whisperkittools-a8c3cdeab8da5d76a7b952aa74ffebfbcd44804b generated files: openai_whisper-tiny.en

openai_whisper-tiny.en/AudioEncoder.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d196c8b7f34488833d89071b82740bceb30133c387bc7d2d5dbaeb9e2b52cd4d
+oid sha256:be62662cccc25b6eca93e6f407e024f90f72cde23a1cda8b0ca753f084274a6e
 size 243

openai_whisper-tiny.en/AudioEncoder.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:92aaedf405bc263b8426a4c2a4d28b568171a6d97fb732a76bd40a6b1b9475a6
+oid sha256:892364c46ef43333af61711e2ef67e32cc6e7040d35263b805cb5795c8d69233
 size 347

openai_whisper-tiny.en/AudioEncoder.mlmodelc/metadata.json

@@ -20,18 +20,16 @@
     "specificationVersion" : 7,
     "mlProgramOperationTypeHistogram" : {
       "Concat" : 28,
-      "Ios16.rsqrt" : 9,
-      "Ios16.mul" : 114,
+      "Ios16.add" : 9,
+      "Ios16.mul" : 96,
       "SliceByIndex" : 168,
-      "Ios16.sub" : 9,
       "Transpose" : 4,
+      "Ios16.batchNorm" : 9,
       "Ios16.einsum" : 192,
-      "Ios16.conv" : 26,
-      "Ios16.add" : 18,
-      "Ios16.reduceMean" : 18,
-      "Ios16.softmax" : 96,
       "Ios16.gelu" : 6,
-      "Ios16.batchNorm" : 9
+      "Ios16.softmax" : 96,
+      "Ios16.layerNorm" : 9,
+      "Ios16.conv" : 26
     },
     "computePrecision" : "Mixed (Float16, Int32)",
     "isUpdatable" : "0",
@@ -49,7 +47,7 @@
     "userDefinedMetadata" : {
       "com.github.apple.coremltools.source_dialect" : "TorchScript",
       "com.github.apple.coremltools.source" : "torch==2.2.2",
-      "com.github.apple.coremltools.version" : "7.1"
+      "com.github.apple.coremltools.version" : "7.2"
     },
     "inputSchema" : [
       {

