compile model

CompiledModel/TinyGPT2.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8eed951770f6ddba60b999585b5de885d80064e66d3f345097a7926aaf73b021
+size 243

CompiledModel/TinyGPT2.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dbf3328fbbf2b00822542623df28b9ba447db58b7a5557e6c9d03fc7df6ccae
+size 317

CompiledModel/TinyGPT2.mlmodelc/metadata.json

@@ -0,0 +1,70 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 5 × 50257)",
+        "shortDescription" : "",
+        "shape" : "[1, 5, 50257]",
+        "name" : "var_289",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.layerNorm" : 5,
+      "Ios17.reshape" : 22,
+      "Ios17.cast" : 1,
+      "Ios17.gather" : 1,
+      "Split" : 2,
+      "Ios17.matmul" : 12,
+      "Ios17.add" : 7,
+      "Ios17.transpose" : 8,
+      "Ios16.softmax" : 2,
+      "Ios16.gelu" : 2,
+      "Ios17.linear" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.6.0",
+      "com.github.apple.coremltools.version" : "8.2"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 5)",
+        "shortDescription" : "",
+        "shape" : "[1, 5]",
+        "name" : "input_ids_1",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "TinyGPT2",
+    "method" : "predict"
+  }
+]

CompiledModel/TinyGPT2.mlmodelc/model.mil

@@ -0,0 +1,163 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3404.16.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})]
+{
+    func main<ios17>(tensor<fp16, [1, 5]> input_ids_1) {
+            tensor<string, []> cast_2_dtype_0 = const()[name = tensor<string, []>("cast_2_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> inputs_embeds_axis_0 = const()[name = tensor<string, []>("inputs_embeds_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> inputs_embeds_batch_dims_0 = const()[name = tensor<string, []>("inputs_embeds_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> inputs_embeds_validate_indices_0 = const()[name = tensor<string, []>("inputs_embeds_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [50257, 2]> model_transformer_wte_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_wte_weight_to_fp16"), val = tensor<fp16, [50257, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<int32, [1, 5]> input_ids_1_to_int32 = cast(dtype = cast_2_dtype_0, x = input_ids_1)[name = tensor<string, []>("cast_32")];
+            tensor<fp16, [1, 5, 2]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids_1_to_int32, validate_indices = inputs_embeds_validate_indices_0, x = model_transformer_wte_weight_to_fp16)[name = tensor<string, []>("inputs_embeds_cast_fp16")];
+            tensor<fp16, [1, 5, 2]> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, [1, 5, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201216)))];
+            tensor<fp16, [1, 5, 2]> input_3_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = const_3_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [2]> model_transformer_h_0_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_ln_1_weight_to_fp16"), val = tensor<fp16, [2]>([0x1p+0, 0x1p+0])];
+            tensor<fp16, [2]> model_transformer_h_0_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_ln_1_bias_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
+            tensor<fp16, []> var_19_to_fp16 = const()[name = tensor<string, []>("op_19_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 5, 2]> x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = model_transformer_h_0_ln_1_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_0_ln_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
+            tensor<int32, [2]> var_88 = const()[name = tensor<string, []>("op_88"), val = tensor<int32, [2]>([-1, 2])];
+            tensor<fp16, [5, 2]> var_89_cast_fp16 = reshape(shape = var_88, x = x_1_cast_fp16)[name = tensor<string, []>("op_89_cast_fp16")];
+            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2, 6]> model_transformer_h_0_attn_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_attn_c_attn_weight_to_fp16"), val = tensor<fp16, [2, 6]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201344)))];
+            tensor<fp16, [5, 6]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = var_89_cast_fp16, y = model_transformer_h_0_attn_c_attn_weight_to_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")];
+            tensor<int32, [3]> var_91 = const()[name = tensor<string, []>("op_91"), val = tensor<int32, [3]>([1, 5, 6])];
