 = "tosa.add"(%603, %630) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%632 = "tosa.reduce_sum"(%631) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%633 = "tosa.mul"(%632, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%634 = "tosa.sub"(%631, %633) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%635 = "tosa.mul"(%634, %634) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%636 = "tosa.reduce_sum"(%635) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%637 = "tosa.mul"(%636, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%638 = "tosa.add"(%637, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%639 = "tosa.rsqrt"(%638) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%640 = "tosa.mul"(%634, %639) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%641 = "tosa.mul"(%640, %89) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%642 = "tosa.add"(%641, %90) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%643 = "tosa.matmul"(%642, %91) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%644 = "tosa.reshape"(%643) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%645 = "tosa.add"(%644, %92) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%646 = "tosa.reshape"(%645) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%647 = "tosa.slice"(%646) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%648 = "tosa.slice"(%646) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%649 = "tosa.slice"(%646) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%650 = "tosa.reshape"(%647) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%651 = "tosa.transpose"(%650, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%652 = "tosa.reshape"(%648) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%653 = "tosa.transpose"(%652, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%654 = "tosa.reshape"(%649) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%655 = "tosa.transpose"(%654, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%656 = "tosa.transpose"(%653, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%657 = "tosa.reshape"(%651) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%658 = "tosa.reshape"(%656) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%659 = "tosa.matmul"(%657, %658) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%660 = "tosa.reshape"(%659) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%661 = "tosa.mul"(%660, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%662 = "tosa.select"(%205, %0, %661) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%663 = "tosa.reduce_max"(%662) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%664 = "tosa.sub"(%662, %663) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%665 = "tosa.exp"(%664) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%666 = "tosa.reduce_sum"(%665) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%667 = "tosa.reciprocal"(%666) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%668 = "tosa.mul"(%665, %667) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%669 = "tosa.reshape"(%668) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%670 = "tosa.reshape"(%655) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%671 = "tosa.matmul"(%669, %670) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%672 = "tosa.reshape"(%671) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%673 = "tosa.transpose"(%672, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%674 = "tosa.reshape"(%673) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%675 = "tosa.matmul"(%674, %93) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%676 = "tosa.reshape"(%675) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%677 = "tosa.add"(%676, %94) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%678 = "tosa.reshape"(%677) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%679 = "tosa.add"(%678, %631) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%680 = "tosa.reduce_sum"(%679) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%681 = "tosa.mul"(%680, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%682 = "tosa.sub"(%679, %681) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%683 = "tosa.mul"(%682, %682) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%684 = "tosa.reduce_sum"(%683) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%685 = "tosa.mul"(%684, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%686 = "tosa.add"(%685, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%687 = "tosa.rsqrt"(%686) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%688 = "tosa.mul"(%682, %687) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%689 = "tosa.mul"(%688, %95) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%690 = "tosa.add"(%689, %96) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%691 = "tosa.matmul"(%690, %97) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%692 = "tosa.reshape"(%691) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%693 = "tosa.add"(%692, %98) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%694 = "tosa.reshape"(%693) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%695 = "tosa.mul"(%694, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%696 = "tosa.pow"(%694, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%697 = "tosa.mul"(%696, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%698 = "tosa.add"(%694, %697) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%699 = "tosa.mul"(%698, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%700 = "tosa.tanh"(%699) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%701 = "tosa.add"(%700, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%702 = "tosa.mul"(%695, %701) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%703 = "tosa.matmul"(%702, %99) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%704 = "tosa.reshape"(%703) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%705 = "tosa.add"(%704, %100) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%706 = "tosa.reshape"(%705) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%707 = "tosa.add"(%679, %706) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%708 = "tosa.reduce_sum"(%707) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%709 = "tosa.mul"(%708, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%710 = "tosa.sub"(%707, %709) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%711 = "tosa.mul"(%710, %710) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%712 = "tosa.reduce_sum"(%711) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%713 = "tosa.mul"(%712, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%714 = "tosa.add"(%713, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%715 = "tosa.rsqrt"(%714) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%716 = "tosa.mul"(%710, %715) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%717 = "tosa.mul"(%716, %101) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%718 = "tosa.add"(%717, %102) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%719 = "tosa.matmul"(%718, %103) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%720 = "tosa.reshape"(%719) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%721 = "tosa.add"(%720, %104) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%722 = "tosa.reshape"(%721) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%723 = "tosa.slice"(%722) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%724 = "tosa.slice"(%722) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%725 = "tosa.slice"(%722) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%726 = "tosa.reshape"(%723) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%727 = "tosa.transpose"(%726, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%728 = "tosa.reshape"(%724) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%729 = "tosa.transpose"(%728, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%730 = "tosa.reshape"(%725) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%731 = "tosa.transpose"(%730, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%732 = "tosa.transpose"(%729, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%733 = "tosa.reshape"(%727) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%734 = "tosa.reshape"(%732) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%735 = "tosa.matmul"(%733, %734) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%736 = "tosa.reshape"(%735) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%737 = "tosa.mul"(%736, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%738 = "tosa.select"(%205, %0, %737) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%739 = "tosa.reduce_max"(%738) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%740 = "tosa.sub"(%738, %739) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%741 = "tosa.exp"(%740) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%742 = "tosa.reduce_sum"(%741) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%743 = "tosa.reciprocal"(%742) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%744 = "tosa.mul"(%741, %743) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%745 = "tosa.reshape"(%744) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%746 = "tosa.reshape"(%731) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%747 = "tosa.matmul"(%745, %746) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%748 = "tosa.reshape"(%747) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%749 = "tosa.transpose"(%748, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%750 = "tosa.reshape"(%749) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%751 = "tosa.matmul"(%750, %105) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%752 = "tosa.reshape"(%751) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%753 = "tosa.add"(%752, %106) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%754 = "tosa.reshape"(%753) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%755 = "tosa.add"(%754, %707) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%756 = "tosa.reduce_sum"(%755) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%757 = "tosa.mul"(%756, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%758 = "tosa.sub"(%755, %757) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%759 = "tosa.mul"(%758, %758) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%760 = "tosa.reduce_sum"(%759) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%761 = "tosa.mul"(%760, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%762 = "tosa.add"(%761, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%763 = "tosa.rsqrt"(%762) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%764 = "tosa.mul"(%758, %763) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%765 = "tosa.mul"(%764, %107) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%766 = "tosa.add"(%765, %108) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%767 = "tosa.matmul"(%766, %109) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%768 = "tosa.reshape"(%767) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%769 = "tosa.add"(%768, %110) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%770 = "tosa.reshape"(%769) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%771 = "tosa.mul"(%770, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%772 = "tosa.pow"(%770, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%773 = "tosa.mul"(%772, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%774 = "tosa.add"(%770, %773) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%775 = "tosa.mul"(%774, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%776 = "tosa.tanh"(%775) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%777 = "tosa.add"(%776, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%778 = "tosa.mul"(%771, %777) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%779 = "tosa.matmul"(%778, %111) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%780 = "tosa.reshape"(%779) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%781 = "tosa.add"(%780, %112) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%782 = "tosa.reshape"(%781) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%783 = "tosa.add"(%755, %782) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%784 = "tosa.reduce_sum"(%783) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%785 = "tosa.mul"(%784, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%786 = "tosa.sub"(%783, %785) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%787 = "tosa.mul"(%786, %786) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%788 = "tosa.reduce_sum"(%787) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%789 = "tosa.mul"(%788, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%790 = "tosa.add"(%789, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%791 = "tosa.rsqrt"(%790) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%792 = "tosa.mul"(%786, %791) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%793 = "tosa.mul"(%792, %113) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%794 = "tosa.add"(%793, %114) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%795 = "tosa.matmul"(%794, %115) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%796 = "tosa.reshape"(%795) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%797 = "tosa.add"(%796, %116) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%798 = "tosa.reshape"(%797) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%799 = "tosa.slice"(%798) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%800 = "tosa.slice"(%798) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%801 = "tosa.slice"(%798) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%802 = "tosa.reshape"(%799) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%803 = "tosa.transpose"(%802, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%804 = "tosa.reshape"(%800) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%805 = "tosa.transpose"(%804, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%806 = "tosa.reshape"(%801) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%807 = "tosa.transpose"(%806, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%808 = "tosa.transpose"(%805, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%809 = "tosa.reshape"(%803) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%810 = "tosa.reshape"(%808) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%811 = "tosa.matmul"(%809, %810) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%812 = "tosa.reshape"(%811) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%813 = "tosa.mul"(%812, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%814 = "tosa.select"(%205, %0, %813) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%815 = "tosa.reduce_max"(%814) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%816 = "tosa.sub"(%814, %815) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%817 = "tosa.exp"(%816) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%818 = "tosa.reduce_sum"(%817) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%819 = "tosa.reciprocal"(%818) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%820 = "tosa.mul"(%817, %819) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%821 = "tosa.reshape"(%820) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%822 = "tosa.reshape"(%807) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%823 = "tosa.matmul"(%821, %822) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%824 = "tosa.reshape"(%823) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%825 = "tosa.transpose"(%824, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%826 = "tosa.reshape"(%825) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%827 = "tosa.matmul"(%826, %117) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%828 = "tosa.reshape"(%827) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%829 = "tosa.add"(%828, %118) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%830 = "tosa.reshape"(%829) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%831 = "tosa.add"(%830, %783) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%832 = "tosa.reduce_sum"(%831) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%833 = "tosa.mul"(%832, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%834 = "tosa.sub"(%831, %833) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%835 = "tosa.mul"(%834, %834) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%836 = "tosa.reduce_sum"(%835) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%837 = "tosa.mul"(%836, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%838 = "tosa.add"(%837, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%839 = "tosa.rsqrt"(%838) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%840 = "tosa.mul"(%834, %839) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%841 = "tosa.mul"(%840, %119) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%842 = "tosa.add"(%841, %120) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%843 = "tosa.matmul"(%842, %121) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%844 = "tosa.reshape"(%843) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%845 = "tosa.add"(%844, %122) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%846 = "tosa.reshape"(%845) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%847 = "tosa.mul"(%846, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%848 = "tosa.pow"(%846, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%849 = "tosa.mul"(%848, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%850 = "tosa.add"(%846, %849) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%851 = "tosa.mul"(%850, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%852 = "tosa.tanh"(%851) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%853 = "tosa.add"(%852, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%854 = "tosa.mul"(%847, %853) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%855 = "tosa.matmul"(%854, %123) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%856 = "tosa.reshape"(%855) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%857 = "tosa.add"(%856, %124) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%858 = "tosa.reshape"(%857) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%859 = "tosa.add"(%831, %858) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%860 = "tosa.reduce_sum"(%859) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%861 = "tosa.mul"(%860, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%862 = "tosa.sub"(%859, %861) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%863 = "tosa.mul"(%862, %862) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%864 = "tosa.reduce_sum"(%863) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%865 = "tosa.mul"(%864, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%866 = "tosa.add"(%865, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%867 = "tosa.rsqrt"(%866) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%868 = "tosa.mul"(%862, %867) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%869 = "tosa.mul"(%868, %125) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%870 = "tosa.add"(%869, %126) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%871 = "tosa.matmul"(%870, %127) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%872 = "tosa.reshape"(%871) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%873 = "tosa.add"(%872, %128) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%874 = "tosa.reshape"(%873) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%875 = "tosa.slice"(%874) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%876 = "tosa.slice"(%874) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%877 = "tosa.slice"(%874) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%878 = "tosa.reshape"(%875) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%879 = "tosa.transpose"(%878, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%880 = "tosa.reshape"(%876) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%881 = "tosa.transpose"(%880, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%882 = "tosa.reshape"(%877) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%883 = "tosa.transpose"(%882, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%884 = "tosa.transpose"(%881, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%885 = "tosa.reshape"(%879) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%886 = "tosa.reshape"(%884) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%887 = "tosa.matmul"(%885, %886) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%888 = "tosa.reshape"(%887) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%889 = "tosa.mul"(%888, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%890 = "tosa.select"(%205, %0, %889) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%891 = "tosa.reduce_max"(%890) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%892 = "tosa.sub"(%890, %891) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%893 = "tosa.exp"(%892) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%894 = "tosa.reduce_sum"(%893) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%895 = "tosa.reciprocal"(%894) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%896 = "tosa.mul"(%893, %895) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%897 = "tosa.reshape"(%896) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%898 = "tosa.reshape"(%883) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%899 = "tosa.matmul"(%897, %898) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%900 = "tosa.reshape"(%899) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%901 = "tosa.transpose"(%900, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%902 = "tosa.reshape"(%901) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%903 = "tosa.matmul"(%902, %129) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%904 = "tosa.reshape"(%903) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%905 = "tosa.add"(%904, %130) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%906 = "tosa.reshape"(%905) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%907 = "tosa.add"(%906, %859) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%908 = "tosa.reduce_sum"(%907) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%909 = "tosa.mul"(%908, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%910 = "tosa.sub"(%907, %909) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%911 = "tosa.mul"(%910, %910) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%912 = "tosa.reduce_sum"(%911) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%913 = "tosa.mul"(%912, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%914 = "tosa.add"(%913, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%915 = "tosa.rsqrt"(%914) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%916 = "tosa.mul"(%910, %915) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%917 = "tosa.mul"(%916, %131) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%918 = "tosa.add"(%917, %132) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%919 = "tosa.matmul"(%918, %133) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%920 = "tosa.reshape"(%919) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%921 = "tosa.add"(%920, %134) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%922 = "tosa.reshape"(%921) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%923 = "tosa.mul"(%922, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%924 = "tosa.pow"(%922, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%925 = "tosa.mul"(%924, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%926 = "tosa.add"(%922, %925) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%927 = "tosa.mul"(%926, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%928 = "tosa.tanh"(%927) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%929 = "tosa.add"(%928, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%930 = "tosa.mul"(%923, %929) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%931 = "tosa.matmul"(%930, %135) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%932 = "tosa.reshape"(%931) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%933 = "tosa.add"(%932, %136) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%934 = "tosa.reshape"(%933) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%935 = "tosa.add"(%907, %934) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%936 = "tosa.reduce_sum"(%935) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%937 = "tosa.mul"(%936, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%938 = "tosa.sub"(%935, %937) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%939 = "tosa.mul"(%938, %938) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%940 = "tosa.reduce_sum"(%939) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%941 = "tosa.mul"(%940, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%942 = "tosa.add"(%941, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%943 = "tosa.rsqrt"(%942) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%944 = "tosa.mul"(%938, %943) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%945 = "tosa.mul"(%944, %137) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%946 = "tosa.add"(%945, %138) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%947 = "tosa.matmul"(%946, %139) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%948 = "tosa.reshape"(%947) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%949 = "tosa.add"(%948, %140) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%950 = "tosa.reshape"(%949) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%951 = "tosa.slice"(%950) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%952 = "tosa.slice"(%950) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%953 = "tosa.slice"(%950) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%954 = "tosa.reshape"(%951) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%955 = "tosa.transpose"(%954, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%956 = "tosa.reshape"(%952) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%957 = "tosa.transpose"(%956, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%958 = "tosa.reshape"(%953) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%959 = "tosa.transpose"(%958, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%960 = "tosa.transpose"(%957, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%961 = "tosa.reshape"(%955) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%962 = "tosa.reshape"(%960) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%963 = "tosa.matmul"(%961, %962) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%964 = "tosa.reshape"(%963) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%965 = "tosa.mul"(%964, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%966 = "tosa.select"(%205, %0, %965) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%967 = "tosa.reduce_max"(%966) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%968 = "tosa.sub"(%966, %967) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%969 = "tosa.exp"(%968) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%970 = "tosa.reduce_sum"(%969) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%971 = "tosa.reciprocal"(%970) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%972 = "tosa.mul"(%969, %971) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%973 = "tosa.reshape"(%972) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%974 = "tosa.reshape"(%959) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%975 = "tosa.matmul"(%973, %974) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%976 = "tosa.reshape"(%975) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%977 = "tosa.transpose"(%976, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%978 = "tosa.reshape"(%977) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%979 = "tosa.matmul"(%978, %141) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%980 = "tosa.reshape"(%979) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%981 = "tosa.add"(%980, %142) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%982 = "tosa.reshape"(%981) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%983 = "tosa.add"(%982, %935) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%984 = "tosa.reduce_sum"(%983) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%985 = "tosa.mul"(%984, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%986 = "tosa.sub"(%983, %985) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%987 = "tosa.mul"(%986, %986) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%988 = "tosa.reduce_sum"(%987) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%989 = "tosa.mul"(%988, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%990 = "tosa.add"(%989, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%991 = "tosa.rsqrt"(%990) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%992 = "tosa.mul"(%986, %991) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%993 = "tosa.mul"(%992, %143) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%994 = "tosa.add"(%993, %144) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%995 = "tosa.matmul"(%994, %145) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%996 = "tosa.reshape"(%995) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%997 = "tosa.add"(%996, %146) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%998 = "tosa.reshape"(%997) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%999 = "tosa.mul"(%998, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1000 = "tosa.pow"(%998, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1001 = "tosa.mul"(%1000, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1002 = "tosa.add"(%998, %1001) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%1003 = "tosa.mul"(%1002, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1004 = "tosa.tanh"(%1003) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%1005 = "tosa.add"(%1004, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1006 = "tosa.mul"(%999, %1005) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%1007 = "tosa.matmul"(%1006, %147) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%1008 = "tosa.reshape"(%1007) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%1009 = "tosa.add"(%1008, %148) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%1010 = "tosa.reshape"(%1009) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%1011 = "tosa.add"(%983, %1010) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%1012 = "tosa.reduce_sum"(%1011) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%1013 = "tosa.mul"(%1012, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1014 = "tosa.sub"(%1011, %1013) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%1015 = "tosa.mul"(%1014, %1014) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%1016 = "tosa.reduce_sum"(%1015) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%1017 = "tosa.mul"(%1016, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1018 = "tosa.add"(%1017, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1019 = "tosa.rsqrt"(%1018) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%1020 = "tosa.mul"(%1014, %1019) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%1021 = "tosa.mul"(%1020, %149) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%1022 = "tosa.add"(%1021, %150) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%1023 = "tosa.matmul"(%1022, %151) : (tensor<1x5x768xf32>, tensor<1x768x2304xf32>) -> tensor<1x5x2304xf32>
%1024 = "tosa.reshape"(%1023) {new_shape = [5, 2304]} : (tensor<1x5x2304xf32>) -> tensor<5x2304xf32>
%1025 = "tosa.add"(%1024, %152) : (tensor<5x2304xf32>, tensor<1x2304xf32>) -> tensor<5x2304xf32>
%1026 = "tosa.reshape"(%1025) {new_shape = [1, 5, 2304]} : (tensor<5x2304xf32>) -> tensor<1x5x2304xf32>
%1027 = "tosa.slice"(%1026) {size = [1, 5, 768], start = [0, 0, 0]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%1028 = "tosa.slice"(%1026) {size = [1, 5, 768], start = [0, 0, 768]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%1029 = "tosa.slice"(%1026) {size = [1, 5, 768], start = [0, 0, 1536]} : (tensor<1x5x2304xf32>) -> tensor<1x5x768xf32>
%1030 = "tosa.reshape"(%1027) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%1031 = "tosa.transpose"(%1030, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%1032 = "tosa.reshape"(%1028) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%1033 = "tosa.transpose"(%1032, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%1034 = "tosa.reshape"(%1029) {new_shape = [1, 5, 12, 64]} : (tensor<1x5x768xf32>) -> tensor<1x5x12x64xf32>
%1035 = "tosa.transpose"(%1034, %5) : (tensor<1x5x12x64xf32>, tensor<4xi64>) -> tensor<1x12x5x64xf32>
%1036 = "tosa.transpose"(%1033, %6) : (tensor<1x12x5x64xf32>, tensor<4xi32>) -> tensor<1x12x64x5xf32>
%1037 = "tosa.reshape"(%1031) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%1038 = "tosa.reshape"(%1036) {new_shape = [12, 64, 5]} : (tensor<1x12x64x5xf32>) -> tensor<12x64x5xf32>
%1039 = "tosa.matmul"(%1037, %1038) : (tensor<12x5x64xf32>, tensor<12x64x5xf32>) -> tensor<12x5x5xf32>
%1040 = "tosa.reshape"(%1039) {new_shape = [1, 12, 5, 5]} : (tensor<12x5x5xf32>) -> tensor<1x12x5x5xf32>
%1041 = "tosa.mul"(%1040, %201) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x1x1x1xf32>) -> tensor<1x12x5x5xf32>