openai_whisper-tiny.en/AudioEncoder.mlmodelc/model.mil

@@ -1,5 +1,5 @@
 program(1.0)
-[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})]
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.2"}})]
 {
     func main<ios16>(tensor<fp16, [1, 80, 1, 3000]> melspectrogram_features) {
             tensor<int32, [2]> var_34 = const()[name = tensor<string, []>("op_34"), val = tensor<int32, [2]>([1, 1])];
@@ -26,18 +26,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1500]> inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_108_to_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
             tensor<int32, []> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_129 = const()[name = tensor<string, []>("op_129"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_130 = const()[name = tensor<string, []>("op_130"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_140 = const()[name = tensor<string, []>("op_140"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_1_cast_fp16 = reduce_mean(axes = var_140, keep_dims = var_130, x = inputs_1_cast_fp16)[name = tensor<string, []>("channels_mean_1_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_1_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_sq_1_cast_fp16")];
-            tensor<int32, [1]> var_144 = const()[name = tensor<string, []>("op_144"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_145_cast_fp16 = reduce_mean(axes = var_144, keep_dims = var_130, x = zero_mean_sq_1_cast_fp16)[name = tensor<string, []>("op_145_cast_fp16")];
+            tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_146_to_fp16 = const()[name = tensor<string, []>("op_146_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_147_cast_fp16 = add(x = var_145_cast_fp16, y = var_146_to_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
-            tensor<fp16, []> denom_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_147_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_146_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
             tensor<fp16, [384]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2222976)))];
             tensor<fp16, [384]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2223808)))];
             tensor<fp16, [384]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2224640)))];
@@ -424,17 +415,9 @@ program(1.0)
             tensor<fp16, [384]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407872)))];
             tensor<fp16, [1, 384, 1, 1500]> obj_3_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_610, groups = var_129, pad = obj_3_pad_0, pad_type = obj_3_pad_type_0, strides = var_608, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_3_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
-            tensor<int32, [1]> var_616 = const()[name = tensor<string, []>("op_616"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_3_cast_fp16 = reduce_mean(axes = var_616, keep_dims = var_130, x = inputs_3_cast_fp16)[name = tensor<string, []>("channels_mean_3_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_3_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_sq_3_cast_fp16")];
-            tensor<int32, [1]> var_620 = const()[name = tensor<string, []>("op_620"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_621_cast_fp16 = reduce_mean(axes = var_620, keep_dims = var_130, x = zero_mean_sq_3_cast_fp16)[name = tensor<string, []>("op_621_cast_fp16")];
+            tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_622_to_fp16 = const()[name = tensor<string, []>("op_622_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_623_cast_fp16 = add(x = var_621_cast_fp16, y = var_622_to_fp16)[name = tensor<string, []>("op_623_cast_fp16")];
-            tensor<fp16, []> denom_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_623_cast_fp16)[name = tensor<string, []>("denom_3_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_622_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
             tensor<fp16, [384]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3408704)))];
             tensor<fp16, [384]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3409536)))];
             tensor<fp16, []> input_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -458,18 +441,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
             tensor<int32, []> var_651 = const()[name = tensor<string, []>("op_651"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_662 = const()[name = tensor<string, []>("op_662"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_663 = const()[name = tensor<string, []>("op_663"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_673 = const()[name = tensor<string, []>("op_673"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_5_cast_fp16 = reduce_mean(axes = var_673, keep_dims = var_663, x = inputs_5_cast_fp16)[name = tensor<string, []>("channels_mean_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_sq_5_cast_fp16")];
-            tensor<int32, [1]> var_677 = const()[name = tensor<string, []>("op_677"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_678_cast_fp16 = reduce_mean(axes = var_677, keep_dims = var_663, x = zero_mean_sq_5_cast_fp16)[name = tensor<string, []>("op_678_cast_fp16")];
+            tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_679_to_fp16 = const()[name = tensor<string, []>("op_679_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_680_cast_fp16 = add(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor<string, []>("op_680_cast_fp16")];
-            tensor<fp16, []> denom_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_680_cast_fp16)[name = tensor<string, []>("denom_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_679_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
             tensor<fp16, [384]> obj_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5773760)))];
             tensor<fp16, [384]> obj_5_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_5_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5774592)))];
             tensor<fp16, []> obj_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -854,17 +828,9 @@ program(1.0)
             tensor<fp16, [384]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6956992)))];
             tensor<fp16, [1, 384, 1, 1500]> obj_7_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_1143, groups = var_662, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_1141, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
-            tensor<int32, [1]> var_1149 = const()[name = tensor<string, []>("op_1149"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_7_cast_fp16 = reduce_mean(axes = var_1149, keep_dims = var_663, x = inputs_7_cast_fp16)[name = tensor<string, []>("channels_mean_7_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_7_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_sq_7_cast_fp16")];
-            tensor<int32, [1]> var_1153 = const()[name = tensor<string, []>("op_1153"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_1154_cast_fp16 = reduce_mean(axes = var_1153, keep_dims = var_663, x = zero_mean_sq_7_cast_fp16)[name = tensor<string, []>("op_1154_cast_fp16")];
+            tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_1155_to_fp16 = const()[name = tensor<string, []>("op_1155_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_1156_cast_fp16 = add(x = var_1154_cast_fp16, y = var_1155_to_fp16)[name = tensor<string, []>("op_1156_cast_fp16")];
-            tensor<fp16, []> denom_7_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_7_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_1156_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_1155_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
             tensor<fp16, [384]> input_11_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_11_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6957824)))];
             tensor<fp16, [384]> input_11_beta_0_to_fp16 = const()[name = tensor<string, []>("input_11_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6958656)))];
             tensor<fp16, []> input_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -888,18 +854,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
             tensor<int32, []> var_1184 = const()[name = tensor<string, []>("op_1184"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_1195 = const()[name = tensor<string, []>("op_1195"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_1196 = const()[name = tensor<string, []>("op_1196"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_1206 = const()[name = tensor<string, []>("op_1206"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_1206, keep_dims = var_1196, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_sq_9_cast_fp16")];
-            tensor<int32, [1]> var_1210 = const()[name = tensor<string, []>("op_1210"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_1211_cast_fp16 = reduce_mean(axes = var_1210, keep_dims = var_1196, x = zero_mean_sq_9_cast_fp16)[name = tensor<string, []>("op_1211_cast_fp16")];
+            tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_1212_to_fp16 = const()[name = tensor<string, []>("op_1212_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_1213_cast_fp16 = add(x = var_1211_cast_fp16, y = var_1212_to_fp16)[name = tensor<string, []>("op_1213_cast_fp16")];
-            tensor<fp16, []> denom_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_1213_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_1212_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
             tensor<fp16, [384]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322880)))];
             tensor<fp16, [384]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9323712)))];
             tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -1284,17 +1241,9 @@ program(1.0)
             tensor<fp16, [384]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506112)))];
             tensor<fp16, [1, 384, 1, 1500]> obj_11_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_1676, groups = var_1195, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_1674, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
-            tensor<int32, [1]> var_1682 = const()[name = tensor<string, []>("op_1682"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_1682, keep_dims = var_1196, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")];
-            tensor<int32, [1]> var_1686 = const()[name = tensor<string, []>("op_1686"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_1687_cast_fp16 = reduce_mean(axes = var_1686, keep_dims = var_1196, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_1687_cast_fp16")];
+            tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_1688_to_fp16 = const()[name = tensor<string, []>("op_1688_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_1689_cast_fp16 = add(x = var_1687_cast_fp16, y = var_1688_to_fp16)[name = tensor<string, []>("op_1689_cast_fp16")];
-            tensor<fp16, []> denom_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_1689_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_1688_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
             tensor<fp16, [384]> input_19_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_19_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506944)))];
             tensor<fp16, [384]> input_19_beta_0_to_fp16 = const()[name = tensor<string, []>("input_19_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10507776)))];
             tensor<fp16, []> input_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -1318,18 +1267,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
             tensor<int32, []> var_1717 = const()[name = tensor<string, []>("op_1717"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_1728 = const()[name = tensor<string, []>("op_1728"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_1729 = const()[name = tensor<string, []>("op_1729"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_1739 = const()[name = tensor<string, []>("op_1739"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_13_cast_fp16 = reduce_mean(axes = var_1739, keep_dims = var_1729, x = inputs_13_cast_fp16)[name = tensor<string, []>("channels_mean_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_sq_13_cast_fp16")];
-            tensor<int32, [1]> var_1743 = const()[name = tensor<string, []>("op_1743"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_1744_cast_fp16 = reduce_mean(axes = var_1743, keep_dims = var_1729, x = zero_mean_sq_13_cast_fp16)[name = tensor<string, []>("op_1744_cast_fp16")];
+            tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_1745_to_fp16 = const()[name = tensor<string, []>("op_1745_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_1746_cast_fp16 = add(x = var_1744_cast_fp16, y = var_1745_to_fp16)[name = tensor<string, []>("op_1746_cast_fp16")];
-            tensor<fp16, []> denom_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_1746_cast_fp16)[name = tensor<string, []>("denom_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_1745_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
             tensor<fp16, [384]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12872000)))];
             tensor<fp16, [384]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12872832)))];
             tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -1714,17 +1654,9 @@ program(1.0)
             tensor<fp16, [384]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14055232)))];
             tensor<fp16, [1, 384, 1, 1500]> obj_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_2209, groups = var_1728, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = var_2207, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
-            tensor<int32, [1]> var_2215 = const()[name = tensor<string, []>("op_2215"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_15_cast_fp16 = reduce_mean(axes = var_2215, keep_dims = var_1729, x = inputs_15_cast_fp16)[name = tensor<string, []>("channels_mean_15_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_15_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_sq_15_cast_fp16")];
-            tensor<int32, [1]> var_2219 = const()[name = tensor<string, []>("op_2219"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_2220_cast_fp16 = reduce_mean(axes = var_2219, keep_dims = var_1729, x = zero_mean_sq_15_cast_fp16)[name = tensor<string, []>("op_2220_cast_fp16")];
+            tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_2221_to_fp16 = const()[name = tensor<string, []>("op_2221_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_2222_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2221_to_fp16)[name = tensor<string, []>("op_2222_cast_fp16")];
-            tensor<fp16, []> denom_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_2222_cast_fp16)[name = tensor<string, []>("denom_15_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_2221_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
             tensor<fp16, [384]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14056064)))];
             tensor<fp16, [384]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14056896)))];
             tensor<fp16, []> input_27_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_27_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -1746,18 +1678,9 @@ program(1.0)
             tensor<fp16, [384]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16420288)))];
             tensor<fp16, [1, 384, 1, 1500]> hidden_states_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_2243, groups = var_1728, pad = hidden_states_pad_0, pad_type = hidden_states_pad_type_0, strides = var_2241, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1500]> inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
-            tensor<bool, []> var_2249 = const()[name = tensor<string, []>("op_2249"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_2253 = const()[name = tensor<string, []>("op_2253"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> channels_mean_cast_fp16 = reduce_mean(axes = var_2253, keep_dims = var_2249, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")];
-            tensor<int32, [1]> var_2257 = const()[name = tensor<string, []>("op_2257"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1500]> var_2258_cast_fp16 = reduce_mean(axes = var_2257, keep_dims = var_2249, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_2258_cast_fp16")];
+            tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_2259_to_fp16 = const()[name = tensor<string, []>("op_2259_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1500]> var_2260_cast_fp16 = add(x = var_2258_cast_fp16, y = var_2259_to_fp16)[name = tensor<string, []>("op_2260_cast_fp16")];
-            tensor<fp16, []> denom_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1500]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_2260_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1500]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1500]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_2259_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
             tensor<fp16, [384]> encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16421120)))];
             tensor<fp16, [384]> encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16421952)))];
             tensor<fp16, []> encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor<string, []>("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];

openai_whisper-tiny.en/AudioEncoder.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:903c9e16f458bac34caf4eacad6bcb4c47e59ab3eead6091bcdd5c0374311d46
+oid sha256:86a53cab0dcd92efa0d89803401347c8119203aac37f336d5a699c0587d01c48
 size 16422784

openai_whisper-tiny.en/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9ffbb7e8703e87c9d31219d6f7535fba26ddf4bb9c76cb7668ba4b7ae936be68
+oid sha256:2fc2d0799af479af957359c81021ff6a464d3251b3415064f8d2c6403cbea68f
 size 243

openai_whisper-tiny.en/MelSpectrogram.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5134062ebf286b76a4482eb57ce0378e1d7b961e3f44b7b4de1dbc6376278fcb
+oid sha256:c977c1199f029235ab96a7dc394e5c5c6d2b606d333f2cab46df750a4df89329
 size 328

openai_whisper-tiny.en/MelSpectrogram.mlmodelc/metadata.json

@@ -50,8 +50,8 @@
     },
     "userDefinedMetadata" : {
       "com.github.apple.coremltools.source_dialect" : "TorchScript",
-      "com.github.apple.coremltools.source" : "torch==2.2.2",
-      "com.github.apple.coremltools.version" : "7.1"
+      "com.github.apple.coremltools.version" : "7.2",
+      "com.github.apple.coremltools.source" : "torch==2.2.2"
     },
     "inputSchema" : [
       {

openai_whisper-tiny.en/MelSpectrogram.mlmodelc/model.mil

@@ -1,13 +1,13 @@
 program(1.0)
-[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})]
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.2"}})]
 {
     func main<ios16>(tensor<fp16, [480000]> audio) {
             tensor<int32, [3]> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, [3]>([1, 1, 480000])];
             tensor<fp16, [1, 1, 480000]> input_1_cast_fp16 = reshape(shape = var_10, x = audio)[name = tensor<string, []>("input_1_cast_fp16")];
             tensor<int32, [6]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 200, 200])];
             tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("reflect")];
-            tensor<fp16, []> input_3_constant_val_0_to_fp16 = const()[name = tensor<string, []>("input_3_constant_val_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
-            tensor<fp16, [1, 1, 480400]> input_3_cast_fp16 = pad(constant_val = input_3_constant_val_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 480400]> input_3_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
             tensor<int32, [1]> var_22 = const()[name = tensor<string, []>("op_22"), val = tensor<int32, [1]>([480400])];
             tensor<fp16, [480400]> input_cast_fp16 = reshape(shape = var_22, x = input_3_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
             tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];

openai_whisper-tiny.en/MelSpectrogram.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7c63f309e59166f308c1243178fd5ff7e5b169c4291c6fafd9cbff98aa831674
+oid sha256:a3afe232689fe92b1958d124f42f9ccf43e611ce1055e584190d40f11dc5a3d6
 size 354080

openai_whisper-tiny.en/TextDecoder.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bf65ffd583b31ee4c16f0c9812e705fbab7f8cb4797ec43210b27ef6fdc43d76
+oid sha256:a3e8ecb0c4fc6f8c91c97b3a8b15fb84715aaa68d64fbe125553224c0c64c743
 size 243

openai_whisper-tiny.en/TextDecoder.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:47c8915461107bdac7eeb0087b0e8830368ec86f79816312ddb6ee690d97137a
+oid sha256:d8bae9187c5fb7e29a7cb0c8aa2c753e82040ba85403c2dbad6dd30f9a6a008d
 size 633

openai_whisper-tiny.en/TextDecoder.mlmodelc/metadata.json

@@ -51,18 +51,18 @@
     "mlProgramOperationTypeHistogram" : {
       "Split" : 2,
       "Concat" : 3,
-      "Ios16.rsqrt" : 13,
-      "Ios16.mul" : 50,
       "Squeeze" : 1,
+      "Ios16.mul" : 24,
+      "Ios16.layerNorm" : 13,
       "SliceByIndex" : 16,
-      "Ios16.sub" : 14,
+      "Ios16.sub" : 1,
       "Transpose" : 1,
       "Ios16.conv" : 40,
-      "Ios16.add" : 38,
+      "Ios16.add" : 25,
       "Ios16.linear" : 1,
       "Ios16.matmul" : 16,
       "Ios16.gelu" : 4,
-      "Ios16.reduceMean" : 27,
+      "Ios16.reduceMean" : 1,
       "ExpandDims" : 6,
       "Ios16.batchNorm" : 13,
       "Ios16.gather" : 2,
@@ -85,7 +85,7 @@
     "userDefinedMetadata" : {
       "com.github.apple.coremltools.source_dialect" : "TorchScript",
       "com.github.apple.coremltools.source" : "torch==2.2.2",
-      "com.github.apple.coremltools.version" : "7.1"
+      "com.github.apple.coremltools.version" : "7.2"
     },
     "inputSchema" : [
       {

openai_whisper-tiny.en/TextDecoder.mlmodelc/model.mil

@@ -1,5 +1,5 @@
 program(1.0)
-[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})]
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.2"}})]
 {
     func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 384, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 1536, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 1536, 1, 448]> value_cache) {
             tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)];
@@ -23,18 +23,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_0, tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_1, tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_2, tensor<fp16, [1, 384, 1, 448]> var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_54_cast_fp16")];
             tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_71 = const()[name = tensor<string, []>("op_71"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_72 = const()[name = tensor<string, []>("op_72"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_84 = const()[name = tensor<string, []>("op_84"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_1_cast_fp16 = reduce_mean(axes = var_84, keep_dims = var_72, x = inputs_1_cast_fp16)[name = tensor<string, []>("channels_mean_1_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_1_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_sq_1_cast_fp16")];
-            tensor<int32, [1]> var_88 = const()[name = tensor<string, []>("op_88"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_89_cast_fp16 = reduce_mean(axes = var_88, keep_dims = var_72, x = zero_mean_sq_1_cast_fp16)[name = tensor<string, []>("op_89_cast_fp16")];
+            tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_90_to_fp16 = const()[name = tensor<string, []>("op_90_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_91_cast_fp16 = add(x = var_89_cast_fp16, y = var_90_to_fp16)[name = tensor<string, []>("op_91_cast_fp16")];
-            tensor<fp16, []> denom_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_91_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_90_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
             tensor<fp16, [384]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40175808)))];
             tensor<fp16, [384]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40176640)))];
             tensor<fp16, [384]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40177472)))];
@@ -103,17 +94,9 @@ program(1.0)
             tensor<fp16, [384]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41360704)))];
             tensor<fp16, [1, 384, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_161, groups = var_71, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_159, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
-            tensor<int32, [1]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_3_cast_fp16 = reduce_mean(axes = var_171, keep_dims = var_72, x = inputs_3_cast_fp16)[name = tensor<string, []>("channels_mean_3_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_3_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_sq_3_cast_fp16")];
-            tensor<int32, [1]> var_175 = const()[name = tensor<string, []>("op_175"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_176_cast_fp16 = reduce_mean(axes = var_175, keep_dims = var_72, x = zero_mean_sq_3_cast_fp16)[name = tensor<string, []>("op_176_cast_fp16")];
+            tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_177_to_fp16 = const()[name = tensor<string, []>("op_177_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_178_cast_fp16 = add(x = var_176_cast_fp16, y = var_177_to_fp16)[name = tensor<string, []>("op_178_cast_fp16")];
-            tensor<fp16, []> denom_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_178_cast_fp16)[name = tensor<string, []>("denom_3_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_177_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
             tensor<fp16, [384]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41361536)))];
             tensor<fp16, [384]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41362368)))];
             tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -163,17 +146,9 @@ program(1.0)
             tensor<fp16, [384]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42544768)))];
             tensor<fp16, [1, 384, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_231, groups = var_71, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_229, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
-            tensor<int32, [1]> var_237 = const()[name = tensor<string, []>("op_237"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_5_cast_fp16 = reduce_mean(axes = var_237, keep_dims = var_72, x = inputs_5_cast_fp16)[name = tensor<string, []>("channels_mean_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_sq_5_cast_fp16")];
-            tensor<int32, [1]> var_241 = const()[name = tensor<string, []>("op_241"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_242_cast_fp16 = reduce_mean(axes = var_241, keep_dims = var_72, x = zero_mean_sq_5_cast_fp16)[name = tensor<string, []>("op_242_cast_fp16")];
+            tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_243_to_fp16 = const()[name = tensor<string, []>("op_243_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_244_cast_fp16 = add(x = var_242_cast_fp16, y = var_243_to_fp16)[name = tensor<string, []>("op_244_cast_fp16")];
-            tensor<fp16, []> denom_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_244_cast_fp16)[name = tensor<string, []>("denom_5_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_243_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
             tensor<fp16, [384]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42545600)))];
             tensor<fp16, [384]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42546432)))];
             tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -197,18 +172,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
             tensor<int32, []> var_278 = const()[name = tensor<string, []>("op_278"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_285 = const()[name = tensor<string, []>("op_285"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_286 = const()[name = tensor<string, []>("op_286"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_298 = const()[name = tensor<string, []>("op_298"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_7_cast_fp16 = reduce_mean(axes = var_298, keep_dims = var_286, x = inputs_7_cast_fp16)[name = tensor<string, []>("channels_mean_7_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_7_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_sq_7_cast_fp16")];
-            tensor<int32, [1]> var_302 = const()[name = tensor<string, []>("op_302"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_303_cast_fp16 = reduce_mean(axes = var_302, keep_dims = var_286, x = zero_mean_sq_7_cast_fp16)[name = tensor<string, []>("op_303_cast_fp16")];
+            tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_304_to_fp16 = const()[name = tensor<string, []>("op_304_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_305_cast_fp16 = add(x = var_303_cast_fp16, y = var_304_to_fp16)[name = tensor<string, []>("op_305_cast_fp16")];
-            tensor<fp16, []> denom_7_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_7_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_305_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_304_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
             tensor<fp16, [384]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44910656)))];
             tensor<fp16, [384]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44911488)))];
             tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -265,17 +231,9 @@ program(1.0)
             tensor<fp16, [384]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46093888)))];
             tensor<fp16, [1, 384, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_375, groups = var_285, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_373, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
-            tensor<int32, [1]> var_385 = const()[name = tensor<string, []>("op_385"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_385, keep_dims = var_286, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_sq_9_cast_fp16")];
-            tensor<int32, [1]> var_389 = const()[name = tensor<string, []>("op_389"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_390_cast_fp16 = reduce_mean(axes = var_389, keep_dims = var_286, x = zero_mean_sq_9_cast_fp16)[name = tensor<string, []>("op_390_cast_fp16")];
+            tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_391_to_fp16 = const()[name = tensor<string, []>("op_391_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_392_cast_fp16 = add(x = var_390_cast_fp16, y = var_391_to_fp16)[name = tensor<string, []>("op_392_cast_fp16")];
-            tensor<fp16, []> denom_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_392_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_391_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
             tensor<fp16, [384]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46094720)))];
             tensor<fp16, [384]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46095552)))];
             tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -325,17 +283,9 @@ program(1.0)
             tensor<fp16, [384]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47277952)))];
             tensor<fp16, [1, 384, 1, 1]> obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_445, groups = var_285, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_443, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
-            tensor<int32, [1]> var_454 = const()[name = tensor<string, []>("op_454"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_454, keep_dims = var_286, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")];
-            tensor<int32, [1]> var_458 = const()[name = tensor<string, []>("op_458"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_459_cast_fp16 = reduce_mean(axes = var_458, keep_dims = var_286, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_459_cast_fp16")];
+            tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_460_to_fp16 = const()[name = tensor<string, []>("op_460_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_461_cast_fp16 = add(x = var_459_cast_fp16, y = var_460_to_fp16)[name = tensor<string, []>("op_461_cast_fp16")];
-            tensor<fp16, []> denom_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_461_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_460_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
             tensor<fp16, [384]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47278784)))];
             tensor<fp16, [384]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47279616)))];
             tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -359,18 +309,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
             tensor<int32, []> var_496 = const()[name = tensor<string, []>("op_496"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_503 = const()[name = tensor<string, []>("op_503"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_504 = const()[name = tensor<string, []>("op_504"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_516 = const()[name = tensor<string, []>("op_516"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_13_cast_fp16 = reduce_mean(axes = var_516, keep_dims = var_504, x = inputs_13_cast_fp16)[name = tensor<string, []>("channels_mean_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor<string, []>("zero_mean_sq_13_cast_fp16")];
-            tensor<int32, [1]> var_520 = const()[name = tensor<string, []>("op_520"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_521_cast_fp16 = reduce_mean(axes = var_520, keep_dims = var_504, x = zero_mean_sq_13_cast_fp16)[name = tensor<string, []>("op_521_cast_fp16")];
+            tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_522_to_fp16 = const()[name = tensor<string, []>("op_522_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_523_cast_fp16 = add(x = var_521_cast_fp16, y = var_522_to_fp16)[name = tensor<string, []>("op_523_cast_fp16")];
-            tensor<fp16, []> denom_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_523_cast_fp16)[name = tensor<string, []>("denom_13_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_522_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
             tensor<fp16, [384]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49643840)))];
             tensor<fp16, [384]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49644672)))];
             tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -427,17 +368,9 @@ program(1.0)
             tensor<fp16, [384]> layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50827072)))];
             tensor<fp16, [1, 384, 1, 1]> obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_593, groups = var_503, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_591, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
-            tensor<int32, [1]> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_15_cast_fp16 = reduce_mean(axes = var_603, keep_dims = var_504, x = inputs_15_cast_fp16)[name = tensor<string, []>("channels_mean_15_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_15_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor<string, []>("zero_mean_sq_15_cast_fp16")];
-            tensor<int32, [1]> var_607 = const()[name = tensor<string, []>("op_607"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_608_cast_fp16 = reduce_mean(axes = var_607, keep_dims = var_504, x = zero_mean_sq_15_cast_fp16)[name = tensor<string, []>("op_608_cast_fp16")];
+            tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_609_to_fp16 = const()[name = tensor<string, []>("op_609_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_610_cast_fp16 = add(x = var_608_cast_fp16, y = var_609_to_fp16)[name = tensor<string, []>("op_610_cast_fp16")];
-            tensor<fp16, []> denom_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_610_cast_fp16)[name = tensor<string, []>("denom_15_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_609_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
             tensor<fp16, [384]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50827904)))];
             tensor<fp16, [384]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50828736)))];
             tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -487,17 +420,9 @@ program(1.0)
             tensor<fp16, [384]> layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52011136)))];
             tensor<fp16, [1, 384, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_663, groups = var_503, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_661, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")];
-            tensor<int32, [1]> var_672 = const()[name = tensor<string, []>("op_672"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_17_cast_fp16 = reduce_mean(axes = var_672, keep_dims = var_504, x = inputs_17_cast_fp16)[name = tensor<string, []>("channels_mean_17_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_17_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor<string, []>("zero_mean_sq_17_cast_fp16")];
-            tensor<int32, [1]> var_676 = const()[name = tensor<string, []>("op_676"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_677_cast_fp16 = reduce_mean(axes = var_676, keep_dims = var_504, x = zero_mean_sq_17_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")];
+            tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_678_to_fp16 = const()[name = tensor<string, []>("op_678_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_679_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_to_fp16)[name = tensor<string, []>("op_679_cast_fp16")];
-            tensor<fp16, []> denom_17_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_17_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_679_cast_fp16)[name = tensor<string, []>("denom_17_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_678_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
             tensor<fp16, [384]> input_25_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_25_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52011968)))];
             tensor<fp16, [384]> input_25_beta_0_to_fp16 = const()[name = tensor<string, []>("input_25_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52012800)))];
             tensor<fp16, []> input_25_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_25_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -521,18 +446,9 @@ program(1.0)
             tensor<fp16, [1, 384, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")];
             tensor<int32, []> var_714 = const()[name = tensor<string, []>("op_714"), val = tensor<int32, []>(3)];
             tensor<int32, []> var_721 = const()[name = tensor<string, []>("op_721"), val = tensor<int32, []>(1)];
-            tensor<bool, []> var_722 = const()[name = tensor<string, []>("op_722"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_734 = const()[name = tensor<string, []>("op_734"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_19_cast_fp16 = reduce_mean(axes = var_734, keep_dims = var_722, x = inputs_19_cast_fp16)[name = tensor<string, []>("channels_mean_19_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_19_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor<string, []>("zero_mean_sq_19_cast_fp16")];
-            tensor<int32, [1]> var_738 = const()[name = tensor<string, []>("op_738"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_739_cast_fp16 = reduce_mean(axes = var_738, keep_dims = var_722, x = zero_mean_sq_19_cast_fp16)[name = tensor<string, []>("op_739_cast_fp16")];
+            tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_740_to_fp16 = const()[name = tensor<string, []>("op_740_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_741_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_to_fp16)[name = tensor<string, []>("op_741_cast_fp16")];
-            tensor<fp16, []> denom_19_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_19_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_741_cast_fp16)[name = tensor<string, []>("denom_19_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_740_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
             tensor<fp16, [384]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54377024)))];
             tensor<fp16, [384]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54377856)))];
             tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -589,17 +505,9 @@ program(1.0)
             tensor<fp16, [384]> layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55560256)))];
             tensor<fp16, [1, 384, 1, 1]> obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_811, groups = var_721, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_809, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")];
-            tensor<int32, [1]> var_821 = const()[name = tensor<string, []>("op_821"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_21_cast_fp16 = reduce_mean(axes = var_821, keep_dims = var_722, x = inputs_21_cast_fp16)[name = tensor<string, []>("channels_mean_21_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_21_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor<string, []>("zero_mean_sq_21_cast_fp16")];
-            tensor<int32, [1]> var_825 = const()[name = tensor<string, []>("op_825"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_826_cast_fp16 = reduce_mean(axes = var_825, keep_dims = var_722, x = zero_mean_sq_21_cast_fp16)[name = tensor<string, []>("op_826_cast_fp16")];
+            tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_827_to_fp16 = const()[name = tensor<string, []>("op_827_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_828_cast_fp16 = add(x = var_826_cast_fp16, y = var_827_to_fp16)[name = tensor<string, []>("op_828_cast_fp16")];
-            tensor<fp16, []> denom_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_828_cast_fp16)[name = tensor<string, []>("denom_21_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_827_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
             tensor<fp16, [384]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55561088)))];
             tensor<fp16, [384]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55561920)))];
             tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -649,17 +557,9 @@ program(1.0)
             tensor<fp16, [384]> layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56744320)))];
             tensor<fp16, [1, 384, 1, 1]> obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_881, groups = var_721, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_879, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")];
-            tensor<int32, [1]> var_890 = const()[name = tensor<string, []>("op_890"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_23_cast_fp16 = reduce_mean(axes = var_890, keep_dims = var_722, x = inputs_23_cast_fp16)[name = tensor<string, []>("channels_mean_23_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_23_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor<string, []>("zero_mean_sq_23_cast_fp16")];
-            tensor<int32, [1]> var_894 = const()[name = tensor<string, []>("op_894"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_895_cast_fp16 = reduce_mean(axes = var_894, keep_dims = var_722, x = zero_mean_sq_23_cast_fp16)[name = tensor<string, []>("op_895_cast_fp16")];
+            tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_896_to_fp16 = const()[name = tensor<string, []>("op_896_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_897_cast_fp16 = add(x = var_895_cast_fp16, y = var_896_to_fp16)[name = tensor<string, []>("op_897_cast_fp16")];
-            tensor<fp16, []> denom_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_897_cast_fp16)[name = tensor<string, []>("denom_23_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_896_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
             tensor<fp16, [384]> input_35_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_35_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56745152)))];
             tensor<fp16, [384]> input_35_beta_0_to_fp16 = const()[name = tensor<string, []>("input_35_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56745984)))];
             tensor<fp16, []> input_35_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_35_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
@@ -681,18 +581,9 @@ program(1.0)
             tensor<fp16, [384]> layers_3_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59109376)))];
             tensor<fp16, [1, 384, 1, 1]> hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_918, groups = var_721, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_916, weight = layers_3_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")];
             tensor<fp16, [1, 384, 1, 1]> inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
-            tensor<bool, []> var_929 = const()[name = tensor<string, []>("op_929"), val = tensor<bool, []>(true)];
-            tensor<int32, [1]> var_933 = const()[name = tensor<string, []>("op_933"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> channels_mean_cast_fp16 = reduce_mean(axes = var_933, keep_dims = var_929, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")];
-            tensor<int32, [1]> var_937 = const()[name = tensor<string, []>("op_937"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 1, 1]> var_938_cast_fp16 = reduce_mean(axes = var_937, keep_dims = var_929, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
+            tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])];
             tensor<fp16, []> var_939_to_fp16 = const()[name = tensor<string, []>("op_939_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
-            tensor<fp16, [1, 1, 1, 1]> var_940_cast_fp16 = add(x = var_938_cast_fp16, y = var_939_to_fp16)[name = tensor<string, []>("op_940_cast_fp16")];
-            tensor<fp16, []> denom_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
-            tensor<fp16, [1, 1, 1, 1]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_940_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
-            tensor<fp16, [1, 384, 1, 1]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
+            tensor<fp16, [1, 384, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_939_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
             tensor<fp16, [384]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59110208)))];
             tensor<fp16, [384]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59111040)))];
             tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];

openai_whisper-tiny.en/TextDecoder.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b9086b2c9f476f0884572aae91e5059c31aeddb4b5658b5f810269aaa0f3c01
+oid sha256:12c57bf53565400bb60c09b13e9a6e31fdaa147585a8e21aeea1b60a96e3400e
 size 59215664