+            tensor<fp16, [1, 5, 6]> var_92_cast_fp16 = reshape(shape = var_91, x = matmul_0_cast_fp16)[name = tensor<string, []>("op_92_cast_fp16")];
+            tensor<int32, [3]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [3]>([2, 2, 2])];
+            tensor<int32, []> var_93_axis_0 = const()[name = tensor<string, []>("op_93_axis_0"), val = tensor<int32, []>(2)];
+            tensor<fp16, [1, 5, 2]> var_93_cast_fp16_0, tensor<fp16, [1, 5, 2]> var_93_cast_fp16_1, tensor<fp16, [1, 5, 2]> var_93_cast_fp16_2 = split(axis = var_93_axis_0, split_sizes = tile_0, x = var_92_cast_fp16)[name = tensor<string, []>("op_93_cast_fp16")];
+            tensor<int32, [4]> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, [4]>([1, 5, -1, 1])];
+            tensor<fp16, [1, 5, 2, 1]> var_102_cast_fp16 = reshape(shape = var_101, x = var_93_cast_fp16_0)[name = tensor<string, []>("op_102_cast_fp16")];
+            tensor<int32, [4]> var_104 = const()[name = tensor<string, []>("op_104"), val = tensor<int32, [4]>([1, 5, -1, 1])];
+            tensor<fp16, [1, 5, 2, 1]> var_105_cast_fp16 = reshape(shape = var_104, x = var_93_cast_fp16_1)[name = tensor<string, []>("op_105_cast_fp16")];
+            tensor<int32, [4]> var_107 = const()[name = tensor<string, []>("op_107"), val = tensor<int32, [4]>([1, 5, -1, 1])];
+            tensor<fp16, [1, 5, 2, 1]> var_108_cast_fp16 = reshape(shape = var_107, x = var_93_cast_fp16_2)[name = tensor<string, []>("op_108_cast_fp16")];
+            tensor<int32, [4]> value_1_perm_0 = const()[name = tensor<string, []>("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_8_perm_0 = const()[name = tensor<string, []>("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_9_perm_0 = const()[name = tensor<string, []>("transpose_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [1, 2, 5, 1]> transpose_9 = transpose(perm = transpose_9_perm_0, x = var_105_cast_fp16)[name = tensor<string, []>("transpose_17")];
+            tensor<fp16, [1, 2, 5, 1]> transpose_8 = transpose(perm = transpose_8_perm_0, x = var_102_cast_fp16)[name = tensor<string, []>("transpose_18")];
+            tensor<fp16, [1, 2, 5, 5]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor<string, []>("matmul_1_cast_fp16")];
+            tensor<fp16, [1, 1, 5, 5]> var_64_to_fp16 = const()[name = tensor<string, []>("op_64_to_fp16"), val = tensor<fp16, [1, 1, 5, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201472)))];
+            tensor<fp16, [1, 2, 5, 5]> add_0_cast_fp16 = add(x = matmul_1_cast_fp16, y = var_64_to_fp16)[name = tensor<string, []>("add_0_cast_fp16")];
+            tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 2, 5, 5]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = tensor<string, []>("softmax_0_cast_fp16")];
+            tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 2, 5, 1]> value_1_cast_fp16 = transpose(perm = value_1_perm_0, x = var_108_cast_fp16)[name = tensor<string, []>("transpose_19")];
+            tensor<fp16, [1, 2, 5, 1]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = tensor<string, []>("attn_output_1_cast_fp16")];
+            tensor<int32, [4]> var_119_perm_0 = const()[name = tensor<string, []>("op_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [2]> var_131 = const()[name = tensor<string, []>("op_131"), val = tensor<int32, [2]>([-1, 2])];
+            tensor<fp16, [1, 5, 2, 1]> var_119_cast_fp16 = transpose(perm = var_119_perm_0, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_16")];
+            tensor<fp16, [5, 2]> var_132_cast_fp16 = reshape(shape = var_131, x = var_119_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")];
+            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2, 2]> model_transformer_h_0_attn_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_attn_c_proj_weight_to_fp16"), val = tensor<fp16, [2, 2]>([[-0x1.54cp-10, -0x1.3d8p-10], [-0x1.d4p-5, -0x1.8p-6]])];
+            tensor<fp16, [5, 2]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = var_132_cast_fp16, y = model_transformer_h_0_attn_c_proj_weight_to_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")];
+            tensor<int32, [3]> var_134 = const()[name = tensor<string, []>("op_134"), val = tensor<int32, [3]>([1, 5, 2])];
+            tensor<fp16, [1, 5, 2]> input_5_cast_fp16 = reshape(shape = var_134, x = matmul_2_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 5, 2]> input_7_cast_fp16 = add(x = input_5_cast_fp16, y = input_3_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [1]> x_9_axes_0 = const()[name = tensor<string, []>("x_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [2]> model_transformer_h_0_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_ln_2_weight_to_fp16"), val = tensor<fp16, [2]>([0x1p+0, 0x1p+0])];
+            tensor<fp16, [2]> model_transformer_h_0_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_ln_2_bias_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
+            tensor<fp16, [1, 5, 2]> x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = model_transformer_h_0_ln_2_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_0_ln_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [2]> var_149 = const()[name = tensor<string, []>("op_149"), val = tensor<int32, [2]>([-1, 2])];
+            tensor<fp16, [5, 2]> var_150_cast_fp16 = reshape(shape = var_149, x = x_9_cast_fp16)[name = tensor<string, []>("op_150_cast_fp16")];
+            tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2, 8]> model_transformer_h_0_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [2, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201600)))];
+            tensor<fp16, [5, 8]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = var_150_cast_fp16, y = model_transformer_h_0_mlp_c_fc_weight_to_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")];
+            tensor<int32, [3]> var_152 = const()[name = tensor<string, []>("op_152"), val = tensor<int32, [3]>([1, 5, 8])];
+            tensor<fp16, [1, 5, 8]> input_9_cast_fp16 = reshape(shape = var_152, x = matmul_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> x_13_mode_0 = const()[name = tensor<string, []>("x_13_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 5, 8]> x_13_cast_fp16 = gelu(mode = x_13_mode_0, x = input_9_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [2]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [2]>([-1, 8])];
+            tensor<fp16, [5, 8]> var_172_cast_fp16 = reshape(shape = var_171, x = x_13_cast_fp16)[name = tensor<string, []>("op_172_cast_fp16")];
+            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [8, 2]> model_transformer_h_0_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_0_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [8, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201728)))];
+            tensor<fp16, [5, 2]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = var_172_cast_fp16, y = model_transformer_h_0_mlp_c_proj_weight_to_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
+            tensor<int32, [3]> var_174 = const()[name = tensor<string, []>("op_174"), val = tensor<int32, [3]>([1, 5, 2])];
+            tensor<fp16, [1, 5, 2]> input_11_cast_fp16 = reshape(shape = var_174, x = matmul_4_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1, 5, 2]> input_13_cast_fp16 = add(x = input_7_cast_fp16, y = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [2]> model_transformer_h_1_ln_1_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_ln_1_weight_to_fp16"), val = tensor<fp16, [2]>([0x1p+0, 0x1p+0])];
+            tensor<fp16, [2]> model_transformer_h_1_ln_1_bias_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_ln_1_bias_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
+            tensor<fp16, [1, 5, 2]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = model_transformer_h_1_ln_1_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_1_ln_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
+            tensor<int32, [2]> var_193 = const()[name = tensor<string, []>("op_193"), val = tensor<int32, [2]>([-1, 2])];
+            tensor<fp16, [5, 2]> var_194_cast_fp16 = reshape(shape = var_193, x = x_17_cast_fp16)[name = tensor<string, []>("op_194_cast_fp16")];
+            tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2, 6]> model_transformer_h_1_attn_c_attn_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_attn_c_attn_weight_to_fp16"), val = tensor<fp16, [2, 6]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201856)))];
+            tensor<fp16, [5, 6]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = var_194_cast_fp16, y = model_transformer_h_1_attn_c_attn_weight_to_fp16)[name = tensor<string, []>("matmul_5_cast_fp16")];
+            tensor<int32, [3]> var_196 = const()[name = tensor<string, []>("op_196"), val = tensor<int32, [3]>([1, 5, 6])];
+            tensor<fp16, [1, 5, 6]> var_197_cast_fp16 = reshape(shape = var_196, x = matmul_5_cast_fp16)[name = tensor<string, []>("op_197_cast_fp16")];
+            tensor<int32, [3]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [3]>([2, 2, 2])];
+            tensor<int32, []> var_198_axis_0 = const()[name = tensor<string, []>("op_198_axis_0"), val = tensor<int32, []>(2)];
+            tensor<fp16, [1, 5, 2]> var_198_cast_fp16_0, tensor<fp16, [1, 5, 2]> var_198_cast_fp16_1, tensor<fp16, [1, 5, 2]> var_198_cast_fp16_2 = split(axis = var_198_axis_0, split_sizes = tile_1, x = var_197_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
+            tensor<int32, [4]> var_206 = const()[name = tensor<string, []>("op_206"), val = tensor<int32, [4]>([1, 5, -1, 1])];
+            tensor<fp16, [1, 5, 2, 1]> var_207_cast_fp16 = reshape(shape = var_206, x = var_198_cast_fp16_0)[name = tensor<string, []>("op_207_cast_fp16")];
+            tensor<int32, [4]> var_209 = const()[name = tensor<string, []>("op_209"), val = tensor<int32, [4]>([1, 5, -1, 1])];
+            tensor<fp16, [1, 5, 2, 1]> var_210_cast_fp16 = reshape(shape = var_209, x = var_198_cast_fp16_1)[name = tensor<string, []>("op_210_cast_fp16")];
+            tensor<int32, [4]> var_212 = const()[name = tensor<string, []>("op_212"), val = tensor<int32, [4]>([1, 5, -1, 1])];
+            tensor<fp16, [1, 5, 2, 1]> var_213_cast_fp16 = reshape(shape = var_212, x = var_198_cast_fp16_2)[name = tensor<string, []>("op_213_cast_fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_10_perm_0 = const()[name = tensor<string, []>("transpose_10_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_11_perm_0 = const()[name = tensor<string, []>("transpose_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [1, 2, 5, 1]> transpose_11 = transpose(perm = transpose_11_perm_0, x = var_210_cast_fp16)[name = tensor<string, []>("transpose_13")];
+            tensor<fp16, [1, 2, 5, 1]> transpose_10 = transpose(perm = transpose_10_perm_0, x = var_207_cast_fp16)[name = tensor<string, []>("transpose_14")];
+            tensor<fp16, [1, 2, 5, 5]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_10, y = transpose_11)[name = tensor<string, []>("matmul_6_cast_fp16")];
+            tensor<fp16, [1, 2, 5, 5]> add_1_cast_fp16 = add(x = matmul_6_cast_fp16, y = var_64_to_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
+            tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 2, 5, 5]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = tensor<string, []>("softmax_1_cast_fp16")];
+            tensor<bool, []> attn_output_7_transpose_x_0 = const()[name = tensor<string, []>("attn_output_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_output_7_transpose_y_0 = const()[name = tensor<string, []>("attn_output_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 2, 5, 1]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = var_213_cast_fp16)[name = tensor<string, []>("transpose_15")];
+            tensor<fp16, [1, 2, 5, 1]> attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = softmax_1_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("attn_output_7_cast_fp16")];
+            tensor<int32, [4]> var_224_perm_0 = const()[name = tensor<string, []>("op_224_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [2]> var_236 = const()[name = tensor<string, []>("op_236"), val = tensor<int32, [2]>([-1, 2])];
+            tensor<fp16, [1, 5, 2, 1]> var_224_cast_fp16 = transpose(perm = var_224_perm_0, x = attn_output_7_cast_fp16)[name = tensor<string, []>("transpose_12")];
+            tensor<fp16, [5, 2]> var_237_cast_fp16 = reshape(shape = var_236, x = var_224_cast_fp16)[name = tensor<string, []>("op_237_cast_fp16")];
+            tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2, 2]> model_transformer_h_1_attn_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_attn_c_proj_weight_to_fp16"), val = tensor<fp16, [2, 2]>([[0x1.e5cp-8, 0x1.54p-7], [0x1.2d8p-6, -0x1.594p-6]])];
+            tensor<fp16, [5, 2]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = var_237_cast_fp16, y = model_transformer_h_1_attn_c_proj_weight_to_fp16)[name = tensor<string, []>("matmul_7_cast_fp16")];
+            tensor<int32, [3]> var_239 = const()[name = tensor<string, []>("op_239"), val = tensor<int32, [3]>([1, 5, 2])];
+            tensor<fp16, [1, 5, 2]> input_15_cast_fp16 = reshape(shape = var_239, x = matmul_7_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<fp16, [1, 5, 2]> input_17_cast_fp16 = add(x = input_15_cast_fp16, y = input_13_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [2]> model_transformer_h_1_ln_2_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_ln_2_weight_to_fp16"), val = tensor<fp16, [2]>([0x1p+0, 0x1p+0])];
+            tensor<fp16, [2]> model_transformer_h_1_ln_2_bias_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_ln_2_bias_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
+            tensor<fp16, [1, 5, 2]> x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = model_transformer_h_1_ln_2_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_h_1_ln_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
+            tensor<int32, [2]> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, [2]>([-1, 2])];
+            tensor<fp16, [5, 2]> var_255_cast_fp16 = reshape(shape = var_254, x = x_25_cast_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
+            tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [2, 8]> model_transformer_h_1_mlp_c_fc_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_mlp_c_fc_weight_to_fp16"), val = tensor<fp16, [2, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201984)))];
+            tensor<fp16, [5, 8]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = var_255_cast_fp16, y = model_transformer_h_1_mlp_c_fc_weight_to_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")];
+            tensor<int32, [3]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [3]>([1, 5, 8])];
+            tensor<fp16, [1, 5, 8]> input_19_cast_fp16 = reshape(shape = var_257, x = matmul_8_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<string, []> x_29_mode_0 = const()[name = tensor<string, []>("x_29_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
+            tensor<fp16, [1, 5, 8]> x_29_cast_fp16 = gelu(mode = x_29_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
+            tensor<int32, [2]> var_276 = const()[name = tensor<string, []>("op_276"), val = tensor<int32, [2]>([-1, 8])];
+            tensor<fp16, [5, 8]> var_277_cast_fp16 = reshape(shape = var_276, x = x_29_cast_fp16)[name = tensor<string, []>("op_277_cast_fp16")];
+            tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [8, 2]> model_transformer_h_1_mlp_c_proj_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_h_1_mlp_c_proj_weight_to_fp16"), val = tensor<fp16, [8, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202112)))];
+            tensor<fp16, [5, 2]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = var_277_cast_fp16, y = model_transformer_h_1_mlp_c_proj_weight_to_fp16)[name = tensor<string, []>("matmul_9_cast_fp16")];
+            tensor<int32, [3]> var_279 = const()[name = tensor<string, []>("op_279"), val = tensor<int32, [3]>([1, 5, 2])];
+            tensor<fp16, [1, 5, 2]> input_21_cast_fp16 = reshape(shape = var_279, x = matmul_9_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [1, 5, 2]> input_23_cast_fp16 = add(x = input_17_cast_fp16, y = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<int32, [1]> hidden_states_axes_0 = const()[name = tensor<string, []>("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [2]> model_transformer_ln_f_weight_to_fp16 = const()[name = tensor<string, []>("model_transformer_ln_f_weight_to_fp16"), val = tensor<fp16, [2]>([0x1p+0, 0x1p+0])];
+            tensor<fp16, [2]> model_transformer_ln_f_bias_to_fp16 = const()[name = tensor<string, []>("model_transformer_ln_f_bias_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
+            tensor<fp16, [1, 5, 2]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, beta = model_transformer_ln_f_bias_to_fp16, epsilon = var_19_to_fp16, gamma = model_transformer_ln_f_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
+            tensor<fp16, [50257]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [50257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202240)))];
+            tensor<fp16, [1, 5, 50257]> var_289 = linear(bias = linear_0_bias_0_to_fp16, weight = model_transformer_wte_weight_to_fp16, x = hidden_states_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+        } -> (var_289);
+}

CompiledModel/TinyGPT2.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:600d58440df9bc1406c9ce36928c1c18ed13b8a7bc51fb2e1157fa0b8ff39b66
+size 302818

TinyGPT2.mlmodelc.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a33d2b992f2c73cbadbe6cfdfbf400ebcc1e144bc5a13933b3200e300ab752b7
+size 191261