%1042 = "tosa.select"(%205, %0, %1041) : (tensor<1x1x5x5xi1>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%1043 = "tosa.reduce_max"(%1042) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%1044 = "tosa.sub"(%1042, %1043) : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%1045 = "tosa.exp"(%1044) : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>
%1046 = "tosa.reduce_sum"(%1045) {axis = 3 : i64} : (tensor<1x12x5x5xf32>) -> tensor<1x12x5x1xf32>
%1047 = "tosa.reciprocal"(%1046) : (tensor<1x12x5x1xf32>) -> tensor<1x12x5x1xf32>
%1048 = "tosa.mul"(%1045, %1047) {shift = 0 : i32} : (tensor<1x12x5x5xf32>, tensor<1x12x5x1xf32>) -> tensor<1x12x5x5xf32>
%1049 = "tosa.reshape"(%1048) {new_shape = [12, 5, 5]} : (tensor<1x12x5x5xf32>) -> tensor<12x5x5xf32>
%1050 = "tosa.reshape"(%1035) {new_shape = [12, 5, 64]} : (tensor<1x12x5x64xf32>) -> tensor<12x5x64xf32>
%1051 = "tosa.matmul"(%1049, %1050) : (tensor<12x5x5xf32>, tensor<12x5x64xf32>) -> tensor<12x5x64xf32>
%1052 = "tosa.reshape"(%1051) {new_shape = [1, 12, 5, 64]} : (tensor<12x5x64xf32>) -> tensor<1x12x5x64xf32>
%1053 = "tosa.transpose"(%1052, %5) : (tensor<1x12x5x64xf32>, tensor<4xi64>) -> tensor<1x5x12x64xf32>
%1054 = "tosa.reshape"(%1053) {new_shape = [1, 5, 768]} : (tensor<1x5x12x64xf32>) -> tensor<1x5x768xf32>
%1055 = "tosa.matmul"(%1054, %153) : (tensor<1x5x768xf32>, tensor<1x768x768xf32>) -> tensor<1x5x768xf32>
%1056 = "tosa.reshape"(%1055) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%1057 = "tosa.add"(%1056, %154) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%1058 = "tosa.reshape"(%1057) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%1059 = "tosa.add"(%1058, %1011) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%1060 = "tosa.reduce_sum"(%1059) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%1061 = "tosa.mul"(%1060, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1062 = "tosa.sub"(%1059, %1061) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%1063 = "tosa.mul"(%1062, %1062) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%1064 = "tosa.reduce_sum"(%1063) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%1065 = "tosa.mul"(%1064, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1066 = "tosa.add"(%1065, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1067 = "tosa.rsqrt"(%1066) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%1068 = "tosa.mul"(%1062, %1067) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%1069 = "tosa.mul"(%1068, %155) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%1070 = "tosa.add"(%1069, %156) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%1071 = "tosa.matmul"(%1070, %157) : (tensor<1x5x768xf32>, tensor<1x768x3072xf32>) -> tensor<1x5x3072xf32>
%1072 = "tosa.reshape"(%1071) {new_shape = [5, 3072]} : (tensor<1x5x3072xf32>) -> tensor<5x3072xf32>
%1073 = "tosa.add"(%1072, %158) : (tensor<5x3072xf32>, tensor<1x3072xf32>) -> tensor<5x3072xf32>
%1074 = "tosa.reshape"(%1073) {new_shape = [1, 5, 3072]} : (tensor<5x3072xf32>) -> tensor<1x5x3072xf32>
%1075 = "tosa.mul"(%1074, %12) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1076 = "tosa.pow"(%1074, %11) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1077 = "tosa.mul"(%1076, %10) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1078 = "tosa.add"(%1074, %1077) : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%1079 = "tosa.mul"(%1078, %9) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1080 = "tosa.tanh"(%1079) : (tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%1081 = "tosa.add"(%1080, %14) : (tensor<1x5x3072xf32>, tensor<1x1x1xf32>) -> tensor<1x5x3072xf32>
%1082 = "tosa.mul"(%1075, %1081) {shift = 0 : i32} : (tensor<1x5x3072xf32>, tensor<1x5x3072xf32>) -> tensor<1x5x3072xf32>
%1083 = "tosa.matmul"(%1082, %159) : (tensor<1x5x3072xf32>, tensor<1x3072x768xf32>) -> tensor<1x5x768xf32>
%1084 = "tosa.reshape"(%1083) {new_shape = [5, 768]} : (tensor<1x5x768xf32>) -> tensor<5x768xf32>
%1085 = "tosa.add"(%1084, %160) : (tensor<5x768xf32>, tensor<1x768xf32>) -> tensor<5x768xf32>
%1086 = "tosa.reshape"(%1085) {new_shape = [1, 5, 768]} : (tensor<5x768xf32>) -> tensor<1x5x768xf32>
%1087 = "tosa.add"(%1059, %1086) : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%1088 = "tosa.reduce_sum"(%1087) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%1089 = "tosa.mul"(%1088, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1090 = "tosa.sub"(%1087, %1089) : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%1091 = "tosa.mul"(%1090, %1090) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x768xf32>) -> tensor<1x5x768xf32>
%1092 = "tosa.reduce_sum"(%1091) {axis = 2 : i64} : (tensor<1x5x768xf32>) -> tensor<1x5x1xf32>
%1093 = "tosa.mul"(%1092, %171) {shift = 0 : i32} : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1094 = "tosa.add"(%1093, %13) : (tensor<1x5x1xf32>, tensor<1x1x1xf32>) -> tensor<1x5x1xf32>
%1095 = "tosa.rsqrt"(%1094) : (tensor<1x5x1xf32>) -> tensor<1x5x1xf32>
%1096 = "tosa.mul"(%1090, %1095) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x5x1xf32>) -> tensor<1x5x768xf32>
%1097 = "tosa.mul"(%1096, %161) {shift = 0 : i32} : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%1098 = "tosa.add"(%1097, %162) : (tensor<1x5x768xf32>, tensor<1x1x768xf32>) -> tensor<1x5x768xf32>
%1099 = "tosa.transpose"(%3, %8) : (tensor<50257x768xf32>, tensor<2xi32>) -> tensor<768x50257xf32>
%1100 = "tosa.reshape"(%1099) {new_shape = [1, 768, 50257]} : (tensor<768x50257xf32>) -> tensor<1x768x50257xf32>
%1101 = "tosa.matmul"(%1098, %1100) : (tensor<1x5x768xf32>, tensor<1x768x50257xf32>) -> tensor<1x5x50257xf32>
return %1101 : tensor<1x5x50257xf32>
}
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment