diff --git a/.DS_Store b/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..7fe8a795c536b259b887a62f7554dd7c2005edfc Binary files /dev/null and b/.DS_Store differ diff --git a/.gitattributes b/.gitattributes index a6344aac8c09253b3b630fb776ae94478aa0275b..901b023557ddecf4cbf7f59443cb48582aba25a1 100644 --- a/.gitattributes +++ b/.gitattributes @@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text *.zst filter=lfs diff=lfs merge=lfs -text *tfevents* filter=lfs diff=lfs merge=lfs -text +*.mlmodelc/* filter=lfs diff=lfs merge=lfs -text diff --git a/base.en/.DS_Store b/base.en/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..2a4d5b10e7293a2c57aef21bfe774209a226db00 Binary files /dev/null and b/base.en/.DS_Store differ diff --git a/base.en/ggml-base.en-encoder.mlmodelc/analytics/coremldata.bin b/base.en/ggml-base.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d0a2a431499e74b9ebb3e9c187d6a8e30ed99faa --- /dev/null +++ b/base.en/ggml-base.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bac2e6295226f3bec590131e715797af43a72773913adb05f1da7c302794858 +size 243 diff --git a/base.en/ggml-base.en-encoder.mlmodelc/coremldata.bin b/base.en/ggml-base.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..208f1486e03a00183268dd0c67064971d3651772 --- /dev/null +++ b/base.en/ggml-base.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dadf937987b5a4db9fa964ba6ddb17c4085b581cf394b8930c13139b79db093 +size 392 diff --git a/base.en/ggml-base.en-encoder.mlmodelc/metadata.json b/base.en/ggml-base.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..e16fecbb963abbb1e5074028019d5caf260f4475 --- /dev/null +++ b/base.en/ggml-base.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,72 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 512)", + "shortDescription" : "", + "shape" : "[1, 1500, 512]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 6, + "Gelu" : 8, + "LayerNorm" : 13, + "Transpose" : 7, + "Softmax" : 48, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 13, + "Einsum" : 96, + "ExpandDims" : 1, + "Split" : 18, + "Conv" : 38 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2025-12-07", + "com.github.apple.coremltools.source" : "torch==2.10.0.dev20251207", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_base_en", + "method" : "predict" + } +] \ No newline at end of file diff --git a/base.en/ggml-base.en-encoder.mlmodelc/model.mil b/base.en/ggml-base.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9809ec63335e5dd4729ed04b904d3020710081e3 --- /dev/null +++ b/base.en/ggml-base.en-encoder.mlmodelc/model.mil @@ -0,0 +1,733 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.10.0.dev20251207"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor logmel_data) { + tensor var_32_pad_type_0 = const()[name = tensor("op_32_pad_type_0"), val = tensor("custom")]; + tensor var_32_pad_0 = const()[name = tensor("op_32_pad_0"), val = tensor([1, 1])]; + tensor var_32_strides_0 = const()[name = tensor("op_32_strides_0"), val = tensor([1])]; + tensor var_32_dilations_0 = const()[name = tensor("op_32_dilations_0"), val = tensor([1])]; + tensor var_32_groups_0 = const()[name = tensor("op_32_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245888)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_27")]; + tensor var_32_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_32_dilations_0, groups = var_32_groups_0, pad = var_32_pad_0, pad_type = var_32_pad_type_0, strides = var_32_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_32_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_32_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_50_pad_type_0 = const()[name = tensor("op_50_pad_type_0"), val = tensor("custom")]; + tensor var_50_pad_0 = const()[name = tensor("op_50_pad_0"), val = tensor([1, 1])]; + tensor var_50_strides_0 = const()[name = tensor("op_50_strides_0"), val = tensor([2])]; + tensor var_50_dilations_0 = const()[name = tensor("op_50_dilations_0"), val = tensor([1])]; + tensor var_50_groups_0 = const()[name = tensor("op_50_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246976)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1819904)))]; + tensor var_50_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_50_dilations_0, groups = var_50_groups_0, pad = var_50_pad_0, pad_type = var_50_pad_type_0, strides = var_50_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_50_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_50_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_55_to_fp16 = const()[name = tensor("op_55_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1820992)))]; + tensor var_57_cast_fp16 = add(x = x_3_cast_fp16, y = var_55_to_fp16)[name = tensor("op_57_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_57_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3357056)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3358144)))]; + tensor var_88_to_fp16 = const()[name = tensor("op_88_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_88_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_123_weight_0_to_fp16 = const()[name = tensor("op_123_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3359232)))]; + tensor var_123_bias_0_to_fp16 = const()[name = tensor("op_123_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3883584)))]; + tensor var_123_cast_fp16 = conv(bias = var_123_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_123_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_123_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3884672)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_121_pad_type_0 = const()[name = tensor("op_121_pad_type_0"), val = tensor("valid")]; + tensor var_121_strides_0 = const()[name = tensor("op_121_strides_0"), val = tensor([1, 1])]; + tensor var_121_pad_0 = const()[name = tensor("op_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_121_dilations_0 = const()[name = tensor("op_121_dilations_0"), val = tensor([1, 1])]; + tensor var_121_groups_0 = const()[name = tensor("op_121_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4409024)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4933376)))]; + tensor var_121_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_121_dilations_0, groups = var_121_groups_0, pad = var_121_pad_0, pad_type = var_121_pad_type_0, strides = var_121_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_121_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_124_axis_0 = const()[name = tensor("op_124_axis_0"), val = tensor(1)]; + tensor var_124_cast_fp16_0, tensor var_124_cast_fp16_1, tensor var_124_cast_fp16_2, tensor var_124_cast_fp16_3, tensor var_124_cast_fp16_4, tensor var_124_cast_fp16_5, tensor var_124_cast_fp16_6, tensor var_124_cast_fp16_7 = split(axis = var_124_axis_0, split_sizes = tile_0, x = var_123_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor var_133_perm_0 = const()[name = tensor("op_133_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_134_axis_0 = const()[name = tensor("op_134_axis_0"), val = tensor(3)]; + tensor var_133_cast_fp16 = transpose(perm = var_133_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_6")]; + tensor var_134_cast_fp16_0, tensor var_134_cast_fp16_1, tensor var_134_cast_fp16_2, tensor var_134_cast_fp16_3, tensor var_134_cast_fp16_4, tensor var_134_cast_fp16_5, tensor var_134_cast_fp16_6, tensor var_134_cast_fp16_7 = split(axis = var_134_axis_0, split_sizes = tile_1, x = var_133_cast_fp16)[name = tensor("op_134_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_143_axis_0 = const()[name = tensor("op_143_axis_0"), val = tensor(1)]; + tensor var_143_cast_fp16_0, tensor var_143_cast_fp16_1, tensor var_143_cast_fp16_2, tensor var_143_cast_fp16_3, tensor var_143_cast_fp16_4, tensor var_143_cast_fp16_5, tensor var_143_cast_fp16_6, tensor var_143_cast_fp16_7 = split(axis = var_143_axis_0, split_sizes = tile_2, x = var_121_cast_fp16)[name = tensor("op_143_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_134_cast_fp16_0, var_124_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_134_cast_fp16_1, var_124_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_134_cast_fp16_2, var_124_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_134_cast_fp16_3, var_124_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_134_cast_fp16_4, var_124_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_134_cast_fp16_5, var_124_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_134_cast_fp16_6, var_124_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_134_cast_fp16_7, var_124_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor var_168_cast_fp16 = softmax(axis = var_72, x = aw_1_cast_fp16)[name = tensor("op_168_cast_fp16")]; + tensor var_169_cast_fp16 = softmax(axis = var_72, x = aw_3_cast_fp16)[name = tensor("op_169_cast_fp16")]; + tensor var_170_cast_fp16 = softmax(axis = var_72, x = aw_5_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor var_171_cast_fp16 = softmax(axis = var_72, x = aw_7_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor var_172_cast_fp16 = softmax(axis = var_72, x = aw_9_cast_fp16)[name = tensor("op_172_cast_fp16")]; + tensor var_173_cast_fp16 = softmax(axis = var_72, x = aw_11_cast_fp16)[name = tensor("op_173_cast_fp16")]; + tensor var_174_cast_fp16 = softmax(axis = var_72, x = aw_13_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_175_cast_fp16 = softmax(axis = var_72, x = aw_15_cast_fp16)[name = tensor("op_175_cast_fp16")]; + tensor var_177_equation_0 = const()[name = tensor("op_177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_177_cast_fp16 = einsum(equation = var_177_equation_0, values = (var_143_cast_fp16_0, var_168_cast_fp16))[name = tensor("op_177_cast_fp16")]; + tensor var_179_equation_0 = const()[name = tensor("op_179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_179_cast_fp16 = einsum(equation = var_179_equation_0, values = (var_143_cast_fp16_1, var_169_cast_fp16))[name = tensor("op_179_cast_fp16")]; + tensor var_181_equation_0 = const()[name = tensor("op_181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_181_cast_fp16 = einsum(equation = var_181_equation_0, values = (var_143_cast_fp16_2, var_170_cast_fp16))[name = tensor("op_181_cast_fp16")]; + tensor var_183_equation_0 = const()[name = tensor("op_183_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_183_cast_fp16 = einsum(equation = var_183_equation_0, values = (var_143_cast_fp16_3, var_171_cast_fp16))[name = tensor("op_183_cast_fp16")]; + tensor var_185_equation_0 = const()[name = tensor("op_185_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_185_cast_fp16 = einsum(equation = var_185_equation_0, values = (var_143_cast_fp16_4, var_172_cast_fp16))[name = tensor("op_185_cast_fp16")]; + tensor var_187_equation_0 = const()[name = tensor("op_187_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_187_cast_fp16 = einsum(equation = var_187_equation_0, values = (var_143_cast_fp16_5, var_173_cast_fp16))[name = tensor("op_187_cast_fp16")]; + tensor var_189_equation_0 = const()[name = tensor("op_189_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_189_cast_fp16 = einsum(equation = var_189_equation_0, values = (var_143_cast_fp16_6, var_174_cast_fp16))[name = tensor("op_189_cast_fp16")]; + tensor var_191_equation_0 = const()[name = tensor("op_191_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_191_cast_fp16 = einsum(equation = var_191_equation_0, values = (var_143_cast_fp16_7, var_175_cast_fp16))[name = tensor("op_191_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_72, interleave = input_5_interleave_0, values = (var_177_cast_fp16, var_179_cast_fp16, var_181_cast_fp16, var_183_cast_fp16, var_185_cast_fp16, var_187_cast_fp16, var_189_cast_fp16, var_191_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_200_pad_type_0 = const()[name = tensor("op_200_pad_type_0"), val = tensor("valid")]; + tensor var_200_strides_0 = const()[name = tensor("op_200_strides_0"), val = tensor([1, 1])]; + tensor var_200_pad_0 = const()[name = tensor("op_200_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_200_dilations_0 = const()[name = tensor("op_200_dilations_0"), val = tensor([1, 1])]; + tensor var_200_groups_0 = const()[name = tensor("op_200_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4934464)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5458816)))]; + tensor var_200_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_200_dilations_0, groups = var_200_groups_0, pad = var_200_pad_0, pad_type = var_200_pad_type_0, strides = var_200_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_200_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5459904)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5460992)))]; + tensor var_210_to_fp16 = const()[name = tensor("op_210_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_210_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5462080)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7559296)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_236_pad_type_0 = const()[name = tensor("op_236_pad_type_0"), val = tensor("valid")]; + tensor var_236_strides_0 = const()[name = tensor("op_236_strides_0"), val = tensor([1, 1])]; + tensor var_236_pad_0 = const()[name = tensor("op_236_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_236_dilations_0 = const()[name = tensor("op_236_dilations_0"), val = tensor([1, 1])]; + tensor var_236_groups_0 = const()[name = tensor("op_236_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7563456)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9660672)))]; + tensor var_236_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_236_dilations_0, groups = var_236_groups_0, pad = var_236_pad_0, pad_type = var_236_pad_type_0, strides = var_236_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_236_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_245 = const()[name = tensor("op_245"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9661760)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9662848)))]; + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_261_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_296_weight_0_to_fp16 = const()[name = tensor("op_296_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9663936)))]; + tensor var_296_bias_0_to_fp16 = const()[name = tensor("op_296_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10188288)))]; + tensor var_296_cast_fp16 = conv(bias = var_296_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_296_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10189376)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_294_pad_type_0 = const()[name = tensor("op_294_pad_type_0"), val = tensor("valid")]; + tensor var_294_strides_0 = const()[name = tensor("op_294_strides_0"), val = tensor([1, 1])]; + tensor var_294_pad_0 = const()[name = tensor("op_294_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_294_dilations_0 = const()[name = tensor("op_294_dilations_0"), val = tensor([1, 1])]; + tensor var_294_groups_0 = const()[name = tensor("op_294_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10713728)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11238080)))]; + tensor var_294_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_294_dilations_0, groups = var_294_groups_0, pad = var_294_pad_0, pad_type = var_294_pad_type_0, strides = var_294_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_297_axis_0 = const()[name = tensor("op_297_axis_0"), val = tensor(1)]; + tensor var_297_cast_fp16_0, tensor var_297_cast_fp16_1, tensor var_297_cast_fp16_2, tensor var_297_cast_fp16_3, tensor var_297_cast_fp16_4, tensor var_297_cast_fp16_5, tensor var_297_cast_fp16_6, tensor var_297_cast_fp16_7 = split(axis = var_297_axis_0, split_sizes = tile_3, x = var_296_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_306_perm_0 = const()[name = tensor("op_306_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_307_axis_0 = const()[name = tensor("op_307_axis_0"), val = tensor(3)]; + tensor var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_5")]; + tensor var_307_cast_fp16_0, tensor var_307_cast_fp16_1, tensor var_307_cast_fp16_2, tensor var_307_cast_fp16_3, tensor var_307_cast_fp16_4, tensor var_307_cast_fp16_5, tensor var_307_cast_fp16_6, tensor var_307_cast_fp16_7 = split(axis = var_307_axis_0, split_sizes = tile_4, x = var_306_cast_fp16)[name = tensor("op_307_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_316_axis_0 = const()[name = tensor("op_316_axis_0"), val = tensor(1)]; + tensor var_316_cast_fp16_0, tensor var_316_cast_fp16_1, tensor var_316_cast_fp16_2, tensor var_316_cast_fp16_3, tensor var_316_cast_fp16_4, tensor var_316_cast_fp16_5, tensor var_316_cast_fp16_6, tensor var_316_cast_fp16_7 = split(axis = var_316_axis_0, split_sizes = tile_5, x = var_294_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_307_cast_fp16_0, var_297_cast_fp16_0))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_307_cast_fp16_1, var_297_cast_fp16_1))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_307_cast_fp16_2, var_297_cast_fp16_2))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_307_cast_fp16_3, var_297_cast_fp16_3))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_307_cast_fp16_4, var_297_cast_fp16_4))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_307_cast_fp16_5, var_297_cast_fp16_5))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_307_cast_fp16_6, var_297_cast_fp16_6))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_307_cast_fp16_7, var_297_cast_fp16_7))[name = tensor("aw_31_cast_fp16")]; + tensor var_341_cast_fp16 = softmax(axis = var_245, x = aw_17_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor var_342_cast_fp16 = softmax(axis = var_245, x = aw_19_cast_fp16)[name = tensor("op_342_cast_fp16")]; + tensor var_343_cast_fp16 = softmax(axis = var_245, x = aw_21_cast_fp16)[name = tensor("op_343_cast_fp16")]; + tensor var_344_cast_fp16 = softmax(axis = var_245, x = aw_23_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = softmax(axis = var_245, x = aw_25_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_346_cast_fp16 = softmax(axis = var_245, x = aw_27_cast_fp16)[name = tensor("op_346_cast_fp16")]; + tensor var_347_cast_fp16 = softmax(axis = var_245, x = aw_29_cast_fp16)[name = tensor("op_347_cast_fp16")]; + tensor var_348_cast_fp16 = softmax(axis = var_245, x = aw_31_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_350_equation_0 = const()[name = tensor("op_350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_350_cast_fp16 = einsum(equation = var_350_equation_0, values = (var_316_cast_fp16_0, var_341_cast_fp16))[name = tensor("op_350_cast_fp16")]; + tensor var_352_equation_0 = const()[name = tensor("op_352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_352_cast_fp16 = einsum(equation = var_352_equation_0, values = (var_316_cast_fp16_1, var_342_cast_fp16))[name = tensor("op_352_cast_fp16")]; + tensor var_354_equation_0 = const()[name = tensor("op_354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_354_cast_fp16 = einsum(equation = var_354_equation_0, values = (var_316_cast_fp16_2, var_343_cast_fp16))[name = tensor("op_354_cast_fp16")]; + tensor var_356_equation_0 = const()[name = tensor("op_356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_356_cast_fp16 = einsum(equation = var_356_equation_0, values = (var_316_cast_fp16_3, var_344_cast_fp16))[name = tensor("op_356_cast_fp16")]; + tensor var_358_equation_0 = const()[name = tensor("op_358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_358_cast_fp16 = einsum(equation = var_358_equation_0, values = (var_316_cast_fp16_4, var_345_cast_fp16))[name = tensor("op_358_cast_fp16")]; + tensor var_360_equation_0 = const()[name = tensor("op_360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_360_cast_fp16 = einsum(equation = var_360_equation_0, values = (var_316_cast_fp16_5, var_346_cast_fp16))[name = tensor("op_360_cast_fp16")]; + tensor var_362_equation_0 = const()[name = tensor("op_362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_362_cast_fp16 = einsum(equation = var_362_equation_0, values = (var_316_cast_fp16_6, var_347_cast_fp16))[name = tensor("op_362_cast_fp16")]; + tensor var_364_equation_0 = const()[name = tensor("op_364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_364_cast_fp16 = einsum(equation = var_364_equation_0, values = (var_316_cast_fp16_7, var_348_cast_fp16))[name = tensor("op_364_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_245, interleave = input_15_interleave_0, values = (var_350_cast_fp16, var_352_cast_fp16, var_354_cast_fp16, var_356_cast_fp16, var_358_cast_fp16, var_360_cast_fp16, var_362_cast_fp16, var_364_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_373_pad_type_0 = const()[name = tensor("op_373_pad_type_0"), val = tensor("valid")]; + tensor var_373_strides_0 = const()[name = tensor("op_373_strides_0"), val = tensor([1, 1])]; + tensor var_373_pad_0 = const()[name = tensor("op_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_373_dilations_0 = const()[name = tensor("op_373_dilations_0"), val = tensor([1, 1])]; + tensor var_373_groups_0 = const()[name = tensor("op_373_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11239168)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11763520)))]; + tensor var_373_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_373_dilations_0, groups = var_373_groups_0, pad = var_373_pad_0, pad_type = var_373_pad_type_0, strides = var_373_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_373_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11764608)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11765696)))]; + tensor var_383_to_fp16 = const()[name = tensor("op_383_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_383_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11766784)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13864000)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_409_pad_type_0 = const()[name = tensor("op_409_pad_type_0"), val = tensor("valid")]; + tensor var_409_strides_0 = const()[name = tensor("op_409_strides_0"), val = tensor([1, 1])]; + tensor var_409_pad_0 = const()[name = tensor("op_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_409_dilations_0 = const()[name = tensor("op_409_dilations_0"), val = tensor([1, 1])]; + tensor var_409_groups_0 = const()[name = tensor("op_409_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13868160)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15965376)))]; + tensor var_409_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_409_dilations_0, groups = var_409_groups_0, pad = var_409_pad_0, pad_type = var_409_pad_type_0, strides = var_409_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_409_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15966464)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15967552)))]; + tensor var_434_to_fp16 = const()[name = tensor("op_434_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_434_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_469_weight_0_to_fp16 = const()[name = tensor("op_469_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15968640)))]; + tensor var_469_bias_0_to_fp16 = const()[name = tensor("op_469_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16492992)))]; + tensor var_469_cast_fp16 = conv(bias = var_469_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_469_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16494080)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_467_pad_type_0 = const()[name = tensor("op_467_pad_type_0"), val = tensor("valid")]; + tensor var_467_strides_0 = const()[name = tensor("op_467_strides_0"), val = tensor([1, 1])]; + tensor var_467_pad_0 = const()[name = tensor("op_467_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_467_dilations_0 = const()[name = tensor("op_467_dilations_0"), val = tensor([1, 1])]; + tensor var_467_groups_0 = const()[name = tensor("op_467_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17018432)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17542784)))]; + tensor var_467_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_467_dilations_0, groups = var_467_groups_0, pad = var_467_pad_0, pad_type = var_467_pad_type_0, strides = var_467_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_467_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_470_axis_0 = const()[name = tensor("op_470_axis_0"), val = tensor(1)]; + tensor var_470_cast_fp16_0, tensor var_470_cast_fp16_1, tensor var_470_cast_fp16_2, tensor var_470_cast_fp16_3, tensor var_470_cast_fp16_4, tensor var_470_cast_fp16_5, tensor var_470_cast_fp16_6, tensor var_470_cast_fp16_7 = split(axis = var_470_axis_0, split_sizes = tile_6, x = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_479_perm_0 = const()[name = tensor("op_479_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_480_axis_0 = const()[name = tensor("op_480_axis_0"), val = tensor(3)]; + tensor var_479_cast_fp16 = transpose(perm = var_479_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_4")]; + tensor var_480_cast_fp16_0, tensor var_480_cast_fp16_1, tensor var_480_cast_fp16_2, tensor var_480_cast_fp16_3, tensor var_480_cast_fp16_4, tensor var_480_cast_fp16_5, tensor var_480_cast_fp16_6, tensor var_480_cast_fp16_7 = split(axis = var_480_axis_0, split_sizes = tile_7, x = var_479_cast_fp16)[name = tensor("op_480_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_489_axis_0 = const()[name = tensor("op_489_axis_0"), val = tensor(1)]; + tensor var_489_cast_fp16_0, tensor var_489_cast_fp16_1, tensor var_489_cast_fp16_2, tensor var_489_cast_fp16_3, tensor var_489_cast_fp16_4, tensor var_489_cast_fp16_5, tensor var_489_cast_fp16_6, tensor var_489_cast_fp16_7 = split(axis = var_489_axis_0, split_sizes = tile_8, x = var_467_cast_fp16)[name = tensor("op_489_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_480_cast_fp16_0, var_470_cast_fp16_0))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_480_cast_fp16_1, var_470_cast_fp16_1))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_480_cast_fp16_2, var_470_cast_fp16_2))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_480_cast_fp16_3, var_470_cast_fp16_3))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_480_cast_fp16_4, var_470_cast_fp16_4))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_480_cast_fp16_5, var_470_cast_fp16_5))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_480_cast_fp16_6, var_470_cast_fp16_6))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_480_cast_fp16_7, var_470_cast_fp16_7))[name = tensor("aw_47_cast_fp16")]; + tensor var_514_cast_fp16 = softmax(axis = var_418, x = aw_33_cast_fp16)[name = tensor("op_514_cast_fp16")]; + tensor var_515_cast_fp16 = softmax(axis = var_418, x = aw_35_cast_fp16)[name = tensor("op_515_cast_fp16")]; + tensor var_516_cast_fp16 = softmax(axis = var_418, x = aw_37_cast_fp16)[name = tensor("op_516_cast_fp16")]; + tensor var_517_cast_fp16 = softmax(axis = var_418, x = aw_39_cast_fp16)[name = tensor("op_517_cast_fp16")]; + tensor var_518_cast_fp16 = softmax(axis = var_418, x = aw_41_cast_fp16)[name = tensor("op_518_cast_fp16")]; + tensor var_519_cast_fp16 = softmax(axis = var_418, x = aw_43_cast_fp16)[name = tensor("op_519_cast_fp16")]; + tensor var_520_cast_fp16 = softmax(axis = var_418, x = aw_45_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor var_521_cast_fp16 = softmax(axis = var_418, x = aw_47_cast_fp16)[name = tensor("op_521_cast_fp16")]; + tensor var_523_equation_0 = const()[name = tensor("op_523_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_523_cast_fp16 = einsum(equation = var_523_equation_0, values = (var_489_cast_fp16_0, var_514_cast_fp16))[name = tensor("op_523_cast_fp16")]; + tensor var_525_equation_0 = const()[name = tensor("op_525_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_525_cast_fp16 = einsum(equation = var_525_equation_0, values = (var_489_cast_fp16_1, var_515_cast_fp16))[name = tensor("op_525_cast_fp16")]; + tensor var_527_equation_0 = const()[name = tensor("op_527_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_527_cast_fp16 = einsum(equation = var_527_equation_0, values = (var_489_cast_fp16_2, var_516_cast_fp16))[name = tensor("op_527_cast_fp16")]; + tensor var_529_equation_0 = const()[name = tensor("op_529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_529_cast_fp16 = einsum(equation = var_529_equation_0, values = (var_489_cast_fp16_3, var_517_cast_fp16))[name = tensor("op_529_cast_fp16")]; + tensor var_531_equation_0 = const()[name = tensor("op_531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_531_cast_fp16 = einsum(equation = var_531_equation_0, values = (var_489_cast_fp16_4, var_518_cast_fp16))[name = tensor("op_531_cast_fp16")]; + tensor var_533_equation_0 = const()[name = tensor("op_533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_533_cast_fp16 = einsum(equation = var_533_equation_0, values = (var_489_cast_fp16_5, var_519_cast_fp16))[name = tensor("op_533_cast_fp16")]; + tensor var_535_equation_0 = const()[name = tensor("op_535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_535_cast_fp16 = einsum(equation = var_535_equation_0, values = (var_489_cast_fp16_6, var_520_cast_fp16))[name = tensor("op_535_cast_fp16")]; + tensor var_537_equation_0 = const()[name = tensor("op_537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_537_cast_fp16 = einsum(equation = var_537_equation_0, values = (var_489_cast_fp16_7, var_521_cast_fp16))[name = tensor("op_537_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_418, interleave = input_25_interleave_0, values = (var_523_cast_fp16, var_525_cast_fp16, var_527_cast_fp16, var_529_cast_fp16, var_531_cast_fp16, var_533_cast_fp16, var_535_cast_fp16, var_537_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_546_pad_type_0 = const()[name = tensor("op_546_pad_type_0"), val = tensor("valid")]; + tensor var_546_strides_0 = const()[name = tensor("op_546_strides_0"), val = tensor([1, 1])]; + tensor var_546_pad_0 = const()[name = tensor("op_546_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_546_dilations_0 = const()[name = tensor("op_546_dilations_0"), val = tensor([1, 1])]; + tensor var_546_groups_0 = const()[name = tensor("op_546_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17543872)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18068224)))]; + tensor var_546_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_546_dilations_0, groups = var_546_groups_0, pad = var_546_pad_0, pad_type = var_546_pad_type_0, strides = var_546_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_546_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_546_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18069312)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18070400)))]; + tensor var_556_to_fp16 = const()[name = tensor("op_556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_556_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18071488)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20168704)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_582_pad_type_0 = const()[name = tensor("op_582_pad_type_0"), val = tensor("valid")]; + tensor var_582_strides_0 = const()[name = tensor("op_582_strides_0"), val = tensor([1, 1])]; + tensor var_582_pad_0 = const()[name = tensor("op_582_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_582_dilations_0 = const()[name = tensor("op_582_dilations_0"), val = tensor([1, 1])]; + tensor var_582_groups_0 = const()[name = tensor("op_582_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20172864)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22270080)))]; + tensor var_582_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_582_dilations_0, groups = var_582_groups_0, pad = var_582_pad_0, pad_type = var_582_pad_type_0, strides = var_582_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_582_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22271168)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22272256)))]; + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_607_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_642_weight_0_to_fp16 = const()[name = tensor("op_642_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22273344)))]; + tensor var_642_bias_0_to_fp16 = const()[name = tensor("op_642_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22797696)))]; + tensor var_642_cast_fp16 = conv(bias = var_642_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_642_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22798784)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_640_pad_type_0 = const()[name = tensor("op_640_pad_type_0"), val = tensor("valid")]; + tensor var_640_strides_0 = const()[name = tensor("op_640_strides_0"), val = tensor([1, 1])]; + tensor var_640_pad_0 = const()[name = tensor("op_640_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_640_dilations_0 = const()[name = tensor("op_640_dilations_0"), val = tensor([1, 1])]; + tensor var_640_groups_0 = const()[name = tensor("op_640_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23323136)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23847488)))]; + tensor var_640_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_640_dilations_0, groups = var_640_groups_0, pad = var_640_pad_0, pad_type = var_640_pad_type_0, strides = var_640_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_640_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_643_axis_0 = const()[name = tensor("op_643_axis_0"), val = tensor(1)]; + tensor var_643_cast_fp16_0, tensor var_643_cast_fp16_1, tensor var_643_cast_fp16_2, tensor var_643_cast_fp16_3, tensor var_643_cast_fp16_4, tensor var_643_cast_fp16_5, tensor var_643_cast_fp16_6, tensor var_643_cast_fp16_7 = split(axis = var_643_axis_0, split_sizes = tile_9, x = var_642_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(3)]; + tensor var_652_cast_fp16 = transpose(perm = var_652_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_3")]; + tensor var_653_cast_fp16_0, tensor var_653_cast_fp16_1, tensor var_653_cast_fp16_2, tensor var_653_cast_fp16_3, tensor var_653_cast_fp16_4, tensor var_653_cast_fp16_5, tensor var_653_cast_fp16_6, tensor var_653_cast_fp16_7 = split(axis = var_653_axis_0, split_sizes = tile_10, x = var_652_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_662_axis_0 = const()[name = tensor("op_662_axis_0"), val = tensor(1)]; + tensor var_662_cast_fp16_0, tensor var_662_cast_fp16_1, tensor var_662_cast_fp16_2, tensor var_662_cast_fp16_3, tensor var_662_cast_fp16_4, tensor var_662_cast_fp16_5, tensor var_662_cast_fp16_6, tensor var_662_cast_fp16_7 = split(axis = var_662_axis_0, split_sizes = tile_11, x = var_640_cast_fp16)[name = tensor("op_662_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_653_cast_fp16_0, var_643_cast_fp16_0))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_653_cast_fp16_1, var_643_cast_fp16_1))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_653_cast_fp16_2, var_643_cast_fp16_2))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_653_cast_fp16_3, var_643_cast_fp16_3))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_653_cast_fp16_4, var_643_cast_fp16_4))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_653_cast_fp16_5, var_643_cast_fp16_5))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_653_cast_fp16_6, var_643_cast_fp16_6))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_653_cast_fp16_7, var_643_cast_fp16_7))[name = tensor("aw_63_cast_fp16")]; + tensor var_687_cast_fp16 = softmax(axis = var_591, x = aw_49_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor var_688_cast_fp16 = softmax(axis = var_591, x = aw_51_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_689_cast_fp16 = softmax(axis = var_591, x = aw_53_cast_fp16)[name = tensor("op_689_cast_fp16")]; + tensor var_690_cast_fp16 = softmax(axis = var_591, x = aw_55_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor var_691_cast_fp16 = softmax(axis = var_591, x = aw_57_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor var_692_cast_fp16 = softmax(axis = var_591, x = aw_59_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_693_cast_fp16 = softmax(axis = var_591, x = aw_61_cast_fp16)[name = tensor("op_693_cast_fp16")]; + tensor var_694_cast_fp16 = softmax(axis = var_591, x = aw_63_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_696_equation_0 = const()[name = tensor("op_696_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_696_cast_fp16 = einsum(equation = var_696_equation_0, values = (var_662_cast_fp16_0, var_687_cast_fp16))[name = tensor("op_696_cast_fp16")]; + tensor var_698_equation_0 = const()[name = tensor("op_698_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_698_cast_fp16 = einsum(equation = var_698_equation_0, values = (var_662_cast_fp16_1, var_688_cast_fp16))[name = tensor("op_698_cast_fp16")]; + tensor var_700_equation_0 = const()[name = tensor("op_700_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_700_cast_fp16 = einsum(equation = var_700_equation_0, values = (var_662_cast_fp16_2, var_689_cast_fp16))[name = tensor("op_700_cast_fp16")]; + tensor var_702_equation_0 = const()[name = tensor("op_702_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_702_cast_fp16 = einsum(equation = var_702_equation_0, values = (var_662_cast_fp16_3, var_690_cast_fp16))[name = tensor("op_702_cast_fp16")]; + tensor var_704_equation_0 = const()[name = tensor("op_704_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_704_cast_fp16 = einsum(equation = var_704_equation_0, values = (var_662_cast_fp16_4, var_691_cast_fp16))[name = tensor("op_704_cast_fp16")]; + tensor var_706_equation_0 = const()[name = tensor("op_706_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_706_cast_fp16 = einsum(equation = var_706_equation_0, values = (var_662_cast_fp16_5, var_692_cast_fp16))[name = tensor("op_706_cast_fp16")]; + tensor var_708_equation_0 = const()[name = tensor("op_708_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_708_cast_fp16 = einsum(equation = var_708_equation_0, values = (var_662_cast_fp16_6, var_693_cast_fp16))[name = tensor("op_708_cast_fp16")]; + tensor var_710_equation_0 = const()[name = tensor("op_710_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_710_cast_fp16 = einsum(equation = var_710_equation_0, values = (var_662_cast_fp16_7, var_694_cast_fp16))[name = tensor("op_710_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_591, interleave = input_35_interleave_0, values = (var_696_cast_fp16, var_698_cast_fp16, var_700_cast_fp16, var_702_cast_fp16, var_704_cast_fp16, var_706_cast_fp16, var_708_cast_fp16, var_710_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_719_pad_type_0 = const()[name = tensor("op_719_pad_type_0"), val = tensor("valid")]; + tensor var_719_strides_0 = const()[name = tensor("op_719_strides_0"), val = tensor([1, 1])]; + tensor var_719_pad_0 = const()[name = tensor("op_719_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_719_dilations_0 = const()[name = tensor("op_719_dilations_0"), val = tensor([1, 1])]; + tensor var_719_groups_0 = const()[name = tensor("op_719_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23848576)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24372928)))]; + tensor var_719_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_719_dilations_0, groups = var_719_groups_0, pad = var_719_pad_0, pad_type = var_719_pad_type_0, strides = var_719_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_719_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_719_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24375104)))]; + tensor var_729_to_fp16 = const()[name = tensor("op_729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_729_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24376192)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_755_pad_type_0 = const()[name = tensor("op_755_pad_type_0"), val = tensor("valid")]; + tensor var_755_strides_0 = const()[name = tensor("op_755_strides_0"), val = tensor([1, 1])]; + tensor var_755_pad_0 = const()[name = tensor("op_755_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_755_dilations_0 = const()[name = tensor("op_755_dilations_0"), val = tensor([1, 1])]; + tensor var_755_groups_0 = const()[name = tensor("op_755_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26477568)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28574784)))]; + tensor var_755_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_755_dilations_0, groups = var_755_groups_0, pad = var_755_pad_0, pad_type = var_755_pad_type_0, strides = var_755_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_755_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_755_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28575872)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28576960)))]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_780_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_815_weight_0_to_fp16 = const()[name = tensor("op_815_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28578048)))]; + tensor var_815_bias_0_to_fp16 = const()[name = tensor("op_815_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29102400)))]; + tensor var_815_cast_fp16 = conv(bias = var_815_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_815_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29103488)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_813_pad_type_0 = const()[name = tensor("op_813_pad_type_0"), val = tensor("valid")]; + tensor var_813_strides_0 = const()[name = tensor("op_813_strides_0"), val = tensor([1, 1])]; + tensor var_813_pad_0 = const()[name = tensor("op_813_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_813_dilations_0 = const()[name = tensor("op_813_dilations_0"), val = tensor([1, 1])]; + tensor var_813_groups_0 = const()[name = tensor("op_813_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29627840)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30152192)))]; + tensor var_813_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_813_dilations_0, groups = var_813_groups_0, pad = var_813_pad_0, pad_type = var_813_pad_type_0, strides = var_813_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_816_axis_0 = const()[name = tensor("op_816_axis_0"), val = tensor(1)]; + tensor var_816_cast_fp16_0, tensor var_816_cast_fp16_1, tensor var_816_cast_fp16_2, tensor var_816_cast_fp16_3, tensor var_816_cast_fp16_4, tensor var_816_cast_fp16_5, tensor var_816_cast_fp16_6, tensor var_816_cast_fp16_7 = split(axis = var_816_axis_0, split_sizes = tile_12, x = var_815_cast_fp16)[name = tensor("op_816_cast_fp16")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_826_axis_0 = const()[name = tensor("op_826_axis_0"), val = tensor(3)]; + tensor var_825_cast_fp16 = transpose(perm = var_825_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_2")]; + tensor var_826_cast_fp16_0, tensor var_826_cast_fp16_1, tensor var_826_cast_fp16_2, tensor var_826_cast_fp16_3, tensor var_826_cast_fp16_4, tensor var_826_cast_fp16_5, tensor var_826_cast_fp16_6, tensor var_826_cast_fp16_7 = split(axis = var_826_axis_0, split_sizes = tile_13, x = var_825_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_835_axis_0 = const()[name = tensor("op_835_axis_0"), val = tensor(1)]; + tensor var_835_cast_fp16_0, tensor var_835_cast_fp16_1, tensor var_835_cast_fp16_2, tensor var_835_cast_fp16_3, tensor var_835_cast_fp16_4, tensor var_835_cast_fp16_5, tensor var_835_cast_fp16_6, tensor var_835_cast_fp16_7 = split(axis = var_835_axis_0, split_sizes = tile_14, x = var_813_cast_fp16)[name = tensor("op_835_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_826_cast_fp16_0, var_816_cast_fp16_0))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_826_cast_fp16_1, var_816_cast_fp16_1))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_826_cast_fp16_2, var_816_cast_fp16_2))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_826_cast_fp16_3, var_816_cast_fp16_3))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_826_cast_fp16_4, var_816_cast_fp16_4))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_826_cast_fp16_5, var_816_cast_fp16_5))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_826_cast_fp16_6, var_816_cast_fp16_6))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_826_cast_fp16_7, var_816_cast_fp16_7))[name = tensor("aw_79_cast_fp16")]; + tensor var_860_cast_fp16 = softmax(axis = var_764, x = aw_65_cast_fp16)[name = tensor("op_860_cast_fp16")]; + tensor var_861_cast_fp16 = softmax(axis = var_764, x = aw_67_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_862_cast_fp16 = softmax(axis = var_764, x = aw_69_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = softmax(axis = var_764, x = aw_71_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = softmax(axis = var_764, x = aw_73_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor var_865_cast_fp16 = softmax(axis = var_764, x = aw_75_cast_fp16)[name = tensor("op_865_cast_fp16")]; + tensor var_866_cast_fp16 = softmax(axis = var_764, x = aw_77_cast_fp16)[name = tensor("op_866_cast_fp16")]; + tensor var_867_cast_fp16 = softmax(axis = var_764, x = aw_79_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_869_equation_0 = const()[name = tensor("op_869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_869_cast_fp16 = einsum(equation = var_869_equation_0, values = (var_835_cast_fp16_0, var_860_cast_fp16))[name = tensor("op_869_cast_fp16")]; + tensor var_871_equation_0 = const()[name = tensor("op_871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_871_cast_fp16 = einsum(equation = var_871_equation_0, values = (var_835_cast_fp16_1, var_861_cast_fp16))[name = tensor("op_871_cast_fp16")]; + tensor var_873_equation_0 = const()[name = tensor("op_873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_873_cast_fp16 = einsum(equation = var_873_equation_0, values = (var_835_cast_fp16_2, var_862_cast_fp16))[name = tensor("op_873_cast_fp16")]; + tensor var_875_equation_0 = const()[name = tensor("op_875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_875_cast_fp16 = einsum(equation = var_875_equation_0, values = (var_835_cast_fp16_3, var_863_cast_fp16))[name = tensor("op_875_cast_fp16")]; + tensor var_877_equation_0 = const()[name = tensor("op_877_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_877_cast_fp16 = einsum(equation = var_877_equation_0, values = (var_835_cast_fp16_4, var_864_cast_fp16))[name = tensor("op_877_cast_fp16")]; + tensor var_879_equation_0 = const()[name = tensor("op_879_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_879_cast_fp16 = einsum(equation = var_879_equation_0, values = (var_835_cast_fp16_5, var_865_cast_fp16))[name = tensor("op_879_cast_fp16")]; + tensor var_881_equation_0 = const()[name = tensor("op_881_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_881_cast_fp16 = einsum(equation = var_881_equation_0, values = (var_835_cast_fp16_6, var_866_cast_fp16))[name = tensor("op_881_cast_fp16")]; + tensor var_883_equation_0 = const()[name = tensor("op_883_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_883_cast_fp16 = einsum(equation = var_883_equation_0, values = (var_835_cast_fp16_7, var_867_cast_fp16))[name = tensor("op_883_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_764, interleave = input_45_interleave_0, values = (var_869_cast_fp16, var_871_cast_fp16, var_873_cast_fp16, var_875_cast_fp16, var_877_cast_fp16, var_879_cast_fp16, var_881_cast_fp16, var_883_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_892_pad_type_0 = const()[name = tensor("op_892_pad_type_0"), val = tensor("valid")]; + tensor var_892_strides_0 = const()[name = tensor("op_892_strides_0"), val = tensor([1, 1])]; + tensor var_892_pad_0 = const()[name = tensor("op_892_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_892_dilations_0 = const()[name = tensor("op_892_dilations_0"), val = tensor([1, 1])]; + tensor var_892_groups_0 = const()[name = tensor("op_892_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30153280)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30677632)))]; + tensor var_892_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_892_dilations_0, groups = var_892_groups_0, pad = var_892_pad_0, pad_type = var_892_pad_type_0, strides = var_892_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_892_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30678720)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30679808)))]; + tensor var_902_to_fp16 = const()[name = tensor("op_902_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_902_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30680896)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32778112)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_928_pad_type_0 = const()[name = tensor("op_928_pad_type_0"), val = tensor("valid")]; + tensor var_928_strides_0 = const()[name = tensor("op_928_strides_0"), val = tensor([1, 1])]; + tensor var_928_pad_0 = const()[name = tensor("op_928_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_928_dilations_0 = const()[name = tensor("op_928_dilations_0"), val = tensor([1, 1])]; + tensor var_928_groups_0 = const()[name = tensor("op_928_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32782272)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34879488)))]; + tensor var_928_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_928_dilations_0, groups = var_928_groups_0, pad = var_928_pad_0, pad_type = var_928_pad_type_0, strides = var_928_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_928_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_937 = const()[name = tensor("op_937"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34880576)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34881664)))]; + tensor var_953_to_fp16 = const()[name = tensor("op_953_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_953_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_988_weight_0_to_fp16 = const()[name = tensor("op_988_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34882752)))]; + tensor var_988_bias_0_to_fp16 = const()[name = tensor("op_988_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35407104)))]; + tensor var_988_cast_fp16 = conv(bias = var_988_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_988_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_988_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35408192)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_986_pad_type_0 = const()[name = tensor("op_986_pad_type_0"), val = tensor("valid")]; + tensor var_986_strides_0 = const()[name = tensor("op_986_strides_0"), val = tensor([1, 1])]; + tensor var_986_pad_0 = const()[name = tensor("op_986_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_986_dilations_0 = const()[name = tensor("op_986_dilations_0"), val = tensor([1, 1])]; + tensor var_986_groups_0 = const()[name = tensor("op_986_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35932544)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36456896)))]; + tensor var_986_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_986_dilations_0, groups = var_986_groups_0, pad = var_986_pad_0, pad_type = var_986_pad_type_0, strides = var_986_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_986_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_989_axis_0 = const()[name = tensor("op_989_axis_0"), val = tensor(1)]; + tensor var_989_cast_fp16_0, tensor var_989_cast_fp16_1, tensor var_989_cast_fp16_2, tensor var_989_cast_fp16_3, tensor var_989_cast_fp16_4, tensor var_989_cast_fp16_5, tensor var_989_cast_fp16_6, tensor var_989_cast_fp16_7 = split(axis = var_989_axis_0, split_sizes = tile_15, x = var_988_cast_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_998_perm_0 = const()[name = tensor("op_998_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(3)]; + tensor var_998_cast_fp16 = transpose(perm = var_998_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_999_cast_fp16_0, tensor var_999_cast_fp16_1, tensor var_999_cast_fp16_2, tensor var_999_cast_fp16_3, tensor var_999_cast_fp16_4, tensor var_999_cast_fp16_5, tensor var_999_cast_fp16_6, tensor var_999_cast_fp16_7 = split(axis = var_999_axis_0, split_sizes = tile_16, x = var_998_cast_fp16)[name = tensor("op_999_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1008_axis_0 = const()[name = tensor("op_1008_axis_0"), val = tensor(1)]; + tensor var_1008_cast_fp16_0, tensor var_1008_cast_fp16_1, tensor var_1008_cast_fp16_2, tensor var_1008_cast_fp16_3, tensor var_1008_cast_fp16_4, tensor var_1008_cast_fp16_5, tensor var_1008_cast_fp16_6, tensor var_1008_cast_fp16_7 = split(axis = var_1008_axis_0, split_sizes = tile_17, x = var_986_cast_fp16)[name = tensor("op_1008_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_999_cast_fp16_0, var_989_cast_fp16_0))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_999_cast_fp16_1, var_989_cast_fp16_1))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_999_cast_fp16_2, var_989_cast_fp16_2))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_999_cast_fp16_3, var_989_cast_fp16_3))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_999_cast_fp16_4, var_989_cast_fp16_4))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_999_cast_fp16_5, var_989_cast_fp16_5))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_999_cast_fp16_6, var_989_cast_fp16_6))[name = tensor("aw_93_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_999_cast_fp16_7, var_989_cast_fp16_7))[name = tensor("aw_cast_fp16")]; + tensor var_1033_cast_fp16 = softmax(axis = var_937, x = aw_81_cast_fp16)[name = tensor("op_1033_cast_fp16")]; + tensor var_1034_cast_fp16 = softmax(axis = var_937, x = aw_83_cast_fp16)[name = tensor("op_1034_cast_fp16")]; + tensor var_1035_cast_fp16 = softmax(axis = var_937, x = aw_85_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor var_1036_cast_fp16 = softmax(axis = var_937, x = aw_87_cast_fp16)[name = tensor("op_1036_cast_fp16")]; + tensor var_1037_cast_fp16 = softmax(axis = var_937, x = aw_89_cast_fp16)[name = tensor("op_1037_cast_fp16")]; + tensor var_1038_cast_fp16 = softmax(axis = var_937, x = aw_91_cast_fp16)[name = tensor("op_1038_cast_fp16")]; + tensor var_1039_cast_fp16 = softmax(axis = var_937, x = aw_93_cast_fp16)[name = tensor("op_1039_cast_fp16")]; + tensor var_1040_cast_fp16 = softmax(axis = var_937, x = aw_cast_fp16)[name = tensor("op_1040_cast_fp16")]; + tensor var_1042_equation_0 = const()[name = tensor("op_1042_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1042_cast_fp16 = einsum(equation = var_1042_equation_0, values = (var_1008_cast_fp16_0, var_1033_cast_fp16))[name = tensor("op_1042_cast_fp16")]; + tensor var_1044_equation_0 = const()[name = tensor("op_1044_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1044_cast_fp16 = einsum(equation = var_1044_equation_0, values = (var_1008_cast_fp16_1, var_1034_cast_fp16))[name = tensor("op_1044_cast_fp16")]; + tensor var_1046_equation_0 = const()[name = tensor("op_1046_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1046_cast_fp16 = einsum(equation = var_1046_equation_0, values = (var_1008_cast_fp16_2, var_1035_cast_fp16))[name = tensor("op_1046_cast_fp16")]; + tensor var_1048_equation_0 = const()[name = tensor("op_1048_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1048_cast_fp16 = einsum(equation = var_1048_equation_0, values = (var_1008_cast_fp16_3, var_1036_cast_fp16))[name = tensor("op_1048_cast_fp16")]; + tensor var_1050_equation_0 = const()[name = tensor("op_1050_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1050_cast_fp16 = einsum(equation = var_1050_equation_0, values = (var_1008_cast_fp16_4, var_1037_cast_fp16))[name = tensor("op_1050_cast_fp16")]; + tensor var_1052_equation_0 = const()[name = tensor("op_1052_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1052_cast_fp16 = einsum(equation = var_1052_equation_0, values = (var_1008_cast_fp16_5, var_1038_cast_fp16))[name = tensor("op_1052_cast_fp16")]; + tensor var_1054_equation_0 = const()[name = tensor("op_1054_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1054_cast_fp16 = einsum(equation = var_1054_equation_0, values = (var_1008_cast_fp16_6, var_1039_cast_fp16))[name = tensor("op_1054_cast_fp16")]; + tensor var_1056_equation_0 = const()[name = tensor("op_1056_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1056_cast_fp16 = einsum(equation = var_1056_equation_0, values = (var_1008_cast_fp16_7, var_1040_cast_fp16))[name = tensor("op_1056_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_937, interleave = input_55_interleave_0, values = (var_1042_cast_fp16, var_1044_cast_fp16, var_1046_cast_fp16, var_1048_cast_fp16, var_1050_cast_fp16, var_1052_cast_fp16, var_1054_cast_fp16, var_1056_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1065_pad_type_0 = const()[name = tensor("op_1065_pad_type_0"), val = tensor("valid")]; + tensor var_1065_strides_0 = const()[name = tensor("op_1065_strides_0"), val = tensor([1, 1])]; + tensor var_1065_pad_0 = const()[name = tensor("op_1065_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1065_dilations_0 = const()[name = tensor("op_1065_dilations_0"), val = tensor([1, 1])]; + tensor var_1065_groups_0 = const()[name = tensor("op_1065_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36457984)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36982336)))]; + tensor var_1065_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1065_dilations_0, groups = var_1065_groups_0, pad = var_1065_pad_0, pad_type = var_1065_pad_type_0, strides = var_1065_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1065_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1065_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36983424)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36984512)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1075_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36985600)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39082816)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_59_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_1101_pad_type_0 = const()[name = tensor("op_1101_pad_type_0"), val = tensor("valid")]; + tensor var_1101_strides_0 = const()[name = tensor("op_1101_strides_0"), val = tensor([1, 1])]; + tensor var_1101_pad_0 = const()[name = tensor("op_1101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1101_dilations_0 = const()[name = tensor("op_1101_dilations_0"), val = tensor([1, 1])]; + tensor var_1101_groups_0 = const()[name = tensor("op_1101_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39086976)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41184192)))]; + tensor var_1101_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1101_dilations_0, groups = var_1101_groups_0, pad = var_1101_pad_0, pad_type = var_1101_pad_type_0, strides = var_1101_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1101_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41185280)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41186368)))]; + tensor var_1115_to_fp16 = const()[name = tensor("op_1115_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_1115_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([2])]; + tensor var_1126_cast_fp16 = squeeze(axes = var_1126_axes_0, x = x_cast_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor var_1129_perm_0 = const()[name = tensor("op_1129_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1129_cast_fp16 = transpose(perm = var_1129_perm_0, x = var_1126_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_1129_cast_fp16_to_fp32_dtype_0, x = var_1129_cast_fp16)[name = tensor("cast_26")]; + } -> (output); +} \ No newline at end of file diff --git a/base.en/ggml-base.en-encoder.mlmodelc/weights/weight.bin b/base.en/ggml-base.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..86cecfd1f63401689bb12a0c1aa24cbb53c0295a --- /dev/null +++ b/base.en/ggml-base.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ddb14d23b665662c6bb22d659e9078e5304e7c4f346b66d66660b2efc07c461 +size 41187456 diff --git a/base.en/ggml-base.en.bin b/base.en/ggml-base.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..87c664c563ef3ff52424dd4fa925cf95b306dba6 --- /dev/null +++ b/base.en/ggml-base.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a03779c86df3323075f5e796cb2ce5029f00ec8869eee3fdfb897afe36c6d002 +size 147964211 diff --git a/base/.DS_Store b/base/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..0cccdecbf552204a0804f7b0a3e84b5e24ee345e Binary files /dev/null and b/base/.DS_Store differ diff --git a/base/ggml-base-encoder.mlmodelc/analytics/coremldata.bin b/base/ggml-base-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4d3ac25920351891c100fec2447ae7af22cd74a5 --- /dev/null +++ b/base/ggml-base-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3336d2fa661565cda8254e2605c4a0d14b30cb7166abd534a442d036dc7210b2 +size 243 diff --git a/base/ggml-base-encoder.mlmodelc/coremldata.bin b/base/ggml-base-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4d9a467f60ab78ce0c7da4e8af0ea14dd7b68774 --- /dev/null +++ b/base/ggml-base-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea2fd8343a60c71e243d887740ea21c619fe40d2bf4b898f3c499df4af43433b +size 320 diff --git a/base/ggml-base-encoder.mlmodelc/metadata.json b/base/ggml-base-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..de25825ed2d1cde0d2a5e4059d73019674f12f4b --- /dev/null +++ b/base/ggml-base-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 512)", + "shortDescription" : "", + "shape" : "[1, 1500, 512]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 6, + "Gelu" : 8, + "LayerNorm" : 13, + "Transpose" : 7, + "Softmax" : 48, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 13, + "Einsum" : 96, + "ExpandDims" : 1, + "Split" : 18, + "Conv" : 38 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_base", + "method" : "predict" + } +] \ No newline at end of file diff --git a/base/ggml-base-encoder.mlmodelc/model.mil b/base/ggml-base-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..19090e632d6d8b0353d98e50f2756e7d5f930331 --- /dev/null +++ b/base/ggml-base-encoder.mlmodelc/model.mil @@ -0,0 +1,733 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_32_pad_type_0 = const()[name = tensor("op_32_pad_type_0"), val = tensor("custom")]; + tensor var_32_pad_0 = const()[name = tensor("op_32_pad_0"), val = tensor([1, 1])]; + tensor var_32_strides_0 = const()[name = tensor("op_32_strides_0"), val = tensor([1])]; + tensor var_32_dilations_0 = const()[name = tensor("op_32_dilations_0"), val = tensor([1])]; + tensor var_32_groups_0 = const()[name = tensor("op_32_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245888)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_28")]; + tensor var_32_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_32_dilations_0, groups = var_32_groups_0, pad = var_32_pad_0, pad_type = var_32_pad_type_0, strides = var_32_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_32_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_32_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_50_pad_type_0 = const()[name = tensor("op_50_pad_type_0"), val = tensor("custom")]; + tensor var_50_pad_0 = const()[name = tensor("op_50_pad_0"), val = tensor([1, 1])]; + tensor var_50_strides_0 = const()[name = tensor("op_50_strides_0"), val = tensor([2])]; + tensor var_50_dilations_0 = const()[name = tensor("op_50_dilations_0"), val = tensor([1])]; + tensor var_50_groups_0 = const()[name = tensor("op_50_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246976)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1819904)))]; + tensor var_50_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_50_dilations_0, groups = var_50_groups_0, pad = var_50_pad_0, pad_type = var_50_pad_type_0, strides = var_50_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_50_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_50_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_55_to_fp16 = const()[name = tensor("op_55_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1820992)))]; + tensor var_57_cast_fp16 = add(x = x_3_cast_fp16, y = var_55_to_fp16)[name = tensor("op_57_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_57_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3357056)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3358144)))]; + tensor var_88_to_fp16 = const()[name = tensor("op_88_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_88_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_123_weight_0_to_fp16 = const()[name = tensor("op_123_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3359232)))]; + tensor var_123_bias_0_to_fp16 = const()[name = tensor("op_123_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3883584)))]; + tensor var_123_cast_fp16 = conv(bias = var_123_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_123_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_123_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3884672)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_121_pad_type_0 = const()[name = tensor("op_121_pad_type_0"), val = tensor("valid")]; + tensor var_121_strides_0 = const()[name = tensor("op_121_strides_0"), val = tensor([1, 1])]; + tensor var_121_pad_0 = const()[name = tensor("op_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_121_dilations_0 = const()[name = tensor("op_121_dilations_0"), val = tensor([1, 1])]; + tensor var_121_groups_0 = const()[name = tensor("op_121_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4409024)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4933376)))]; + tensor var_121_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_121_dilations_0, groups = var_121_groups_0, pad = var_121_pad_0, pad_type = var_121_pad_type_0, strides = var_121_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_121_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_124_axis_0 = const()[name = tensor("op_124_axis_0"), val = tensor(1)]; + tensor var_124_cast_fp16_0, tensor var_124_cast_fp16_1, tensor var_124_cast_fp16_2, tensor var_124_cast_fp16_3, tensor var_124_cast_fp16_4, tensor var_124_cast_fp16_5, tensor var_124_cast_fp16_6, tensor var_124_cast_fp16_7 = split(axis = var_124_axis_0, split_sizes = tile_0, x = var_123_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor var_133_perm_0 = const()[name = tensor("op_133_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_134_axis_0 = const()[name = tensor("op_134_axis_0"), val = tensor(3)]; + tensor var_133_cast_fp16 = transpose(perm = var_133_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_6")]; + tensor var_134_cast_fp16_0, tensor var_134_cast_fp16_1, tensor var_134_cast_fp16_2, tensor var_134_cast_fp16_3, tensor var_134_cast_fp16_4, tensor var_134_cast_fp16_5, tensor var_134_cast_fp16_6, tensor var_134_cast_fp16_7 = split(axis = var_134_axis_0, split_sizes = tile_1, x = var_133_cast_fp16)[name = tensor("op_134_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_143_axis_0 = const()[name = tensor("op_143_axis_0"), val = tensor(1)]; + tensor var_143_cast_fp16_0, tensor var_143_cast_fp16_1, tensor var_143_cast_fp16_2, tensor var_143_cast_fp16_3, tensor var_143_cast_fp16_4, tensor var_143_cast_fp16_5, tensor var_143_cast_fp16_6, tensor var_143_cast_fp16_7 = split(axis = var_143_axis_0, split_sizes = tile_2, x = var_121_cast_fp16)[name = tensor("op_143_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_134_cast_fp16_0, var_124_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_134_cast_fp16_1, var_124_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_134_cast_fp16_2, var_124_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_134_cast_fp16_3, var_124_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_134_cast_fp16_4, var_124_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_134_cast_fp16_5, var_124_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_134_cast_fp16_6, var_124_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_134_cast_fp16_7, var_124_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor var_168_cast_fp16 = softmax(axis = var_72, x = aw_1_cast_fp16)[name = tensor("op_168_cast_fp16")]; + tensor var_169_cast_fp16 = softmax(axis = var_72, x = aw_3_cast_fp16)[name = tensor("op_169_cast_fp16")]; + tensor var_170_cast_fp16 = softmax(axis = var_72, x = aw_5_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor var_171_cast_fp16 = softmax(axis = var_72, x = aw_7_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor var_172_cast_fp16 = softmax(axis = var_72, x = aw_9_cast_fp16)[name = tensor("op_172_cast_fp16")]; + tensor var_173_cast_fp16 = softmax(axis = var_72, x = aw_11_cast_fp16)[name = tensor("op_173_cast_fp16")]; + tensor var_174_cast_fp16 = softmax(axis = var_72, x = aw_13_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_175_cast_fp16 = softmax(axis = var_72, x = aw_15_cast_fp16)[name = tensor("op_175_cast_fp16")]; + tensor var_177_equation_0 = const()[name = tensor("op_177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_177_cast_fp16 = einsum(equation = var_177_equation_0, values = (var_143_cast_fp16_0, var_168_cast_fp16))[name = tensor("op_177_cast_fp16")]; + tensor var_179_equation_0 = const()[name = tensor("op_179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_179_cast_fp16 = einsum(equation = var_179_equation_0, values = (var_143_cast_fp16_1, var_169_cast_fp16))[name = tensor("op_179_cast_fp16")]; + tensor var_181_equation_0 = const()[name = tensor("op_181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_181_cast_fp16 = einsum(equation = var_181_equation_0, values = (var_143_cast_fp16_2, var_170_cast_fp16))[name = tensor("op_181_cast_fp16")]; + tensor var_183_equation_0 = const()[name = tensor("op_183_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_183_cast_fp16 = einsum(equation = var_183_equation_0, values = (var_143_cast_fp16_3, var_171_cast_fp16))[name = tensor("op_183_cast_fp16")]; + tensor var_185_equation_0 = const()[name = tensor("op_185_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_185_cast_fp16 = einsum(equation = var_185_equation_0, values = (var_143_cast_fp16_4, var_172_cast_fp16))[name = tensor("op_185_cast_fp16")]; + tensor var_187_equation_0 = const()[name = tensor("op_187_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_187_cast_fp16 = einsum(equation = var_187_equation_0, values = (var_143_cast_fp16_5, var_173_cast_fp16))[name = tensor("op_187_cast_fp16")]; + tensor var_189_equation_0 = const()[name = tensor("op_189_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_189_cast_fp16 = einsum(equation = var_189_equation_0, values = (var_143_cast_fp16_6, var_174_cast_fp16))[name = tensor("op_189_cast_fp16")]; + tensor var_191_equation_0 = const()[name = tensor("op_191_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_191_cast_fp16 = einsum(equation = var_191_equation_0, values = (var_143_cast_fp16_7, var_175_cast_fp16))[name = tensor("op_191_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_72, interleave = input_5_interleave_0, values = (var_177_cast_fp16, var_179_cast_fp16, var_181_cast_fp16, var_183_cast_fp16, var_185_cast_fp16, var_187_cast_fp16, var_189_cast_fp16, var_191_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_200_pad_type_0 = const()[name = tensor("op_200_pad_type_0"), val = tensor("valid")]; + tensor var_200_strides_0 = const()[name = tensor("op_200_strides_0"), val = tensor([1, 1])]; + tensor var_200_pad_0 = const()[name = tensor("op_200_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_200_dilations_0 = const()[name = tensor("op_200_dilations_0"), val = tensor([1, 1])]; + tensor var_200_groups_0 = const()[name = tensor("op_200_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4934464)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5458816)))]; + tensor var_200_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_200_dilations_0, groups = var_200_groups_0, pad = var_200_pad_0, pad_type = var_200_pad_type_0, strides = var_200_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_200_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5459904)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5460992)))]; + tensor var_210_to_fp16 = const()[name = tensor("op_210_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_210_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5462080)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7559296)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_236_pad_type_0 = const()[name = tensor("op_236_pad_type_0"), val = tensor("valid")]; + tensor var_236_strides_0 = const()[name = tensor("op_236_strides_0"), val = tensor([1, 1])]; + tensor var_236_pad_0 = const()[name = tensor("op_236_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_236_dilations_0 = const()[name = tensor("op_236_dilations_0"), val = tensor([1, 1])]; + tensor var_236_groups_0 = const()[name = tensor("op_236_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7563456)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9660672)))]; + tensor var_236_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_236_dilations_0, groups = var_236_groups_0, pad = var_236_pad_0, pad_type = var_236_pad_type_0, strides = var_236_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_236_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_245 = const()[name = tensor("op_245"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9661760)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9662848)))]; + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_261_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_296_weight_0_to_fp16 = const()[name = tensor("op_296_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9663936)))]; + tensor var_296_bias_0_to_fp16 = const()[name = tensor("op_296_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10188288)))]; + tensor var_296_cast_fp16 = conv(bias = var_296_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_296_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10189376)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_294_pad_type_0 = const()[name = tensor("op_294_pad_type_0"), val = tensor("valid")]; + tensor var_294_strides_0 = const()[name = tensor("op_294_strides_0"), val = tensor([1, 1])]; + tensor var_294_pad_0 = const()[name = tensor("op_294_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_294_dilations_0 = const()[name = tensor("op_294_dilations_0"), val = tensor([1, 1])]; + tensor var_294_groups_0 = const()[name = tensor("op_294_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10713728)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11238080)))]; + tensor var_294_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_294_dilations_0, groups = var_294_groups_0, pad = var_294_pad_0, pad_type = var_294_pad_type_0, strides = var_294_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_297_axis_0 = const()[name = tensor("op_297_axis_0"), val = tensor(1)]; + tensor var_297_cast_fp16_0, tensor var_297_cast_fp16_1, tensor var_297_cast_fp16_2, tensor var_297_cast_fp16_3, tensor var_297_cast_fp16_4, tensor var_297_cast_fp16_5, tensor var_297_cast_fp16_6, tensor var_297_cast_fp16_7 = split(axis = var_297_axis_0, split_sizes = tile_3, x = var_296_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_306_perm_0 = const()[name = tensor("op_306_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_307_axis_0 = const()[name = tensor("op_307_axis_0"), val = tensor(3)]; + tensor var_306_cast_fp16 = transpose(perm = var_306_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_5")]; + tensor var_307_cast_fp16_0, tensor var_307_cast_fp16_1, tensor var_307_cast_fp16_2, tensor var_307_cast_fp16_3, tensor var_307_cast_fp16_4, tensor var_307_cast_fp16_5, tensor var_307_cast_fp16_6, tensor var_307_cast_fp16_7 = split(axis = var_307_axis_0, split_sizes = tile_4, x = var_306_cast_fp16)[name = tensor("op_307_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_316_axis_0 = const()[name = tensor("op_316_axis_0"), val = tensor(1)]; + tensor var_316_cast_fp16_0, tensor var_316_cast_fp16_1, tensor var_316_cast_fp16_2, tensor var_316_cast_fp16_3, tensor var_316_cast_fp16_4, tensor var_316_cast_fp16_5, tensor var_316_cast_fp16_6, tensor var_316_cast_fp16_7 = split(axis = var_316_axis_0, split_sizes = tile_5, x = var_294_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_307_cast_fp16_0, var_297_cast_fp16_0))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_307_cast_fp16_1, var_297_cast_fp16_1))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_307_cast_fp16_2, var_297_cast_fp16_2))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_307_cast_fp16_3, var_297_cast_fp16_3))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_307_cast_fp16_4, var_297_cast_fp16_4))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_307_cast_fp16_5, var_297_cast_fp16_5))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_307_cast_fp16_6, var_297_cast_fp16_6))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_307_cast_fp16_7, var_297_cast_fp16_7))[name = tensor("aw_31_cast_fp16")]; + tensor var_341_cast_fp16 = softmax(axis = var_245, x = aw_17_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor var_342_cast_fp16 = softmax(axis = var_245, x = aw_19_cast_fp16)[name = tensor("op_342_cast_fp16")]; + tensor var_343_cast_fp16 = softmax(axis = var_245, x = aw_21_cast_fp16)[name = tensor("op_343_cast_fp16")]; + tensor var_344_cast_fp16 = softmax(axis = var_245, x = aw_23_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = softmax(axis = var_245, x = aw_25_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_346_cast_fp16 = softmax(axis = var_245, x = aw_27_cast_fp16)[name = tensor("op_346_cast_fp16")]; + tensor var_347_cast_fp16 = softmax(axis = var_245, x = aw_29_cast_fp16)[name = tensor("op_347_cast_fp16")]; + tensor var_348_cast_fp16 = softmax(axis = var_245, x = aw_31_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_350_equation_0 = const()[name = tensor("op_350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_350_cast_fp16 = einsum(equation = var_350_equation_0, values = (var_316_cast_fp16_0, var_341_cast_fp16))[name = tensor("op_350_cast_fp16")]; + tensor var_352_equation_0 = const()[name = tensor("op_352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_352_cast_fp16 = einsum(equation = var_352_equation_0, values = (var_316_cast_fp16_1, var_342_cast_fp16))[name = tensor("op_352_cast_fp16")]; + tensor var_354_equation_0 = const()[name = tensor("op_354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_354_cast_fp16 = einsum(equation = var_354_equation_0, values = (var_316_cast_fp16_2, var_343_cast_fp16))[name = tensor("op_354_cast_fp16")]; + tensor var_356_equation_0 = const()[name = tensor("op_356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_356_cast_fp16 = einsum(equation = var_356_equation_0, values = (var_316_cast_fp16_3, var_344_cast_fp16))[name = tensor("op_356_cast_fp16")]; + tensor var_358_equation_0 = const()[name = tensor("op_358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_358_cast_fp16 = einsum(equation = var_358_equation_0, values = (var_316_cast_fp16_4, var_345_cast_fp16))[name = tensor("op_358_cast_fp16")]; + tensor var_360_equation_0 = const()[name = tensor("op_360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_360_cast_fp16 = einsum(equation = var_360_equation_0, values = (var_316_cast_fp16_5, var_346_cast_fp16))[name = tensor("op_360_cast_fp16")]; + tensor var_362_equation_0 = const()[name = tensor("op_362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_362_cast_fp16 = einsum(equation = var_362_equation_0, values = (var_316_cast_fp16_6, var_347_cast_fp16))[name = tensor("op_362_cast_fp16")]; + tensor var_364_equation_0 = const()[name = tensor("op_364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_364_cast_fp16 = einsum(equation = var_364_equation_0, values = (var_316_cast_fp16_7, var_348_cast_fp16))[name = tensor("op_364_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_245, interleave = input_15_interleave_0, values = (var_350_cast_fp16, var_352_cast_fp16, var_354_cast_fp16, var_356_cast_fp16, var_358_cast_fp16, var_360_cast_fp16, var_362_cast_fp16, var_364_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_373_pad_type_0 = const()[name = tensor("op_373_pad_type_0"), val = tensor("valid")]; + tensor var_373_strides_0 = const()[name = tensor("op_373_strides_0"), val = tensor([1, 1])]; + tensor var_373_pad_0 = const()[name = tensor("op_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_373_dilations_0 = const()[name = tensor("op_373_dilations_0"), val = tensor([1, 1])]; + tensor var_373_groups_0 = const()[name = tensor("op_373_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11239168)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11763520)))]; + tensor var_373_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_373_dilations_0, groups = var_373_groups_0, pad = var_373_pad_0, pad_type = var_373_pad_type_0, strides = var_373_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_373_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11764608)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11765696)))]; + tensor var_383_to_fp16 = const()[name = tensor("op_383_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_383_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11766784)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13864000)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_409_pad_type_0 = const()[name = tensor("op_409_pad_type_0"), val = tensor("valid")]; + tensor var_409_strides_0 = const()[name = tensor("op_409_strides_0"), val = tensor([1, 1])]; + tensor var_409_pad_0 = const()[name = tensor("op_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_409_dilations_0 = const()[name = tensor("op_409_dilations_0"), val = tensor([1, 1])]; + tensor var_409_groups_0 = const()[name = tensor("op_409_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13868160)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15965376)))]; + tensor var_409_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_409_dilations_0, groups = var_409_groups_0, pad = var_409_pad_0, pad_type = var_409_pad_type_0, strides = var_409_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_409_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15966464)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15967552)))]; + tensor var_434_to_fp16 = const()[name = tensor("op_434_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_434_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_469_weight_0_to_fp16 = const()[name = tensor("op_469_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15968640)))]; + tensor var_469_bias_0_to_fp16 = const()[name = tensor("op_469_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16492992)))]; + tensor var_469_cast_fp16 = conv(bias = var_469_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_469_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16494080)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_467_pad_type_0 = const()[name = tensor("op_467_pad_type_0"), val = tensor("valid")]; + tensor var_467_strides_0 = const()[name = tensor("op_467_strides_0"), val = tensor([1, 1])]; + tensor var_467_pad_0 = const()[name = tensor("op_467_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_467_dilations_0 = const()[name = tensor("op_467_dilations_0"), val = tensor([1, 1])]; + tensor var_467_groups_0 = const()[name = tensor("op_467_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17018432)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17542784)))]; + tensor var_467_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_467_dilations_0, groups = var_467_groups_0, pad = var_467_pad_0, pad_type = var_467_pad_type_0, strides = var_467_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_467_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_470_axis_0 = const()[name = tensor("op_470_axis_0"), val = tensor(1)]; + tensor var_470_cast_fp16_0, tensor var_470_cast_fp16_1, tensor var_470_cast_fp16_2, tensor var_470_cast_fp16_3, tensor var_470_cast_fp16_4, tensor var_470_cast_fp16_5, tensor var_470_cast_fp16_6, tensor var_470_cast_fp16_7 = split(axis = var_470_axis_0, split_sizes = tile_6, x = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_479_perm_0 = const()[name = tensor("op_479_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_480_axis_0 = const()[name = tensor("op_480_axis_0"), val = tensor(3)]; + tensor var_479_cast_fp16 = transpose(perm = var_479_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_4")]; + tensor var_480_cast_fp16_0, tensor var_480_cast_fp16_1, tensor var_480_cast_fp16_2, tensor var_480_cast_fp16_3, tensor var_480_cast_fp16_4, tensor var_480_cast_fp16_5, tensor var_480_cast_fp16_6, tensor var_480_cast_fp16_7 = split(axis = var_480_axis_0, split_sizes = tile_7, x = var_479_cast_fp16)[name = tensor("op_480_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_489_axis_0 = const()[name = tensor("op_489_axis_0"), val = tensor(1)]; + tensor var_489_cast_fp16_0, tensor var_489_cast_fp16_1, tensor var_489_cast_fp16_2, tensor var_489_cast_fp16_3, tensor var_489_cast_fp16_4, tensor var_489_cast_fp16_5, tensor var_489_cast_fp16_6, tensor var_489_cast_fp16_7 = split(axis = var_489_axis_0, split_sizes = tile_8, x = var_467_cast_fp16)[name = tensor("op_489_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_480_cast_fp16_0, var_470_cast_fp16_0))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_480_cast_fp16_1, var_470_cast_fp16_1))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_480_cast_fp16_2, var_470_cast_fp16_2))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_480_cast_fp16_3, var_470_cast_fp16_3))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_480_cast_fp16_4, var_470_cast_fp16_4))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_480_cast_fp16_5, var_470_cast_fp16_5))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_480_cast_fp16_6, var_470_cast_fp16_6))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_480_cast_fp16_7, var_470_cast_fp16_7))[name = tensor("aw_47_cast_fp16")]; + tensor var_514_cast_fp16 = softmax(axis = var_418, x = aw_33_cast_fp16)[name = tensor("op_514_cast_fp16")]; + tensor var_515_cast_fp16 = softmax(axis = var_418, x = aw_35_cast_fp16)[name = tensor("op_515_cast_fp16")]; + tensor var_516_cast_fp16 = softmax(axis = var_418, x = aw_37_cast_fp16)[name = tensor("op_516_cast_fp16")]; + tensor var_517_cast_fp16 = softmax(axis = var_418, x = aw_39_cast_fp16)[name = tensor("op_517_cast_fp16")]; + tensor var_518_cast_fp16 = softmax(axis = var_418, x = aw_41_cast_fp16)[name = tensor("op_518_cast_fp16")]; + tensor var_519_cast_fp16 = softmax(axis = var_418, x = aw_43_cast_fp16)[name = tensor("op_519_cast_fp16")]; + tensor var_520_cast_fp16 = softmax(axis = var_418, x = aw_45_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor var_521_cast_fp16 = softmax(axis = var_418, x = aw_47_cast_fp16)[name = tensor("op_521_cast_fp16")]; + tensor var_523_equation_0 = const()[name = tensor("op_523_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_523_cast_fp16 = einsum(equation = var_523_equation_0, values = (var_489_cast_fp16_0, var_514_cast_fp16))[name = tensor("op_523_cast_fp16")]; + tensor var_525_equation_0 = const()[name = tensor("op_525_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_525_cast_fp16 = einsum(equation = var_525_equation_0, values = (var_489_cast_fp16_1, var_515_cast_fp16))[name = tensor("op_525_cast_fp16")]; + tensor var_527_equation_0 = const()[name = tensor("op_527_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_527_cast_fp16 = einsum(equation = var_527_equation_0, values = (var_489_cast_fp16_2, var_516_cast_fp16))[name = tensor("op_527_cast_fp16")]; + tensor var_529_equation_0 = const()[name = tensor("op_529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_529_cast_fp16 = einsum(equation = var_529_equation_0, values = (var_489_cast_fp16_3, var_517_cast_fp16))[name = tensor("op_529_cast_fp16")]; + tensor var_531_equation_0 = const()[name = tensor("op_531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_531_cast_fp16 = einsum(equation = var_531_equation_0, values = (var_489_cast_fp16_4, var_518_cast_fp16))[name = tensor("op_531_cast_fp16")]; + tensor var_533_equation_0 = const()[name = tensor("op_533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_533_cast_fp16 = einsum(equation = var_533_equation_0, values = (var_489_cast_fp16_5, var_519_cast_fp16))[name = tensor("op_533_cast_fp16")]; + tensor var_535_equation_0 = const()[name = tensor("op_535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_535_cast_fp16 = einsum(equation = var_535_equation_0, values = (var_489_cast_fp16_6, var_520_cast_fp16))[name = tensor("op_535_cast_fp16")]; + tensor var_537_equation_0 = const()[name = tensor("op_537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_537_cast_fp16 = einsum(equation = var_537_equation_0, values = (var_489_cast_fp16_7, var_521_cast_fp16))[name = tensor("op_537_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_418, interleave = input_25_interleave_0, values = (var_523_cast_fp16, var_525_cast_fp16, var_527_cast_fp16, var_529_cast_fp16, var_531_cast_fp16, var_533_cast_fp16, var_535_cast_fp16, var_537_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_546_pad_type_0 = const()[name = tensor("op_546_pad_type_0"), val = tensor("valid")]; + tensor var_546_strides_0 = const()[name = tensor("op_546_strides_0"), val = tensor([1, 1])]; + tensor var_546_pad_0 = const()[name = tensor("op_546_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_546_dilations_0 = const()[name = tensor("op_546_dilations_0"), val = tensor([1, 1])]; + tensor var_546_groups_0 = const()[name = tensor("op_546_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17543872)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18068224)))]; + tensor var_546_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_546_dilations_0, groups = var_546_groups_0, pad = var_546_pad_0, pad_type = var_546_pad_type_0, strides = var_546_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_546_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_546_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18069312)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18070400)))]; + tensor var_556_to_fp16 = const()[name = tensor("op_556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_556_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18071488)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20168704)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_582_pad_type_0 = const()[name = tensor("op_582_pad_type_0"), val = tensor("valid")]; + tensor var_582_strides_0 = const()[name = tensor("op_582_strides_0"), val = tensor([1, 1])]; + tensor var_582_pad_0 = const()[name = tensor("op_582_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_582_dilations_0 = const()[name = tensor("op_582_dilations_0"), val = tensor([1, 1])]; + tensor var_582_groups_0 = const()[name = tensor("op_582_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20172864)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22270080)))]; + tensor var_582_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_582_dilations_0, groups = var_582_groups_0, pad = var_582_pad_0, pad_type = var_582_pad_type_0, strides = var_582_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_582_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22271168)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22272256)))]; + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_607_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_642_weight_0_to_fp16 = const()[name = tensor("op_642_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22273344)))]; + tensor var_642_bias_0_to_fp16 = const()[name = tensor("op_642_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22797696)))]; + tensor var_642_cast_fp16 = conv(bias = var_642_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_642_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22798784)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_640_pad_type_0 = const()[name = tensor("op_640_pad_type_0"), val = tensor("valid")]; + tensor var_640_strides_0 = const()[name = tensor("op_640_strides_0"), val = tensor([1, 1])]; + tensor var_640_pad_0 = const()[name = tensor("op_640_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_640_dilations_0 = const()[name = tensor("op_640_dilations_0"), val = tensor([1, 1])]; + tensor var_640_groups_0 = const()[name = tensor("op_640_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23323136)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23847488)))]; + tensor var_640_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_640_dilations_0, groups = var_640_groups_0, pad = var_640_pad_0, pad_type = var_640_pad_type_0, strides = var_640_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_640_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_643_axis_0 = const()[name = tensor("op_643_axis_0"), val = tensor(1)]; + tensor var_643_cast_fp16_0, tensor var_643_cast_fp16_1, tensor var_643_cast_fp16_2, tensor var_643_cast_fp16_3, tensor var_643_cast_fp16_4, tensor var_643_cast_fp16_5, tensor var_643_cast_fp16_6, tensor var_643_cast_fp16_7 = split(axis = var_643_axis_0, split_sizes = tile_9, x = var_642_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(3)]; + tensor var_652_cast_fp16 = transpose(perm = var_652_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_3")]; + tensor var_653_cast_fp16_0, tensor var_653_cast_fp16_1, tensor var_653_cast_fp16_2, tensor var_653_cast_fp16_3, tensor var_653_cast_fp16_4, tensor var_653_cast_fp16_5, tensor var_653_cast_fp16_6, tensor var_653_cast_fp16_7 = split(axis = var_653_axis_0, split_sizes = tile_10, x = var_652_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_662_axis_0 = const()[name = tensor("op_662_axis_0"), val = tensor(1)]; + tensor var_662_cast_fp16_0, tensor var_662_cast_fp16_1, tensor var_662_cast_fp16_2, tensor var_662_cast_fp16_3, tensor var_662_cast_fp16_4, tensor var_662_cast_fp16_5, tensor var_662_cast_fp16_6, tensor var_662_cast_fp16_7 = split(axis = var_662_axis_0, split_sizes = tile_11, x = var_640_cast_fp16)[name = tensor("op_662_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_653_cast_fp16_0, var_643_cast_fp16_0))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_653_cast_fp16_1, var_643_cast_fp16_1))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_653_cast_fp16_2, var_643_cast_fp16_2))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_653_cast_fp16_3, var_643_cast_fp16_3))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_653_cast_fp16_4, var_643_cast_fp16_4))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_653_cast_fp16_5, var_643_cast_fp16_5))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_653_cast_fp16_6, var_643_cast_fp16_6))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_653_cast_fp16_7, var_643_cast_fp16_7))[name = tensor("aw_63_cast_fp16")]; + tensor var_687_cast_fp16 = softmax(axis = var_591, x = aw_49_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor var_688_cast_fp16 = softmax(axis = var_591, x = aw_51_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_689_cast_fp16 = softmax(axis = var_591, x = aw_53_cast_fp16)[name = tensor("op_689_cast_fp16")]; + tensor var_690_cast_fp16 = softmax(axis = var_591, x = aw_55_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor var_691_cast_fp16 = softmax(axis = var_591, x = aw_57_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor var_692_cast_fp16 = softmax(axis = var_591, x = aw_59_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_693_cast_fp16 = softmax(axis = var_591, x = aw_61_cast_fp16)[name = tensor("op_693_cast_fp16")]; + tensor var_694_cast_fp16 = softmax(axis = var_591, x = aw_63_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_696_equation_0 = const()[name = tensor("op_696_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_696_cast_fp16 = einsum(equation = var_696_equation_0, values = (var_662_cast_fp16_0, var_687_cast_fp16))[name = tensor("op_696_cast_fp16")]; + tensor var_698_equation_0 = const()[name = tensor("op_698_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_698_cast_fp16 = einsum(equation = var_698_equation_0, values = (var_662_cast_fp16_1, var_688_cast_fp16))[name = tensor("op_698_cast_fp16")]; + tensor var_700_equation_0 = const()[name = tensor("op_700_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_700_cast_fp16 = einsum(equation = var_700_equation_0, values = (var_662_cast_fp16_2, var_689_cast_fp16))[name = tensor("op_700_cast_fp16")]; + tensor var_702_equation_0 = const()[name = tensor("op_702_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_702_cast_fp16 = einsum(equation = var_702_equation_0, values = (var_662_cast_fp16_3, var_690_cast_fp16))[name = tensor("op_702_cast_fp16")]; + tensor var_704_equation_0 = const()[name = tensor("op_704_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_704_cast_fp16 = einsum(equation = var_704_equation_0, values = (var_662_cast_fp16_4, var_691_cast_fp16))[name = tensor("op_704_cast_fp16")]; + tensor var_706_equation_0 = const()[name = tensor("op_706_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_706_cast_fp16 = einsum(equation = var_706_equation_0, values = (var_662_cast_fp16_5, var_692_cast_fp16))[name = tensor("op_706_cast_fp16")]; + tensor var_708_equation_0 = const()[name = tensor("op_708_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_708_cast_fp16 = einsum(equation = var_708_equation_0, values = (var_662_cast_fp16_6, var_693_cast_fp16))[name = tensor("op_708_cast_fp16")]; + tensor var_710_equation_0 = const()[name = tensor("op_710_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_710_cast_fp16 = einsum(equation = var_710_equation_0, values = (var_662_cast_fp16_7, var_694_cast_fp16))[name = tensor("op_710_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_591, interleave = input_35_interleave_0, values = (var_696_cast_fp16, var_698_cast_fp16, var_700_cast_fp16, var_702_cast_fp16, var_704_cast_fp16, var_706_cast_fp16, var_708_cast_fp16, var_710_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_719_pad_type_0 = const()[name = tensor("op_719_pad_type_0"), val = tensor("valid")]; + tensor var_719_strides_0 = const()[name = tensor("op_719_strides_0"), val = tensor([1, 1])]; + tensor var_719_pad_0 = const()[name = tensor("op_719_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_719_dilations_0 = const()[name = tensor("op_719_dilations_0"), val = tensor([1, 1])]; + tensor var_719_groups_0 = const()[name = tensor("op_719_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23848576)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24372928)))]; + tensor var_719_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_719_dilations_0, groups = var_719_groups_0, pad = var_719_pad_0, pad_type = var_719_pad_type_0, strides = var_719_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_719_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_719_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24375104)))]; + tensor var_729_to_fp16 = const()[name = tensor("op_729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_729_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24376192)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_755_pad_type_0 = const()[name = tensor("op_755_pad_type_0"), val = tensor("valid")]; + tensor var_755_strides_0 = const()[name = tensor("op_755_strides_0"), val = tensor([1, 1])]; + tensor var_755_pad_0 = const()[name = tensor("op_755_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_755_dilations_0 = const()[name = tensor("op_755_dilations_0"), val = tensor([1, 1])]; + tensor var_755_groups_0 = const()[name = tensor("op_755_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26477568)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28574784)))]; + tensor var_755_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_755_dilations_0, groups = var_755_groups_0, pad = var_755_pad_0, pad_type = var_755_pad_type_0, strides = var_755_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_755_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_755_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28575872)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28576960)))]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_780_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_815_weight_0_to_fp16 = const()[name = tensor("op_815_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28578048)))]; + tensor var_815_bias_0_to_fp16 = const()[name = tensor("op_815_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29102400)))]; + tensor var_815_cast_fp16 = conv(bias = var_815_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_815_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29103488)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_813_pad_type_0 = const()[name = tensor("op_813_pad_type_0"), val = tensor("valid")]; + tensor var_813_strides_0 = const()[name = tensor("op_813_strides_0"), val = tensor([1, 1])]; + tensor var_813_pad_0 = const()[name = tensor("op_813_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_813_dilations_0 = const()[name = tensor("op_813_dilations_0"), val = tensor([1, 1])]; + tensor var_813_groups_0 = const()[name = tensor("op_813_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29627840)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30152192)))]; + tensor var_813_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_813_dilations_0, groups = var_813_groups_0, pad = var_813_pad_0, pad_type = var_813_pad_type_0, strides = var_813_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_816_axis_0 = const()[name = tensor("op_816_axis_0"), val = tensor(1)]; + tensor var_816_cast_fp16_0, tensor var_816_cast_fp16_1, tensor var_816_cast_fp16_2, tensor var_816_cast_fp16_3, tensor var_816_cast_fp16_4, tensor var_816_cast_fp16_5, tensor var_816_cast_fp16_6, tensor var_816_cast_fp16_7 = split(axis = var_816_axis_0, split_sizes = tile_12, x = var_815_cast_fp16)[name = tensor("op_816_cast_fp16")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_826_axis_0 = const()[name = tensor("op_826_axis_0"), val = tensor(3)]; + tensor var_825_cast_fp16 = transpose(perm = var_825_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_2")]; + tensor var_826_cast_fp16_0, tensor var_826_cast_fp16_1, tensor var_826_cast_fp16_2, tensor var_826_cast_fp16_3, tensor var_826_cast_fp16_4, tensor var_826_cast_fp16_5, tensor var_826_cast_fp16_6, tensor var_826_cast_fp16_7 = split(axis = var_826_axis_0, split_sizes = tile_13, x = var_825_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_835_axis_0 = const()[name = tensor("op_835_axis_0"), val = tensor(1)]; + tensor var_835_cast_fp16_0, tensor var_835_cast_fp16_1, tensor var_835_cast_fp16_2, tensor var_835_cast_fp16_3, tensor var_835_cast_fp16_4, tensor var_835_cast_fp16_5, tensor var_835_cast_fp16_6, tensor var_835_cast_fp16_7 = split(axis = var_835_axis_0, split_sizes = tile_14, x = var_813_cast_fp16)[name = tensor("op_835_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_826_cast_fp16_0, var_816_cast_fp16_0))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_826_cast_fp16_1, var_816_cast_fp16_1))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_826_cast_fp16_2, var_816_cast_fp16_2))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_826_cast_fp16_3, var_816_cast_fp16_3))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_826_cast_fp16_4, var_816_cast_fp16_4))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_826_cast_fp16_5, var_816_cast_fp16_5))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_826_cast_fp16_6, var_816_cast_fp16_6))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_826_cast_fp16_7, var_816_cast_fp16_7))[name = tensor("aw_79_cast_fp16")]; + tensor var_860_cast_fp16 = softmax(axis = var_764, x = aw_65_cast_fp16)[name = tensor("op_860_cast_fp16")]; + tensor var_861_cast_fp16 = softmax(axis = var_764, x = aw_67_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_862_cast_fp16 = softmax(axis = var_764, x = aw_69_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = softmax(axis = var_764, x = aw_71_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = softmax(axis = var_764, x = aw_73_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor var_865_cast_fp16 = softmax(axis = var_764, x = aw_75_cast_fp16)[name = tensor("op_865_cast_fp16")]; + tensor var_866_cast_fp16 = softmax(axis = var_764, x = aw_77_cast_fp16)[name = tensor("op_866_cast_fp16")]; + tensor var_867_cast_fp16 = softmax(axis = var_764, x = aw_79_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_869_equation_0 = const()[name = tensor("op_869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_869_cast_fp16 = einsum(equation = var_869_equation_0, values = (var_835_cast_fp16_0, var_860_cast_fp16))[name = tensor("op_869_cast_fp16")]; + tensor var_871_equation_0 = const()[name = tensor("op_871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_871_cast_fp16 = einsum(equation = var_871_equation_0, values = (var_835_cast_fp16_1, var_861_cast_fp16))[name = tensor("op_871_cast_fp16")]; + tensor var_873_equation_0 = const()[name = tensor("op_873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_873_cast_fp16 = einsum(equation = var_873_equation_0, values = (var_835_cast_fp16_2, var_862_cast_fp16))[name = tensor("op_873_cast_fp16")]; + tensor var_875_equation_0 = const()[name = tensor("op_875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_875_cast_fp16 = einsum(equation = var_875_equation_0, values = (var_835_cast_fp16_3, var_863_cast_fp16))[name = tensor("op_875_cast_fp16")]; + tensor var_877_equation_0 = const()[name = tensor("op_877_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_877_cast_fp16 = einsum(equation = var_877_equation_0, values = (var_835_cast_fp16_4, var_864_cast_fp16))[name = tensor("op_877_cast_fp16")]; + tensor var_879_equation_0 = const()[name = tensor("op_879_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_879_cast_fp16 = einsum(equation = var_879_equation_0, values = (var_835_cast_fp16_5, var_865_cast_fp16))[name = tensor("op_879_cast_fp16")]; + tensor var_881_equation_0 = const()[name = tensor("op_881_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_881_cast_fp16 = einsum(equation = var_881_equation_0, values = (var_835_cast_fp16_6, var_866_cast_fp16))[name = tensor("op_881_cast_fp16")]; + tensor var_883_equation_0 = const()[name = tensor("op_883_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_883_cast_fp16 = einsum(equation = var_883_equation_0, values = (var_835_cast_fp16_7, var_867_cast_fp16))[name = tensor("op_883_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_764, interleave = input_45_interleave_0, values = (var_869_cast_fp16, var_871_cast_fp16, var_873_cast_fp16, var_875_cast_fp16, var_877_cast_fp16, var_879_cast_fp16, var_881_cast_fp16, var_883_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_892_pad_type_0 = const()[name = tensor("op_892_pad_type_0"), val = tensor("valid")]; + tensor var_892_strides_0 = const()[name = tensor("op_892_strides_0"), val = tensor([1, 1])]; + tensor var_892_pad_0 = const()[name = tensor("op_892_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_892_dilations_0 = const()[name = tensor("op_892_dilations_0"), val = tensor([1, 1])]; + tensor var_892_groups_0 = const()[name = tensor("op_892_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30153280)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30677632)))]; + tensor var_892_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_892_dilations_0, groups = var_892_groups_0, pad = var_892_pad_0, pad_type = var_892_pad_type_0, strides = var_892_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_892_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30678720)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30679808)))]; + tensor var_902_to_fp16 = const()[name = tensor("op_902_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_902_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30680896)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32778112)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_928_pad_type_0 = const()[name = tensor("op_928_pad_type_0"), val = tensor("valid")]; + tensor var_928_strides_0 = const()[name = tensor("op_928_strides_0"), val = tensor([1, 1])]; + tensor var_928_pad_0 = const()[name = tensor("op_928_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_928_dilations_0 = const()[name = tensor("op_928_dilations_0"), val = tensor([1, 1])]; + tensor var_928_groups_0 = const()[name = tensor("op_928_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32782272)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34879488)))]; + tensor var_928_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_928_dilations_0, groups = var_928_groups_0, pad = var_928_pad_0, pad_type = var_928_pad_type_0, strides = var_928_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_928_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_937 = const()[name = tensor("op_937"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34880576)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34881664)))]; + tensor var_953_to_fp16 = const()[name = tensor("op_953_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_953_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_988_weight_0_to_fp16 = const()[name = tensor("op_988_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34882752)))]; + tensor var_988_bias_0_to_fp16 = const()[name = tensor("op_988_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35407104)))]; + tensor var_988_cast_fp16 = conv(bias = var_988_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_988_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_988_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35408192)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_986_pad_type_0 = const()[name = tensor("op_986_pad_type_0"), val = tensor("valid")]; + tensor var_986_strides_0 = const()[name = tensor("op_986_strides_0"), val = tensor([1, 1])]; + tensor var_986_pad_0 = const()[name = tensor("op_986_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_986_dilations_0 = const()[name = tensor("op_986_dilations_0"), val = tensor([1, 1])]; + tensor var_986_groups_0 = const()[name = tensor("op_986_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35932544)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36456896)))]; + tensor var_986_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_986_dilations_0, groups = var_986_groups_0, pad = var_986_pad_0, pad_type = var_986_pad_type_0, strides = var_986_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_986_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_989_axis_0 = const()[name = tensor("op_989_axis_0"), val = tensor(1)]; + tensor var_989_cast_fp16_0, tensor var_989_cast_fp16_1, tensor var_989_cast_fp16_2, tensor var_989_cast_fp16_3, tensor var_989_cast_fp16_4, tensor var_989_cast_fp16_5, tensor var_989_cast_fp16_6, tensor var_989_cast_fp16_7 = split(axis = var_989_axis_0, split_sizes = tile_15, x = var_988_cast_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_998_perm_0 = const()[name = tensor("op_998_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(3)]; + tensor var_998_cast_fp16 = transpose(perm = var_998_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_999_cast_fp16_0, tensor var_999_cast_fp16_1, tensor var_999_cast_fp16_2, tensor var_999_cast_fp16_3, tensor var_999_cast_fp16_4, tensor var_999_cast_fp16_5, tensor var_999_cast_fp16_6, tensor var_999_cast_fp16_7 = split(axis = var_999_axis_0, split_sizes = tile_16, x = var_998_cast_fp16)[name = tensor("op_999_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1008_axis_0 = const()[name = tensor("op_1008_axis_0"), val = tensor(1)]; + tensor var_1008_cast_fp16_0, tensor var_1008_cast_fp16_1, tensor var_1008_cast_fp16_2, tensor var_1008_cast_fp16_3, tensor var_1008_cast_fp16_4, tensor var_1008_cast_fp16_5, tensor var_1008_cast_fp16_6, tensor var_1008_cast_fp16_7 = split(axis = var_1008_axis_0, split_sizes = tile_17, x = var_986_cast_fp16)[name = tensor("op_1008_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_999_cast_fp16_0, var_989_cast_fp16_0))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_999_cast_fp16_1, var_989_cast_fp16_1))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_999_cast_fp16_2, var_989_cast_fp16_2))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_999_cast_fp16_3, var_989_cast_fp16_3))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_999_cast_fp16_4, var_989_cast_fp16_4))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_999_cast_fp16_5, var_989_cast_fp16_5))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_999_cast_fp16_6, var_989_cast_fp16_6))[name = tensor("aw_93_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_999_cast_fp16_7, var_989_cast_fp16_7))[name = tensor("aw_cast_fp16")]; + tensor var_1033_cast_fp16 = softmax(axis = var_937, x = aw_81_cast_fp16)[name = tensor("op_1033_cast_fp16")]; + tensor var_1034_cast_fp16 = softmax(axis = var_937, x = aw_83_cast_fp16)[name = tensor("op_1034_cast_fp16")]; + tensor var_1035_cast_fp16 = softmax(axis = var_937, x = aw_85_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor var_1036_cast_fp16 = softmax(axis = var_937, x = aw_87_cast_fp16)[name = tensor("op_1036_cast_fp16")]; + tensor var_1037_cast_fp16 = softmax(axis = var_937, x = aw_89_cast_fp16)[name = tensor("op_1037_cast_fp16")]; + tensor var_1038_cast_fp16 = softmax(axis = var_937, x = aw_91_cast_fp16)[name = tensor("op_1038_cast_fp16")]; + tensor var_1039_cast_fp16 = softmax(axis = var_937, x = aw_93_cast_fp16)[name = tensor("op_1039_cast_fp16")]; + tensor var_1040_cast_fp16 = softmax(axis = var_937, x = aw_cast_fp16)[name = tensor("op_1040_cast_fp16")]; + tensor var_1042_equation_0 = const()[name = tensor("op_1042_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1042_cast_fp16 = einsum(equation = var_1042_equation_0, values = (var_1008_cast_fp16_0, var_1033_cast_fp16))[name = tensor("op_1042_cast_fp16")]; + tensor var_1044_equation_0 = const()[name = tensor("op_1044_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1044_cast_fp16 = einsum(equation = var_1044_equation_0, values = (var_1008_cast_fp16_1, var_1034_cast_fp16))[name = tensor("op_1044_cast_fp16")]; + tensor var_1046_equation_0 = const()[name = tensor("op_1046_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1046_cast_fp16 = einsum(equation = var_1046_equation_0, values = (var_1008_cast_fp16_2, var_1035_cast_fp16))[name = tensor("op_1046_cast_fp16")]; + tensor var_1048_equation_0 = const()[name = tensor("op_1048_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1048_cast_fp16 = einsum(equation = var_1048_equation_0, values = (var_1008_cast_fp16_3, var_1036_cast_fp16))[name = tensor("op_1048_cast_fp16")]; + tensor var_1050_equation_0 = const()[name = tensor("op_1050_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1050_cast_fp16 = einsum(equation = var_1050_equation_0, values = (var_1008_cast_fp16_4, var_1037_cast_fp16))[name = tensor("op_1050_cast_fp16")]; + tensor var_1052_equation_0 = const()[name = tensor("op_1052_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1052_cast_fp16 = einsum(equation = var_1052_equation_0, values = (var_1008_cast_fp16_5, var_1038_cast_fp16))[name = tensor("op_1052_cast_fp16")]; + tensor var_1054_equation_0 = const()[name = tensor("op_1054_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1054_cast_fp16 = einsum(equation = var_1054_equation_0, values = (var_1008_cast_fp16_6, var_1039_cast_fp16))[name = tensor("op_1054_cast_fp16")]; + tensor var_1056_equation_0 = const()[name = tensor("op_1056_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1056_cast_fp16 = einsum(equation = var_1056_equation_0, values = (var_1008_cast_fp16_7, var_1040_cast_fp16))[name = tensor("op_1056_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_937, interleave = input_55_interleave_0, values = (var_1042_cast_fp16, var_1044_cast_fp16, var_1046_cast_fp16, var_1048_cast_fp16, var_1050_cast_fp16, var_1052_cast_fp16, var_1054_cast_fp16, var_1056_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1065_pad_type_0 = const()[name = tensor("op_1065_pad_type_0"), val = tensor("valid")]; + tensor var_1065_strides_0 = const()[name = tensor("op_1065_strides_0"), val = tensor([1, 1])]; + tensor var_1065_pad_0 = const()[name = tensor("op_1065_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1065_dilations_0 = const()[name = tensor("op_1065_dilations_0"), val = tensor([1, 1])]; + tensor var_1065_groups_0 = const()[name = tensor("op_1065_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36457984)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36982336)))]; + tensor var_1065_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1065_dilations_0, groups = var_1065_groups_0, pad = var_1065_pad_0, pad_type = var_1065_pad_type_0, strides = var_1065_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1065_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1065_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36983424)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36984512)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1075_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36985600)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39082816)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_59_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_1101_pad_type_0 = const()[name = tensor("op_1101_pad_type_0"), val = tensor("valid")]; + tensor var_1101_strides_0 = const()[name = tensor("op_1101_strides_0"), val = tensor([1, 1])]; + tensor var_1101_pad_0 = const()[name = tensor("op_1101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1101_dilations_0 = const()[name = tensor("op_1101_dilations_0"), val = tensor([1, 1])]; + tensor var_1101_groups_0 = const()[name = tensor("op_1101_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39086976)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41184192)))]; + tensor var_1101_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1101_dilations_0, groups = var_1101_groups_0, pad = var_1101_pad_0, pad_type = var_1101_pad_type_0, strides = var_1101_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1101_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41185280)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41186368)))]; + tensor var_1115_to_fp16 = const()[name = tensor("op_1115_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_1115_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([2])]; + tensor var_1126_cast_fp16 = squeeze(axes = var_1126_axes_0, x = x_cast_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor var_1129_perm_0 = const()[name = tensor("op_1129_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1129_cast_fp16 = transpose(perm = var_1129_perm_0, x = var_1126_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_1129_cast_fp16_to_fp32_dtype_0, x = var_1129_cast_fp16)[name = tensor("cast_27")]; + } -> (output); +} \ No newline at end of file diff --git a/base/ggml-base-encoder.mlmodelc/weights/weight.bin b/base/ggml-base-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cb3d24c42f71f1a8da10dec6dd90d8268c837658 --- /dev/null +++ b/base/ggml-base-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7f7b1747c1873721448575a03baf7cf56b0bc72f6e04df7d1f9ce305688b9a2 +size 41187456 diff --git a/base/ggml-base.bin b/base/ggml-base.bin new file mode 100644 index 0000000000000000000000000000000000000000..17993a254ac4e9db7642a261cf53af3dc446145c --- /dev/null +++ b/base/ggml-base.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe +size 147951465 diff --git a/large-v1/.DS_Store b/large-v1/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..0585c5ab640da137e4708c84a4b7e0d63c6360af Binary files /dev/null and b/large-v1/.DS_Store differ diff --git a/large-v1/ggml-large-v1-encoder.mlmodelc/analytics/coremldata.bin b/large-v1/ggml-large-v1-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6dc984a46d0a814f3f51e1b9a397639a2cae2e9e --- /dev/null +++ b/large-v1/ggml-large-v1-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ab6e6c98293af2c12bf584cdda19a464709cef895b408c568445cd111d206b9 +size 243 diff --git a/large-v1/ggml-large-v1-encoder.mlmodelc/coremldata.bin b/large-v1/ggml-large-v1-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d3539e433ce80b83110a6d06e68290711f92c478 --- /dev/null +++ b/large-v1/ggml-large-v1-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bc95fdd4e4c14ec1ad848ab88c742c79010ffa3969276143703ffed87fc8763 +size 320 diff --git a/large-v1/ggml-large-v1-encoder.mlmodelc/metadata.json b/large-v1/ggml-large-v1-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..a01304a9126b85f6f579e0334bb2dce26cd657d9 --- /dev/null +++ b/large-v1/ggml-large-v1-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 1280)", + "shortDescription" : "", + "shape" : "[1, 1500, 1280]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 32, + "Gelu" : 34, + "LayerNorm" : 65, + "Transpose" : 33, + "Softmax" : 640, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 65, + "Einsum" : 1280, + "ExpandDims" : 1, + "Split" : 96, + "Conv" : 194 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.source" : "torch==2.2.2", + "com.github.apple.coremltools.version" : "8.3.0" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_large_v1", + "method" : "predict" + } +] \ No newline at end of file diff --git a/large-v1/ggml-large-v1-encoder.mlmodelc/model.mil b/large-v1/ggml-large-v1-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8ab8cd9e6674b7e4122ebb511e9de5d9978c2bb6 --- /dev/null +++ b/large-v1/ggml-large-v1-encoder.mlmodelc/model.mil @@ -0,0 +1,5643 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_84_pad_type_0 = const()[name = tensor("op_84_pad_type_0"), val = tensor("custom")]; + tensor var_84_pad_0 = const()[name = tensor("op_84_pad_0"), val = tensor([1, 1])]; + tensor var_84_strides_0 = const()[name = tensor("op_84_strides_0"), val = tensor([1])]; + tensor var_84_dilations_0 = const()[name = tensor("op_84_dilations_0"), val = tensor([1])]; + tensor var_84_groups_0 = const()[name = tensor("op_84_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614528)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_132")]; + tensor var_84_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_84_dilations_0, groups = var_84_groups_0, pad = var_84_pad_0, pad_type = var_84_pad_type_0, strides = var_84_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_84_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_102_pad_type_0 = const()[name = tensor("op_102_pad_type_0"), val = tensor("custom")]; + tensor var_102_pad_0 = const()[name = tensor("op_102_pad_0"), val = tensor([1, 1])]; + tensor var_102_strides_0 = const()[name = tensor("op_102_strides_0"), val = tensor([2])]; + tensor var_102_dilations_0 = const()[name = tensor("op_102_dilations_0"), val = tensor([1])]; + tensor var_102_groups_0 = const()[name = tensor("op_102_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617152)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10447616)))]; + tensor var_102_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_102_dilations_0, groups = var_102_groups_0, pad = var_102_pad_0, pad_type = var_102_pad_type_0, strides = var_102_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_102_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_102_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_107_to_fp16 = const()[name = tensor("op_107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10450240)))]; + tensor var_109_cast_fp16 = add(x = x_3_cast_fp16, y = var_107_to_fp16)[name = tensor("op_109_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_109_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_124 = const()[name = tensor("op_124"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14290304)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14292928)))]; + tensor var_140_to_fp16 = const()[name = tensor("op_140_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_140_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_175_weight_0_to_fp16 = const()[name = tensor("op_175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14295552)))]; + tensor var_175_bias_0_to_fp16 = const()[name = tensor("op_175_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17572416)))]; + tensor var_175_cast_fp16 = conv(bias = var_175_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_175_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_175_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17575040)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_173_pad_type_0 = const()[name = tensor("op_173_pad_type_0"), val = tensor("valid")]; + tensor var_173_strides_0 = const()[name = tensor("op_173_strides_0"), val = tensor([1, 1])]; + tensor var_173_pad_0 = const()[name = tensor("op_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_173_dilations_0 = const()[name = tensor("op_173_dilations_0"), val = tensor([1, 1])]; + tensor var_173_groups_0 = const()[name = tensor("op_173_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20851904)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24128768)))]; + tensor var_173_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_173_dilations_0, groups = var_173_groups_0, pad = var_173_pad_0, pad_type = var_173_pad_type_0, strides = var_173_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_173_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_176_axis_0 = const()[name = tensor("op_176_axis_0"), val = tensor(1)]; + tensor var_176_cast_fp16_0, tensor var_176_cast_fp16_1, tensor var_176_cast_fp16_2, tensor var_176_cast_fp16_3, tensor var_176_cast_fp16_4, tensor var_176_cast_fp16_5, tensor var_176_cast_fp16_6, tensor var_176_cast_fp16_7, tensor var_176_cast_fp16_8, tensor var_176_cast_fp16_9, tensor var_176_cast_fp16_10, tensor var_176_cast_fp16_11, tensor var_176_cast_fp16_12, tensor var_176_cast_fp16_13, tensor var_176_cast_fp16_14, tensor var_176_cast_fp16_15, tensor var_176_cast_fp16_16, tensor var_176_cast_fp16_17, tensor var_176_cast_fp16_18, tensor var_176_cast_fp16_19 = split(axis = var_176_axis_0, split_sizes = tile_0, x = var_175_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor var_197_perm_0 = const()[name = tensor("op_197_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_198_axis_0 = const()[name = tensor("op_198_axis_0"), val = tensor(3)]; + tensor var_197_cast_fp16 = transpose(perm = var_197_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_32")]; + tensor var_198_cast_fp16_0, tensor var_198_cast_fp16_1, tensor var_198_cast_fp16_2, tensor var_198_cast_fp16_3, tensor var_198_cast_fp16_4, tensor var_198_cast_fp16_5, tensor var_198_cast_fp16_6, tensor var_198_cast_fp16_7, tensor var_198_cast_fp16_8, tensor var_198_cast_fp16_9, tensor var_198_cast_fp16_10, tensor var_198_cast_fp16_11, tensor var_198_cast_fp16_12, tensor var_198_cast_fp16_13, tensor var_198_cast_fp16_14, tensor var_198_cast_fp16_15, tensor var_198_cast_fp16_16, tensor var_198_cast_fp16_17, tensor var_198_cast_fp16_18, tensor var_198_cast_fp16_19 = split(axis = var_198_axis_0, split_sizes = tile_1, x = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_219_axis_0 = const()[name = tensor("op_219_axis_0"), val = tensor(1)]; + tensor var_219_cast_fp16_0, tensor var_219_cast_fp16_1, tensor var_219_cast_fp16_2, tensor var_219_cast_fp16_3, tensor var_219_cast_fp16_4, tensor var_219_cast_fp16_5, tensor var_219_cast_fp16_6, tensor var_219_cast_fp16_7, tensor var_219_cast_fp16_8, tensor var_219_cast_fp16_9, tensor var_219_cast_fp16_10, tensor var_219_cast_fp16_11, tensor var_219_cast_fp16_12, tensor var_219_cast_fp16_13, tensor var_219_cast_fp16_14, tensor var_219_cast_fp16_15, tensor var_219_cast_fp16_16, tensor var_219_cast_fp16_17, tensor var_219_cast_fp16_18, tensor var_219_cast_fp16_19 = split(axis = var_219_axis_0, split_sizes = tile_2, x = var_173_cast_fp16)[name = tensor("op_219_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_198_cast_fp16_0, var_176_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_198_cast_fp16_1, var_176_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_198_cast_fp16_2, var_176_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_198_cast_fp16_3, var_176_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_198_cast_fp16_4, var_176_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_198_cast_fp16_5, var_176_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_198_cast_fp16_6, var_176_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_198_cast_fp16_7, var_176_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_198_cast_fp16_8, var_176_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_198_cast_fp16_9, var_176_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_198_cast_fp16_10, var_176_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_198_cast_fp16_11, var_176_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_198_cast_fp16_12, var_176_cast_fp16_12))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_198_cast_fp16_13, var_176_cast_fp16_13))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_198_cast_fp16_14, var_176_cast_fp16_14))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_198_cast_fp16_15, var_176_cast_fp16_15))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_198_cast_fp16_16, var_176_cast_fp16_16))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_198_cast_fp16_17, var_176_cast_fp16_17))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_198_cast_fp16_18, var_176_cast_fp16_18))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_198_cast_fp16_19, var_176_cast_fp16_19))[name = tensor("aw_39_cast_fp16")]; + tensor var_280_cast_fp16 = softmax(axis = var_124, x = aw_1_cast_fp16)[name = tensor("op_280_cast_fp16")]; + tensor var_281_cast_fp16 = softmax(axis = var_124, x = aw_3_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor var_282_cast_fp16 = softmax(axis = var_124, x = aw_5_cast_fp16)[name = tensor("op_282_cast_fp16")]; + tensor var_283_cast_fp16 = softmax(axis = var_124, x = aw_7_cast_fp16)[name = tensor("op_283_cast_fp16")]; + tensor var_284_cast_fp16 = softmax(axis = var_124, x = aw_9_cast_fp16)[name = tensor("op_284_cast_fp16")]; + tensor var_285_cast_fp16 = softmax(axis = var_124, x = aw_11_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_286_cast_fp16 = softmax(axis = var_124, x = aw_13_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = softmax(axis = var_124, x = aw_15_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor var_288_cast_fp16 = softmax(axis = var_124, x = aw_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_289_cast_fp16 = softmax(axis = var_124, x = aw_19_cast_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_290_cast_fp16 = softmax(axis = var_124, x = aw_21_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = softmax(axis = var_124, x = aw_23_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor var_292_cast_fp16 = softmax(axis = var_124, x = aw_25_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_293_cast_fp16 = softmax(axis = var_124, x = aw_27_cast_fp16)[name = tensor("op_293_cast_fp16")]; + tensor var_294_cast_fp16 = softmax(axis = var_124, x = aw_29_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_295_cast_fp16 = softmax(axis = var_124, x = aw_31_cast_fp16)[name = tensor("op_295_cast_fp16")]; + tensor var_296_cast_fp16 = softmax(axis = var_124, x = aw_33_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_297_cast_fp16 = softmax(axis = var_124, x = aw_35_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_298_cast_fp16 = softmax(axis = var_124, x = aw_37_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_299_cast_fp16 = softmax(axis = var_124, x = aw_39_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_equation_0 = const()[name = tensor("op_301_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_301_cast_fp16 = einsum(equation = var_301_equation_0, values = (var_219_cast_fp16_0, var_280_cast_fp16))[name = tensor("op_301_cast_fp16")]; + tensor var_303_equation_0 = const()[name = tensor("op_303_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_303_cast_fp16 = einsum(equation = var_303_equation_0, values = (var_219_cast_fp16_1, var_281_cast_fp16))[name = tensor("op_303_cast_fp16")]; + tensor var_305_equation_0 = const()[name = tensor("op_305_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_305_cast_fp16 = einsum(equation = var_305_equation_0, values = (var_219_cast_fp16_2, var_282_cast_fp16))[name = tensor("op_305_cast_fp16")]; + tensor var_307_equation_0 = const()[name = tensor("op_307_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_307_cast_fp16 = einsum(equation = var_307_equation_0, values = (var_219_cast_fp16_3, var_283_cast_fp16))[name = tensor("op_307_cast_fp16")]; + tensor var_309_equation_0 = const()[name = tensor("op_309_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_309_cast_fp16 = einsum(equation = var_309_equation_0, values = (var_219_cast_fp16_4, var_284_cast_fp16))[name = tensor("op_309_cast_fp16")]; + tensor var_311_equation_0 = const()[name = tensor("op_311_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_311_cast_fp16 = einsum(equation = var_311_equation_0, values = (var_219_cast_fp16_5, var_285_cast_fp16))[name = tensor("op_311_cast_fp16")]; + tensor var_313_equation_0 = const()[name = tensor("op_313_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_313_cast_fp16 = einsum(equation = var_313_equation_0, values = (var_219_cast_fp16_6, var_286_cast_fp16))[name = tensor("op_313_cast_fp16")]; + tensor var_315_equation_0 = const()[name = tensor("op_315_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_315_cast_fp16 = einsum(equation = var_315_equation_0, values = (var_219_cast_fp16_7, var_287_cast_fp16))[name = tensor("op_315_cast_fp16")]; + tensor var_317_equation_0 = const()[name = tensor("op_317_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_317_cast_fp16 = einsum(equation = var_317_equation_0, values = (var_219_cast_fp16_8, var_288_cast_fp16))[name = tensor("op_317_cast_fp16")]; + tensor var_319_equation_0 = const()[name = tensor("op_319_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_319_cast_fp16 = einsum(equation = var_319_equation_0, values = (var_219_cast_fp16_9, var_289_cast_fp16))[name = tensor("op_319_cast_fp16")]; + tensor var_321_equation_0 = const()[name = tensor("op_321_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_321_cast_fp16 = einsum(equation = var_321_equation_0, values = (var_219_cast_fp16_10, var_290_cast_fp16))[name = tensor("op_321_cast_fp16")]; + tensor var_323_equation_0 = const()[name = tensor("op_323_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_323_cast_fp16 = einsum(equation = var_323_equation_0, values = (var_219_cast_fp16_11, var_291_cast_fp16))[name = tensor("op_323_cast_fp16")]; + tensor var_325_equation_0 = const()[name = tensor("op_325_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_325_cast_fp16 = einsum(equation = var_325_equation_0, values = (var_219_cast_fp16_12, var_292_cast_fp16))[name = tensor("op_325_cast_fp16")]; + tensor var_327_equation_0 = const()[name = tensor("op_327_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_327_cast_fp16 = einsum(equation = var_327_equation_0, values = (var_219_cast_fp16_13, var_293_cast_fp16))[name = tensor("op_327_cast_fp16")]; + tensor var_329_equation_0 = const()[name = tensor("op_329_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_329_cast_fp16 = einsum(equation = var_329_equation_0, values = (var_219_cast_fp16_14, var_294_cast_fp16))[name = tensor("op_329_cast_fp16")]; + tensor var_331_equation_0 = const()[name = tensor("op_331_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_331_cast_fp16 = einsum(equation = var_331_equation_0, values = (var_219_cast_fp16_15, var_295_cast_fp16))[name = tensor("op_331_cast_fp16")]; + tensor var_333_equation_0 = const()[name = tensor("op_333_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_333_cast_fp16 = einsum(equation = var_333_equation_0, values = (var_219_cast_fp16_16, var_296_cast_fp16))[name = tensor("op_333_cast_fp16")]; + tensor var_335_equation_0 = const()[name = tensor("op_335_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_335_cast_fp16 = einsum(equation = var_335_equation_0, values = (var_219_cast_fp16_17, var_297_cast_fp16))[name = tensor("op_335_cast_fp16")]; + tensor var_337_equation_0 = const()[name = tensor("op_337_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_337_cast_fp16 = einsum(equation = var_337_equation_0, values = (var_219_cast_fp16_18, var_298_cast_fp16))[name = tensor("op_337_cast_fp16")]; + tensor var_339_equation_0 = const()[name = tensor("op_339_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_339_cast_fp16 = einsum(equation = var_339_equation_0, values = (var_219_cast_fp16_19, var_299_cast_fp16))[name = tensor("op_339_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_124, interleave = input_5_interleave_0, values = (var_301_cast_fp16, var_303_cast_fp16, var_305_cast_fp16, var_307_cast_fp16, var_309_cast_fp16, var_311_cast_fp16, var_313_cast_fp16, var_315_cast_fp16, var_317_cast_fp16, var_319_cast_fp16, var_321_cast_fp16, var_323_cast_fp16, var_325_cast_fp16, var_327_cast_fp16, var_329_cast_fp16, var_331_cast_fp16, var_333_cast_fp16, var_335_cast_fp16, var_337_cast_fp16, var_339_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_348_pad_type_0 = const()[name = tensor("op_348_pad_type_0"), val = tensor("valid")]; + tensor var_348_strides_0 = const()[name = tensor("op_348_strides_0"), val = tensor([1, 1])]; + tensor var_348_pad_0 = const()[name = tensor("op_348_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_348_dilations_0 = const()[name = tensor("op_348_dilations_0"), val = tensor([1, 1])]; + tensor var_348_groups_0 = const()[name = tensor("op_348_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24131392)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27408256)))]; + tensor var_348_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_348_dilations_0, groups = var_348_groups_0, pad = var_348_pad_0, pad_type = var_348_pad_type_0, strides = var_348_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_348_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27410880)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27413504)))]; + tensor var_358_to_fp16 = const()[name = tensor("op_358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_358_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27416128)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40523392)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_384_pad_type_0 = const()[name = tensor("op_384_pad_type_0"), val = tensor("valid")]; + tensor var_384_strides_0 = const()[name = tensor("op_384_strides_0"), val = tensor([1, 1])]; + tensor var_384_pad_0 = const()[name = tensor("op_384_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_384_dilations_0 = const()[name = tensor("op_384_dilations_0"), val = tensor([1, 1])]; + tensor var_384_groups_0 = const()[name = tensor("op_384_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40533696)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53640960)))]; + tensor var_384_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_384_dilations_0, groups = var_384_groups_0, pad = var_384_pad_0, pad_type = var_384_pad_type_0, strides = var_384_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_384_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53643584)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53646208)))]; + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_409_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_444_weight_0_to_fp16 = const()[name = tensor("op_444_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53648832)))]; + tensor var_444_bias_0_to_fp16 = const()[name = tensor("op_444_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56925696)))]; + tensor var_444_cast_fp16 = conv(bias = var_444_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_444_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_444_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56928320)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_442_pad_type_0 = const()[name = tensor("op_442_pad_type_0"), val = tensor("valid")]; + tensor var_442_strides_0 = const()[name = tensor("op_442_strides_0"), val = tensor([1, 1])]; + tensor var_442_pad_0 = const()[name = tensor("op_442_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_442_dilations_0 = const()[name = tensor("op_442_dilations_0"), val = tensor([1, 1])]; + tensor var_442_groups_0 = const()[name = tensor("op_442_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60205184)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63482048)))]; + tensor var_442_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_442_dilations_0, groups = var_442_groups_0, pad = var_442_pad_0, pad_type = var_442_pad_type_0, strides = var_442_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_cast_fp16_0, tensor var_445_cast_fp16_1, tensor var_445_cast_fp16_2, tensor var_445_cast_fp16_3, tensor var_445_cast_fp16_4, tensor var_445_cast_fp16_5, tensor var_445_cast_fp16_6, tensor var_445_cast_fp16_7, tensor var_445_cast_fp16_8, tensor var_445_cast_fp16_9, tensor var_445_cast_fp16_10, tensor var_445_cast_fp16_11, tensor var_445_cast_fp16_12, tensor var_445_cast_fp16_13, tensor var_445_cast_fp16_14, tensor var_445_cast_fp16_15, tensor var_445_cast_fp16_16, tensor var_445_cast_fp16_17, tensor var_445_cast_fp16_18, tensor var_445_cast_fp16_19 = split(axis = var_445_axis_0, split_sizes = tile_3, x = var_444_cast_fp16)[name = tensor("op_445_cast_fp16")]; + tensor var_466_perm_0 = const()[name = tensor("op_466_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_467_axis_0 = const()[name = tensor("op_467_axis_0"), val = tensor(3)]; + tensor var_466_cast_fp16 = transpose(perm = var_466_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_31")]; + tensor var_467_cast_fp16_0, tensor var_467_cast_fp16_1, tensor var_467_cast_fp16_2, tensor var_467_cast_fp16_3, tensor var_467_cast_fp16_4, tensor var_467_cast_fp16_5, tensor var_467_cast_fp16_6, tensor var_467_cast_fp16_7, tensor var_467_cast_fp16_8, tensor var_467_cast_fp16_9, tensor var_467_cast_fp16_10, tensor var_467_cast_fp16_11, tensor var_467_cast_fp16_12, tensor var_467_cast_fp16_13, tensor var_467_cast_fp16_14, tensor var_467_cast_fp16_15, tensor var_467_cast_fp16_16, tensor var_467_cast_fp16_17, tensor var_467_cast_fp16_18, tensor var_467_cast_fp16_19 = split(axis = var_467_axis_0, split_sizes = tile_4, x = var_466_cast_fp16)[name = tensor("op_467_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_488_axis_0 = const()[name = tensor("op_488_axis_0"), val = tensor(1)]; + tensor var_488_cast_fp16_0, tensor var_488_cast_fp16_1, tensor var_488_cast_fp16_2, tensor var_488_cast_fp16_3, tensor var_488_cast_fp16_4, tensor var_488_cast_fp16_5, tensor var_488_cast_fp16_6, tensor var_488_cast_fp16_7, tensor var_488_cast_fp16_8, tensor var_488_cast_fp16_9, tensor var_488_cast_fp16_10, tensor var_488_cast_fp16_11, tensor var_488_cast_fp16_12, tensor var_488_cast_fp16_13, tensor var_488_cast_fp16_14, tensor var_488_cast_fp16_15, tensor var_488_cast_fp16_16, tensor var_488_cast_fp16_17, tensor var_488_cast_fp16_18, tensor var_488_cast_fp16_19 = split(axis = var_488_axis_0, split_sizes = tile_5, x = var_442_cast_fp16)[name = tensor("op_488_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_467_cast_fp16_0, var_445_cast_fp16_0))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_467_cast_fp16_1, var_445_cast_fp16_1))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_467_cast_fp16_2, var_445_cast_fp16_2))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_467_cast_fp16_3, var_445_cast_fp16_3))[name = tensor("aw_47_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_467_cast_fp16_4, var_445_cast_fp16_4))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_467_cast_fp16_5, var_445_cast_fp16_5))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_467_cast_fp16_6, var_445_cast_fp16_6))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_467_cast_fp16_7, var_445_cast_fp16_7))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_467_cast_fp16_8, var_445_cast_fp16_8))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_467_cast_fp16_9, var_445_cast_fp16_9))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_467_cast_fp16_10, var_445_cast_fp16_10))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_467_cast_fp16_11, var_445_cast_fp16_11))[name = tensor("aw_63_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_467_cast_fp16_12, var_445_cast_fp16_12))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_467_cast_fp16_13, var_445_cast_fp16_13))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_467_cast_fp16_14, var_445_cast_fp16_14))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_467_cast_fp16_15, var_445_cast_fp16_15))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_467_cast_fp16_16, var_445_cast_fp16_16))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_467_cast_fp16_17, var_445_cast_fp16_17))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_467_cast_fp16_18, var_445_cast_fp16_18))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_467_cast_fp16_19, var_445_cast_fp16_19))[name = tensor("aw_79_cast_fp16")]; + tensor var_549_cast_fp16 = softmax(axis = var_393, x = aw_41_cast_fp16)[name = tensor("op_549_cast_fp16")]; + tensor var_550_cast_fp16 = softmax(axis = var_393, x = aw_43_cast_fp16)[name = tensor("op_550_cast_fp16")]; + tensor var_551_cast_fp16 = softmax(axis = var_393, x = aw_45_cast_fp16)[name = tensor("op_551_cast_fp16")]; + tensor var_552_cast_fp16 = softmax(axis = var_393, x = aw_47_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = softmax(axis = var_393, x = aw_49_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor var_554_cast_fp16 = softmax(axis = var_393, x = aw_51_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_555_cast_fp16 = softmax(axis = var_393, x = aw_53_cast_fp16)[name = tensor("op_555_cast_fp16")]; + tensor var_556_cast_fp16 = softmax(axis = var_393, x = aw_55_cast_fp16)[name = tensor("op_556_cast_fp16")]; + tensor var_557_cast_fp16 = softmax(axis = var_393, x = aw_57_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_558_cast_fp16 = softmax(axis = var_393, x = aw_59_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor var_559_cast_fp16 = softmax(axis = var_393, x = aw_61_cast_fp16)[name = tensor("op_559_cast_fp16")]; + tensor var_560_cast_fp16 = softmax(axis = var_393, x = aw_63_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_561_cast_fp16 = softmax(axis = var_393, x = aw_65_cast_fp16)[name = tensor("op_561_cast_fp16")]; + tensor var_562_cast_fp16 = softmax(axis = var_393, x = aw_67_cast_fp16)[name = tensor("op_562_cast_fp16")]; + tensor var_563_cast_fp16 = softmax(axis = var_393, x = aw_69_cast_fp16)[name = tensor("op_563_cast_fp16")]; + tensor var_564_cast_fp16 = softmax(axis = var_393, x = aw_71_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor var_565_cast_fp16 = softmax(axis = var_393, x = aw_73_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_566_cast_fp16 = softmax(axis = var_393, x = aw_75_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = softmax(axis = var_393, x = aw_77_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor var_568_cast_fp16 = softmax(axis = var_393, x = aw_79_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor var_570_equation_0 = const()[name = tensor("op_570_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_570_cast_fp16 = einsum(equation = var_570_equation_0, values = (var_488_cast_fp16_0, var_549_cast_fp16))[name = tensor("op_570_cast_fp16")]; + tensor var_572_equation_0 = const()[name = tensor("op_572_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_572_cast_fp16 = einsum(equation = var_572_equation_0, values = (var_488_cast_fp16_1, var_550_cast_fp16))[name = tensor("op_572_cast_fp16")]; + tensor var_574_equation_0 = const()[name = tensor("op_574_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_574_cast_fp16 = einsum(equation = var_574_equation_0, values = (var_488_cast_fp16_2, var_551_cast_fp16))[name = tensor("op_574_cast_fp16")]; + tensor var_576_equation_0 = const()[name = tensor("op_576_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_576_cast_fp16 = einsum(equation = var_576_equation_0, values = (var_488_cast_fp16_3, var_552_cast_fp16))[name = tensor("op_576_cast_fp16")]; + tensor var_578_equation_0 = const()[name = tensor("op_578_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_578_cast_fp16 = einsum(equation = var_578_equation_0, values = (var_488_cast_fp16_4, var_553_cast_fp16))[name = tensor("op_578_cast_fp16")]; + tensor var_580_equation_0 = const()[name = tensor("op_580_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_580_cast_fp16 = einsum(equation = var_580_equation_0, values = (var_488_cast_fp16_5, var_554_cast_fp16))[name = tensor("op_580_cast_fp16")]; + tensor var_582_equation_0 = const()[name = tensor("op_582_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_582_cast_fp16 = einsum(equation = var_582_equation_0, values = (var_488_cast_fp16_6, var_555_cast_fp16))[name = tensor("op_582_cast_fp16")]; + tensor var_584_equation_0 = const()[name = tensor("op_584_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_584_cast_fp16 = einsum(equation = var_584_equation_0, values = (var_488_cast_fp16_7, var_556_cast_fp16))[name = tensor("op_584_cast_fp16")]; + tensor var_586_equation_0 = const()[name = tensor("op_586_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_586_cast_fp16 = einsum(equation = var_586_equation_0, values = (var_488_cast_fp16_8, var_557_cast_fp16))[name = tensor("op_586_cast_fp16")]; + tensor var_588_equation_0 = const()[name = tensor("op_588_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_588_cast_fp16 = einsum(equation = var_588_equation_0, values = (var_488_cast_fp16_9, var_558_cast_fp16))[name = tensor("op_588_cast_fp16")]; + tensor var_590_equation_0 = const()[name = tensor("op_590_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_590_cast_fp16 = einsum(equation = var_590_equation_0, values = (var_488_cast_fp16_10, var_559_cast_fp16))[name = tensor("op_590_cast_fp16")]; + tensor var_592_equation_0 = const()[name = tensor("op_592_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_592_cast_fp16 = einsum(equation = var_592_equation_0, values = (var_488_cast_fp16_11, var_560_cast_fp16))[name = tensor("op_592_cast_fp16")]; + tensor var_594_equation_0 = const()[name = tensor("op_594_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_594_cast_fp16 = einsum(equation = var_594_equation_0, values = (var_488_cast_fp16_12, var_561_cast_fp16))[name = tensor("op_594_cast_fp16")]; + tensor var_596_equation_0 = const()[name = tensor("op_596_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_596_cast_fp16 = einsum(equation = var_596_equation_0, values = (var_488_cast_fp16_13, var_562_cast_fp16))[name = tensor("op_596_cast_fp16")]; + tensor var_598_equation_0 = const()[name = tensor("op_598_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_598_cast_fp16 = einsum(equation = var_598_equation_0, values = (var_488_cast_fp16_14, var_563_cast_fp16))[name = tensor("op_598_cast_fp16")]; + tensor var_600_equation_0 = const()[name = tensor("op_600_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_600_cast_fp16 = einsum(equation = var_600_equation_0, values = (var_488_cast_fp16_15, var_564_cast_fp16))[name = tensor("op_600_cast_fp16")]; + tensor var_602_equation_0 = const()[name = tensor("op_602_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_602_cast_fp16 = einsum(equation = var_602_equation_0, values = (var_488_cast_fp16_16, var_565_cast_fp16))[name = tensor("op_602_cast_fp16")]; + tensor var_604_equation_0 = const()[name = tensor("op_604_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_604_cast_fp16 = einsum(equation = var_604_equation_0, values = (var_488_cast_fp16_17, var_566_cast_fp16))[name = tensor("op_604_cast_fp16")]; + tensor var_606_equation_0 = const()[name = tensor("op_606_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_606_cast_fp16 = einsum(equation = var_606_equation_0, values = (var_488_cast_fp16_18, var_567_cast_fp16))[name = tensor("op_606_cast_fp16")]; + tensor var_608_equation_0 = const()[name = tensor("op_608_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_608_cast_fp16 = einsum(equation = var_608_equation_0, values = (var_488_cast_fp16_19, var_568_cast_fp16))[name = tensor("op_608_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_393, interleave = input_15_interleave_0, values = (var_570_cast_fp16, var_572_cast_fp16, var_574_cast_fp16, var_576_cast_fp16, var_578_cast_fp16, var_580_cast_fp16, var_582_cast_fp16, var_584_cast_fp16, var_586_cast_fp16, var_588_cast_fp16, var_590_cast_fp16, var_592_cast_fp16, var_594_cast_fp16, var_596_cast_fp16, var_598_cast_fp16, var_600_cast_fp16, var_602_cast_fp16, var_604_cast_fp16, var_606_cast_fp16, var_608_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_617_pad_type_0 = const()[name = tensor("op_617_pad_type_0"), val = tensor("valid")]; + tensor var_617_strides_0 = const()[name = tensor("op_617_strides_0"), val = tensor([1, 1])]; + tensor var_617_pad_0 = const()[name = tensor("op_617_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_617_dilations_0 = const()[name = tensor("op_617_dilations_0"), val = tensor([1, 1])]; + tensor var_617_groups_0 = const()[name = tensor("op_617_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63484672)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66761536)))]; + tensor var_617_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_617_dilations_0, groups = var_617_groups_0, pad = var_617_pad_0, pad_type = var_617_pad_type_0, strides = var_617_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_617_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66764160)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66766784)))]; + tensor var_627_to_fp16 = const()[name = tensor("op_627_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_627_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66769408)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79876672)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_653_pad_type_0 = const()[name = tensor("op_653_pad_type_0"), val = tensor("valid")]; + tensor var_653_strides_0 = const()[name = tensor("op_653_strides_0"), val = tensor([1, 1])]; + tensor var_653_pad_0 = const()[name = tensor("op_653_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_653_dilations_0 = const()[name = tensor("op_653_dilations_0"), val = tensor([1, 1])]; + tensor var_653_groups_0 = const()[name = tensor("op_653_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79886976)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92994240)))]; + tensor var_653_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_653_dilations_0, groups = var_653_groups_0, pad = var_653_pad_0, pad_type = var_653_pad_type_0, strides = var_653_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_653_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_662 = const()[name = tensor("op_662"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92996864)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92999488)))]; + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_678_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_713_weight_0_to_fp16 = const()[name = tensor("op_713_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93002112)))]; + tensor var_713_bias_0_to_fp16 = const()[name = tensor("op_713_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96278976)))]; + tensor var_713_cast_fp16 = conv(bias = var_713_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_713_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96281600)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_711_pad_type_0 = const()[name = tensor("op_711_pad_type_0"), val = tensor("valid")]; + tensor var_711_strides_0 = const()[name = tensor("op_711_strides_0"), val = tensor([1, 1])]; + tensor var_711_pad_0 = const()[name = tensor("op_711_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_711_dilations_0 = const()[name = tensor("op_711_dilations_0"), val = tensor([1, 1])]; + tensor var_711_groups_0 = const()[name = tensor("op_711_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99558464)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102835328)))]; + tensor var_711_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_711_dilations_0, groups = var_711_groups_0, pad = var_711_pad_0, pad_type = var_711_pad_type_0, strides = var_711_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_711_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_714_axis_0 = const()[name = tensor("op_714_axis_0"), val = tensor(1)]; + tensor var_714_cast_fp16_0, tensor var_714_cast_fp16_1, tensor var_714_cast_fp16_2, tensor var_714_cast_fp16_3, tensor var_714_cast_fp16_4, tensor var_714_cast_fp16_5, tensor var_714_cast_fp16_6, tensor var_714_cast_fp16_7, tensor var_714_cast_fp16_8, tensor var_714_cast_fp16_9, tensor var_714_cast_fp16_10, tensor var_714_cast_fp16_11, tensor var_714_cast_fp16_12, tensor var_714_cast_fp16_13, tensor var_714_cast_fp16_14, tensor var_714_cast_fp16_15, tensor var_714_cast_fp16_16, tensor var_714_cast_fp16_17, tensor var_714_cast_fp16_18, tensor var_714_cast_fp16_19 = split(axis = var_714_axis_0, split_sizes = tile_6, x = var_713_cast_fp16)[name = tensor("op_714_cast_fp16")]; + tensor var_735_perm_0 = const()[name = tensor("op_735_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_736_axis_0 = const()[name = tensor("op_736_axis_0"), val = tensor(3)]; + tensor var_735_cast_fp16 = transpose(perm = var_735_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_30")]; + tensor var_736_cast_fp16_0, tensor var_736_cast_fp16_1, tensor var_736_cast_fp16_2, tensor var_736_cast_fp16_3, tensor var_736_cast_fp16_4, tensor var_736_cast_fp16_5, tensor var_736_cast_fp16_6, tensor var_736_cast_fp16_7, tensor var_736_cast_fp16_8, tensor var_736_cast_fp16_9, tensor var_736_cast_fp16_10, tensor var_736_cast_fp16_11, tensor var_736_cast_fp16_12, tensor var_736_cast_fp16_13, tensor var_736_cast_fp16_14, tensor var_736_cast_fp16_15, tensor var_736_cast_fp16_16, tensor var_736_cast_fp16_17, tensor var_736_cast_fp16_18, tensor var_736_cast_fp16_19 = split(axis = var_736_axis_0, split_sizes = tile_7, x = var_735_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_cast_fp16_0, tensor var_757_cast_fp16_1, tensor var_757_cast_fp16_2, tensor var_757_cast_fp16_3, tensor var_757_cast_fp16_4, tensor var_757_cast_fp16_5, tensor var_757_cast_fp16_6, tensor var_757_cast_fp16_7, tensor var_757_cast_fp16_8, tensor var_757_cast_fp16_9, tensor var_757_cast_fp16_10, tensor var_757_cast_fp16_11, tensor var_757_cast_fp16_12, tensor var_757_cast_fp16_13, tensor var_757_cast_fp16_14, tensor var_757_cast_fp16_15, tensor var_757_cast_fp16_16, tensor var_757_cast_fp16_17, tensor var_757_cast_fp16_18, tensor var_757_cast_fp16_19 = split(axis = var_757_axis_0, split_sizes = tile_8, x = var_711_cast_fp16)[name = tensor("op_757_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_736_cast_fp16_0, var_714_cast_fp16_0))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_736_cast_fp16_1, var_714_cast_fp16_1))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_736_cast_fp16_2, var_714_cast_fp16_2))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_736_cast_fp16_3, var_714_cast_fp16_3))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_736_cast_fp16_4, var_714_cast_fp16_4))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_736_cast_fp16_5, var_714_cast_fp16_5))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_736_cast_fp16_6, var_714_cast_fp16_6))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_736_cast_fp16_7, var_714_cast_fp16_7))[name = tensor("aw_95_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_736_cast_fp16_8, var_714_cast_fp16_8))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_736_cast_fp16_9, var_714_cast_fp16_9))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_736_cast_fp16_10, var_714_cast_fp16_10))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_736_cast_fp16_11, var_714_cast_fp16_11))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_736_cast_fp16_12, var_714_cast_fp16_12))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_736_cast_fp16_13, var_714_cast_fp16_13))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_736_cast_fp16_14, var_714_cast_fp16_14))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_736_cast_fp16_15, var_714_cast_fp16_15))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_736_cast_fp16_16, var_714_cast_fp16_16))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_736_cast_fp16_17, var_714_cast_fp16_17))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_736_cast_fp16_18, var_714_cast_fp16_18))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_736_cast_fp16_19, var_714_cast_fp16_19))[name = tensor("aw_119_cast_fp16")]; + tensor var_818_cast_fp16 = softmax(axis = var_662, x = aw_81_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_819_cast_fp16 = softmax(axis = var_662, x = aw_83_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820_cast_fp16 = softmax(axis = var_662, x = aw_85_cast_fp16)[name = tensor("op_820_cast_fp16")]; + tensor var_821_cast_fp16 = softmax(axis = var_662, x = aw_87_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = softmax(axis = var_662, x = aw_89_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor var_823_cast_fp16 = softmax(axis = var_662, x = aw_91_cast_fp16)[name = tensor("op_823_cast_fp16")]; + tensor var_824_cast_fp16 = softmax(axis = var_662, x = aw_93_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor var_825_cast_fp16 = softmax(axis = var_662, x = aw_95_cast_fp16)[name = tensor("op_825_cast_fp16")]; + tensor var_826_cast_fp16 = softmax(axis = var_662, x = aw_97_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_827_cast_fp16 = softmax(axis = var_662, x = aw_99_cast_fp16)[name = tensor("op_827_cast_fp16")]; + tensor var_828_cast_fp16 = softmax(axis = var_662, x = aw_101_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_cast_fp16 = softmax(axis = var_662, x = aw_103_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_830_cast_fp16 = softmax(axis = var_662, x = aw_105_cast_fp16)[name = tensor("op_830_cast_fp16")]; + tensor var_831_cast_fp16 = softmax(axis = var_662, x = aw_107_cast_fp16)[name = tensor("op_831_cast_fp16")]; + tensor var_832_cast_fp16 = softmax(axis = var_662, x = aw_109_cast_fp16)[name = tensor("op_832_cast_fp16")]; + tensor var_833_cast_fp16 = softmax(axis = var_662, x = aw_111_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_834_cast_fp16 = softmax(axis = var_662, x = aw_113_cast_fp16)[name = tensor("op_834_cast_fp16")]; + tensor var_835_cast_fp16 = softmax(axis = var_662, x = aw_115_cast_fp16)[name = tensor("op_835_cast_fp16")]; + tensor var_836_cast_fp16 = softmax(axis = var_662, x = aw_117_cast_fp16)[name = tensor("op_836_cast_fp16")]; + tensor var_837_cast_fp16 = softmax(axis = var_662, x = aw_119_cast_fp16)[name = tensor("op_837_cast_fp16")]; + tensor var_839_equation_0 = const()[name = tensor("op_839_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_839_cast_fp16 = einsum(equation = var_839_equation_0, values = (var_757_cast_fp16_0, var_818_cast_fp16))[name = tensor("op_839_cast_fp16")]; + tensor var_841_equation_0 = const()[name = tensor("op_841_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_841_cast_fp16 = einsum(equation = var_841_equation_0, values = (var_757_cast_fp16_1, var_819_cast_fp16))[name = tensor("op_841_cast_fp16")]; + tensor var_843_equation_0 = const()[name = tensor("op_843_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_843_cast_fp16 = einsum(equation = var_843_equation_0, values = (var_757_cast_fp16_2, var_820_cast_fp16))[name = tensor("op_843_cast_fp16")]; + tensor var_845_equation_0 = const()[name = tensor("op_845_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_845_cast_fp16 = einsum(equation = var_845_equation_0, values = (var_757_cast_fp16_3, var_821_cast_fp16))[name = tensor("op_845_cast_fp16")]; + tensor var_847_equation_0 = const()[name = tensor("op_847_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_847_cast_fp16 = einsum(equation = var_847_equation_0, values = (var_757_cast_fp16_4, var_822_cast_fp16))[name = tensor("op_847_cast_fp16")]; + tensor var_849_equation_0 = const()[name = tensor("op_849_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_849_cast_fp16 = einsum(equation = var_849_equation_0, values = (var_757_cast_fp16_5, var_823_cast_fp16))[name = tensor("op_849_cast_fp16")]; + tensor var_851_equation_0 = const()[name = tensor("op_851_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_851_cast_fp16 = einsum(equation = var_851_equation_0, values = (var_757_cast_fp16_6, var_824_cast_fp16))[name = tensor("op_851_cast_fp16")]; + tensor var_853_equation_0 = const()[name = tensor("op_853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_853_cast_fp16 = einsum(equation = var_853_equation_0, values = (var_757_cast_fp16_7, var_825_cast_fp16))[name = tensor("op_853_cast_fp16")]; + tensor var_855_equation_0 = const()[name = tensor("op_855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_855_cast_fp16 = einsum(equation = var_855_equation_0, values = (var_757_cast_fp16_8, var_826_cast_fp16))[name = tensor("op_855_cast_fp16")]; + tensor var_857_equation_0 = const()[name = tensor("op_857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_857_cast_fp16 = einsum(equation = var_857_equation_0, values = (var_757_cast_fp16_9, var_827_cast_fp16))[name = tensor("op_857_cast_fp16")]; + tensor var_859_equation_0 = const()[name = tensor("op_859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_859_cast_fp16 = einsum(equation = var_859_equation_0, values = (var_757_cast_fp16_10, var_828_cast_fp16))[name = tensor("op_859_cast_fp16")]; + tensor var_861_equation_0 = const()[name = tensor("op_861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_861_cast_fp16 = einsum(equation = var_861_equation_0, values = (var_757_cast_fp16_11, var_829_cast_fp16))[name = tensor("op_861_cast_fp16")]; + tensor var_863_equation_0 = const()[name = tensor("op_863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_863_cast_fp16 = einsum(equation = var_863_equation_0, values = (var_757_cast_fp16_12, var_830_cast_fp16))[name = tensor("op_863_cast_fp16")]; + tensor var_865_equation_0 = const()[name = tensor("op_865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_865_cast_fp16 = einsum(equation = var_865_equation_0, values = (var_757_cast_fp16_13, var_831_cast_fp16))[name = tensor("op_865_cast_fp16")]; + tensor var_867_equation_0 = const()[name = tensor("op_867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_867_cast_fp16 = einsum(equation = var_867_equation_0, values = (var_757_cast_fp16_14, var_832_cast_fp16))[name = tensor("op_867_cast_fp16")]; + tensor var_869_equation_0 = const()[name = tensor("op_869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_869_cast_fp16 = einsum(equation = var_869_equation_0, values = (var_757_cast_fp16_15, var_833_cast_fp16))[name = tensor("op_869_cast_fp16")]; + tensor var_871_equation_0 = const()[name = tensor("op_871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_871_cast_fp16 = einsum(equation = var_871_equation_0, values = (var_757_cast_fp16_16, var_834_cast_fp16))[name = tensor("op_871_cast_fp16")]; + tensor var_873_equation_0 = const()[name = tensor("op_873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_873_cast_fp16 = einsum(equation = var_873_equation_0, values = (var_757_cast_fp16_17, var_835_cast_fp16))[name = tensor("op_873_cast_fp16")]; + tensor var_875_equation_0 = const()[name = tensor("op_875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_875_cast_fp16 = einsum(equation = var_875_equation_0, values = (var_757_cast_fp16_18, var_836_cast_fp16))[name = tensor("op_875_cast_fp16")]; + tensor var_877_equation_0 = const()[name = tensor("op_877_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_877_cast_fp16 = einsum(equation = var_877_equation_0, values = (var_757_cast_fp16_19, var_837_cast_fp16))[name = tensor("op_877_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_662, interleave = input_25_interleave_0, values = (var_839_cast_fp16, var_841_cast_fp16, var_843_cast_fp16, var_845_cast_fp16, var_847_cast_fp16, var_849_cast_fp16, var_851_cast_fp16, var_853_cast_fp16, var_855_cast_fp16, var_857_cast_fp16, var_859_cast_fp16, var_861_cast_fp16, var_863_cast_fp16, var_865_cast_fp16, var_867_cast_fp16, var_869_cast_fp16, var_871_cast_fp16, var_873_cast_fp16, var_875_cast_fp16, var_877_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_886_pad_type_0 = const()[name = tensor("op_886_pad_type_0"), val = tensor("valid")]; + tensor var_886_strides_0 = const()[name = tensor("op_886_strides_0"), val = tensor([1, 1])]; + tensor var_886_pad_0 = const()[name = tensor("op_886_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_886_dilations_0 = const()[name = tensor("op_886_dilations_0"), val = tensor([1, 1])]; + tensor var_886_groups_0 = const()[name = tensor("op_886_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102837952)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106114816)))]; + tensor var_886_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_886_dilations_0, groups = var_886_groups_0, pad = var_886_pad_0, pad_type = var_886_pad_type_0, strides = var_886_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_886_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106117440)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106120064)))]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_896_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106122688)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119229952)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_922_pad_type_0 = const()[name = tensor("op_922_pad_type_0"), val = tensor("valid")]; + tensor var_922_strides_0 = const()[name = tensor("op_922_strides_0"), val = tensor([1, 1])]; + tensor var_922_pad_0 = const()[name = tensor("op_922_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_922_dilations_0 = const()[name = tensor("op_922_dilations_0"), val = tensor([1, 1])]; + tensor var_922_groups_0 = const()[name = tensor("op_922_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119240256)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132347520)))]; + tensor var_922_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_922_dilations_0, groups = var_922_groups_0, pad = var_922_pad_0, pad_type = var_922_pad_type_0, strides = var_922_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_922_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132350144)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132352768)))]; + tensor var_947_to_fp16 = const()[name = tensor("op_947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_947_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_982_weight_0_to_fp16 = const()[name = tensor("op_982_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132355392)))]; + tensor var_982_bias_0_to_fp16 = const()[name = tensor("op_982_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135632256)))]; + tensor var_982_cast_fp16 = conv(bias = var_982_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_982_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_982_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135634880)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_980_pad_type_0 = const()[name = tensor("op_980_pad_type_0"), val = tensor("valid")]; + tensor var_980_strides_0 = const()[name = tensor("op_980_strides_0"), val = tensor([1, 1])]; + tensor var_980_pad_0 = const()[name = tensor("op_980_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_980_dilations_0 = const()[name = tensor("op_980_dilations_0"), val = tensor([1, 1])]; + tensor var_980_groups_0 = const()[name = tensor("op_980_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138911744)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142188608)))]; + tensor var_980_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_980_dilations_0, groups = var_980_groups_0, pad = var_980_pad_0, pad_type = var_980_pad_type_0, strides = var_980_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_980_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_983_axis_0 = const()[name = tensor("op_983_axis_0"), val = tensor(1)]; + tensor var_983_cast_fp16_0, tensor var_983_cast_fp16_1, tensor var_983_cast_fp16_2, tensor var_983_cast_fp16_3, tensor var_983_cast_fp16_4, tensor var_983_cast_fp16_5, tensor var_983_cast_fp16_6, tensor var_983_cast_fp16_7, tensor var_983_cast_fp16_8, tensor var_983_cast_fp16_9, tensor var_983_cast_fp16_10, tensor var_983_cast_fp16_11, tensor var_983_cast_fp16_12, tensor var_983_cast_fp16_13, tensor var_983_cast_fp16_14, tensor var_983_cast_fp16_15, tensor var_983_cast_fp16_16, tensor var_983_cast_fp16_17, tensor var_983_cast_fp16_18, tensor var_983_cast_fp16_19 = split(axis = var_983_axis_0, split_sizes = tile_9, x = var_982_cast_fp16)[name = tensor("op_983_cast_fp16")]; + tensor var_1004_perm_0 = const()[name = tensor("op_1004_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1005_axis_0 = const()[name = tensor("op_1005_axis_0"), val = tensor(3)]; + tensor var_1004_cast_fp16 = transpose(perm = var_1004_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_29")]; + tensor var_1005_cast_fp16_0, tensor var_1005_cast_fp16_1, tensor var_1005_cast_fp16_2, tensor var_1005_cast_fp16_3, tensor var_1005_cast_fp16_4, tensor var_1005_cast_fp16_5, tensor var_1005_cast_fp16_6, tensor var_1005_cast_fp16_7, tensor var_1005_cast_fp16_8, tensor var_1005_cast_fp16_9, tensor var_1005_cast_fp16_10, tensor var_1005_cast_fp16_11, tensor var_1005_cast_fp16_12, tensor var_1005_cast_fp16_13, tensor var_1005_cast_fp16_14, tensor var_1005_cast_fp16_15, tensor var_1005_cast_fp16_16, tensor var_1005_cast_fp16_17, tensor var_1005_cast_fp16_18, tensor var_1005_cast_fp16_19 = split(axis = var_1005_axis_0, split_sizes = tile_10, x = var_1004_cast_fp16)[name = tensor("op_1005_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1026_axis_0 = const()[name = tensor("op_1026_axis_0"), val = tensor(1)]; + tensor var_1026_cast_fp16_0, tensor var_1026_cast_fp16_1, tensor var_1026_cast_fp16_2, tensor var_1026_cast_fp16_3, tensor var_1026_cast_fp16_4, tensor var_1026_cast_fp16_5, tensor var_1026_cast_fp16_6, tensor var_1026_cast_fp16_7, tensor var_1026_cast_fp16_8, tensor var_1026_cast_fp16_9, tensor var_1026_cast_fp16_10, tensor var_1026_cast_fp16_11, tensor var_1026_cast_fp16_12, tensor var_1026_cast_fp16_13, tensor var_1026_cast_fp16_14, tensor var_1026_cast_fp16_15, tensor var_1026_cast_fp16_16, tensor var_1026_cast_fp16_17, tensor var_1026_cast_fp16_18, tensor var_1026_cast_fp16_19 = split(axis = var_1026_axis_0, split_sizes = tile_11, x = var_980_cast_fp16)[name = tensor("op_1026_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_1005_cast_fp16_0, var_983_cast_fp16_0))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_1005_cast_fp16_1, var_983_cast_fp16_1))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_1005_cast_fp16_2, var_983_cast_fp16_2))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_1005_cast_fp16_3, var_983_cast_fp16_3))[name = tensor("aw_127_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1005_cast_fp16_4, var_983_cast_fp16_4))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1005_cast_fp16_5, var_983_cast_fp16_5))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1005_cast_fp16_6, var_983_cast_fp16_6))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1005_cast_fp16_7, var_983_cast_fp16_7))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1005_cast_fp16_8, var_983_cast_fp16_8))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1005_cast_fp16_9, var_983_cast_fp16_9))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1005_cast_fp16_10, var_983_cast_fp16_10))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1005_cast_fp16_11, var_983_cast_fp16_11))[name = tensor("aw_143_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1005_cast_fp16_12, var_983_cast_fp16_12))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1005_cast_fp16_13, var_983_cast_fp16_13))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1005_cast_fp16_14, var_983_cast_fp16_14))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1005_cast_fp16_15, var_983_cast_fp16_15))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1005_cast_fp16_16, var_983_cast_fp16_16))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1005_cast_fp16_17, var_983_cast_fp16_17))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1005_cast_fp16_18, var_983_cast_fp16_18))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1005_cast_fp16_19, var_983_cast_fp16_19))[name = tensor("aw_159_cast_fp16")]; + tensor var_1087_cast_fp16 = softmax(axis = var_931, x = aw_121_cast_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor var_1088_cast_fp16 = softmax(axis = var_931, x = aw_123_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = softmax(axis = var_931, x = aw_125_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor var_1090_cast_fp16 = softmax(axis = var_931, x = aw_127_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1091_cast_fp16 = softmax(axis = var_931, x = aw_129_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor var_1092_cast_fp16 = softmax(axis = var_931, x = aw_131_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = softmax(axis = var_931, x = aw_133_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor var_1094_cast_fp16 = softmax(axis = var_931, x = aw_135_cast_fp16)[name = tensor("op_1094_cast_fp16")]; + tensor var_1095_cast_fp16 = softmax(axis = var_931, x = aw_137_cast_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = softmax(axis = var_931, x = aw_139_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = softmax(axis = var_931, x = aw_141_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor var_1098_cast_fp16 = softmax(axis = var_931, x = aw_143_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = softmax(axis = var_931, x = aw_145_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor var_1100_cast_fp16 = softmax(axis = var_931, x = aw_147_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = softmax(axis = var_931, x = aw_149_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = softmax(axis = var_931, x = aw_151_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor var_1103_cast_fp16 = softmax(axis = var_931, x = aw_153_cast_fp16)[name = tensor("op_1103_cast_fp16")]; + tensor var_1104_cast_fp16 = softmax(axis = var_931, x = aw_155_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = softmax(axis = var_931, x = aw_157_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = softmax(axis = var_931, x = aw_159_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor var_1108_equation_0 = const()[name = tensor("op_1108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1108_cast_fp16 = einsum(equation = var_1108_equation_0, values = (var_1026_cast_fp16_0, var_1087_cast_fp16))[name = tensor("op_1108_cast_fp16")]; + tensor var_1110_equation_0 = const()[name = tensor("op_1110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1110_cast_fp16 = einsum(equation = var_1110_equation_0, values = (var_1026_cast_fp16_1, var_1088_cast_fp16))[name = tensor("op_1110_cast_fp16")]; + tensor var_1112_equation_0 = const()[name = tensor("op_1112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1112_cast_fp16 = einsum(equation = var_1112_equation_0, values = (var_1026_cast_fp16_2, var_1089_cast_fp16))[name = tensor("op_1112_cast_fp16")]; + tensor var_1114_equation_0 = const()[name = tensor("op_1114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1114_cast_fp16 = einsum(equation = var_1114_equation_0, values = (var_1026_cast_fp16_3, var_1090_cast_fp16))[name = tensor("op_1114_cast_fp16")]; + tensor var_1116_equation_0 = const()[name = tensor("op_1116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1116_cast_fp16 = einsum(equation = var_1116_equation_0, values = (var_1026_cast_fp16_4, var_1091_cast_fp16))[name = tensor("op_1116_cast_fp16")]; + tensor var_1118_equation_0 = const()[name = tensor("op_1118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1118_cast_fp16 = einsum(equation = var_1118_equation_0, values = (var_1026_cast_fp16_5, var_1092_cast_fp16))[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_equation_0 = const()[name = tensor("op_1120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1120_cast_fp16 = einsum(equation = var_1120_equation_0, values = (var_1026_cast_fp16_6, var_1093_cast_fp16))[name = tensor("op_1120_cast_fp16")]; + tensor var_1122_equation_0 = const()[name = tensor("op_1122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1122_cast_fp16 = einsum(equation = var_1122_equation_0, values = (var_1026_cast_fp16_7, var_1094_cast_fp16))[name = tensor("op_1122_cast_fp16")]; + tensor var_1124_equation_0 = const()[name = tensor("op_1124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1124_cast_fp16 = einsum(equation = var_1124_equation_0, values = (var_1026_cast_fp16_8, var_1095_cast_fp16))[name = tensor("op_1124_cast_fp16")]; + tensor var_1126_equation_0 = const()[name = tensor("op_1126_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1126_cast_fp16 = einsum(equation = var_1126_equation_0, values = (var_1026_cast_fp16_9, var_1096_cast_fp16))[name = tensor("op_1126_cast_fp16")]; + tensor var_1128_equation_0 = const()[name = tensor("op_1128_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1128_cast_fp16 = einsum(equation = var_1128_equation_0, values = (var_1026_cast_fp16_10, var_1097_cast_fp16))[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_equation_0 = const()[name = tensor("op_1130_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1130_cast_fp16 = einsum(equation = var_1130_equation_0, values = (var_1026_cast_fp16_11, var_1098_cast_fp16))[name = tensor("op_1130_cast_fp16")]; + tensor var_1132_equation_0 = const()[name = tensor("op_1132_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1132_cast_fp16 = einsum(equation = var_1132_equation_0, values = (var_1026_cast_fp16_12, var_1099_cast_fp16))[name = tensor("op_1132_cast_fp16")]; + tensor var_1134_equation_0 = const()[name = tensor("op_1134_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1134_cast_fp16 = einsum(equation = var_1134_equation_0, values = (var_1026_cast_fp16_13, var_1100_cast_fp16))[name = tensor("op_1134_cast_fp16")]; + tensor var_1136_equation_0 = const()[name = tensor("op_1136_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1136_cast_fp16 = einsum(equation = var_1136_equation_0, values = (var_1026_cast_fp16_14, var_1101_cast_fp16))[name = tensor("op_1136_cast_fp16")]; + tensor var_1138_equation_0 = const()[name = tensor("op_1138_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1138_cast_fp16 = einsum(equation = var_1138_equation_0, values = (var_1026_cast_fp16_15, var_1102_cast_fp16))[name = tensor("op_1138_cast_fp16")]; + tensor var_1140_equation_0 = const()[name = tensor("op_1140_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1140_cast_fp16 = einsum(equation = var_1140_equation_0, values = (var_1026_cast_fp16_16, var_1103_cast_fp16))[name = tensor("op_1140_cast_fp16")]; + tensor var_1142_equation_0 = const()[name = tensor("op_1142_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1142_cast_fp16 = einsum(equation = var_1142_equation_0, values = (var_1026_cast_fp16_17, var_1104_cast_fp16))[name = tensor("op_1142_cast_fp16")]; + tensor var_1144_equation_0 = const()[name = tensor("op_1144_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1144_cast_fp16 = einsum(equation = var_1144_equation_0, values = (var_1026_cast_fp16_18, var_1105_cast_fp16))[name = tensor("op_1144_cast_fp16")]; + tensor var_1146_equation_0 = const()[name = tensor("op_1146_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1146_cast_fp16 = einsum(equation = var_1146_equation_0, values = (var_1026_cast_fp16_19, var_1106_cast_fp16))[name = tensor("op_1146_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_931, interleave = input_35_interleave_0, values = (var_1108_cast_fp16, var_1110_cast_fp16, var_1112_cast_fp16, var_1114_cast_fp16, var_1116_cast_fp16, var_1118_cast_fp16, var_1120_cast_fp16, var_1122_cast_fp16, var_1124_cast_fp16, var_1126_cast_fp16, var_1128_cast_fp16, var_1130_cast_fp16, var_1132_cast_fp16, var_1134_cast_fp16, var_1136_cast_fp16, var_1138_cast_fp16, var_1140_cast_fp16, var_1142_cast_fp16, var_1144_cast_fp16, var_1146_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_1155_pad_type_0 = const()[name = tensor("op_1155_pad_type_0"), val = tensor("valid")]; + tensor var_1155_strides_0 = const()[name = tensor("op_1155_strides_0"), val = tensor([1, 1])]; + tensor var_1155_pad_0 = const()[name = tensor("op_1155_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1155_dilations_0 = const()[name = tensor("op_1155_dilations_0"), val = tensor([1, 1])]; + tensor var_1155_groups_0 = const()[name = tensor("op_1155_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142191232)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145468096)))]; + tensor var_1155_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_1155_dilations_0, groups = var_1155_groups_0, pad = var_1155_pad_0, pad_type = var_1155_pad_type_0, strides = var_1155_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_1155_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145470720)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145473344)))]; + tensor var_1165_to_fp16 = const()[name = tensor("op_1165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_1165_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145475968)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158583232)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1191_pad_type_0 = const()[name = tensor("op_1191_pad_type_0"), val = tensor("valid")]; + tensor var_1191_strides_0 = const()[name = tensor("op_1191_strides_0"), val = tensor([1, 1])]; + tensor var_1191_pad_0 = const()[name = tensor("op_1191_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1191_dilations_0 = const()[name = tensor("op_1191_dilations_0"), val = tensor([1, 1])]; + tensor var_1191_groups_0 = const()[name = tensor("op_1191_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158593536)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171700800)))]; + tensor var_1191_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_1191_dilations_0, groups = var_1191_groups_0, pad = var_1191_pad_0, pad_type = var_1191_pad_type_0, strides = var_1191_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_1191_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171703424)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171706048)))]; + tensor var_1216_to_fp16 = const()[name = tensor("op_1216_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_1216_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_1251_weight_0_to_fp16 = const()[name = tensor("op_1251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171708672)))]; + tensor var_1251_bias_0_to_fp16 = const()[name = tensor("op_1251_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174985536)))]; + tensor var_1251_cast_fp16 = conv(bias = var_1251_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_1251_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1251_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174988160)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_1249_pad_type_0 = const()[name = tensor("op_1249_pad_type_0"), val = tensor("valid")]; + tensor var_1249_strides_0 = const()[name = tensor("op_1249_strides_0"), val = tensor([1, 1])]; + tensor var_1249_pad_0 = const()[name = tensor("op_1249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1249_dilations_0 = const()[name = tensor("op_1249_dilations_0"), val = tensor([1, 1])]; + tensor var_1249_groups_0 = const()[name = tensor("op_1249_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178265024)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181541888)))]; + tensor var_1249_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_1249_dilations_0, groups = var_1249_groups_0, pad = var_1249_pad_0, pad_type = var_1249_pad_type_0, strides = var_1249_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1249_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1252_axis_0 = const()[name = tensor("op_1252_axis_0"), val = tensor(1)]; + tensor var_1252_cast_fp16_0, tensor var_1252_cast_fp16_1, tensor var_1252_cast_fp16_2, tensor var_1252_cast_fp16_3, tensor var_1252_cast_fp16_4, tensor var_1252_cast_fp16_5, tensor var_1252_cast_fp16_6, tensor var_1252_cast_fp16_7, tensor var_1252_cast_fp16_8, tensor var_1252_cast_fp16_9, tensor var_1252_cast_fp16_10, tensor var_1252_cast_fp16_11, tensor var_1252_cast_fp16_12, tensor var_1252_cast_fp16_13, tensor var_1252_cast_fp16_14, tensor var_1252_cast_fp16_15, tensor var_1252_cast_fp16_16, tensor var_1252_cast_fp16_17, tensor var_1252_cast_fp16_18, tensor var_1252_cast_fp16_19 = split(axis = var_1252_axis_0, split_sizes = tile_12, x = var_1251_cast_fp16)[name = tensor("op_1252_cast_fp16")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(3)]; + tensor var_1273_cast_fp16 = transpose(perm = var_1273_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_28")]; + tensor var_1274_cast_fp16_0, tensor var_1274_cast_fp16_1, tensor var_1274_cast_fp16_2, tensor var_1274_cast_fp16_3, tensor var_1274_cast_fp16_4, tensor var_1274_cast_fp16_5, tensor var_1274_cast_fp16_6, tensor var_1274_cast_fp16_7, tensor var_1274_cast_fp16_8, tensor var_1274_cast_fp16_9, tensor var_1274_cast_fp16_10, tensor var_1274_cast_fp16_11, tensor var_1274_cast_fp16_12, tensor var_1274_cast_fp16_13, tensor var_1274_cast_fp16_14, tensor var_1274_cast_fp16_15, tensor var_1274_cast_fp16_16, tensor var_1274_cast_fp16_17, tensor var_1274_cast_fp16_18, tensor var_1274_cast_fp16_19 = split(axis = var_1274_axis_0, split_sizes = tile_13, x = var_1273_cast_fp16)[name = tensor("op_1274_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1295_axis_0 = const()[name = tensor("op_1295_axis_0"), val = tensor(1)]; + tensor var_1295_cast_fp16_0, tensor var_1295_cast_fp16_1, tensor var_1295_cast_fp16_2, tensor var_1295_cast_fp16_3, tensor var_1295_cast_fp16_4, tensor var_1295_cast_fp16_5, tensor var_1295_cast_fp16_6, tensor var_1295_cast_fp16_7, tensor var_1295_cast_fp16_8, tensor var_1295_cast_fp16_9, tensor var_1295_cast_fp16_10, tensor var_1295_cast_fp16_11, tensor var_1295_cast_fp16_12, tensor var_1295_cast_fp16_13, tensor var_1295_cast_fp16_14, tensor var_1295_cast_fp16_15, tensor var_1295_cast_fp16_16, tensor var_1295_cast_fp16_17, tensor var_1295_cast_fp16_18, tensor var_1295_cast_fp16_19 = split(axis = var_1295_axis_0, split_sizes = tile_14, x = var_1249_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1274_cast_fp16_0, var_1252_cast_fp16_0))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1274_cast_fp16_1, var_1252_cast_fp16_1))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1274_cast_fp16_2, var_1252_cast_fp16_2))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1274_cast_fp16_3, var_1252_cast_fp16_3))[name = tensor("aw_167_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1274_cast_fp16_4, var_1252_cast_fp16_4))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1274_cast_fp16_5, var_1252_cast_fp16_5))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1274_cast_fp16_6, var_1252_cast_fp16_6))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1274_cast_fp16_7, var_1252_cast_fp16_7))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1274_cast_fp16_8, var_1252_cast_fp16_8))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1274_cast_fp16_9, var_1252_cast_fp16_9))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1274_cast_fp16_10, var_1252_cast_fp16_10))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1274_cast_fp16_11, var_1252_cast_fp16_11))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1274_cast_fp16_12, var_1252_cast_fp16_12))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1274_cast_fp16_13, var_1252_cast_fp16_13))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1274_cast_fp16_14, var_1252_cast_fp16_14))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1274_cast_fp16_15, var_1252_cast_fp16_15))[name = tensor("aw_191_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1274_cast_fp16_16, var_1252_cast_fp16_16))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1274_cast_fp16_17, var_1252_cast_fp16_17))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1274_cast_fp16_18, var_1252_cast_fp16_18))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1274_cast_fp16_19, var_1252_cast_fp16_19))[name = tensor("aw_199_cast_fp16")]; + tensor var_1356_cast_fp16 = softmax(axis = var_1200, x = aw_161_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor var_1357_cast_fp16 = softmax(axis = var_1200, x = aw_163_cast_fp16)[name = tensor("op_1357_cast_fp16")]; + tensor var_1358_cast_fp16 = softmax(axis = var_1200, x = aw_165_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = softmax(axis = var_1200, x = aw_167_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor var_1360_cast_fp16 = softmax(axis = var_1200, x = aw_169_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor var_1361_cast_fp16 = softmax(axis = var_1200, x = aw_171_cast_fp16)[name = tensor("op_1361_cast_fp16")]; + tensor var_1362_cast_fp16 = softmax(axis = var_1200, x = aw_173_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = softmax(axis = var_1200, x = aw_175_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = softmax(axis = var_1200, x = aw_177_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor var_1365_cast_fp16 = softmax(axis = var_1200, x = aw_179_cast_fp16)[name = tensor("op_1365_cast_fp16")]; + tensor var_1366_cast_fp16 = softmax(axis = var_1200, x = aw_181_cast_fp16)[name = tensor("op_1366_cast_fp16")]; + tensor var_1367_cast_fp16 = softmax(axis = var_1200, x = aw_183_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = softmax(axis = var_1200, x = aw_185_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor var_1369_cast_fp16 = softmax(axis = var_1200, x = aw_187_cast_fp16)[name = tensor("op_1369_cast_fp16")]; + tensor var_1370_cast_fp16 = softmax(axis = var_1200, x = aw_189_cast_fp16)[name = tensor("op_1370_cast_fp16")]; + tensor var_1371_cast_fp16 = softmax(axis = var_1200, x = aw_191_cast_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = softmax(axis = var_1200, x = aw_193_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor var_1373_cast_fp16 = softmax(axis = var_1200, x = aw_195_cast_fp16)[name = tensor("op_1373_cast_fp16")]; + tensor var_1374_cast_fp16 = softmax(axis = var_1200, x = aw_197_cast_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor var_1375_cast_fp16 = softmax(axis = var_1200, x = aw_199_cast_fp16)[name = tensor("op_1375_cast_fp16")]; + tensor var_1377_equation_0 = const()[name = tensor("op_1377_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1377_cast_fp16 = einsum(equation = var_1377_equation_0, values = (var_1295_cast_fp16_0, var_1356_cast_fp16))[name = tensor("op_1377_cast_fp16")]; + tensor var_1379_equation_0 = const()[name = tensor("op_1379_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1379_cast_fp16 = einsum(equation = var_1379_equation_0, values = (var_1295_cast_fp16_1, var_1357_cast_fp16))[name = tensor("op_1379_cast_fp16")]; + tensor var_1381_equation_0 = const()[name = tensor("op_1381_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1381_cast_fp16 = einsum(equation = var_1381_equation_0, values = (var_1295_cast_fp16_2, var_1358_cast_fp16))[name = tensor("op_1381_cast_fp16")]; + tensor var_1383_equation_0 = const()[name = tensor("op_1383_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1383_cast_fp16 = einsum(equation = var_1383_equation_0, values = (var_1295_cast_fp16_3, var_1359_cast_fp16))[name = tensor("op_1383_cast_fp16")]; + tensor var_1385_equation_0 = const()[name = tensor("op_1385_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1385_cast_fp16 = einsum(equation = var_1385_equation_0, values = (var_1295_cast_fp16_4, var_1360_cast_fp16))[name = tensor("op_1385_cast_fp16")]; + tensor var_1387_equation_0 = const()[name = tensor("op_1387_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1387_cast_fp16 = einsum(equation = var_1387_equation_0, values = (var_1295_cast_fp16_5, var_1361_cast_fp16))[name = tensor("op_1387_cast_fp16")]; + tensor var_1389_equation_0 = const()[name = tensor("op_1389_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1389_cast_fp16 = einsum(equation = var_1389_equation_0, values = (var_1295_cast_fp16_6, var_1362_cast_fp16))[name = tensor("op_1389_cast_fp16")]; + tensor var_1391_equation_0 = const()[name = tensor("op_1391_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1391_cast_fp16 = einsum(equation = var_1391_equation_0, values = (var_1295_cast_fp16_7, var_1363_cast_fp16))[name = tensor("op_1391_cast_fp16")]; + tensor var_1393_equation_0 = const()[name = tensor("op_1393_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1393_cast_fp16 = einsum(equation = var_1393_equation_0, values = (var_1295_cast_fp16_8, var_1364_cast_fp16))[name = tensor("op_1393_cast_fp16")]; + tensor var_1395_equation_0 = const()[name = tensor("op_1395_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1395_cast_fp16 = einsum(equation = var_1395_equation_0, values = (var_1295_cast_fp16_9, var_1365_cast_fp16))[name = tensor("op_1395_cast_fp16")]; + tensor var_1397_equation_0 = const()[name = tensor("op_1397_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1397_cast_fp16 = einsum(equation = var_1397_equation_0, values = (var_1295_cast_fp16_10, var_1366_cast_fp16))[name = tensor("op_1397_cast_fp16")]; + tensor var_1399_equation_0 = const()[name = tensor("op_1399_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1399_cast_fp16 = einsum(equation = var_1399_equation_0, values = (var_1295_cast_fp16_11, var_1367_cast_fp16))[name = tensor("op_1399_cast_fp16")]; + tensor var_1401_equation_0 = const()[name = tensor("op_1401_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1401_cast_fp16 = einsum(equation = var_1401_equation_0, values = (var_1295_cast_fp16_12, var_1368_cast_fp16))[name = tensor("op_1401_cast_fp16")]; + tensor var_1403_equation_0 = const()[name = tensor("op_1403_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1403_cast_fp16 = einsum(equation = var_1403_equation_0, values = (var_1295_cast_fp16_13, var_1369_cast_fp16))[name = tensor("op_1403_cast_fp16")]; + tensor var_1405_equation_0 = const()[name = tensor("op_1405_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1405_cast_fp16 = einsum(equation = var_1405_equation_0, values = (var_1295_cast_fp16_14, var_1370_cast_fp16))[name = tensor("op_1405_cast_fp16")]; + tensor var_1407_equation_0 = const()[name = tensor("op_1407_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1407_cast_fp16 = einsum(equation = var_1407_equation_0, values = (var_1295_cast_fp16_15, var_1371_cast_fp16))[name = tensor("op_1407_cast_fp16")]; + tensor var_1409_equation_0 = const()[name = tensor("op_1409_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1409_cast_fp16 = einsum(equation = var_1409_equation_0, values = (var_1295_cast_fp16_16, var_1372_cast_fp16))[name = tensor("op_1409_cast_fp16")]; + tensor var_1411_equation_0 = const()[name = tensor("op_1411_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1411_cast_fp16 = einsum(equation = var_1411_equation_0, values = (var_1295_cast_fp16_17, var_1373_cast_fp16))[name = tensor("op_1411_cast_fp16")]; + tensor var_1413_equation_0 = const()[name = tensor("op_1413_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1413_cast_fp16 = einsum(equation = var_1413_equation_0, values = (var_1295_cast_fp16_18, var_1374_cast_fp16))[name = tensor("op_1413_cast_fp16")]; + tensor var_1415_equation_0 = const()[name = tensor("op_1415_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1415_cast_fp16 = einsum(equation = var_1415_equation_0, values = (var_1295_cast_fp16_19, var_1375_cast_fp16))[name = tensor("op_1415_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_1200, interleave = input_45_interleave_0, values = (var_1377_cast_fp16, var_1379_cast_fp16, var_1381_cast_fp16, var_1383_cast_fp16, var_1385_cast_fp16, var_1387_cast_fp16, var_1389_cast_fp16, var_1391_cast_fp16, var_1393_cast_fp16, var_1395_cast_fp16, var_1397_cast_fp16, var_1399_cast_fp16, var_1401_cast_fp16, var_1403_cast_fp16, var_1405_cast_fp16, var_1407_cast_fp16, var_1409_cast_fp16, var_1411_cast_fp16, var_1413_cast_fp16, var_1415_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1424_pad_type_0 = const()[name = tensor("op_1424_pad_type_0"), val = tensor("valid")]; + tensor var_1424_strides_0 = const()[name = tensor("op_1424_strides_0"), val = tensor([1, 1])]; + tensor var_1424_pad_0 = const()[name = tensor("op_1424_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1424_dilations_0 = const()[name = tensor("op_1424_dilations_0"), val = tensor([1, 1])]; + tensor var_1424_groups_0 = const()[name = tensor("op_1424_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181544512)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184821376)))]; + tensor var_1424_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1424_dilations_0, groups = var_1424_groups_0, pad = var_1424_pad_0, pad_type = var_1424_pad_type_0, strides = var_1424_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1424_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184824000)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184826624)))]; + tensor var_1434_to_fp16 = const()[name = tensor("op_1434_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1434_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184829248)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197936512)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1460_pad_type_0 = const()[name = tensor("op_1460_pad_type_0"), val = tensor("valid")]; + tensor var_1460_strides_0 = const()[name = tensor("op_1460_strides_0"), val = tensor([1, 1])]; + tensor var_1460_pad_0 = const()[name = tensor("op_1460_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1460_dilations_0 = const()[name = tensor("op_1460_dilations_0"), val = tensor([1, 1])]; + tensor var_1460_groups_0 = const()[name = tensor("op_1460_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197946816)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211054080)))]; + tensor var_1460_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1460_dilations_0, groups = var_1460_groups_0, pad = var_1460_pad_0, pad_type = var_1460_pad_type_0, strides = var_1460_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1460_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1469 = const()[name = tensor("op_1469"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211056704)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211059328)))]; + tensor var_1485_to_fp16 = const()[name = tensor("op_1485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1485_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1520_weight_0_to_fp16 = const()[name = tensor("op_1520_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211061952)))]; + tensor var_1520_bias_0_to_fp16 = const()[name = tensor("op_1520_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214338816)))]; + tensor var_1520_cast_fp16 = conv(bias = var_1520_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1520_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214341440)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1518_pad_type_0 = const()[name = tensor("op_1518_pad_type_0"), val = tensor("valid")]; + tensor var_1518_strides_0 = const()[name = tensor("op_1518_strides_0"), val = tensor([1, 1])]; + tensor var_1518_pad_0 = const()[name = tensor("op_1518_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1518_dilations_0 = const()[name = tensor("op_1518_dilations_0"), val = tensor([1, 1])]; + tensor var_1518_groups_0 = const()[name = tensor("op_1518_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217618304)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220895168)))]; + tensor var_1518_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1518_dilations_0, groups = var_1518_groups_0, pad = var_1518_pad_0, pad_type = var_1518_pad_type_0, strides = var_1518_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1518_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1521_axis_0 = const()[name = tensor("op_1521_axis_0"), val = tensor(1)]; + tensor var_1521_cast_fp16_0, tensor var_1521_cast_fp16_1, tensor var_1521_cast_fp16_2, tensor var_1521_cast_fp16_3, tensor var_1521_cast_fp16_4, tensor var_1521_cast_fp16_5, tensor var_1521_cast_fp16_6, tensor var_1521_cast_fp16_7, tensor var_1521_cast_fp16_8, tensor var_1521_cast_fp16_9, tensor var_1521_cast_fp16_10, tensor var_1521_cast_fp16_11, tensor var_1521_cast_fp16_12, tensor var_1521_cast_fp16_13, tensor var_1521_cast_fp16_14, tensor var_1521_cast_fp16_15, tensor var_1521_cast_fp16_16, tensor var_1521_cast_fp16_17, tensor var_1521_cast_fp16_18, tensor var_1521_cast_fp16_19 = split(axis = var_1521_axis_0, split_sizes = tile_15, x = var_1520_cast_fp16)[name = tensor("op_1521_cast_fp16")]; + tensor var_1542_perm_0 = const()[name = tensor("op_1542_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1543_axis_0 = const()[name = tensor("op_1543_axis_0"), val = tensor(3)]; + tensor var_1542_cast_fp16 = transpose(perm = var_1542_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1543_cast_fp16_0, tensor var_1543_cast_fp16_1, tensor var_1543_cast_fp16_2, tensor var_1543_cast_fp16_3, tensor var_1543_cast_fp16_4, tensor var_1543_cast_fp16_5, tensor var_1543_cast_fp16_6, tensor var_1543_cast_fp16_7, tensor var_1543_cast_fp16_8, tensor var_1543_cast_fp16_9, tensor var_1543_cast_fp16_10, tensor var_1543_cast_fp16_11, tensor var_1543_cast_fp16_12, tensor var_1543_cast_fp16_13, tensor var_1543_cast_fp16_14, tensor var_1543_cast_fp16_15, tensor var_1543_cast_fp16_16, tensor var_1543_cast_fp16_17, tensor var_1543_cast_fp16_18, tensor var_1543_cast_fp16_19 = split(axis = var_1543_axis_0, split_sizes = tile_16, x = var_1542_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1564_axis_0 = const()[name = tensor("op_1564_axis_0"), val = tensor(1)]; + tensor var_1564_cast_fp16_0, tensor var_1564_cast_fp16_1, tensor var_1564_cast_fp16_2, tensor var_1564_cast_fp16_3, tensor var_1564_cast_fp16_4, tensor var_1564_cast_fp16_5, tensor var_1564_cast_fp16_6, tensor var_1564_cast_fp16_7, tensor var_1564_cast_fp16_8, tensor var_1564_cast_fp16_9, tensor var_1564_cast_fp16_10, tensor var_1564_cast_fp16_11, tensor var_1564_cast_fp16_12, tensor var_1564_cast_fp16_13, tensor var_1564_cast_fp16_14, tensor var_1564_cast_fp16_15, tensor var_1564_cast_fp16_16, tensor var_1564_cast_fp16_17, tensor var_1564_cast_fp16_18, tensor var_1564_cast_fp16_19 = split(axis = var_1564_axis_0, split_sizes = tile_17, x = var_1518_cast_fp16)[name = tensor("op_1564_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1543_cast_fp16_0, var_1521_cast_fp16_0))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1543_cast_fp16_1, var_1521_cast_fp16_1))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1543_cast_fp16_2, var_1521_cast_fp16_2))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1543_cast_fp16_3, var_1521_cast_fp16_3))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1543_cast_fp16_4, var_1521_cast_fp16_4))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1543_cast_fp16_5, var_1521_cast_fp16_5))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1543_cast_fp16_6, var_1521_cast_fp16_6))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1543_cast_fp16_7, var_1521_cast_fp16_7))[name = tensor("aw_215_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1543_cast_fp16_8, var_1521_cast_fp16_8))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1543_cast_fp16_9, var_1521_cast_fp16_9))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1543_cast_fp16_10, var_1521_cast_fp16_10))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1543_cast_fp16_11, var_1521_cast_fp16_11))[name = tensor("aw_223_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1543_cast_fp16_12, var_1521_cast_fp16_12))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1543_cast_fp16_13, var_1521_cast_fp16_13))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1543_cast_fp16_14, var_1521_cast_fp16_14))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1543_cast_fp16_15, var_1521_cast_fp16_15))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1543_cast_fp16_16, var_1521_cast_fp16_16))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1543_cast_fp16_17, var_1521_cast_fp16_17))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1543_cast_fp16_18, var_1521_cast_fp16_18))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1543_cast_fp16_19, var_1521_cast_fp16_19))[name = tensor("aw_239_cast_fp16")]; + tensor var_1625_cast_fp16 = softmax(axis = var_1469, x = aw_201_cast_fp16)[name = tensor("op_1625_cast_fp16")]; + tensor var_1626_cast_fp16 = softmax(axis = var_1469, x = aw_203_cast_fp16)[name = tensor("op_1626_cast_fp16")]; + tensor var_1627_cast_fp16 = softmax(axis = var_1469, x = aw_205_cast_fp16)[name = tensor("op_1627_cast_fp16")]; + tensor var_1628_cast_fp16 = softmax(axis = var_1469, x = aw_207_cast_fp16)[name = tensor("op_1628_cast_fp16")]; + tensor var_1629_cast_fp16 = softmax(axis = var_1469, x = aw_209_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1630_cast_fp16 = softmax(axis = var_1469, x = aw_211_cast_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = softmax(axis = var_1469, x = aw_213_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_cast_fp16 = softmax(axis = var_1469, x = aw_215_cast_fp16)[name = tensor("op_1632_cast_fp16")]; + tensor var_1633_cast_fp16 = softmax(axis = var_1469, x = aw_217_cast_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor var_1634_cast_fp16 = softmax(axis = var_1469, x = aw_219_cast_fp16)[name = tensor("op_1634_cast_fp16")]; + tensor var_1635_cast_fp16 = softmax(axis = var_1469, x = aw_221_cast_fp16)[name = tensor("op_1635_cast_fp16")]; + tensor var_1636_cast_fp16 = softmax(axis = var_1469, x = aw_223_cast_fp16)[name = tensor("op_1636_cast_fp16")]; + tensor var_1637_cast_fp16 = softmax(axis = var_1469, x = aw_225_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor var_1638_cast_fp16 = softmax(axis = var_1469, x = aw_227_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_cast_fp16 = softmax(axis = var_1469, x = aw_229_cast_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1640_cast_fp16 = softmax(axis = var_1469, x = aw_231_cast_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor var_1641_cast_fp16 = softmax(axis = var_1469, x = aw_233_cast_fp16)[name = tensor("op_1641_cast_fp16")]; + tensor var_1642_cast_fp16 = softmax(axis = var_1469, x = aw_235_cast_fp16)[name = tensor("op_1642_cast_fp16")]; + tensor var_1643_cast_fp16 = softmax(axis = var_1469, x = aw_237_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor var_1644_cast_fp16 = softmax(axis = var_1469, x = aw_239_cast_fp16)[name = tensor("op_1644_cast_fp16")]; + tensor var_1646_equation_0 = const()[name = tensor("op_1646_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1646_cast_fp16 = einsum(equation = var_1646_equation_0, values = (var_1564_cast_fp16_0, var_1625_cast_fp16))[name = tensor("op_1646_cast_fp16")]; + tensor var_1648_equation_0 = const()[name = tensor("op_1648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1648_cast_fp16 = einsum(equation = var_1648_equation_0, values = (var_1564_cast_fp16_1, var_1626_cast_fp16))[name = tensor("op_1648_cast_fp16")]; + tensor var_1650_equation_0 = const()[name = tensor("op_1650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1650_cast_fp16 = einsum(equation = var_1650_equation_0, values = (var_1564_cast_fp16_2, var_1627_cast_fp16))[name = tensor("op_1650_cast_fp16")]; + tensor var_1652_equation_0 = const()[name = tensor("op_1652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1652_cast_fp16 = einsum(equation = var_1652_equation_0, values = (var_1564_cast_fp16_3, var_1628_cast_fp16))[name = tensor("op_1652_cast_fp16")]; + tensor var_1654_equation_0 = const()[name = tensor("op_1654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1654_cast_fp16 = einsum(equation = var_1654_equation_0, values = (var_1564_cast_fp16_4, var_1629_cast_fp16))[name = tensor("op_1654_cast_fp16")]; + tensor var_1656_equation_0 = const()[name = tensor("op_1656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1656_cast_fp16 = einsum(equation = var_1656_equation_0, values = (var_1564_cast_fp16_5, var_1630_cast_fp16))[name = tensor("op_1656_cast_fp16")]; + tensor var_1658_equation_0 = const()[name = tensor("op_1658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1658_cast_fp16 = einsum(equation = var_1658_equation_0, values = (var_1564_cast_fp16_6, var_1631_cast_fp16))[name = tensor("op_1658_cast_fp16")]; + tensor var_1660_equation_0 = const()[name = tensor("op_1660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1660_cast_fp16 = einsum(equation = var_1660_equation_0, values = (var_1564_cast_fp16_7, var_1632_cast_fp16))[name = tensor("op_1660_cast_fp16")]; + tensor var_1662_equation_0 = const()[name = tensor("op_1662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1662_cast_fp16 = einsum(equation = var_1662_equation_0, values = (var_1564_cast_fp16_8, var_1633_cast_fp16))[name = tensor("op_1662_cast_fp16")]; + tensor var_1664_equation_0 = const()[name = tensor("op_1664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1664_cast_fp16 = einsum(equation = var_1664_equation_0, values = (var_1564_cast_fp16_9, var_1634_cast_fp16))[name = tensor("op_1664_cast_fp16")]; + tensor var_1666_equation_0 = const()[name = tensor("op_1666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1666_cast_fp16 = einsum(equation = var_1666_equation_0, values = (var_1564_cast_fp16_10, var_1635_cast_fp16))[name = tensor("op_1666_cast_fp16")]; + tensor var_1668_equation_0 = const()[name = tensor("op_1668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1668_cast_fp16 = einsum(equation = var_1668_equation_0, values = (var_1564_cast_fp16_11, var_1636_cast_fp16))[name = tensor("op_1668_cast_fp16")]; + tensor var_1670_equation_0 = const()[name = tensor("op_1670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1670_cast_fp16 = einsum(equation = var_1670_equation_0, values = (var_1564_cast_fp16_12, var_1637_cast_fp16))[name = tensor("op_1670_cast_fp16")]; + tensor var_1672_equation_0 = const()[name = tensor("op_1672_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1672_cast_fp16 = einsum(equation = var_1672_equation_0, values = (var_1564_cast_fp16_13, var_1638_cast_fp16))[name = tensor("op_1672_cast_fp16")]; + tensor var_1674_equation_0 = const()[name = tensor("op_1674_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1674_cast_fp16 = einsum(equation = var_1674_equation_0, values = (var_1564_cast_fp16_14, var_1639_cast_fp16))[name = tensor("op_1674_cast_fp16")]; + tensor var_1676_equation_0 = const()[name = tensor("op_1676_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1676_cast_fp16 = einsum(equation = var_1676_equation_0, values = (var_1564_cast_fp16_15, var_1640_cast_fp16))[name = tensor("op_1676_cast_fp16")]; + tensor var_1678_equation_0 = const()[name = tensor("op_1678_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1678_cast_fp16 = einsum(equation = var_1678_equation_0, values = (var_1564_cast_fp16_16, var_1641_cast_fp16))[name = tensor("op_1678_cast_fp16")]; + tensor var_1680_equation_0 = const()[name = tensor("op_1680_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1680_cast_fp16 = einsum(equation = var_1680_equation_0, values = (var_1564_cast_fp16_17, var_1642_cast_fp16))[name = tensor("op_1680_cast_fp16")]; + tensor var_1682_equation_0 = const()[name = tensor("op_1682_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1682_cast_fp16 = einsum(equation = var_1682_equation_0, values = (var_1564_cast_fp16_18, var_1643_cast_fp16))[name = tensor("op_1682_cast_fp16")]; + tensor var_1684_equation_0 = const()[name = tensor("op_1684_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1684_cast_fp16 = einsum(equation = var_1684_equation_0, values = (var_1564_cast_fp16_19, var_1644_cast_fp16))[name = tensor("op_1684_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1469, interleave = input_55_interleave_0, values = (var_1646_cast_fp16, var_1648_cast_fp16, var_1650_cast_fp16, var_1652_cast_fp16, var_1654_cast_fp16, var_1656_cast_fp16, var_1658_cast_fp16, var_1660_cast_fp16, var_1662_cast_fp16, var_1664_cast_fp16, var_1666_cast_fp16, var_1668_cast_fp16, var_1670_cast_fp16, var_1672_cast_fp16, var_1674_cast_fp16, var_1676_cast_fp16, var_1678_cast_fp16, var_1680_cast_fp16, var_1682_cast_fp16, var_1684_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1693_pad_type_0 = const()[name = tensor("op_1693_pad_type_0"), val = tensor("valid")]; + tensor var_1693_strides_0 = const()[name = tensor("op_1693_strides_0"), val = tensor([1, 1])]; + tensor var_1693_pad_0 = const()[name = tensor("op_1693_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1693_dilations_0 = const()[name = tensor("op_1693_dilations_0"), val = tensor([1, 1])]; + tensor var_1693_groups_0 = const()[name = tensor("op_1693_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220897792)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224174656)))]; + tensor var_1693_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1693_dilations_0, groups = var_1693_groups_0, pad = var_1693_pad_0, pad_type = var_1693_pad_type_0, strides = var_1693_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1693_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1693_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224177280)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224179904)))]; + tensor var_1703_to_fp16 = const()[name = tensor("op_1703_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1703_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224182528)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237289792)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1729_pad_type_0 = const()[name = tensor("op_1729_pad_type_0"), val = tensor("valid")]; + tensor var_1729_strides_0 = const()[name = tensor("op_1729_strides_0"), val = tensor([1, 1])]; + tensor var_1729_pad_0 = const()[name = tensor("op_1729_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1729_dilations_0 = const()[name = tensor("op_1729_dilations_0"), val = tensor([1, 1])]; + tensor var_1729_groups_0 = const()[name = tensor("op_1729_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237300096)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250407360)))]; + tensor var_1729_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1729_dilations_0, groups = var_1729_groups_0, pad = var_1729_pad_0, pad_type = var_1729_pad_type_0, strides = var_1729_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1729_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1729_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1738 = const()[name = tensor("op_1738"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250409984)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250412608)))]; + tensor var_1754_to_fp16 = const()[name = tensor("op_1754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1754_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1789_weight_0_to_fp16 = const()[name = tensor("op_1789_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250415232)))]; + tensor var_1789_bias_0_to_fp16 = const()[name = tensor("op_1789_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253692096)))]; + tensor var_1789_cast_fp16 = conv(bias = var_1789_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1789_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1789_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253694720)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1787_pad_type_0 = const()[name = tensor("op_1787_pad_type_0"), val = tensor("valid")]; + tensor var_1787_strides_0 = const()[name = tensor("op_1787_strides_0"), val = tensor([1, 1])]; + tensor var_1787_pad_0 = const()[name = tensor("op_1787_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1787_dilations_0 = const()[name = tensor("op_1787_dilations_0"), val = tensor([1, 1])]; + tensor var_1787_groups_0 = const()[name = tensor("op_1787_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256971584)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260248448)))]; + tensor var_1787_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1787_dilations_0, groups = var_1787_groups_0, pad = var_1787_pad_0, pad_type = var_1787_pad_type_0, strides = var_1787_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1790_axis_0 = const()[name = tensor("op_1790_axis_0"), val = tensor(1)]; + tensor var_1790_cast_fp16_0, tensor var_1790_cast_fp16_1, tensor var_1790_cast_fp16_2, tensor var_1790_cast_fp16_3, tensor var_1790_cast_fp16_4, tensor var_1790_cast_fp16_5, tensor var_1790_cast_fp16_6, tensor var_1790_cast_fp16_7, tensor var_1790_cast_fp16_8, tensor var_1790_cast_fp16_9, tensor var_1790_cast_fp16_10, tensor var_1790_cast_fp16_11, tensor var_1790_cast_fp16_12, tensor var_1790_cast_fp16_13, tensor var_1790_cast_fp16_14, tensor var_1790_cast_fp16_15, tensor var_1790_cast_fp16_16, tensor var_1790_cast_fp16_17, tensor var_1790_cast_fp16_18, tensor var_1790_cast_fp16_19 = split(axis = var_1790_axis_0, split_sizes = tile_18, x = var_1789_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1811_perm_0 = const()[name = tensor("op_1811_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1812_axis_0 = const()[name = tensor("op_1812_axis_0"), val = tensor(3)]; + tensor var_1811_cast_fp16 = transpose(perm = var_1811_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_26")]; + tensor var_1812_cast_fp16_0, tensor var_1812_cast_fp16_1, tensor var_1812_cast_fp16_2, tensor var_1812_cast_fp16_3, tensor var_1812_cast_fp16_4, tensor var_1812_cast_fp16_5, tensor var_1812_cast_fp16_6, tensor var_1812_cast_fp16_7, tensor var_1812_cast_fp16_8, tensor var_1812_cast_fp16_9, tensor var_1812_cast_fp16_10, tensor var_1812_cast_fp16_11, tensor var_1812_cast_fp16_12, tensor var_1812_cast_fp16_13, tensor var_1812_cast_fp16_14, tensor var_1812_cast_fp16_15, tensor var_1812_cast_fp16_16, tensor var_1812_cast_fp16_17, tensor var_1812_cast_fp16_18, tensor var_1812_cast_fp16_19 = split(axis = var_1812_axis_0, split_sizes = tile_19, x = var_1811_cast_fp16)[name = tensor("op_1812_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1833_axis_0 = const()[name = tensor("op_1833_axis_0"), val = tensor(1)]; + tensor var_1833_cast_fp16_0, tensor var_1833_cast_fp16_1, tensor var_1833_cast_fp16_2, tensor var_1833_cast_fp16_3, tensor var_1833_cast_fp16_4, tensor var_1833_cast_fp16_5, tensor var_1833_cast_fp16_6, tensor var_1833_cast_fp16_7, tensor var_1833_cast_fp16_8, tensor var_1833_cast_fp16_9, tensor var_1833_cast_fp16_10, tensor var_1833_cast_fp16_11, tensor var_1833_cast_fp16_12, tensor var_1833_cast_fp16_13, tensor var_1833_cast_fp16_14, tensor var_1833_cast_fp16_15, tensor var_1833_cast_fp16_16, tensor var_1833_cast_fp16_17, tensor var_1833_cast_fp16_18, tensor var_1833_cast_fp16_19 = split(axis = var_1833_axis_0, split_sizes = tile_20, x = var_1787_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_1812_cast_fp16_0, var_1790_cast_fp16_0))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_1812_cast_fp16_1, var_1790_cast_fp16_1))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_1812_cast_fp16_2, var_1790_cast_fp16_2))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_1812_cast_fp16_3, var_1790_cast_fp16_3))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_1812_cast_fp16_4, var_1790_cast_fp16_4))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_1812_cast_fp16_5, var_1790_cast_fp16_5))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_1812_cast_fp16_6, var_1790_cast_fp16_6))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_1812_cast_fp16_7, var_1790_cast_fp16_7))[name = tensor("aw_255_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_1812_cast_fp16_8, var_1790_cast_fp16_8))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_1812_cast_fp16_9, var_1790_cast_fp16_9))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_1812_cast_fp16_10, var_1790_cast_fp16_10))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_1812_cast_fp16_11, var_1790_cast_fp16_11))[name = tensor("aw_263_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_1812_cast_fp16_12, var_1790_cast_fp16_12))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_1812_cast_fp16_13, var_1790_cast_fp16_13))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_1812_cast_fp16_14, var_1790_cast_fp16_14))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_1812_cast_fp16_15, var_1790_cast_fp16_15))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_1812_cast_fp16_16, var_1790_cast_fp16_16))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_1812_cast_fp16_17, var_1790_cast_fp16_17))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_1812_cast_fp16_18, var_1790_cast_fp16_18))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_1812_cast_fp16_19, var_1790_cast_fp16_19))[name = tensor("aw_279_cast_fp16")]; + tensor var_1894_cast_fp16 = softmax(axis = var_1738, x = aw_241_cast_fp16)[name = tensor("op_1894_cast_fp16")]; + tensor var_1895_cast_fp16 = softmax(axis = var_1738, x = aw_243_cast_fp16)[name = tensor("op_1895_cast_fp16")]; + tensor var_1896_cast_fp16 = softmax(axis = var_1738, x = aw_245_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1897_cast_fp16 = softmax(axis = var_1738, x = aw_247_cast_fp16)[name = tensor("op_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = softmax(axis = var_1738, x = aw_249_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_cast_fp16 = softmax(axis = var_1738, x = aw_251_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor var_1900_cast_fp16 = softmax(axis = var_1738, x = aw_253_cast_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor var_1901_cast_fp16 = softmax(axis = var_1738, x = aw_255_cast_fp16)[name = tensor("op_1901_cast_fp16")]; + tensor var_1902_cast_fp16 = softmax(axis = var_1738, x = aw_257_cast_fp16)[name = tensor("op_1902_cast_fp16")]; + tensor var_1903_cast_fp16 = softmax(axis = var_1738, x = aw_259_cast_fp16)[name = tensor("op_1903_cast_fp16")]; + tensor var_1904_cast_fp16 = softmax(axis = var_1738, x = aw_261_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1905_cast_fp16 = softmax(axis = var_1738, x = aw_263_cast_fp16)[name = tensor("op_1905_cast_fp16")]; + tensor var_1906_cast_fp16 = softmax(axis = var_1738, x = aw_265_cast_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1907_cast_fp16 = softmax(axis = var_1738, x = aw_267_cast_fp16)[name = tensor("op_1907_cast_fp16")]; + tensor var_1908_cast_fp16 = softmax(axis = var_1738, x = aw_269_cast_fp16)[name = tensor("op_1908_cast_fp16")]; + tensor var_1909_cast_fp16 = softmax(axis = var_1738, x = aw_271_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor var_1910_cast_fp16 = softmax(axis = var_1738, x = aw_273_cast_fp16)[name = tensor("op_1910_cast_fp16")]; + tensor var_1911_cast_fp16 = softmax(axis = var_1738, x = aw_275_cast_fp16)[name = tensor("op_1911_cast_fp16")]; + tensor var_1912_cast_fp16 = softmax(axis = var_1738, x = aw_277_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1913_cast_fp16 = softmax(axis = var_1738, x = aw_279_cast_fp16)[name = tensor("op_1913_cast_fp16")]; + tensor var_1915_equation_0 = const()[name = tensor("op_1915_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1915_cast_fp16 = einsum(equation = var_1915_equation_0, values = (var_1833_cast_fp16_0, var_1894_cast_fp16))[name = tensor("op_1915_cast_fp16")]; + tensor var_1917_equation_0 = const()[name = tensor("op_1917_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1917_cast_fp16 = einsum(equation = var_1917_equation_0, values = (var_1833_cast_fp16_1, var_1895_cast_fp16))[name = tensor("op_1917_cast_fp16")]; + tensor var_1919_equation_0 = const()[name = tensor("op_1919_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1919_cast_fp16 = einsum(equation = var_1919_equation_0, values = (var_1833_cast_fp16_2, var_1896_cast_fp16))[name = tensor("op_1919_cast_fp16")]; + tensor var_1921_equation_0 = const()[name = tensor("op_1921_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1921_cast_fp16 = einsum(equation = var_1921_equation_0, values = (var_1833_cast_fp16_3, var_1897_cast_fp16))[name = tensor("op_1921_cast_fp16")]; + tensor var_1923_equation_0 = const()[name = tensor("op_1923_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1923_cast_fp16 = einsum(equation = var_1923_equation_0, values = (var_1833_cast_fp16_4, var_1898_cast_fp16))[name = tensor("op_1923_cast_fp16")]; + tensor var_1925_equation_0 = const()[name = tensor("op_1925_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1925_cast_fp16 = einsum(equation = var_1925_equation_0, values = (var_1833_cast_fp16_5, var_1899_cast_fp16))[name = tensor("op_1925_cast_fp16")]; + tensor var_1927_equation_0 = const()[name = tensor("op_1927_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1927_cast_fp16 = einsum(equation = var_1927_equation_0, values = (var_1833_cast_fp16_6, var_1900_cast_fp16))[name = tensor("op_1927_cast_fp16")]; + tensor var_1929_equation_0 = const()[name = tensor("op_1929_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1929_cast_fp16 = einsum(equation = var_1929_equation_0, values = (var_1833_cast_fp16_7, var_1901_cast_fp16))[name = tensor("op_1929_cast_fp16")]; + tensor var_1931_equation_0 = const()[name = tensor("op_1931_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1931_cast_fp16 = einsum(equation = var_1931_equation_0, values = (var_1833_cast_fp16_8, var_1902_cast_fp16))[name = tensor("op_1931_cast_fp16")]; + tensor var_1933_equation_0 = const()[name = tensor("op_1933_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1933_cast_fp16 = einsum(equation = var_1933_equation_0, values = (var_1833_cast_fp16_9, var_1903_cast_fp16))[name = tensor("op_1933_cast_fp16")]; + tensor var_1935_equation_0 = const()[name = tensor("op_1935_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1935_cast_fp16 = einsum(equation = var_1935_equation_0, values = (var_1833_cast_fp16_10, var_1904_cast_fp16))[name = tensor("op_1935_cast_fp16")]; + tensor var_1937_equation_0 = const()[name = tensor("op_1937_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1937_cast_fp16 = einsum(equation = var_1937_equation_0, values = (var_1833_cast_fp16_11, var_1905_cast_fp16))[name = tensor("op_1937_cast_fp16")]; + tensor var_1939_equation_0 = const()[name = tensor("op_1939_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1939_cast_fp16 = einsum(equation = var_1939_equation_0, values = (var_1833_cast_fp16_12, var_1906_cast_fp16))[name = tensor("op_1939_cast_fp16")]; + tensor var_1941_equation_0 = const()[name = tensor("op_1941_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1941_cast_fp16 = einsum(equation = var_1941_equation_0, values = (var_1833_cast_fp16_13, var_1907_cast_fp16))[name = tensor("op_1941_cast_fp16")]; + tensor var_1943_equation_0 = const()[name = tensor("op_1943_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1943_cast_fp16 = einsum(equation = var_1943_equation_0, values = (var_1833_cast_fp16_14, var_1908_cast_fp16))[name = tensor("op_1943_cast_fp16")]; + tensor var_1945_equation_0 = const()[name = tensor("op_1945_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1945_cast_fp16 = einsum(equation = var_1945_equation_0, values = (var_1833_cast_fp16_15, var_1909_cast_fp16))[name = tensor("op_1945_cast_fp16")]; + tensor var_1947_equation_0 = const()[name = tensor("op_1947_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1947_cast_fp16 = einsum(equation = var_1947_equation_0, values = (var_1833_cast_fp16_16, var_1910_cast_fp16))[name = tensor("op_1947_cast_fp16")]; + tensor var_1949_equation_0 = const()[name = tensor("op_1949_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1949_cast_fp16 = einsum(equation = var_1949_equation_0, values = (var_1833_cast_fp16_17, var_1911_cast_fp16))[name = tensor("op_1949_cast_fp16")]; + tensor var_1951_equation_0 = const()[name = tensor("op_1951_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1951_cast_fp16 = einsum(equation = var_1951_equation_0, values = (var_1833_cast_fp16_18, var_1912_cast_fp16))[name = tensor("op_1951_cast_fp16")]; + tensor var_1953_equation_0 = const()[name = tensor("op_1953_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1953_cast_fp16 = einsum(equation = var_1953_equation_0, values = (var_1833_cast_fp16_19, var_1913_cast_fp16))[name = tensor("op_1953_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1738, interleave = input_65_interleave_0, values = (var_1915_cast_fp16, var_1917_cast_fp16, var_1919_cast_fp16, var_1921_cast_fp16, var_1923_cast_fp16, var_1925_cast_fp16, var_1927_cast_fp16, var_1929_cast_fp16, var_1931_cast_fp16, var_1933_cast_fp16, var_1935_cast_fp16, var_1937_cast_fp16, var_1939_cast_fp16, var_1941_cast_fp16, var_1943_cast_fp16, var_1945_cast_fp16, var_1947_cast_fp16, var_1949_cast_fp16, var_1951_cast_fp16, var_1953_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1962_pad_type_0 = const()[name = tensor("op_1962_pad_type_0"), val = tensor("valid")]; + tensor var_1962_strides_0 = const()[name = tensor("op_1962_strides_0"), val = tensor([1, 1])]; + tensor var_1962_pad_0 = const()[name = tensor("op_1962_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1962_dilations_0 = const()[name = tensor("op_1962_dilations_0"), val = tensor([1, 1])]; + tensor var_1962_groups_0 = const()[name = tensor("op_1962_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260251072)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263527936)))]; + tensor var_1962_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1962_dilations_0, groups = var_1962_groups_0, pad = var_1962_pad_0, pad_type = var_1962_pad_type_0, strides = var_1962_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1962_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263530560)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263533184)))]; + tensor var_1972_to_fp16 = const()[name = tensor("op_1972_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1972_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263535808)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276643072)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1998_pad_type_0 = const()[name = tensor("op_1998_pad_type_0"), val = tensor("valid")]; + tensor var_1998_strides_0 = const()[name = tensor("op_1998_strides_0"), val = tensor([1, 1])]; + tensor var_1998_pad_0 = const()[name = tensor("op_1998_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1998_dilations_0 = const()[name = tensor("op_1998_dilations_0"), val = tensor([1, 1])]; + tensor var_1998_groups_0 = const()[name = tensor("op_1998_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276653376)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289760640)))]; + tensor var_1998_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1998_dilations_0, groups = var_1998_groups_0, pad = var_1998_pad_0, pad_type = var_1998_pad_type_0, strides = var_1998_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1998_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_2007 = const()[name = tensor("op_2007"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289763264)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289765888)))]; + tensor var_2023_to_fp16 = const()[name = tensor("op_2023_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_2023_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_2058_weight_0_to_fp16 = const()[name = tensor("op_2058_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289768512)))]; + tensor var_2058_bias_0_to_fp16 = const()[name = tensor("op_2058_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293045376)))]; + tensor var_2058_cast_fp16 = conv(bias = var_2058_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_2058_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_2058_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293048000)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_2056_pad_type_0 = const()[name = tensor("op_2056_pad_type_0"), val = tensor("valid")]; + tensor var_2056_strides_0 = const()[name = tensor("op_2056_strides_0"), val = tensor([1, 1])]; + tensor var_2056_pad_0 = const()[name = tensor("op_2056_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2056_dilations_0 = const()[name = tensor("op_2056_dilations_0"), val = tensor([1, 1])]; + tensor var_2056_groups_0 = const()[name = tensor("op_2056_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296324864)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299601728)))]; + tensor var_2056_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_2056_dilations_0, groups = var_2056_groups_0, pad = var_2056_pad_0, pad_type = var_2056_pad_type_0, strides = var_2056_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2059_axis_0 = const()[name = tensor("op_2059_axis_0"), val = tensor(1)]; + tensor var_2059_cast_fp16_0, tensor var_2059_cast_fp16_1, tensor var_2059_cast_fp16_2, tensor var_2059_cast_fp16_3, tensor var_2059_cast_fp16_4, tensor var_2059_cast_fp16_5, tensor var_2059_cast_fp16_6, tensor var_2059_cast_fp16_7, tensor var_2059_cast_fp16_8, tensor var_2059_cast_fp16_9, tensor var_2059_cast_fp16_10, tensor var_2059_cast_fp16_11, tensor var_2059_cast_fp16_12, tensor var_2059_cast_fp16_13, tensor var_2059_cast_fp16_14, tensor var_2059_cast_fp16_15, tensor var_2059_cast_fp16_16, tensor var_2059_cast_fp16_17, tensor var_2059_cast_fp16_18, tensor var_2059_cast_fp16_19 = split(axis = var_2059_axis_0, split_sizes = tile_21, x = var_2058_cast_fp16)[name = tensor("op_2059_cast_fp16")]; + tensor var_2080_perm_0 = const()[name = tensor("op_2080_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2081_axis_0 = const()[name = tensor("op_2081_axis_0"), val = tensor(3)]; + tensor var_2080_cast_fp16 = transpose(perm = var_2080_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_25")]; + tensor var_2081_cast_fp16_0, tensor var_2081_cast_fp16_1, tensor var_2081_cast_fp16_2, tensor var_2081_cast_fp16_3, tensor var_2081_cast_fp16_4, tensor var_2081_cast_fp16_5, tensor var_2081_cast_fp16_6, tensor var_2081_cast_fp16_7, tensor var_2081_cast_fp16_8, tensor var_2081_cast_fp16_9, tensor var_2081_cast_fp16_10, tensor var_2081_cast_fp16_11, tensor var_2081_cast_fp16_12, tensor var_2081_cast_fp16_13, tensor var_2081_cast_fp16_14, tensor var_2081_cast_fp16_15, tensor var_2081_cast_fp16_16, tensor var_2081_cast_fp16_17, tensor var_2081_cast_fp16_18, tensor var_2081_cast_fp16_19 = split(axis = var_2081_axis_0, split_sizes = tile_22, x = var_2080_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2102_axis_0 = const()[name = tensor("op_2102_axis_0"), val = tensor(1)]; + tensor var_2102_cast_fp16_0, tensor var_2102_cast_fp16_1, tensor var_2102_cast_fp16_2, tensor var_2102_cast_fp16_3, tensor var_2102_cast_fp16_4, tensor var_2102_cast_fp16_5, tensor var_2102_cast_fp16_6, tensor var_2102_cast_fp16_7, tensor var_2102_cast_fp16_8, tensor var_2102_cast_fp16_9, tensor var_2102_cast_fp16_10, tensor var_2102_cast_fp16_11, tensor var_2102_cast_fp16_12, tensor var_2102_cast_fp16_13, tensor var_2102_cast_fp16_14, tensor var_2102_cast_fp16_15, tensor var_2102_cast_fp16_16, tensor var_2102_cast_fp16_17, tensor var_2102_cast_fp16_18, tensor var_2102_cast_fp16_19 = split(axis = var_2102_axis_0, split_sizes = tile_23, x = var_2056_cast_fp16)[name = tensor("op_2102_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2081_cast_fp16_0, var_2059_cast_fp16_0))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2081_cast_fp16_1, var_2059_cast_fp16_1))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2081_cast_fp16_2, var_2059_cast_fp16_2))[name = tensor("aw_285_cast_fp16")]; + tensor aw_287_equation_0 = const()[name = tensor("aw_287_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_287_cast_fp16 = einsum(equation = aw_287_equation_0, values = (var_2081_cast_fp16_3, var_2059_cast_fp16_3))[name = tensor("aw_287_cast_fp16")]; + tensor aw_289_equation_0 = const()[name = tensor("aw_289_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_289_cast_fp16 = einsum(equation = aw_289_equation_0, values = (var_2081_cast_fp16_4, var_2059_cast_fp16_4))[name = tensor("aw_289_cast_fp16")]; + tensor aw_291_equation_0 = const()[name = tensor("aw_291_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_291_cast_fp16 = einsum(equation = aw_291_equation_0, values = (var_2081_cast_fp16_5, var_2059_cast_fp16_5))[name = tensor("aw_291_cast_fp16")]; + tensor aw_293_equation_0 = const()[name = tensor("aw_293_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_293_cast_fp16 = einsum(equation = aw_293_equation_0, values = (var_2081_cast_fp16_6, var_2059_cast_fp16_6))[name = tensor("aw_293_cast_fp16")]; + tensor aw_295_equation_0 = const()[name = tensor("aw_295_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_295_cast_fp16 = einsum(equation = aw_295_equation_0, values = (var_2081_cast_fp16_7, var_2059_cast_fp16_7))[name = tensor("aw_295_cast_fp16")]; + tensor aw_297_equation_0 = const()[name = tensor("aw_297_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_297_cast_fp16 = einsum(equation = aw_297_equation_0, values = (var_2081_cast_fp16_8, var_2059_cast_fp16_8))[name = tensor("aw_297_cast_fp16")]; + tensor aw_299_equation_0 = const()[name = tensor("aw_299_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_299_cast_fp16 = einsum(equation = aw_299_equation_0, values = (var_2081_cast_fp16_9, var_2059_cast_fp16_9))[name = tensor("aw_299_cast_fp16")]; + tensor aw_301_equation_0 = const()[name = tensor("aw_301_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_301_cast_fp16 = einsum(equation = aw_301_equation_0, values = (var_2081_cast_fp16_10, var_2059_cast_fp16_10))[name = tensor("aw_301_cast_fp16")]; + tensor aw_303_equation_0 = const()[name = tensor("aw_303_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_303_cast_fp16 = einsum(equation = aw_303_equation_0, values = (var_2081_cast_fp16_11, var_2059_cast_fp16_11))[name = tensor("aw_303_cast_fp16")]; + tensor aw_305_equation_0 = const()[name = tensor("aw_305_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_305_cast_fp16 = einsum(equation = aw_305_equation_0, values = (var_2081_cast_fp16_12, var_2059_cast_fp16_12))[name = tensor("aw_305_cast_fp16")]; + tensor aw_307_equation_0 = const()[name = tensor("aw_307_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_307_cast_fp16 = einsum(equation = aw_307_equation_0, values = (var_2081_cast_fp16_13, var_2059_cast_fp16_13))[name = tensor("aw_307_cast_fp16")]; + tensor aw_309_equation_0 = const()[name = tensor("aw_309_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_309_cast_fp16 = einsum(equation = aw_309_equation_0, values = (var_2081_cast_fp16_14, var_2059_cast_fp16_14))[name = tensor("aw_309_cast_fp16")]; + tensor aw_311_equation_0 = const()[name = tensor("aw_311_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_311_cast_fp16 = einsum(equation = aw_311_equation_0, values = (var_2081_cast_fp16_15, var_2059_cast_fp16_15))[name = tensor("aw_311_cast_fp16")]; + tensor aw_313_equation_0 = const()[name = tensor("aw_313_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_313_cast_fp16 = einsum(equation = aw_313_equation_0, values = (var_2081_cast_fp16_16, var_2059_cast_fp16_16))[name = tensor("aw_313_cast_fp16")]; + tensor aw_315_equation_0 = const()[name = tensor("aw_315_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_315_cast_fp16 = einsum(equation = aw_315_equation_0, values = (var_2081_cast_fp16_17, var_2059_cast_fp16_17))[name = tensor("aw_315_cast_fp16")]; + tensor aw_317_equation_0 = const()[name = tensor("aw_317_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_317_cast_fp16 = einsum(equation = aw_317_equation_0, values = (var_2081_cast_fp16_18, var_2059_cast_fp16_18))[name = tensor("aw_317_cast_fp16")]; + tensor aw_319_equation_0 = const()[name = tensor("aw_319_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_319_cast_fp16 = einsum(equation = aw_319_equation_0, values = (var_2081_cast_fp16_19, var_2059_cast_fp16_19))[name = tensor("aw_319_cast_fp16")]; + tensor var_2163_cast_fp16 = softmax(axis = var_2007, x = aw_281_cast_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = softmax(axis = var_2007, x = aw_283_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor var_2165_cast_fp16 = softmax(axis = var_2007, x = aw_285_cast_fp16)[name = tensor("op_2165_cast_fp16")]; + tensor var_2166_cast_fp16 = softmax(axis = var_2007, x = aw_287_cast_fp16)[name = tensor("op_2166_cast_fp16")]; + tensor var_2167_cast_fp16 = softmax(axis = var_2007, x = aw_289_cast_fp16)[name = tensor("op_2167_cast_fp16")]; + tensor var_2168_cast_fp16 = softmax(axis = var_2007, x = aw_291_cast_fp16)[name = tensor("op_2168_cast_fp16")]; + tensor var_2169_cast_fp16 = softmax(axis = var_2007, x = aw_293_cast_fp16)[name = tensor("op_2169_cast_fp16")]; + tensor var_2170_cast_fp16 = softmax(axis = var_2007, x = aw_295_cast_fp16)[name = tensor("op_2170_cast_fp16")]; + tensor var_2171_cast_fp16 = softmax(axis = var_2007, x = aw_297_cast_fp16)[name = tensor("op_2171_cast_fp16")]; + tensor var_2172_cast_fp16 = softmax(axis = var_2007, x = aw_299_cast_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor var_2173_cast_fp16 = softmax(axis = var_2007, x = aw_301_cast_fp16)[name = tensor("op_2173_cast_fp16")]; + tensor var_2174_cast_fp16 = softmax(axis = var_2007, x = aw_303_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor var_2175_cast_fp16 = softmax(axis = var_2007, x = aw_305_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor var_2176_cast_fp16 = softmax(axis = var_2007, x = aw_307_cast_fp16)[name = tensor("op_2176_cast_fp16")]; + tensor var_2177_cast_fp16 = softmax(axis = var_2007, x = aw_309_cast_fp16)[name = tensor("op_2177_cast_fp16")]; + tensor var_2178_cast_fp16 = softmax(axis = var_2007, x = aw_311_cast_fp16)[name = tensor("op_2178_cast_fp16")]; + tensor var_2179_cast_fp16 = softmax(axis = var_2007, x = aw_313_cast_fp16)[name = tensor("op_2179_cast_fp16")]; + tensor var_2180_cast_fp16 = softmax(axis = var_2007, x = aw_315_cast_fp16)[name = tensor("op_2180_cast_fp16")]; + tensor var_2181_cast_fp16 = softmax(axis = var_2007, x = aw_317_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2182_cast_fp16 = softmax(axis = var_2007, x = aw_319_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor var_2184_equation_0 = const()[name = tensor("op_2184_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2184_cast_fp16 = einsum(equation = var_2184_equation_0, values = (var_2102_cast_fp16_0, var_2163_cast_fp16))[name = tensor("op_2184_cast_fp16")]; + tensor var_2186_equation_0 = const()[name = tensor("op_2186_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2186_cast_fp16 = einsum(equation = var_2186_equation_0, values = (var_2102_cast_fp16_1, var_2164_cast_fp16))[name = tensor("op_2186_cast_fp16")]; + tensor var_2188_equation_0 = const()[name = tensor("op_2188_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2188_cast_fp16 = einsum(equation = var_2188_equation_0, values = (var_2102_cast_fp16_2, var_2165_cast_fp16))[name = tensor("op_2188_cast_fp16")]; + tensor var_2190_equation_0 = const()[name = tensor("op_2190_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2190_cast_fp16 = einsum(equation = var_2190_equation_0, values = (var_2102_cast_fp16_3, var_2166_cast_fp16))[name = tensor("op_2190_cast_fp16")]; + tensor var_2192_equation_0 = const()[name = tensor("op_2192_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2192_cast_fp16 = einsum(equation = var_2192_equation_0, values = (var_2102_cast_fp16_4, var_2167_cast_fp16))[name = tensor("op_2192_cast_fp16")]; + tensor var_2194_equation_0 = const()[name = tensor("op_2194_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2194_cast_fp16 = einsum(equation = var_2194_equation_0, values = (var_2102_cast_fp16_5, var_2168_cast_fp16))[name = tensor("op_2194_cast_fp16")]; + tensor var_2196_equation_0 = const()[name = tensor("op_2196_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2196_cast_fp16 = einsum(equation = var_2196_equation_0, values = (var_2102_cast_fp16_6, var_2169_cast_fp16))[name = tensor("op_2196_cast_fp16")]; + tensor var_2198_equation_0 = const()[name = tensor("op_2198_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2198_cast_fp16 = einsum(equation = var_2198_equation_0, values = (var_2102_cast_fp16_7, var_2170_cast_fp16))[name = tensor("op_2198_cast_fp16")]; + tensor var_2200_equation_0 = const()[name = tensor("op_2200_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2200_cast_fp16 = einsum(equation = var_2200_equation_0, values = (var_2102_cast_fp16_8, var_2171_cast_fp16))[name = tensor("op_2200_cast_fp16")]; + tensor var_2202_equation_0 = const()[name = tensor("op_2202_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2202_cast_fp16 = einsum(equation = var_2202_equation_0, values = (var_2102_cast_fp16_9, var_2172_cast_fp16))[name = tensor("op_2202_cast_fp16")]; + tensor var_2204_equation_0 = const()[name = tensor("op_2204_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2204_cast_fp16 = einsum(equation = var_2204_equation_0, values = (var_2102_cast_fp16_10, var_2173_cast_fp16))[name = tensor("op_2204_cast_fp16")]; + tensor var_2206_equation_0 = const()[name = tensor("op_2206_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2206_cast_fp16 = einsum(equation = var_2206_equation_0, values = (var_2102_cast_fp16_11, var_2174_cast_fp16))[name = tensor("op_2206_cast_fp16")]; + tensor var_2208_equation_0 = const()[name = tensor("op_2208_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2208_cast_fp16 = einsum(equation = var_2208_equation_0, values = (var_2102_cast_fp16_12, var_2175_cast_fp16))[name = tensor("op_2208_cast_fp16")]; + tensor var_2210_equation_0 = const()[name = tensor("op_2210_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2210_cast_fp16 = einsum(equation = var_2210_equation_0, values = (var_2102_cast_fp16_13, var_2176_cast_fp16))[name = tensor("op_2210_cast_fp16")]; + tensor var_2212_equation_0 = const()[name = tensor("op_2212_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2212_cast_fp16 = einsum(equation = var_2212_equation_0, values = (var_2102_cast_fp16_14, var_2177_cast_fp16))[name = tensor("op_2212_cast_fp16")]; + tensor var_2214_equation_0 = const()[name = tensor("op_2214_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2214_cast_fp16 = einsum(equation = var_2214_equation_0, values = (var_2102_cast_fp16_15, var_2178_cast_fp16))[name = tensor("op_2214_cast_fp16")]; + tensor var_2216_equation_0 = const()[name = tensor("op_2216_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2216_cast_fp16 = einsum(equation = var_2216_equation_0, values = (var_2102_cast_fp16_16, var_2179_cast_fp16))[name = tensor("op_2216_cast_fp16")]; + tensor var_2218_equation_0 = const()[name = tensor("op_2218_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2218_cast_fp16 = einsum(equation = var_2218_equation_0, values = (var_2102_cast_fp16_17, var_2180_cast_fp16))[name = tensor("op_2218_cast_fp16")]; + tensor var_2220_equation_0 = const()[name = tensor("op_2220_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2220_cast_fp16 = einsum(equation = var_2220_equation_0, values = (var_2102_cast_fp16_18, var_2181_cast_fp16))[name = tensor("op_2220_cast_fp16")]; + tensor var_2222_equation_0 = const()[name = tensor("op_2222_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2222_cast_fp16 = einsum(equation = var_2222_equation_0, values = (var_2102_cast_fp16_19, var_2182_cast_fp16))[name = tensor("op_2222_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_2007, interleave = input_75_interleave_0, values = (var_2184_cast_fp16, var_2186_cast_fp16, var_2188_cast_fp16, var_2190_cast_fp16, var_2192_cast_fp16, var_2194_cast_fp16, var_2196_cast_fp16, var_2198_cast_fp16, var_2200_cast_fp16, var_2202_cast_fp16, var_2204_cast_fp16, var_2206_cast_fp16, var_2208_cast_fp16, var_2210_cast_fp16, var_2212_cast_fp16, var_2214_cast_fp16, var_2216_cast_fp16, var_2218_cast_fp16, var_2220_cast_fp16, var_2222_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_2231_pad_type_0 = const()[name = tensor("op_2231_pad_type_0"), val = tensor("valid")]; + tensor var_2231_strides_0 = const()[name = tensor("op_2231_strides_0"), val = tensor([1, 1])]; + tensor var_2231_pad_0 = const()[name = tensor("op_2231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2231_dilations_0 = const()[name = tensor("op_2231_dilations_0"), val = tensor([1, 1])]; + tensor var_2231_groups_0 = const()[name = tensor("op_2231_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299604352)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302881216)))]; + tensor var_2231_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_2231_dilations_0, groups = var_2231_groups_0, pad = var_2231_pad_0, pad_type = var_2231_pad_type_0, strides = var_2231_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_2231_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302883840)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302886464)))]; + tensor var_2241_to_fp16 = const()[name = tensor("op_2241_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_2241_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302889088)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315996352)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_2267_pad_type_0 = const()[name = tensor("op_2267_pad_type_0"), val = tensor("valid")]; + tensor var_2267_strides_0 = const()[name = tensor("op_2267_strides_0"), val = tensor([1, 1])]; + tensor var_2267_pad_0 = const()[name = tensor("op_2267_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2267_dilations_0 = const()[name = tensor("op_2267_dilations_0"), val = tensor([1, 1])]; + tensor var_2267_groups_0 = const()[name = tensor("op_2267_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316006656)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329113920)))]; + tensor var_2267_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_2267_dilations_0, groups = var_2267_groups_0, pad = var_2267_pad_0, pad_type = var_2267_pad_type_0, strides = var_2267_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_2267_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_2276 = const()[name = tensor("op_2276"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329116544)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329119168)))]; + tensor var_2292_to_fp16 = const()[name = tensor("op_2292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_2292_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_2327_weight_0_to_fp16 = const()[name = tensor("op_2327_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329121792)))]; + tensor var_2327_bias_0_to_fp16 = const()[name = tensor("op_2327_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332398656)))]; + tensor var_2327_cast_fp16 = conv(bias = var_2327_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_2327_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2327_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332401280)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_2325_pad_type_0 = const()[name = tensor("op_2325_pad_type_0"), val = tensor("valid")]; + tensor var_2325_strides_0 = const()[name = tensor("op_2325_strides_0"), val = tensor([1, 1])]; + tensor var_2325_pad_0 = const()[name = tensor("op_2325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2325_dilations_0 = const()[name = tensor("op_2325_dilations_0"), val = tensor([1, 1])]; + tensor var_2325_groups_0 = const()[name = tensor("op_2325_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335678144)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338955008)))]; + tensor var_2325_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_2325_dilations_0, groups = var_2325_groups_0, pad = var_2325_pad_0, pad_type = var_2325_pad_type_0, strides = var_2325_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2328_axis_0 = const()[name = tensor("op_2328_axis_0"), val = tensor(1)]; + tensor var_2328_cast_fp16_0, tensor var_2328_cast_fp16_1, tensor var_2328_cast_fp16_2, tensor var_2328_cast_fp16_3, tensor var_2328_cast_fp16_4, tensor var_2328_cast_fp16_5, tensor var_2328_cast_fp16_6, tensor var_2328_cast_fp16_7, tensor var_2328_cast_fp16_8, tensor var_2328_cast_fp16_9, tensor var_2328_cast_fp16_10, tensor var_2328_cast_fp16_11, tensor var_2328_cast_fp16_12, tensor var_2328_cast_fp16_13, tensor var_2328_cast_fp16_14, tensor var_2328_cast_fp16_15, tensor var_2328_cast_fp16_16, tensor var_2328_cast_fp16_17, tensor var_2328_cast_fp16_18, tensor var_2328_cast_fp16_19 = split(axis = var_2328_axis_0, split_sizes = tile_24, x = var_2327_cast_fp16)[name = tensor("op_2328_cast_fp16")]; + tensor var_2349_perm_0 = const()[name = tensor("op_2349_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2350_axis_0 = const()[name = tensor("op_2350_axis_0"), val = tensor(3)]; + tensor var_2349_cast_fp16 = transpose(perm = var_2349_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_24")]; + tensor var_2350_cast_fp16_0, tensor var_2350_cast_fp16_1, tensor var_2350_cast_fp16_2, tensor var_2350_cast_fp16_3, tensor var_2350_cast_fp16_4, tensor var_2350_cast_fp16_5, tensor var_2350_cast_fp16_6, tensor var_2350_cast_fp16_7, tensor var_2350_cast_fp16_8, tensor var_2350_cast_fp16_9, tensor var_2350_cast_fp16_10, tensor var_2350_cast_fp16_11, tensor var_2350_cast_fp16_12, tensor var_2350_cast_fp16_13, tensor var_2350_cast_fp16_14, tensor var_2350_cast_fp16_15, tensor var_2350_cast_fp16_16, tensor var_2350_cast_fp16_17, tensor var_2350_cast_fp16_18, tensor var_2350_cast_fp16_19 = split(axis = var_2350_axis_0, split_sizes = tile_25, x = var_2349_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2371_axis_0 = const()[name = tensor("op_2371_axis_0"), val = tensor(1)]; + tensor var_2371_cast_fp16_0, tensor var_2371_cast_fp16_1, tensor var_2371_cast_fp16_2, tensor var_2371_cast_fp16_3, tensor var_2371_cast_fp16_4, tensor var_2371_cast_fp16_5, tensor var_2371_cast_fp16_6, tensor var_2371_cast_fp16_7, tensor var_2371_cast_fp16_8, tensor var_2371_cast_fp16_9, tensor var_2371_cast_fp16_10, tensor var_2371_cast_fp16_11, tensor var_2371_cast_fp16_12, tensor var_2371_cast_fp16_13, tensor var_2371_cast_fp16_14, tensor var_2371_cast_fp16_15, tensor var_2371_cast_fp16_16, tensor var_2371_cast_fp16_17, tensor var_2371_cast_fp16_18, tensor var_2371_cast_fp16_19 = split(axis = var_2371_axis_0, split_sizes = tile_26, x = var_2325_cast_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor aw_321_equation_0 = const()[name = tensor("aw_321_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_321_cast_fp16 = einsum(equation = aw_321_equation_0, values = (var_2350_cast_fp16_0, var_2328_cast_fp16_0))[name = tensor("aw_321_cast_fp16")]; + tensor aw_323_equation_0 = const()[name = tensor("aw_323_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_323_cast_fp16 = einsum(equation = aw_323_equation_0, values = (var_2350_cast_fp16_1, var_2328_cast_fp16_1))[name = tensor("aw_323_cast_fp16")]; + tensor aw_325_equation_0 = const()[name = tensor("aw_325_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_325_cast_fp16 = einsum(equation = aw_325_equation_0, values = (var_2350_cast_fp16_2, var_2328_cast_fp16_2))[name = tensor("aw_325_cast_fp16")]; + tensor aw_327_equation_0 = const()[name = tensor("aw_327_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_327_cast_fp16 = einsum(equation = aw_327_equation_0, values = (var_2350_cast_fp16_3, var_2328_cast_fp16_3))[name = tensor("aw_327_cast_fp16")]; + tensor aw_329_equation_0 = const()[name = tensor("aw_329_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_329_cast_fp16 = einsum(equation = aw_329_equation_0, values = (var_2350_cast_fp16_4, var_2328_cast_fp16_4))[name = tensor("aw_329_cast_fp16")]; + tensor aw_331_equation_0 = const()[name = tensor("aw_331_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_331_cast_fp16 = einsum(equation = aw_331_equation_0, values = (var_2350_cast_fp16_5, var_2328_cast_fp16_5))[name = tensor("aw_331_cast_fp16")]; + tensor aw_333_equation_0 = const()[name = tensor("aw_333_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_333_cast_fp16 = einsum(equation = aw_333_equation_0, values = (var_2350_cast_fp16_6, var_2328_cast_fp16_6))[name = tensor("aw_333_cast_fp16")]; + tensor aw_335_equation_0 = const()[name = tensor("aw_335_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_335_cast_fp16 = einsum(equation = aw_335_equation_0, values = (var_2350_cast_fp16_7, var_2328_cast_fp16_7))[name = tensor("aw_335_cast_fp16")]; + tensor aw_337_equation_0 = const()[name = tensor("aw_337_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_337_cast_fp16 = einsum(equation = aw_337_equation_0, values = (var_2350_cast_fp16_8, var_2328_cast_fp16_8))[name = tensor("aw_337_cast_fp16")]; + tensor aw_339_equation_0 = const()[name = tensor("aw_339_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_339_cast_fp16 = einsum(equation = aw_339_equation_0, values = (var_2350_cast_fp16_9, var_2328_cast_fp16_9))[name = tensor("aw_339_cast_fp16")]; + tensor aw_341_equation_0 = const()[name = tensor("aw_341_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_341_cast_fp16 = einsum(equation = aw_341_equation_0, values = (var_2350_cast_fp16_10, var_2328_cast_fp16_10))[name = tensor("aw_341_cast_fp16")]; + tensor aw_343_equation_0 = const()[name = tensor("aw_343_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_343_cast_fp16 = einsum(equation = aw_343_equation_0, values = (var_2350_cast_fp16_11, var_2328_cast_fp16_11))[name = tensor("aw_343_cast_fp16")]; + tensor aw_345_equation_0 = const()[name = tensor("aw_345_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_345_cast_fp16 = einsum(equation = aw_345_equation_0, values = (var_2350_cast_fp16_12, var_2328_cast_fp16_12))[name = tensor("aw_345_cast_fp16")]; + tensor aw_347_equation_0 = const()[name = tensor("aw_347_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_347_cast_fp16 = einsum(equation = aw_347_equation_0, values = (var_2350_cast_fp16_13, var_2328_cast_fp16_13))[name = tensor("aw_347_cast_fp16")]; + tensor aw_349_equation_0 = const()[name = tensor("aw_349_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_349_cast_fp16 = einsum(equation = aw_349_equation_0, values = (var_2350_cast_fp16_14, var_2328_cast_fp16_14))[name = tensor("aw_349_cast_fp16")]; + tensor aw_351_equation_0 = const()[name = tensor("aw_351_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_351_cast_fp16 = einsum(equation = aw_351_equation_0, values = (var_2350_cast_fp16_15, var_2328_cast_fp16_15))[name = tensor("aw_351_cast_fp16")]; + tensor aw_353_equation_0 = const()[name = tensor("aw_353_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_353_cast_fp16 = einsum(equation = aw_353_equation_0, values = (var_2350_cast_fp16_16, var_2328_cast_fp16_16))[name = tensor("aw_353_cast_fp16")]; + tensor aw_355_equation_0 = const()[name = tensor("aw_355_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_355_cast_fp16 = einsum(equation = aw_355_equation_0, values = (var_2350_cast_fp16_17, var_2328_cast_fp16_17))[name = tensor("aw_355_cast_fp16")]; + tensor aw_357_equation_0 = const()[name = tensor("aw_357_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_357_cast_fp16 = einsum(equation = aw_357_equation_0, values = (var_2350_cast_fp16_18, var_2328_cast_fp16_18))[name = tensor("aw_357_cast_fp16")]; + tensor aw_359_equation_0 = const()[name = tensor("aw_359_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_359_cast_fp16 = einsum(equation = aw_359_equation_0, values = (var_2350_cast_fp16_19, var_2328_cast_fp16_19))[name = tensor("aw_359_cast_fp16")]; + tensor var_2432_cast_fp16 = softmax(axis = var_2276, x = aw_321_cast_fp16)[name = tensor("op_2432_cast_fp16")]; + tensor var_2433_cast_fp16 = softmax(axis = var_2276, x = aw_323_cast_fp16)[name = tensor("op_2433_cast_fp16")]; + tensor var_2434_cast_fp16 = softmax(axis = var_2276, x = aw_325_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = softmax(axis = var_2276, x = aw_327_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = softmax(axis = var_2276, x = aw_329_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2437_cast_fp16 = softmax(axis = var_2276, x = aw_331_cast_fp16)[name = tensor("op_2437_cast_fp16")]; + tensor var_2438_cast_fp16 = softmax(axis = var_2276, x = aw_333_cast_fp16)[name = tensor("op_2438_cast_fp16")]; + tensor var_2439_cast_fp16 = softmax(axis = var_2276, x = aw_335_cast_fp16)[name = tensor("op_2439_cast_fp16")]; + tensor var_2440_cast_fp16 = softmax(axis = var_2276, x = aw_337_cast_fp16)[name = tensor("op_2440_cast_fp16")]; + tensor var_2441_cast_fp16 = softmax(axis = var_2276, x = aw_339_cast_fp16)[name = tensor("op_2441_cast_fp16")]; + tensor var_2442_cast_fp16 = softmax(axis = var_2276, x = aw_341_cast_fp16)[name = tensor("op_2442_cast_fp16")]; + tensor var_2443_cast_fp16 = softmax(axis = var_2276, x = aw_343_cast_fp16)[name = tensor("op_2443_cast_fp16")]; + tensor var_2444_cast_fp16 = softmax(axis = var_2276, x = aw_345_cast_fp16)[name = tensor("op_2444_cast_fp16")]; + tensor var_2445_cast_fp16 = softmax(axis = var_2276, x = aw_347_cast_fp16)[name = tensor("op_2445_cast_fp16")]; + tensor var_2446_cast_fp16 = softmax(axis = var_2276, x = aw_349_cast_fp16)[name = tensor("op_2446_cast_fp16")]; + tensor var_2447_cast_fp16 = softmax(axis = var_2276, x = aw_351_cast_fp16)[name = tensor("op_2447_cast_fp16")]; + tensor var_2448_cast_fp16 = softmax(axis = var_2276, x = aw_353_cast_fp16)[name = tensor("op_2448_cast_fp16")]; + tensor var_2449_cast_fp16 = softmax(axis = var_2276, x = aw_355_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor var_2450_cast_fp16 = softmax(axis = var_2276, x = aw_357_cast_fp16)[name = tensor("op_2450_cast_fp16")]; + tensor var_2451_cast_fp16 = softmax(axis = var_2276, x = aw_359_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2453_equation_0 = const()[name = tensor("op_2453_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2453_cast_fp16 = einsum(equation = var_2453_equation_0, values = (var_2371_cast_fp16_0, var_2432_cast_fp16))[name = tensor("op_2453_cast_fp16")]; + tensor var_2455_equation_0 = const()[name = tensor("op_2455_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2455_cast_fp16 = einsum(equation = var_2455_equation_0, values = (var_2371_cast_fp16_1, var_2433_cast_fp16))[name = tensor("op_2455_cast_fp16")]; + tensor var_2457_equation_0 = const()[name = tensor("op_2457_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2457_cast_fp16 = einsum(equation = var_2457_equation_0, values = (var_2371_cast_fp16_2, var_2434_cast_fp16))[name = tensor("op_2457_cast_fp16")]; + tensor var_2459_equation_0 = const()[name = tensor("op_2459_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2459_cast_fp16 = einsum(equation = var_2459_equation_0, values = (var_2371_cast_fp16_3, var_2435_cast_fp16))[name = tensor("op_2459_cast_fp16")]; + tensor var_2461_equation_0 = const()[name = tensor("op_2461_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2461_cast_fp16 = einsum(equation = var_2461_equation_0, values = (var_2371_cast_fp16_4, var_2436_cast_fp16))[name = tensor("op_2461_cast_fp16")]; + tensor var_2463_equation_0 = const()[name = tensor("op_2463_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2463_cast_fp16 = einsum(equation = var_2463_equation_0, values = (var_2371_cast_fp16_5, var_2437_cast_fp16))[name = tensor("op_2463_cast_fp16")]; + tensor var_2465_equation_0 = const()[name = tensor("op_2465_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2465_cast_fp16 = einsum(equation = var_2465_equation_0, values = (var_2371_cast_fp16_6, var_2438_cast_fp16))[name = tensor("op_2465_cast_fp16")]; + tensor var_2467_equation_0 = const()[name = tensor("op_2467_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2467_cast_fp16 = einsum(equation = var_2467_equation_0, values = (var_2371_cast_fp16_7, var_2439_cast_fp16))[name = tensor("op_2467_cast_fp16")]; + tensor var_2469_equation_0 = const()[name = tensor("op_2469_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2469_cast_fp16 = einsum(equation = var_2469_equation_0, values = (var_2371_cast_fp16_8, var_2440_cast_fp16))[name = tensor("op_2469_cast_fp16")]; + tensor var_2471_equation_0 = const()[name = tensor("op_2471_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2471_cast_fp16 = einsum(equation = var_2471_equation_0, values = (var_2371_cast_fp16_9, var_2441_cast_fp16))[name = tensor("op_2471_cast_fp16")]; + tensor var_2473_equation_0 = const()[name = tensor("op_2473_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2473_cast_fp16 = einsum(equation = var_2473_equation_0, values = (var_2371_cast_fp16_10, var_2442_cast_fp16))[name = tensor("op_2473_cast_fp16")]; + tensor var_2475_equation_0 = const()[name = tensor("op_2475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2475_cast_fp16 = einsum(equation = var_2475_equation_0, values = (var_2371_cast_fp16_11, var_2443_cast_fp16))[name = tensor("op_2475_cast_fp16")]; + tensor var_2477_equation_0 = const()[name = tensor("op_2477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2477_cast_fp16 = einsum(equation = var_2477_equation_0, values = (var_2371_cast_fp16_12, var_2444_cast_fp16))[name = tensor("op_2477_cast_fp16")]; + tensor var_2479_equation_0 = const()[name = tensor("op_2479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2479_cast_fp16 = einsum(equation = var_2479_equation_0, values = (var_2371_cast_fp16_13, var_2445_cast_fp16))[name = tensor("op_2479_cast_fp16")]; + tensor var_2481_equation_0 = const()[name = tensor("op_2481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2481_cast_fp16 = einsum(equation = var_2481_equation_0, values = (var_2371_cast_fp16_14, var_2446_cast_fp16))[name = tensor("op_2481_cast_fp16")]; + tensor var_2483_equation_0 = const()[name = tensor("op_2483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2483_cast_fp16 = einsum(equation = var_2483_equation_0, values = (var_2371_cast_fp16_15, var_2447_cast_fp16))[name = tensor("op_2483_cast_fp16")]; + tensor var_2485_equation_0 = const()[name = tensor("op_2485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2485_cast_fp16 = einsum(equation = var_2485_equation_0, values = (var_2371_cast_fp16_16, var_2448_cast_fp16))[name = tensor("op_2485_cast_fp16")]; + tensor var_2487_equation_0 = const()[name = tensor("op_2487_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2487_cast_fp16 = einsum(equation = var_2487_equation_0, values = (var_2371_cast_fp16_17, var_2449_cast_fp16))[name = tensor("op_2487_cast_fp16")]; + tensor var_2489_equation_0 = const()[name = tensor("op_2489_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2489_cast_fp16 = einsum(equation = var_2489_equation_0, values = (var_2371_cast_fp16_18, var_2450_cast_fp16))[name = tensor("op_2489_cast_fp16")]; + tensor var_2491_equation_0 = const()[name = tensor("op_2491_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2491_cast_fp16 = einsum(equation = var_2491_equation_0, values = (var_2371_cast_fp16_19, var_2451_cast_fp16))[name = tensor("op_2491_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_2276, interleave = input_85_interleave_0, values = (var_2453_cast_fp16, var_2455_cast_fp16, var_2457_cast_fp16, var_2459_cast_fp16, var_2461_cast_fp16, var_2463_cast_fp16, var_2465_cast_fp16, var_2467_cast_fp16, var_2469_cast_fp16, var_2471_cast_fp16, var_2473_cast_fp16, var_2475_cast_fp16, var_2477_cast_fp16, var_2479_cast_fp16, var_2481_cast_fp16, var_2483_cast_fp16, var_2485_cast_fp16, var_2487_cast_fp16, var_2489_cast_fp16, var_2491_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_2500_pad_type_0 = const()[name = tensor("op_2500_pad_type_0"), val = tensor("valid")]; + tensor var_2500_strides_0 = const()[name = tensor("op_2500_strides_0"), val = tensor([1, 1])]; + tensor var_2500_pad_0 = const()[name = tensor("op_2500_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2500_dilations_0 = const()[name = tensor("op_2500_dilations_0"), val = tensor([1, 1])]; + tensor var_2500_groups_0 = const()[name = tensor("op_2500_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338957632)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342234496)))]; + tensor var_2500_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_2500_dilations_0, groups = var_2500_groups_0, pad = var_2500_pad_0, pad_type = var_2500_pad_type_0, strides = var_2500_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_2500_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342237120)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342239744)))]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_2510_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342242368)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355349632)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2536_pad_type_0 = const()[name = tensor("op_2536_pad_type_0"), val = tensor("valid")]; + tensor var_2536_strides_0 = const()[name = tensor("op_2536_strides_0"), val = tensor([1, 1])]; + tensor var_2536_pad_0 = const()[name = tensor("op_2536_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2536_dilations_0 = const()[name = tensor("op_2536_dilations_0"), val = tensor([1, 1])]; + tensor var_2536_groups_0 = const()[name = tensor("op_2536_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355359936)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368467200)))]; + tensor var_2536_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_2536_dilations_0, groups = var_2536_groups_0, pad = var_2536_pad_0, pad_type = var_2536_pad_type_0, strides = var_2536_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_2536_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2545 = const()[name = tensor("op_2545"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368469824)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368472448)))]; + tensor var_2561_to_fp16 = const()[name = tensor("op_2561_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_2561_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_2596_weight_0_to_fp16 = const()[name = tensor("op_2596_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368475072)))]; + tensor var_2596_bias_0_to_fp16 = const()[name = tensor("op_2596_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371751936)))]; + tensor var_2596_cast_fp16 = conv(bias = var_2596_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_2596_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371754560)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_2594_pad_type_0 = const()[name = tensor("op_2594_pad_type_0"), val = tensor("valid")]; + tensor var_2594_strides_0 = const()[name = tensor("op_2594_strides_0"), val = tensor([1, 1])]; + tensor var_2594_pad_0 = const()[name = tensor("op_2594_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2594_dilations_0 = const()[name = tensor("op_2594_dilations_0"), val = tensor([1, 1])]; + tensor var_2594_groups_0 = const()[name = tensor("op_2594_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375031424)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378308288)))]; + tensor var_2594_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_2594_dilations_0, groups = var_2594_groups_0, pad = var_2594_pad_0, pad_type = var_2594_pad_type_0, strides = var_2594_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2594_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2597_axis_0 = const()[name = tensor("op_2597_axis_0"), val = tensor(1)]; + tensor var_2597_cast_fp16_0, tensor var_2597_cast_fp16_1, tensor var_2597_cast_fp16_2, tensor var_2597_cast_fp16_3, tensor var_2597_cast_fp16_4, tensor var_2597_cast_fp16_5, tensor var_2597_cast_fp16_6, tensor var_2597_cast_fp16_7, tensor var_2597_cast_fp16_8, tensor var_2597_cast_fp16_9, tensor var_2597_cast_fp16_10, tensor var_2597_cast_fp16_11, tensor var_2597_cast_fp16_12, tensor var_2597_cast_fp16_13, tensor var_2597_cast_fp16_14, tensor var_2597_cast_fp16_15, tensor var_2597_cast_fp16_16, tensor var_2597_cast_fp16_17, tensor var_2597_cast_fp16_18, tensor var_2597_cast_fp16_19 = split(axis = var_2597_axis_0, split_sizes = tile_27, x = var_2596_cast_fp16)[name = tensor("op_2597_cast_fp16")]; + tensor var_2618_perm_0 = const()[name = tensor("op_2618_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2619_axis_0 = const()[name = tensor("op_2619_axis_0"), val = tensor(3)]; + tensor var_2618_cast_fp16 = transpose(perm = var_2618_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_23")]; + tensor var_2619_cast_fp16_0, tensor var_2619_cast_fp16_1, tensor var_2619_cast_fp16_2, tensor var_2619_cast_fp16_3, tensor var_2619_cast_fp16_4, tensor var_2619_cast_fp16_5, tensor var_2619_cast_fp16_6, tensor var_2619_cast_fp16_7, tensor var_2619_cast_fp16_8, tensor var_2619_cast_fp16_9, tensor var_2619_cast_fp16_10, tensor var_2619_cast_fp16_11, tensor var_2619_cast_fp16_12, tensor var_2619_cast_fp16_13, tensor var_2619_cast_fp16_14, tensor var_2619_cast_fp16_15, tensor var_2619_cast_fp16_16, tensor var_2619_cast_fp16_17, tensor var_2619_cast_fp16_18, tensor var_2619_cast_fp16_19 = split(axis = var_2619_axis_0, split_sizes = tile_28, x = var_2618_cast_fp16)[name = tensor("op_2619_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2640_axis_0 = const()[name = tensor("op_2640_axis_0"), val = tensor(1)]; + tensor var_2640_cast_fp16_0, tensor var_2640_cast_fp16_1, tensor var_2640_cast_fp16_2, tensor var_2640_cast_fp16_3, tensor var_2640_cast_fp16_4, tensor var_2640_cast_fp16_5, tensor var_2640_cast_fp16_6, tensor var_2640_cast_fp16_7, tensor var_2640_cast_fp16_8, tensor var_2640_cast_fp16_9, tensor var_2640_cast_fp16_10, tensor var_2640_cast_fp16_11, tensor var_2640_cast_fp16_12, tensor var_2640_cast_fp16_13, tensor var_2640_cast_fp16_14, tensor var_2640_cast_fp16_15, tensor var_2640_cast_fp16_16, tensor var_2640_cast_fp16_17, tensor var_2640_cast_fp16_18, tensor var_2640_cast_fp16_19 = split(axis = var_2640_axis_0, split_sizes = tile_29, x = var_2594_cast_fp16)[name = tensor("op_2640_cast_fp16")]; + tensor aw_361_equation_0 = const()[name = tensor("aw_361_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_361_cast_fp16 = einsum(equation = aw_361_equation_0, values = (var_2619_cast_fp16_0, var_2597_cast_fp16_0))[name = tensor("aw_361_cast_fp16")]; + tensor aw_363_equation_0 = const()[name = tensor("aw_363_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_363_cast_fp16 = einsum(equation = aw_363_equation_0, values = (var_2619_cast_fp16_1, var_2597_cast_fp16_1))[name = tensor("aw_363_cast_fp16")]; + tensor aw_365_equation_0 = const()[name = tensor("aw_365_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_365_cast_fp16 = einsum(equation = aw_365_equation_0, values = (var_2619_cast_fp16_2, var_2597_cast_fp16_2))[name = tensor("aw_365_cast_fp16")]; + tensor aw_367_equation_0 = const()[name = tensor("aw_367_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_367_cast_fp16 = einsum(equation = aw_367_equation_0, values = (var_2619_cast_fp16_3, var_2597_cast_fp16_3))[name = tensor("aw_367_cast_fp16")]; + tensor aw_369_equation_0 = const()[name = tensor("aw_369_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_369_cast_fp16 = einsum(equation = aw_369_equation_0, values = (var_2619_cast_fp16_4, var_2597_cast_fp16_4))[name = tensor("aw_369_cast_fp16")]; + tensor aw_371_equation_0 = const()[name = tensor("aw_371_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_371_cast_fp16 = einsum(equation = aw_371_equation_0, values = (var_2619_cast_fp16_5, var_2597_cast_fp16_5))[name = tensor("aw_371_cast_fp16")]; + tensor aw_373_equation_0 = const()[name = tensor("aw_373_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_373_cast_fp16 = einsum(equation = aw_373_equation_0, values = (var_2619_cast_fp16_6, var_2597_cast_fp16_6))[name = tensor("aw_373_cast_fp16")]; + tensor aw_375_equation_0 = const()[name = tensor("aw_375_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_375_cast_fp16 = einsum(equation = aw_375_equation_0, values = (var_2619_cast_fp16_7, var_2597_cast_fp16_7))[name = tensor("aw_375_cast_fp16")]; + tensor aw_377_equation_0 = const()[name = tensor("aw_377_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_377_cast_fp16 = einsum(equation = aw_377_equation_0, values = (var_2619_cast_fp16_8, var_2597_cast_fp16_8))[name = tensor("aw_377_cast_fp16")]; + tensor aw_379_equation_0 = const()[name = tensor("aw_379_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_379_cast_fp16 = einsum(equation = aw_379_equation_0, values = (var_2619_cast_fp16_9, var_2597_cast_fp16_9))[name = tensor("aw_379_cast_fp16")]; + tensor aw_381_equation_0 = const()[name = tensor("aw_381_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_381_cast_fp16 = einsum(equation = aw_381_equation_0, values = (var_2619_cast_fp16_10, var_2597_cast_fp16_10))[name = tensor("aw_381_cast_fp16")]; + tensor aw_383_equation_0 = const()[name = tensor("aw_383_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_383_cast_fp16 = einsum(equation = aw_383_equation_0, values = (var_2619_cast_fp16_11, var_2597_cast_fp16_11))[name = tensor("aw_383_cast_fp16")]; + tensor aw_385_equation_0 = const()[name = tensor("aw_385_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_385_cast_fp16 = einsum(equation = aw_385_equation_0, values = (var_2619_cast_fp16_12, var_2597_cast_fp16_12))[name = tensor("aw_385_cast_fp16")]; + tensor aw_387_equation_0 = const()[name = tensor("aw_387_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_387_cast_fp16 = einsum(equation = aw_387_equation_0, values = (var_2619_cast_fp16_13, var_2597_cast_fp16_13))[name = tensor("aw_387_cast_fp16")]; + tensor aw_389_equation_0 = const()[name = tensor("aw_389_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_389_cast_fp16 = einsum(equation = aw_389_equation_0, values = (var_2619_cast_fp16_14, var_2597_cast_fp16_14))[name = tensor("aw_389_cast_fp16")]; + tensor aw_391_equation_0 = const()[name = tensor("aw_391_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_391_cast_fp16 = einsum(equation = aw_391_equation_0, values = (var_2619_cast_fp16_15, var_2597_cast_fp16_15))[name = tensor("aw_391_cast_fp16")]; + tensor aw_393_equation_0 = const()[name = tensor("aw_393_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_393_cast_fp16 = einsum(equation = aw_393_equation_0, values = (var_2619_cast_fp16_16, var_2597_cast_fp16_16))[name = tensor("aw_393_cast_fp16")]; + tensor aw_395_equation_0 = const()[name = tensor("aw_395_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_395_cast_fp16 = einsum(equation = aw_395_equation_0, values = (var_2619_cast_fp16_17, var_2597_cast_fp16_17))[name = tensor("aw_395_cast_fp16")]; + tensor aw_397_equation_0 = const()[name = tensor("aw_397_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_397_cast_fp16 = einsum(equation = aw_397_equation_0, values = (var_2619_cast_fp16_18, var_2597_cast_fp16_18))[name = tensor("aw_397_cast_fp16")]; + tensor aw_399_equation_0 = const()[name = tensor("aw_399_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_399_cast_fp16 = einsum(equation = aw_399_equation_0, values = (var_2619_cast_fp16_19, var_2597_cast_fp16_19))[name = tensor("aw_399_cast_fp16")]; + tensor var_2701_cast_fp16 = softmax(axis = var_2545, x = aw_361_cast_fp16)[name = tensor("op_2701_cast_fp16")]; + tensor var_2702_cast_fp16 = softmax(axis = var_2545, x = aw_363_cast_fp16)[name = tensor("op_2702_cast_fp16")]; + tensor var_2703_cast_fp16 = softmax(axis = var_2545, x = aw_365_cast_fp16)[name = tensor("op_2703_cast_fp16")]; + tensor var_2704_cast_fp16 = softmax(axis = var_2545, x = aw_367_cast_fp16)[name = tensor("op_2704_cast_fp16")]; + tensor var_2705_cast_fp16 = softmax(axis = var_2545, x = aw_369_cast_fp16)[name = tensor("op_2705_cast_fp16")]; + tensor var_2706_cast_fp16 = softmax(axis = var_2545, x = aw_371_cast_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = softmax(axis = var_2545, x = aw_373_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor var_2708_cast_fp16 = softmax(axis = var_2545, x = aw_375_cast_fp16)[name = tensor("op_2708_cast_fp16")]; + tensor var_2709_cast_fp16 = softmax(axis = var_2545, x = aw_377_cast_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor var_2710_cast_fp16 = softmax(axis = var_2545, x = aw_379_cast_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor var_2711_cast_fp16 = softmax(axis = var_2545, x = aw_381_cast_fp16)[name = tensor("op_2711_cast_fp16")]; + tensor var_2712_cast_fp16 = softmax(axis = var_2545, x = aw_383_cast_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = softmax(axis = var_2545, x = aw_385_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor var_2714_cast_fp16 = softmax(axis = var_2545, x = aw_387_cast_fp16)[name = tensor("op_2714_cast_fp16")]; + tensor var_2715_cast_fp16 = softmax(axis = var_2545, x = aw_389_cast_fp16)[name = tensor("op_2715_cast_fp16")]; + tensor var_2716_cast_fp16 = softmax(axis = var_2545, x = aw_391_cast_fp16)[name = tensor("op_2716_cast_fp16")]; + tensor var_2717_cast_fp16 = softmax(axis = var_2545, x = aw_393_cast_fp16)[name = tensor("op_2717_cast_fp16")]; + tensor var_2718_cast_fp16 = softmax(axis = var_2545, x = aw_395_cast_fp16)[name = tensor("op_2718_cast_fp16")]; + tensor var_2719_cast_fp16 = softmax(axis = var_2545, x = aw_397_cast_fp16)[name = tensor("op_2719_cast_fp16")]; + tensor var_2720_cast_fp16 = softmax(axis = var_2545, x = aw_399_cast_fp16)[name = tensor("op_2720_cast_fp16")]; + tensor var_2722_equation_0 = const()[name = tensor("op_2722_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2722_cast_fp16 = einsum(equation = var_2722_equation_0, values = (var_2640_cast_fp16_0, var_2701_cast_fp16))[name = tensor("op_2722_cast_fp16")]; + tensor var_2724_equation_0 = const()[name = tensor("op_2724_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2724_cast_fp16 = einsum(equation = var_2724_equation_0, values = (var_2640_cast_fp16_1, var_2702_cast_fp16))[name = tensor("op_2724_cast_fp16")]; + tensor var_2726_equation_0 = const()[name = tensor("op_2726_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2726_cast_fp16 = einsum(equation = var_2726_equation_0, values = (var_2640_cast_fp16_2, var_2703_cast_fp16))[name = tensor("op_2726_cast_fp16")]; + tensor var_2728_equation_0 = const()[name = tensor("op_2728_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2728_cast_fp16 = einsum(equation = var_2728_equation_0, values = (var_2640_cast_fp16_3, var_2704_cast_fp16))[name = tensor("op_2728_cast_fp16")]; + tensor var_2730_equation_0 = const()[name = tensor("op_2730_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2730_cast_fp16 = einsum(equation = var_2730_equation_0, values = (var_2640_cast_fp16_4, var_2705_cast_fp16))[name = tensor("op_2730_cast_fp16")]; + tensor var_2732_equation_0 = const()[name = tensor("op_2732_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2732_cast_fp16 = einsum(equation = var_2732_equation_0, values = (var_2640_cast_fp16_5, var_2706_cast_fp16))[name = tensor("op_2732_cast_fp16")]; + tensor var_2734_equation_0 = const()[name = tensor("op_2734_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2734_cast_fp16 = einsum(equation = var_2734_equation_0, values = (var_2640_cast_fp16_6, var_2707_cast_fp16))[name = tensor("op_2734_cast_fp16")]; + tensor var_2736_equation_0 = const()[name = tensor("op_2736_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2736_cast_fp16 = einsum(equation = var_2736_equation_0, values = (var_2640_cast_fp16_7, var_2708_cast_fp16))[name = tensor("op_2736_cast_fp16")]; + tensor var_2738_equation_0 = const()[name = tensor("op_2738_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2738_cast_fp16 = einsum(equation = var_2738_equation_0, values = (var_2640_cast_fp16_8, var_2709_cast_fp16))[name = tensor("op_2738_cast_fp16")]; + tensor var_2740_equation_0 = const()[name = tensor("op_2740_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2740_cast_fp16 = einsum(equation = var_2740_equation_0, values = (var_2640_cast_fp16_9, var_2710_cast_fp16))[name = tensor("op_2740_cast_fp16")]; + tensor var_2742_equation_0 = const()[name = tensor("op_2742_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2742_cast_fp16 = einsum(equation = var_2742_equation_0, values = (var_2640_cast_fp16_10, var_2711_cast_fp16))[name = tensor("op_2742_cast_fp16")]; + tensor var_2744_equation_0 = const()[name = tensor("op_2744_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2744_cast_fp16 = einsum(equation = var_2744_equation_0, values = (var_2640_cast_fp16_11, var_2712_cast_fp16))[name = tensor("op_2744_cast_fp16")]; + tensor var_2746_equation_0 = const()[name = tensor("op_2746_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2746_cast_fp16 = einsum(equation = var_2746_equation_0, values = (var_2640_cast_fp16_12, var_2713_cast_fp16))[name = tensor("op_2746_cast_fp16")]; + tensor var_2748_equation_0 = const()[name = tensor("op_2748_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2748_cast_fp16 = einsum(equation = var_2748_equation_0, values = (var_2640_cast_fp16_13, var_2714_cast_fp16))[name = tensor("op_2748_cast_fp16")]; + tensor var_2750_equation_0 = const()[name = tensor("op_2750_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2750_cast_fp16 = einsum(equation = var_2750_equation_0, values = (var_2640_cast_fp16_14, var_2715_cast_fp16))[name = tensor("op_2750_cast_fp16")]; + tensor var_2752_equation_0 = const()[name = tensor("op_2752_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2752_cast_fp16 = einsum(equation = var_2752_equation_0, values = (var_2640_cast_fp16_15, var_2716_cast_fp16))[name = tensor("op_2752_cast_fp16")]; + tensor var_2754_equation_0 = const()[name = tensor("op_2754_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2754_cast_fp16 = einsum(equation = var_2754_equation_0, values = (var_2640_cast_fp16_16, var_2717_cast_fp16))[name = tensor("op_2754_cast_fp16")]; + tensor var_2756_equation_0 = const()[name = tensor("op_2756_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2756_cast_fp16 = einsum(equation = var_2756_equation_0, values = (var_2640_cast_fp16_17, var_2718_cast_fp16))[name = tensor("op_2756_cast_fp16")]; + tensor var_2758_equation_0 = const()[name = tensor("op_2758_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2758_cast_fp16 = einsum(equation = var_2758_equation_0, values = (var_2640_cast_fp16_18, var_2719_cast_fp16))[name = tensor("op_2758_cast_fp16")]; + tensor var_2760_equation_0 = const()[name = tensor("op_2760_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2760_cast_fp16 = einsum(equation = var_2760_equation_0, values = (var_2640_cast_fp16_19, var_2720_cast_fp16))[name = tensor("op_2760_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_2545, interleave = input_95_interleave_0, values = (var_2722_cast_fp16, var_2724_cast_fp16, var_2726_cast_fp16, var_2728_cast_fp16, var_2730_cast_fp16, var_2732_cast_fp16, var_2734_cast_fp16, var_2736_cast_fp16, var_2738_cast_fp16, var_2740_cast_fp16, var_2742_cast_fp16, var_2744_cast_fp16, var_2746_cast_fp16, var_2748_cast_fp16, var_2750_cast_fp16, var_2752_cast_fp16, var_2754_cast_fp16, var_2756_cast_fp16, var_2758_cast_fp16, var_2760_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2769_pad_type_0 = const()[name = tensor("op_2769_pad_type_0"), val = tensor("valid")]; + tensor var_2769_strides_0 = const()[name = tensor("op_2769_strides_0"), val = tensor([1, 1])]; + tensor var_2769_pad_0 = const()[name = tensor("op_2769_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2769_dilations_0 = const()[name = tensor("op_2769_dilations_0"), val = tensor([1, 1])]; + tensor var_2769_groups_0 = const()[name = tensor("op_2769_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378310912)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381587776)))]; + tensor var_2769_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2769_dilations_0, groups = var_2769_groups_0, pad = var_2769_pad_0, pad_type = var_2769_pad_type_0, strides = var_2769_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2769_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2769_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381590400)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381593024)))]; + tensor var_2779_to_fp16 = const()[name = tensor("op_2779_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2779_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381595648)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394702912)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2805_pad_type_0 = const()[name = tensor("op_2805_pad_type_0"), val = tensor("valid")]; + tensor var_2805_strides_0 = const()[name = tensor("op_2805_strides_0"), val = tensor([1, 1])]; + tensor var_2805_pad_0 = const()[name = tensor("op_2805_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2805_dilations_0 = const()[name = tensor("op_2805_dilations_0"), val = tensor([1, 1])]; + tensor var_2805_groups_0 = const()[name = tensor("op_2805_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394713216)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407820480)))]; + tensor var_2805_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2805_dilations_0, groups = var_2805_groups_0, pad = var_2805_pad_0, pad_type = var_2805_pad_type_0, strides = var_2805_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2805_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2805_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407823104)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407825728)))]; + tensor var_2830_to_fp16 = const()[name = tensor("op_2830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2830_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2865_weight_0_to_fp16 = const()[name = tensor("op_2865_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407828352)))]; + tensor var_2865_bias_0_to_fp16 = const()[name = tensor("op_2865_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411105216)))]; + tensor var_2865_cast_fp16 = conv(bias = var_2865_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2865_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411107840)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2863_pad_type_0 = const()[name = tensor("op_2863_pad_type_0"), val = tensor("valid")]; + tensor var_2863_strides_0 = const()[name = tensor("op_2863_strides_0"), val = tensor([1, 1])]; + tensor var_2863_pad_0 = const()[name = tensor("op_2863_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2863_dilations_0 = const()[name = tensor("op_2863_dilations_0"), val = tensor([1, 1])]; + tensor var_2863_groups_0 = const()[name = tensor("op_2863_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414384704)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417661568)))]; + tensor var_2863_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2863_dilations_0, groups = var_2863_groups_0, pad = var_2863_pad_0, pad_type = var_2863_pad_type_0, strides = var_2863_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2863_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2866_axis_0 = const()[name = tensor("op_2866_axis_0"), val = tensor(1)]; + tensor var_2866_cast_fp16_0, tensor var_2866_cast_fp16_1, tensor var_2866_cast_fp16_2, tensor var_2866_cast_fp16_3, tensor var_2866_cast_fp16_4, tensor var_2866_cast_fp16_5, tensor var_2866_cast_fp16_6, tensor var_2866_cast_fp16_7, tensor var_2866_cast_fp16_8, tensor var_2866_cast_fp16_9, tensor var_2866_cast_fp16_10, tensor var_2866_cast_fp16_11, tensor var_2866_cast_fp16_12, tensor var_2866_cast_fp16_13, tensor var_2866_cast_fp16_14, tensor var_2866_cast_fp16_15, tensor var_2866_cast_fp16_16, tensor var_2866_cast_fp16_17, tensor var_2866_cast_fp16_18, tensor var_2866_cast_fp16_19 = split(axis = var_2866_axis_0, split_sizes = tile_30, x = var_2865_cast_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor var_2887_perm_0 = const()[name = tensor("op_2887_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2888_axis_0 = const()[name = tensor("op_2888_axis_0"), val = tensor(3)]; + tensor var_2887_cast_fp16 = transpose(perm = var_2887_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_22")]; + tensor var_2888_cast_fp16_0, tensor var_2888_cast_fp16_1, tensor var_2888_cast_fp16_2, tensor var_2888_cast_fp16_3, tensor var_2888_cast_fp16_4, tensor var_2888_cast_fp16_5, tensor var_2888_cast_fp16_6, tensor var_2888_cast_fp16_7, tensor var_2888_cast_fp16_8, tensor var_2888_cast_fp16_9, tensor var_2888_cast_fp16_10, tensor var_2888_cast_fp16_11, tensor var_2888_cast_fp16_12, tensor var_2888_cast_fp16_13, tensor var_2888_cast_fp16_14, tensor var_2888_cast_fp16_15, tensor var_2888_cast_fp16_16, tensor var_2888_cast_fp16_17, tensor var_2888_cast_fp16_18, tensor var_2888_cast_fp16_19 = split(axis = var_2888_axis_0, split_sizes = tile_31, x = var_2887_cast_fp16)[name = tensor("op_2888_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2909_axis_0 = const()[name = tensor("op_2909_axis_0"), val = tensor(1)]; + tensor var_2909_cast_fp16_0, tensor var_2909_cast_fp16_1, tensor var_2909_cast_fp16_2, tensor var_2909_cast_fp16_3, tensor var_2909_cast_fp16_4, tensor var_2909_cast_fp16_5, tensor var_2909_cast_fp16_6, tensor var_2909_cast_fp16_7, tensor var_2909_cast_fp16_8, tensor var_2909_cast_fp16_9, tensor var_2909_cast_fp16_10, tensor var_2909_cast_fp16_11, tensor var_2909_cast_fp16_12, tensor var_2909_cast_fp16_13, tensor var_2909_cast_fp16_14, tensor var_2909_cast_fp16_15, tensor var_2909_cast_fp16_16, tensor var_2909_cast_fp16_17, tensor var_2909_cast_fp16_18, tensor var_2909_cast_fp16_19 = split(axis = var_2909_axis_0, split_sizes = tile_32, x = var_2863_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor aw_401_equation_0 = const()[name = tensor("aw_401_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_401_cast_fp16 = einsum(equation = aw_401_equation_0, values = (var_2888_cast_fp16_0, var_2866_cast_fp16_0))[name = tensor("aw_401_cast_fp16")]; + tensor aw_403_equation_0 = const()[name = tensor("aw_403_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_403_cast_fp16 = einsum(equation = aw_403_equation_0, values = (var_2888_cast_fp16_1, var_2866_cast_fp16_1))[name = tensor("aw_403_cast_fp16")]; + tensor aw_405_equation_0 = const()[name = tensor("aw_405_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_405_cast_fp16 = einsum(equation = aw_405_equation_0, values = (var_2888_cast_fp16_2, var_2866_cast_fp16_2))[name = tensor("aw_405_cast_fp16")]; + tensor aw_407_equation_0 = const()[name = tensor("aw_407_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_407_cast_fp16 = einsum(equation = aw_407_equation_0, values = (var_2888_cast_fp16_3, var_2866_cast_fp16_3))[name = tensor("aw_407_cast_fp16")]; + tensor aw_409_equation_0 = const()[name = tensor("aw_409_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_409_cast_fp16 = einsum(equation = aw_409_equation_0, values = (var_2888_cast_fp16_4, var_2866_cast_fp16_4))[name = tensor("aw_409_cast_fp16")]; + tensor aw_411_equation_0 = const()[name = tensor("aw_411_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_411_cast_fp16 = einsum(equation = aw_411_equation_0, values = (var_2888_cast_fp16_5, var_2866_cast_fp16_5))[name = tensor("aw_411_cast_fp16")]; + tensor aw_413_equation_0 = const()[name = tensor("aw_413_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_413_cast_fp16 = einsum(equation = aw_413_equation_0, values = (var_2888_cast_fp16_6, var_2866_cast_fp16_6))[name = tensor("aw_413_cast_fp16")]; + tensor aw_415_equation_0 = const()[name = tensor("aw_415_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_415_cast_fp16 = einsum(equation = aw_415_equation_0, values = (var_2888_cast_fp16_7, var_2866_cast_fp16_7))[name = tensor("aw_415_cast_fp16")]; + tensor aw_417_equation_0 = const()[name = tensor("aw_417_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_417_cast_fp16 = einsum(equation = aw_417_equation_0, values = (var_2888_cast_fp16_8, var_2866_cast_fp16_8))[name = tensor("aw_417_cast_fp16")]; + tensor aw_419_equation_0 = const()[name = tensor("aw_419_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_419_cast_fp16 = einsum(equation = aw_419_equation_0, values = (var_2888_cast_fp16_9, var_2866_cast_fp16_9))[name = tensor("aw_419_cast_fp16")]; + tensor aw_421_equation_0 = const()[name = tensor("aw_421_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_421_cast_fp16 = einsum(equation = aw_421_equation_0, values = (var_2888_cast_fp16_10, var_2866_cast_fp16_10))[name = tensor("aw_421_cast_fp16")]; + tensor aw_423_equation_0 = const()[name = tensor("aw_423_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_423_cast_fp16 = einsum(equation = aw_423_equation_0, values = (var_2888_cast_fp16_11, var_2866_cast_fp16_11))[name = tensor("aw_423_cast_fp16")]; + tensor aw_425_equation_0 = const()[name = tensor("aw_425_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_425_cast_fp16 = einsum(equation = aw_425_equation_0, values = (var_2888_cast_fp16_12, var_2866_cast_fp16_12))[name = tensor("aw_425_cast_fp16")]; + tensor aw_427_equation_0 = const()[name = tensor("aw_427_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_427_cast_fp16 = einsum(equation = aw_427_equation_0, values = (var_2888_cast_fp16_13, var_2866_cast_fp16_13))[name = tensor("aw_427_cast_fp16")]; + tensor aw_429_equation_0 = const()[name = tensor("aw_429_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_429_cast_fp16 = einsum(equation = aw_429_equation_0, values = (var_2888_cast_fp16_14, var_2866_cast_fp16_14))[name = tensor("aw_429_cast_fp16")]; + tensor aw_431_equation_0 = const()[name = tensor("aw_431_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_431_cast_fp16 = einsum(equation = aw_431_equation_0, values = (var_2888_cast_fp16_15, var_2866_cast_fp16_15))[name = tensor("aw_431_cast_fp16")]; + tensor aw_433_equation_0 = const()[name = tensor("aw_433_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_433_cast_fp16 = einsum(equation = aw_433_equation_0, values = (var_2888_cast_fp16_16, var_2866_cast_fp16_16))[name = tensor("aw_433_cast_fp16")]; + tensor aw_435_equation_0 = const()[name = tensor("aw_435_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_435_cast_fp16 = einsum(equation = aw_435_equation_0, values = (var_2888_cast_fp16_17, var_2866_cast_fp16_17))[name = tensor("aw_435_cast_fp16")]; + tensor aw_437_equation_0 = const()[name = tensor("aw_437_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_437_cast_fp16 = einsum(equation = aw_437_equation_0, values = (var_2888_cast_fp16_18, var_2866_cast_fp16_18))[name = tensor("aw_437_cast_fp16")]; + tensor aw_439_equation_0 = const()[name = tensor("aw_439_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_439_cast_fp16 = einsum(equation = aw_439_equation_0, values = (var_2888_cast_fp16_19, var_2866_cast_fp16_19))[name = tensor("aw_439_cast_fp16")]; + tensor var_2970_cast_fp16 = softmax(axis = var_2814, x = aw_401_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = softmax(axis = var_2814, x = aw_403_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_cast_fp16 = softmax(axis = var_2814, x = aw_405_cast_fp16)[name = tensor("op_2972_cast_fp16")]; + tensor var_2973_cast_fp16 = softmax(axis = var_2814, x = aw_407_cast_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor var_2974_cast_fp16 = softmax(axis = var_2814, x = aw_409_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = softmax(axis = var_2814, x = aw_411_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor var_2976_cast_fp16 = softmax(axis = var_2814, x = aw_413_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor var_2977_cast_fp16 = softmax(axis = var_2814, x = aw_415_cast_fp16)[name = tensor("op_2977_cast_fp16")]; + tensor var_2978_cast_fp16 = softmax(axis = var_2814, x = aw_417_cast_fp16)[name = tensor("op_2978_cast_fp16")]; + tensor var_2979_cast_fp16 = softmax(axis = var_2814, x = aw_419_cast_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor var_2980_cast_fp16 = softmax(axis = var_2814, x = aw_421_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor var_2981_cast_fp16 = softmax(axis = var_2814, x = aw_423_cast_fp16)[name = tensor("op_2981_cast_fp16")]; + tensor var_2982_cast_fp16 = softmax(axis = var_2814, x = aw_425_cast_fp16)[name = tensor("op_2982_cast_fp16")]; + tensor var_2983_cast_fp16 = softmax(axis = var_2814, x = aw_427_cast_fp16)[name = tensor("op_2983_cast_fp16")]; + tensor var_2984_cast_fp16 = softmax(axis = var_2814, x = aw_429_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor var_2985_cast_fp16 = softmax(axis = var_2814, x = aw_431_cast_fp16)[name = tensor("op_2985_cast_fp16")]; + tensor var_2986_cast_fp16 = softmax(axis = var_2814, x = aw_433_cast_fp16)[name = tensor("op_2986_cast_fp16")]; + tensor var_2987_cast_fp16 = softmax(axis = var_2814, x = aw_435_cast_fp16)[name = tensor("op_2987_cast_fp16")]; + tensor var_2988_cast_fp16 = softmax(axis = var_2814, x = aw_437_cast_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2989_cast_fp16 = softmax(axis = var_2814, x = aw_439_cast_fp16)[name = tensor("op_2989_cast_fp16")]; + tensor var_2991_equation_0 = const()[name = tensor("op_2991_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2991_cast_fp16 = einsum(equation = var_2991_equation_0, values = (var_2909_cast_fp16_0, var_2970_cast_fp16))[name = tensor("op_2991_cast_fp16")]; + tensor var_2993_equation_0 = const()[name = tensor("op_2993_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2993_cast_fp16 = einsum(equation = var_2993_equation_0, values = (var_2909_cast_fp16_1, var_2971_cast_fp16))[name = tensor("op_2993_cast_fp16")]; + tensor var_2995_equation_0 = const()[name = tensor("op_2995_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2995_cast_fp16 = einsum(equation = var_2995_equation_0, values = (var_2909_cast_fp16_2, var_2972_cast_fp16))[name = tensor("op_2995_cast_fp16")]; + tensor var_2997_equation_0 = const()[name = tensor("op_2997_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2997_cast_fp16 = einsum(equation = var_2997_equation_0, values = (var_2909_cast_fp16_3, var_2973_cast_fp16))[name = tensor("op_2997_cast_fp16")]; + tensor var_2999_equation_0 = const()[name = tensor("op_2999_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2999_cast_fp16 = einsum(equation = var_2999_equation_0, values = (var_2909_cast_fp16_4, var_2974_cast_fp16))[name = tensor("op_2999_cast_fp16")]; + tensor var_3001_equation_0 = const()[name = tensor("op_3001_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3001_cast_fp16 = einsum(equation = var_3001_equation_0, values = (var_2909_cast_fp16_5, var_2975_cast_fp16))[name = tensor("op_3001_cast_fp16")]; + tensor var_3003_equation_0 = const()[name = tensor("op_3003_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3003_cast_fp16 = einsum(equation = var_3003_equation_0, values = (var_2909_cast_fp16_6, var_2976_cast_fp16))[name = tensor("op_3003_cast_fp16")]; + tensor var_3005_equation_0 = const()[name = tensor("op_3005_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3005_cast_fp16 = einsum(equation = var_3005_equation_0, values = (var_2909_cast_fp16_7, var_2977_cast_fp16))[name = tensor("op_3005_cast_fp16")]; + tensor var_3007_equation_0 = const()[name = tensor("op_3007_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3007_cast_fp16 = einsum(equation = var_3007_equation_0, values = (var_2909_cast_fp16_8, var_2978_cast_fp16))[name = tensor("op_3007_cast_fp16")]; + tensor var_3009_equation_0 = const()[name = tensor("op_3009_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3009_cast_fp16 = einsum(equation = var_3009_equation_0, values = (var_2909_cast_fp16_9, var_2979_cast_fp16))[name = tensor("op_3009_cast_fp16")]; + tensor var_3011_equation_0 = const()[name = tensor("op_3011_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3011_cast_fp16 = einsum(equation = var_3011_equation_0, values = (var_2909_cast_fp16_10, var_2980_cast_fp16))[name = tensor("op_3011_cast_fp16")]; + tensor var_3013_equation_0 = const()[name = tensor("op_3013_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3013_cast_fp16 = einsum(equation = var_3013_equation_0, values = (var_2909_cast_fp16_11, var_2981_cast_fp16))[name = tensor("op_3013_cast_fp16")]; + tensor var_3015_equation_0 = const()[name = tensor("op_3015_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3015_cast_fp16 = einsum(equation = var_3015_equation_0, values = (var_2909_cast_fp16_12, var_2982_cast_fp16))[name = tensor("op_3015_cast_fp16")]; + tensor var_3017_equation_0 = const()[name = tensor("op_3017_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3017_cast_fp16 = einsum(equation = var_3017_equation_0, values = (var_2909_cast_fp16_13, var_2983_cast_fp16))[name = tensor("op_3017_cast_fp16")]; + tensor var_3019_equation_0 = const()[name = tensor("op_3019_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3019_cast_fp16 = einsum(equation = var_3019_equation_0, values = (var_2909_cast_fp16_14, var_2984_cast_fp16))[name = tensor("op_3019_cast_fp16")]; + tensor var_3021_equation_0 = const()[name = tensor("op_3021_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3021_cast_fp16 = einsum(equation = var_3021_equation_0, values = (var_2909_cast_fp16_15, var_2985_cast_fp16))[name = tensor("op_3021_cast_fp16")]; + tensor var_3023_equation_0 = const()[name = tensor("op_3023_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3023_cast_fp16 = einsum(equation = var_3023_equation_0, values = (var_2909_cast_fp16_16, var_2986_cast_fp16))[name = tensor("op_3023_cast_fp16")]; + tensor var_3025_equation_0 = const()[name = tensor("op_3025_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3025_cast_fp16 = einsum(equation = var_3025_equation_0, values = (var_2909_cast_fp16_17, var_2987_cast_fp16))[name = tensor("op_3025_cast_fp16")]; + tensor var_3027_equation_0 = const()[name = tensor("op_3027_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3027_cast_fp16 = einsum(equation = var_3027_equation_0, values = (var_2909_cast_fp16_18, var_2988_cast_fp16))[name = tensor("op_3027_cast_fp16")]; + tensor var_3029_equation_0 = const()[name = tensor("op_3029_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3029_cast_fp16 = einsum(equation = var_3029_equation_0, values = (var_2909_cast_fp16_19, var_2989_cast_fp16))[name = tensor("op_3029_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2814, interleave = input_105_interleave_0, values = (var_2991_cast_fp16, var_2993_cast_fp16, var_2995_cast_fp16, var_2997_cast_fp16, var_2999_cast_fp16, var_3001_cast_fp16, var_3003_cast_fp16, var_3005_cast_fp16, var_3007_cast_fp16, var_3009_cast_fp16, var_3011_cast_fp16, var_3013_cast_fp16, var_3015_cast_fp16, var_3017_cast_fp16, var_3019_cast_fp16, var_3021_cast_fp16, var_3023_cast_fp16, var_3025_cast_fp16, var_3027_cast_fp16, var_3029_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_3038_pad_type_0 = const()[name = tensor("op_3038_pad_type_0"), val = tensor("valid")]; + tensor var_3038_strides_0 = const()[name = tensor("op_3038_strides_0"), val = tensor([1, 1])]; + tensor var_3038_pad_0 = const()[name = tensor("op_3038_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3038_dilations_0 = const()[name = tensor("op_3038_dilations_0"), val = tensor([1, 1])]; + tensor var_3038_groups_0 = const()[name = tensor("op_3038_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417664192)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420941056)))]; + tensor var_3038_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_3038_dilations_0, groups = var_3038_groups_0, pad = var_3038_pad_0, pad_type = var_3038_pad_type_0, strides = var_3038_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_3038_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420943680)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420946304)))]; + tensor var_3048_to_fp16 = const()[name = tensor("op_3048_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_3048_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420948928)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434056192)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_3074_pad_type_0 = const()[name = tensor("op_3074_pad_type_0"), val = tensor("valid")]; + tensor var_3074_strides_0 = const()[name = tensor("op_3074_strides_0"), val = tensor([1, 1])]; + tensor var_3074_pad_0 = const()[name = tensor("op_3074_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3074_dilations_0 = const()[name = tensor("op_3074_dilations_0"), val = tensor([1, 1])]; + tensor var_3074_groups_0 = const()[name = tensor("op_3074_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434066496)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447173760)))]; + tensor var_3074_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_3074_dilations_0, groups = var_3074_groups_0, pad = var_3074_pad_0, pad_type = var_3074_pad_type_0, strides = var_3074_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_3074_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_3074_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447176384)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447179008)))]; + tensor var_3099_to_fp16 = const()[name = tensor("op_3099_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_3099_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_23_pad_type_0 = const()[name = tensor("q_23_pad_type_0"), val = tensor("valid")]; + tensor q_23_strides_0 = const()[name = tensor("q_23_strides_0"), val = tensor([1, 1])]; + tensor q_23_pad_0 = const()[name = tensor("q_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_23_dilations_0 = const()[name = tensor("q_23_dilations_0"), val = tensor([1, 1])]; + tensor q_23_groups_0 = const()[name = tensor("q_23_groups_0"), val = tensor(1)]; + tensor var_3134_weight_0_to_fp16 = const()[name = tensor("op_3134_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447181632)))]; + tensor var_3134_bias_0_to_fp16 = const()[name = tensor("op_3134_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450458496)))]; + tensor var_3134_cast_fp16 = conv(bias = var_3134_bias_0_to_fp16, dilations = q_23_dilations_0, groups = q_23_groups_0, pad = q_23_pad_0, pad_type = q_23_pad_type_0, strides = q_23_strides_0, weight = var_3134_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor k_23_pad_type_0 = const()[name = tensor("k_23_pad_type_0"), val = tensor("valid")]; + tensor k_23_strides_0 = const()[name = tensor("k_23_strides_0"), val = tensor([1, 1])]; + tensor k_23_pad_0 = const()[name = tensor("k_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_23_dilations_0 = const()[name = tensor("k_23_dilations_0"), val = tensor([1, 1])]; + tensor k_23_groups_0 = const()[name = tensor("k_23_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450461120)))]; + tensor k_23_cast_fp16 = conv(dilations = k_23_dilations_0, groups = k_23_groups_0, pad = k_23_pad_0, pad_type = k_23_pad_type_0, strides = k_23_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_3132_pad_type_0 = const()[name = tensor("op_3132_pad_type_0"), val = tensor("valid")]; + tensor var_3132_strides_0 = const()[name = tensor("op_3132_strides_0"), val = tensor([1, 1])]; + tensor var_3132_pad_0 = const()[name = tensor("op_3132_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3132_dilations_0 = const()[name = tensor("op_3132_dilations_0"), val = tensor([1, 1])]; + tensor var_3132_groups_0 = const()[name = tensor("op_3132_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453737984)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457014848)))]; + tensor var_3132_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_3132_dilations_0, groups = var_3132_groups_0, pad = var_3132_pad_0, pad_type = var_3132_pad_type_0, strides = var_3132_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_3132_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3135_axis_0 = const()[name = tensor("op_3135_axis_0"), val = tensor(1)]; + tensor var_3135_cast_fp16_0, tensor var_3135_cast_fp16_1, tensor var_3135_cast_fp16_2, tensor var_3135_cast_fp16_3, tensor var_3135_cast_fp16_4, tensor var_3135_cast_fp16_5, tensor var_3135_cast_fp16_6, tensor var_3135_cast_fp16_7, tensor var_3135_cast_fp16_8, tensor var_3135_cast_fp16_9, tensor var_3135_cast_fp16_10, tensor var_3135_cast_fp16_11, tensor var_3135_cast_fp16_12, tensor var_3135_cast_fp16_13, tensor var_3135_cast_fp16_14, tensor var_3135_cast_fp16_15, tensor var_3135_cast_fp16_16, tensor var_3135_cast_fp16_17, tensor var_3135_cast_fp16_18, tensor var_3135_cast_fp16_19 = split(axis = var_3135_axis_0, split_sizes = tile_33, x = var_3134_cast_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3156_perm_0 = const()[name = tensor("op_3156_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3157_axis_0 = const()[name = tensor("op_3157_axis_0"), val = tensor(3)]; + tensor var_3156_cast_fp16 = transpose(perm = var_3156_perm_0, x = k_23_cast_fp16)[name = tensor("transpose_21")]; + tensor var_3157_cast_fp16_0, tensor var_3157_cast_fp16_1, tensor var_3157_cast_fp16_2, tensor var_3157_cast_fp16_3, tensor var_3157_cast_fp16_4, tensor var_3157_cast_fp16_5, tensor var_3157_cast_fp16_6, tensor var_3157_cast_fp16_7, tensor var_3157_cast_fp16_8, tensor var_3157_cast_fp16_9, tensor var_3157_cast_fp16_10, tensor var_3157_cast_fp16_11, tensor var_3157_cast_fp16_12, tensor var_3157_cast_fp16_13, tensor var_3157_cast_fp16_14, tensor var_3157_cast_fp16_15, tensor var_3157_cast_fp16_16, tensor var_3157_cast_fp16_17, tensor var_3157_cast_fp16_18, tensor var_3157_cast_fp16_19 = split(axis = var_3157_axis_0, split_sizes = tile_34, x = var_3156_cast_fp16)[name = tensor("op_3157_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3178_axis_0 = const()[name = tensor("op_3178_axis_0"), val = tensor(1)]; + tensor var_3178_cast_fp16_0, tensor var_3178_cast_fp16_1, tensor var_3178_cast_fp16_2, tensor var_3178_cast_fp16_3, tensor var_3178_cast_fp16_4, tensor var_3178_cast_fp16_5, tensor var_3178_cast_fp16_6, tensor var_3178_cast_fp16_7, tensor var_3178_cast_fp16_8, tensor var_3178_cast_fp16_9, tensor var_3178_cast_fp16_10, tensor var_3178_cast_fp16_11, tensor var_3178_cast_fp16_12, tensor var_3178_cast_fp16_13, tensor var_3178_cast_fp16_14, tensor var_3178_cast_fp16_15, tensor var_3178_cast_fp16_16, tensor var_3178_cast_fp16_17, tensor var_3178_cast_fp16_18, tensor var_3178_cast_fp16_19 = split(axis = var_3178_axis_0, split_sizes = tile_35, x = var_3132_cast_fp16)[name = tensor("op_3178_cast_fp16")]; + tensor aw_441_equation_0 = const()[name = tensor("aw_441_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_441_cast_fp16 = einsum(equation = aw_441_equation_0, values = (var_3157_cast_fp16_0, var_3135_cast_fp16_0))[name = tensor("aw_441_cast_fp16")]; + tensor aw_443_equation_0 = const()[name = tensor("aw_443_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_443_cast_fp16 = einsum(equation = aw_443_equation_0, values = (var_3157_cast_fp16_1, var_3135_cast_fp16_1))[name = tensor("aw_443_cast_fp16")]; + tensor aw_445_equation_0 = const()[name = tensor("aw_445_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_445_cast_fp16 = einsum(equation = aw_445_equation_0, values = (var_3157_cast_fp16_2, var_3135_cast_fp16_2))[name = tensor("aw_445_cast_fp16")]; + tensor aw_447_equation_0 = const()[name = tensor("aw_447_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_447_cast_fp16 = einsum(equation = aw_447_equation_0, values = (var_3157_cast_fp16_3, var_3135_cast_fp16_3))[name = tensor("aw_447_cast_fp16")]; + tensor aw_449_equation_0 = const()[name = tensor("aw_449_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_449_cast_fp16 = einsum(equation = aw_449_equation_0, values = (var_3157_cast_fp16_4, var_3135_cast_fp16_4))[name = tensor("aw_449_cast_fp16")]; + tensor aw_451_equation_0 = const()[name = tensor("aw_451_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_451_cast_fp16 = einsum(equation = aw_451_equation_0, values = (var_3157_cast_fp16_5, var_3135_cast_fp16_5))[name = tensor("aw_451_cast_fp16")]; + tensor aw_453_equation_0 = const()[name = tensor("aw_453_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_453_cast_fp16 = einsum(equation = aw_453_equation_0, values = (var_3157_cast_fp16_6, var_3135_cast_fp16_6))[name = tensor("aw_453_cast_fp16")]; + tensor aw_455_equation_0 = const()[name = tensor("aw_455_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_455_cast_fp16 = einsum(equation = aw_455_equation_0, values = (var_3157_cast_fp16_7, var_3135_cast_fp16_7))[name = tensor("aw_455_cast_fp16")]; + tensor aw_457_equation_0 = const()[name = tensor("aw_457_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_457_cast_fp16 = einsum(equation = aw_457_equation_0, values = (var_3157_cast_fp16_8, var_3135_cast_fp16_8))[name = tensor("aw_457_cast_fp16")]; + tensor aw_459_equation_0 = const()[name = tensor("aw_459_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_459_cast_fp16 = einsum(equation = aw_459_equation_0, values = (var_3157_cast_fp16_9, var_3135_cast_fp16_9))[name = tensor("aw_459_cast_fp16")]; + tensor aw_461_equation_0 = const()[name = tensor("aw_461_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_461_cast_fp16 = einsum(equation = aw_461_equation_0, values = (var_3157_cast_fp16_10, var_3135_cast_fp16_10))[name = tensor("aw_461_cast_fp16")]; + tensor aw_463_equation_0 = const()[name = tensor("aw_463_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_463_cast_fp16 = einsum(equation = aw_463_equation_0, values = (var_3157_cast_fp16_11, var_3135_cast_fp16_11))[name = tensor("aw_463_cast_fp16")]; + tensor aw_465_equation_0 = const()[name = tensor("aw_465_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_465_cast_fp16 = einsum(equation = aw_465_equation_0, values = (var_3157_cast_fp16_12, var_3135_cast_fp16_12))[name = tensor("aw_465_cast_fp16")]; + tensor aw_467_equation_0 = const()[name = tensor("aw_467_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_467_cast_fp16 = einsum(equation = aw_467_equation_0, values = (var_3157_cast_fp16_13, var_3135_cast_fp16_13))[name = tensor("aw_467_cast_fp16")]; + tensor aw_469_equation_0 = const()[name = tensor("aw_469_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_469_cast_fp16 = einsum(equation = aw_469_equation_0, values = (var_3157_cast_fp16_14, var_3135_cast_fp16_14))[name = tensor("aw_469_cast_fp16")]; + tensor aw_471_equation_0 = const()[name = tensor("aw_471_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_471_cast_fp16 = einsum(equation = aw_471_equation_0, values = (var_3157_cast_fp16_15, var_3135_cast_fp16_15))[name = tensor("aw_471_cast_fp16")]; + tensor aw_473_equation_0 = const()[name = tensor("aw_473_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_473_cast_fp16 = einsum(equation = aw_473_equation_0, values = (var_3157_cast_fp16_16, var_3135_cast_fp16_16))[name = tensor("aw_473_cast_fp16")]; + tensor aw_475_equation_0 = const()[name = tensor("aw_475_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_475_cast_fp16 = einsum(equation = aw_475_equation_0, values = (var_3157_cast_fp16_17, var_3135_cast_fp16_17))[name = tensor("aw_475_cast_fp16")]; + tensor aw_477_equation_0 = const()[name = tensor("aw_477_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_477_cast_fp16 = einsum(equation = aw_477_equation_0, values = (var_3157_cast_fp16_18, var_3135_cast_fp16_18))[name = tensor("aw_477_cast_fp16")]; + tensor aw_479_equation_0 = const()[name = tensor("aw_479_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_479_cast_fp16 = einsum(equation = aw_479_equation_0, values = (var_3157_cast_fp16_19, var_3135_cast_fp16_19))[name = tensor("aw_479_cast_fp16")]; + tensor var_3239_cast_fp16 = softmax(axis = var_3083, x = aw_441_cast_fp16)[name = tensor("op_3239_cast_fp16")]; + tensor var_3240_cast_fp16 = softmax(axis = var_3083, x = aw_443_cast_fp16)[name = tensor("op_3240_cast_fp16")]; + tensor var_3241_cast_fp16 = softmax(axis = var_3083, x = aw_445_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor var_3242_cast_fp16 = softmax(axis = var_3083, x = aw_447_cast_fp16)[name = tensor("op_3242_cast_fp16")]; + tensor var_3243_cast_fp16 = softmax(axis = var_3083, x = aw_449_cast_fp16)[name = tensor("op_3243_cast_fp16")]; + tensor var_3244_cast_fp16 = softmax(axis = var_3083, x = aw_451_cast_fp16)[name = tensor("op_3244_cast_fp16")]; + tensor var_3245_cast_fp16 = softmax(axis = var_3083, x = aw_453_cast_fp16)[name = tensor("op_3245_cast_fp16")]; + tensor var_3246_cast_fp16 = softmax(axis = var_3083, x = aw_455_cast_fp16)[name = tensor("op_3246_cast_fp16")]; + tensor var_3247_cast_fp16 = softmax(axis = var_3083, x = aw_457_cast_fp16)[name = tensor("op_3247_cast_fp16")]; + tensor var_3248_cast_fp16 = softmax(axis = var_3083, x = aw_459_cast_fp16)[name = tensor("op_3248_cast_fp16")]; + tensor var_3249_cast_fp16 = softmax(axis = var_3083, x = aw_461_cast_fp16)[name = tensor("op_3249_cast_fp16")]; + tensor var_3250_cast_fp16 = softmax(axis = var_3083, x = aw_463_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = softmax(axis = var_3083, x = aw_465_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor var_3252_cast_fp16 = softmax(axis = var_3083, x = aw_467_cast_fp16)[name = tensor("op_3252_cast_fp16")]; + tensor var_3253_cast_fp16 = softmax(axis = var_3083, x = aw_469_cast_fp16)[name = tensor("op_3253_cast_fp16")]; + tensor var_3254_cast_fp16 = softmax(axis = var_3083, x = aw_471_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = softmax(axis = var_3083, x = aw_473_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor var_3256_cast_fp16 = softmax(axis = var_3083, x = aw_475_cast_fp16)[name = tensor("op_3256_cast_fp16")]; + tensor var_3257_cast_fp16 = softmax(axis = var_3083, x = aw_477_cast_fp16)[name = tensor("op_3257_cast_fp16")]; + tensor var_3258_cast_fp16 = softmax(axis = var_3083, x = aw_479_cast_fp16)[name = tensor("op_3258_cast_fp16")]; + tensor var_3260_equation_0 = const()[name = tensor("op_3260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3260_cast_fp16 = einsum(equation = var_3260_equation_0, values = (var_3178_cast_fp16_0, var_3239_cast_fp16))[name = tensor("op_3260_cast_fp16")]; + tensor var_3262_equation_0 = const()[name = tensor("op_3262_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3262_cast_fp16 = einsum(equation = var_3262_equation_0, values = (var_3178_cast_fp16_1, var_3240_cast_fp16))[name = tensor("op_3262_cast_fp16")]; + tensor var_3264_equation_0 = const()[name = tensor("op_3264_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3264_cast_fp16 = einsum(equation = var_3264_equation_0, values = (var_3178_cast_fp16_2, var_3241_cast_fp16))[name = tensor("op_3264_cast_fp16")]; + tensor var_3266_equation_0 = const()[name = tensor("op_3266_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3266_cast_fp16 = einsum(equation = var_3266_equation_0, values = (var_3178_cast_fp16_3, var_3242_cast_fp16))[name = tensor("op_3266_cast_fp16")]; + tensor var_3268_equation_0 = const()[name = tensor("op_3268_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3268_cast_fp16 = einsum(equation = var_3268_equation_0, values = (var_3178_cast_fp16_4, var_3243_cast_fp16))[name = tensor("op_3268_cast_fp16")]; + tensor var_3270_equation_0 = const()[name = tensor("op_3270_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3270_cast_fp16 = einsum(equation = var_3270_equation_0, values = (var_3178_cast_fp16_5, var_3244_cast_fp16))[name = tensor("op_3270_cast_fp16")]; + tensor var_3272_equation_0 = const()[name = tensor("op_3272_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3272_cast_fp16 = einsum(equation = var_3272_equation_0, values = (var_3178_cast_fp16_6, var_3245_cast_fp16))[name = tensor("op_3272_cast_fp16")]; + tensor var_3274_equation_0 = const()[name = tensor("op_3274_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3274_cast_fp16 = einsum(equation = var_3274_equation_0, values = (var_3178_cast_fp16_7, var_3246_cast_fp16))[name = tensor("op_3274_cast_fp16")]; + tensor var_3276_equation_0 = const()[name = tensor("op_3276_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3276_cast_fp16 = einsum(equation = var_3276_equation_0, values = (var_3178_cast_fp16_8, var_3247_cast_fp16))[name = tensor("op_3276_cast_fp16")]; + tensor var_3278_equation_0 = const()[name = tensor("op_3278_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3278_cast_fp16 = einsum(equation = var_3278_equation_0, values = (var_3178_cast_fp16_9, var_3248_cast_fp16))[name = tensor("op_3278_cast_fp16")]; + tensor var_3280_equation_0 = const()[name = tensor("op_3280_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3280_cast_fp16 = einsum(equation = var_3280_equation_0, values = (var_3178_cast_fp16_10, var_3249_cast_fp16))[name = tensor("op_3280_cast_fp16")]; + tensor var_3282_equation_0 = const()[name = tensor("op_3282_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3282_cast_fp16 = einsum(equation = var_3282_equation_0, values = (var_3178_cast_fp16_11, var_3250_cast_fp16))[name = tensor("op_3282_cast_fp16")]; + tensor var_3284_equation_0 = const()[name = tensor("op_3284_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3284_cast_fp16 = einsum(equation = var_3284_equation_0, values = (var_3178_cast_fp16_12, var_3251_cast_fp16))[name = tensor("op_3284_cast_fp16")]; + tensor var_3286_equation_0 = const()[name = tensor("op_3286_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3286_cast_fp16 = einsum(equation = var_3286_equation_0, values = (var_3178_cast_fp16_13, var_3252_cast_fp16))[name = tensor("op_3286_cast_fp16")]; + tensor var_3288_equation_0 = const()[name = tensor("op_3288_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3288_cast_fp16 = einsum(equation = var_3288_equation_0, values = (var_3178_cast_fp16_14, var_3253_cast_fp16))[name = tensor("op_3288_cast_fp16")]; + tensor var_3290_equation_0 = const()[name = tensor("op_3290_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3290_cast_fp16 = einsum(equation = var_3290_equation_0, values = (var_3178_cast_fp16_15, var_3254_cast_fp16))[name = tensor("op_3290_cast_fp16")]; + tensor var_3292_equation_0 = const()[name = tensor("op_3292_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3292_cast_fp16 = einsum(equation = var_3292_equation_0, values = (var_3178_cast_fp16_16, var_3255_cast_fp16))[name = tensor("op_3292_cast_fp16")]; + tensor var_3294_equation_0 = const()[name = tensor("op_3294_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3294_cast_fp16 = einsum(equation = var_3294_equation_0, values = (var_3178_cast_fp16_17, var_3256_cast_fp16))[name = tensor("op_3294_cast_fp16")]; + tensor var_3296_equation_0 = const()[name = tensor("op_3296_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3296_cast_fp16 = einsum(equation = var_3296_equation_0, values = (var_3178_cast_fp16_18, var_3257_cast_fp16))[name = tensor("op_3296_cast_fp16")]; + tensor var_3298_equation_0 = const()[name = tensor("op_3298_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3298_cast_fp16 = einsum(equation = var_3298_equation_0, values = (var_3178_cast_fp16_19, var_3258_cast_fp16))[name = tensor("op_3298_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_3083, interleave = input_115_interleave_0, values = (var_3260_cast_fp16, var_3262_cast_fp16, var_3264_cast_fp16, var_3266_cast_fp16, var_3268_cast_fp16, var_3270_cast_fp16, var_3272_cast_fp16, var_3274_cast_fp16, var_3276_cast_fp16, var_3278_cast_fp16, var_3280_cast_fp16, var_3282_cast_fp16, var_3284_cast_fp16, var_3286_cast_fp16, var_3288_cast_fp16, var_3290_cast_fp16, var_3292_cast_fp16, var_3294_cast_fp16, var_3296_cast_fp16, var_3298_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_3307_pad_type_0 = const()[name = tensor("op_3307_pad_type_0"), val = tensor("valid")]; + tensor var_3307_strides_0 = const()[name = tensor("op_3307_strides_0"), val = tensor([1, 1])]; + tensor var_3307_pad_0 = const()[name = tensor("op_3307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3307_dilations_0 = const()[name = tensor("op_3307_dilations_0"), val = tensor([1, 1])]; + tensor var_3307_groups_0 = const()[name = tensor("op_3307_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457017472)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460294336)))]; + tensor var_3307_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_3307_dilations_0, groups = var_3307_groups_0, pad = var_3307_pad_0, pad_type = var_3307_pad_type_0, strides = var_3307_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_3307_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460296960)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460299584)))]; + tensor var_3317_to_fp16 = const()[name = tensor("op_3317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_3317_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460302208)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473409472)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_3343_pad_type_0 = const()[name = tensor("op_3343_pad_type_0"), val = tensor("valid")]; + tensor var_3343_strides_0 = const()[name = tensor("op_3343_strides_0"), val = tensor([1, 1])]; + tensor var_3343_pad_0 = const()[name = tensor("op_3343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3343_dilations_0 = const()[name = tensor("op_3343_dilations_0"), val = tensor([1, 1])]; + tensor var_3343_groups_0 = const()[name = tensor("op_3343_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473419776)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486527040)))]; + tensor var_3343_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_3343_dilations_0, groups = var_3343_groups_0, pad = var_3343_pad_0, pad_type = var_3343_pad_type_0, strides = var_3343_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("op_3343_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_3343_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor(1)]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([1])]; + tensor input_123_gamma_0_to_fp16 = const()[name = tensor("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486529664)))]; + tensor input_123_beta_0_to_fp16 = const()[name = tensor("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486532288)))]; + tensor var_3368_to_fp16 = const()[name = tensor("op_3368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = input_123_beta_0_to_fp16, epsilon = var_3368_to_fp16, gamma = input_123_gamma_0_to_fp16, x = inputs_49_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor q_25_pad_type_0 = const()[name = tensor("q_25_pad_type_0"), val = tensor("valid")]; + tensor q_25_strides_0 = const()[name = tensor("q_25_strides_0"), val = tensor([1, 1])]; + tensor q_25_pad_0 = const()[name = tensor("q_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_25_dilations_0 = const()[name = tensor("q_25_dilations_0"), val = tensor([1, 1])]; + tensor q_25_groups_0 = const()[name = tensor("q_25_groups_0"), val = tensor(1)]; + tensor var_3403_weight_0_to_fp16 = const()[name = tensor("op_3403_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486534912)))]; + tensor var_3403_bias_0_to_fp16 = const()[name = tensor("op_3403_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489811776)))]; + tensor var_3403_cast_fp16 = conv(bias = var_3403_bias_0_to_fp16, dilations = q_25_dilations_0, groups = q_25_groups_0, pad = q_25_pad_0, pad_type = q_25_pad_type_0, strides = q_25_strides_0, weight = var_3403_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3403_cast_fp16")]; + tensor k_25_pad_type_0 = const()[name = tensor("k_25_pad_type_0"), val = tensor("valid")]; + tensor k_25_strides_0 = const()[name = tensor("k_25_strides_0"), val = tensor([1, 1])]; + tensor k_25_pad_0 = const()[name = tensor("k_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_25_dilations_0 = const()[name = tensor("k_25_dilations_0"), val = tensor([1, 1])]; + tensor k_25_groups_0 = const()[name = tensor("k_25_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_key_weight_to_fp16 = const()[name = tensor("blocks_12_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489814400)))]; + tensor k_25_cast_fp16 = conv(dilations = k_25_dilations_0, groups = k_25_groups_0, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = k_25_strides_0, weight = blocks_12_attn_key_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor var_3401_pad_type_0 = const()[name = tensor("op_3401_pad_type_0"), val = tensor("valid")]; + tensor var_3401_strides_0 = const()[name = tensor("op_3401_strides_0"), val = tensor([1, 1])]; + tensor var_3401_pad_0 = const()[name = tensor("op_3401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3401_dilations_0 = const()[name = tensor("op_3401_dilations_0"), val = tensor([1, 1])]; + tensor var_3401_groups_0 = const()[name = tensor("op_3401_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_value_weight_to_fp16 = const()[name = tensor("blocks_12_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493091264)))]; + tensor blocks_12_attn_value_bias_to_fp16 = const()[name = tensor("blocks_12_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496368128)))]; + tensor var_3401_cast_fp16 = conv(bias = blocks_12_attn_value_bias_to_fp16, dilations = var_3401_dilations_0, groups = var_3401_groups_0, pad = var_3401_pad_0, pad_type = var_3401_pad_type_0, strides = var_3401_strides_0, weight = blocks_12_attn_value_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3401_cast_fp16")]; + tensor tile_36 = const()[name = tensor("tile_36"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3404_axis_0 = const()[name = tensor("op_3404_axis_0"), val = tensor(1)]; + tensor var_3404_cast_fp16_0, tensor var_3404_cast_fp16_1, tensor var_3404_cast_fp16_2, tensor var_3404_cast_fp16_3, tensor var_3404_cast_fp16_4, tensor var_3404_cast_fp16_5, tensor var_3404_cast_fp16_6, tensor var_3404_cast_fp16_7, tensor var_3404_cast_fp16_8, tensor var_3404_cast_fp16_9, tensor var_3404_cast_fp16_10, tensor var_3404_cast_fp16_11, tensor var_3404_cast_fp16_12, tensor var_3404_cast_fp16_13, tensor var_3404_cast_fp16_14, tensor var_3404_cast_fp16_15, tensor var_3404_cast_fp16_16, tensor var_3404_cast_fp16_17, tensor var_3404_cast_fp16_18, tensor var_3404_cast_fp16_19 = split(axis = var_3404_axis_0, split_sizes = tile_36, x = var_3403_cast_fp16)[name = tensor("op_3404_cast_fp16")]; + tensor var_3425_perm_0 = const()[name = tensor("op_3425_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_37 = const()[name = tensor("tile_37"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3426_axis_0 = const()[name = tensor("op_3426_axis_0"), val = tensor(3)]; + tensor var_3425_cast_fp16 = transpose(perm = var_3425_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_20")]; + tensor var_3426_cast_fp16_0, tensor var_3426_cast_fp16_1, tensor var_3426_cast_fp16_2, tensor var_3426_cast_fp16_3, tensor var_3426_cast_fp16_4, tensor var_3426_cast_fp16_5, tensor var_3426_cast_fp16_6, tensor var_3426_cast_fp16_7, tensor var_3426_cast_fp16_8, tensor var_3426_cast_fp16_9, tensor var_3426_cast_fp16_10, tensor var_3426_cast_fp16_11, tensor var_3426_cast_fp16_12, tensor var_3426_cast_fp16_13, tensor var_3426_cast_fp16_14, tensor var_3426_cast_fp16_15, tensor var_3426_cast_fp16_16, tensor var_3426_cast_fp16_17, tensor var_3426_cast_fp16_18, tensor var_3426_cast_fp16_19 = split(axis = var_3426_axis_0, split_sizes = tile_37, x = var_3425_cast_fp16)[name = tensor("op_3426_cast_fp16")]; + tensor tile_38 = const()[name = tensor("tile_38"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3447_axis_0 = const()[name = tensor("op_3447_axis_0"), val = tensor(1)]; + tensor var_3447_cast_fp16_0, tensor var_3447_cast_fp16_1, tensor var_3447_cast_fp16_2, tensor var_3447_cast_fp16_3, tensor var_3447_cast_fp16_4, tensor var_3447_cast_fp16_5, tensor var_3447_cast_fp16_6, tensor var_3447_cast_fp16_7, tensor var_3447_cast_fp16_8, tensor var_3447_cast_fp16_9, tensor var_3447_cast_fp16_10, tensor var_3447_cast_fp16_11, tensor var_3447_cast_fp16_12, tensor var_3447_cast_fp16_13, tensor var_3447_cast_fp16_14, tensor var_3447_cast_fp16_15, tensor var_3447_cast_fp16_16, tensor var_3447_cast_fp16_17, tensor var_3447_cast_fp16_18, tensor var_3447_cast_fp16_19 = split(axis = var_3447_axis_0, split_sizes = tile_38, x = var_3401_cast_fp16)[name = tensor("op_3447_cast_fp16")]; + tensor aw_481_equation_0 = const()[name = tensor("aw_481_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_481_cast_fp16 = einsum(equation = aw_481_equation_0, values = (var_3426_cast_fp16_0, var_3404_cast_fp16_0))[name = tensor("aw_481_cast_fp16")]; + tensor aw_483_equation_0 = const()[name = tensor("aw_483_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_483_cast_fp16 = einsum(equation = aw_483_equation_0, values = (var_3426_cast_fp16_1, var_3404_cast_fp16_1))[name = tensor("aw_483_cast_fp16")]; + tensor aw_485_equation_0 = const()[name = tensor("aw_485_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_485_cast_fp16 = einsum(equation = aw_485_equation_0, values = (var_3426_cast_fp16_2, var_3404_cast_fp16_2))[name = tensor("aw_485_cast_fp16")]; + tensor aw_487_equation_0 = const()[name = tensor("aw_487_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_487_cast_fp16 = einsum(equation = aw_487_equation_0, values = (var_3426_cast_fp16_3, var_3404_cast_fp16_3))[name = tensor("aw_487_cast_fp16")]; + tensor aw_489_equation_0 = const()[name = tensor("aw_489_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_489_cast_fp16 = einsum(equation = aw_489_equation_0, values = (var_3426_cast_fp16_4, var_3404_cast_fp16_4))[name = tensor("aw_489_cast_fp16")]; + tensor aw_491_equation_0 = const()[name = tensor("aw_491_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_491_cast_fp16 = einsum(equation = aw_491_equation_0, values = (var_3426_cast_fp16_5, var_3404_cast_fp16_5))[name = tensor("aw_491_cast_fp16")]; + tensor aw_493_equation_0 = const()[name = tensor("aw_493_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_493_cast_fp16 = einsum(equation = aw_493_equation_0, values = (var_3426_cast_fp16_6, var_3404_cast_fp16_6))[name = tensor("aw_493_cast_fp16")]; + tensor aw_495_equation_0 = const()[name = tensor("aw_495_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_495_cast_fp16 = einsum(equation = aw_495_equation_0, values = (var_3426_cast_fp16_7, var_3404_cast_fp16_7))[name = tensor("aw_495_cast_fp16")]; + tensor aw_497_equation_0 = const()[name = tensor("aw_497_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_497_cast_fp16 = einsum(equation = aw_497_equation_0, values = (var_3426_cast_fp16_8, var_3404_cast_fp16_8))[name = tensor("aw_497_cast_fp16")]; + tensor aw_499_equation_0 = const()[name = tensor("aw_499_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_499_cast_fp16 = einsum(equation = aw_499_equation_0, values = (var_3426_cast_fp16_9, var_3404_cast_fp16_9))[name = tensor("aw_499_cast_fp16")]; + tensor aw_501_equation_0 = const()[name = tensor("aw_501_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_501_cast_fp16 = einsum(equation = aw_501_equation_0, values = (var_3426_cast_fp16_10, var_3404_cast_fp16_10))[name = tensor("aw_501_cast_fp16")]; + tensor aw_503_equation_0 = const()[name = tensor("aw_503_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_503_cast_fp16 = einsum(equation = aw_503_equation_0, values = (var_3426_cast_fp16_11, var_3404_cast_fp16_11))[name = tensor("aw_503_cast_fp16")]; + tensor aw_505_equation_0 = const()[name = tensor("aw_505_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_505_cast_fp16 = einsum(equation = aw_505_equation_0, values = (var_3426_cast_fp16_12, var_3404_cast_fp16_12))[name = tensor("aw_505_cast_fp16")]; + tensor aw_507_equation_0 = const()[name = tensor("aw_507_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_507_cast_fp16 = einsum(equation = aw_507_equation_0, values = (var_3426_cast_fp16_13, var_3404_cast_fp16_13))[name = tensor("aw_507_cast_fp16")]; + tensor aw_509_equation_0 = const()[name = tensor("aw_509_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_509_cast_fp16 = einsum(equation = aw_509_equation_0, values = (var_3426_cast_fp16_14, var_3404_cast_fp16_14))[name = tensor("aw_509_cast_fp16")]; + tensor aw_511_equation_0 = const()[name = tensor("aw_511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_511_cast_fp16 = einsum(equation = aw_511_equation_0, values = (var_3426_cast_fp16_15, var_3404_cast_fp16_15))[name = tensor("aw_511_cast_fp16")]; + tensor aw_513_equation_0 = const()[name = tensor("aw_513_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_513_cast_fp16 = einsum(equation = aw_513_equation_0, values = (var_3426_cast_fp16_16, var_3404_cast_fp16_16))[name = tensor("aw_513_cast_fp16")]; + tensor aw_515_equation_0 = const()[name = tensor("aw_515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_515_cast_fp16 = einsum(equation = aw_515_equation_0, values = (var_3426_cast_fp16_17, var_3404_cast_fp16_17))[name = tensor("aw_515_cast_fp16")]; + tensor aw_517_equation_0 = const()[name = tensor("aw_517_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_517_cast_fp16 = einsum(equation = aw_517_equation_0, values = (var_3426_cast_fp16_18, var_3404_cast_fp16_18))[name = tensor("aw_517_cast_fp16")]; + tensor aw_519_equation_0 = const()[name = tensor("aw_519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_519_cast_fp16 = einsum(equation = aw_519_equation_0, values = (var_3426_cast_fp16_19, var_3404_cast_fp16_19))[name = tensor("aw_519_cast_fp16")]; + tensor var_3508_cast_fp16 = softmax(axis = var_3352, x = aw_481_cast_fp16)[name = tensor("op_3508_cast_fp16")]; + tensor var_3509_cast_fp16 = softmax(axis = var_3352, x = aw_483_cast_fp16)[name = tensor("op_3509_cast_fp16")]; + tensor var_3510_cast_fp16 = softmax(axis = var_3352, x = aw_485_cast_fp16)[name = tensor("op_3510_cast_fp16")]; + tensor var_3511_cast_fp16 = softmax(axis = var_3352, x = aw_487_cast_fp16)[name = tensor("op_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = softmax(axis = var_3352, x = aw_489_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_cast_fp16 = softmax(axis = var_3352, x = aw_491_cast_fp16)[name = tensor("op_3513_cast_fp16")]; + tensor var_3514_cast_fp16 = softmax(axis = var_3352, x = aw_493_cast_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor var_3515_cast_fp16 = softmax(axis = var_3352, x = aw_495_cast_fp16)[name = tensor("op_3515_cast_fp16")]; + tensor var_3516_cast_fp16 = softmax(axis = var_3352, x = aw_497_cast_fp16)[name = tensor("op_3516_cast_fp16")]; + tensor var_3517_cast_fp16 = softmax(axis = var_3352, x = aw_499_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3518_cast_fp16 = softmax(axis = var_3352, x = aw_501_cast_fp16)[name = tensor("op_3518_cast_fp16")]; + tensor var_3519_cast_fp16 = softmax(axis = var_3352, x = aw_503_cast_fp16)[name = tensor("op_3519_cast_fp16")]; + tensor var_3520_cast_fp16 = softmax(axis = var_3352, x = aw_505_cast_fp16)[name = tensor("op_3520_cast_fp16")]; + tensor var_3521_cast_fp16 = softmax(axis = var_3352, x = aw_507_cast_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3522_cast_fp16 = softmax(axis = var_3352, x = aw_509_cast_fp16)[name = tensor("op_3522_cast_fp16")]; + tensor var_3523_cast_fp16 = softmax(axis = var_3352, x = aw_511_cast_fp16)[name = tensor("op_3523_cast_fp16")]; + tensor var_3524_cast_fp16 = softmax(axis = var_3352, x = aw_513_cast_fp16)[name = tensor("op_3524_cast_fp16")]; + tensor var_3525_cast_fp16 = softmax(axis = var_3352, x = aw_515_cast_fp16)[name = tensor("op_3525_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_3352, x = aw_517_cast_fp16)[name = tensor("op_3526_cast_fp16")]; + tensor var_3527_cast_fp16 = softmax(axis = var_3352, x = aw_519_cast_fp16)[name = tensor("op_3527_cast_fp16")]; + tensor var_3529_equation_0 = const()[name = tensor("op_3529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3529_cast_fp16 = einsum(equation = var_3529_equation_0, values = (var_3447_cast_fp16_0, var_3508_cast_fp16))[name = tensor("op_3529_cast_fp16")]; + tensor var_3531_equation_0 = const()[name = tensor("op_3531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3531_cast_fp16 = einsum(equation = var_3531_equation_0, values = (var_3447_cast_fp16_1, var_3509_cast_fp16))[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_equation_0 = const()[name = tensor("op_3533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3533_cast_fp16 = einsum(equation = var_3533_equation_0, values = (var_3447_cast_fp16_2, var_3510_cast_fp16))[name = tensor("op_3533_cast_fp16")]; + tensor var_3535_equation_0 = const()[name = tensor("op_3535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3535_cast_fp16 = einsum(equation = var_3535_equation_0, values = (var_3447_cast_fp16_3, var_3511_cast_fp16))[name = tensor("op_3535_cast_fp16")]; + tensor var_3537_equation_0 = const()[name = tensor("op_3537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3537_cast_fp16 = einsum(equation = var_3537_equation_0, values = (var_3447_cast_fp16_4, var_3512_cast_fp16))[name = tensor("op_3537_cast_fp16")]; + tensor var_3539_equation_0 = const()[name = tensor("op_3539_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3539_cast_fp16 = einsum(equation = var_3539_equation_0, values = (var_3447_cast_fp16_5, var_3513_cast_fp16))[name = tensor("op_3539_cast_fp16")]; + tensor var_3541_equation_0 = const()[name = tensor("op_3541_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3541_cast_fp16 = einsum(equation = var_3541_equation_0, values = (var_3447_cast_fp16_6, var_3514_cast_fp16))[name = tensor("op_3541_cast_fp16")]; + tensor var_3543_equation_0 = const()[name = tensor("op_3543_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3543_cast_fp16 = einsum(equation = var_3543_equation_0, values = (var_3447_cast_fp16_7, var_3515_cast_fp16))[name = tensor("op_3543_cast_fp16")]; + tensor var_3545_equation_0 = const()[name = tensor("op_3545_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3545_cast_fp16 = einsum(equation = var_3545_equation_0, values = (var_3447_cast_fp16_8, var_3516_cast_fp16))[name = tensor("op_3545_cast_fp16")]; + tensor var_3547_equation_0 = const()[name = tensor("op_3547_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3547_cast_fp16 = einsum(equation = var_3547_equation_0, values = (var_3447_cast_fp16_9, var_3517_cast_fp16))[name = tensor("op_3547_cast_fp16")]; + tensor var_3549_equation_0 = const()[name = tensor("op_3549_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3549_cast_fp16 = einsum(equation = var_3549_equation_0, values = (var_3447_cast_fp16_10, var_3518_cast_fp16))[name = tensor("op_3549_cast_fp16")]; + tensor var_3551_equation_0 = const()[name = tensor("op_3551_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3551_cast_fp16 = einsum(equation = var_3551_equation_0, values = (var_3447_cast_fp16_11, var_3519_cast_fp16))[name = tensor("op_3551_cast_fp16")]; + tensor var_3553_equation_0 = const()[name = tensor("op_3553_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3553_cast_fp16 = einsum(equation = var_3553_equation_0, values = (var_3447_cast_fp16_12, var_3520_cast_fp16))[name = tensor("op_3553_cast_fp16")]; + tensor var_3555_equation_0 = const()[name = tensor("op_3555_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3555_cast_fp16 = einsum(equation = var_3555_equation_0, values = (var_3447_cast_fp16_13, var_3521_cast_fp16))[name = tensor("op_3555_cast_fp16")]; + tensor var_3557_equation_0 = const()[name = tensor("op_3557_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3557_cast_fp16 = einsum(equation = var_3557_equation_0, values = (var_3447_cast_fp16_14, var_3522_cast_fp16))[name = tensor("op_3557_cast_fp16")]; + tensor var_3559_equation_0 = const()[name = tensor("op_3559_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3559_cast_fp16 = einsum(equation = var_3559_equation_0, values = (var_3447_cast_fp16_15, var_3523_cast_fp16))[name = tensor("op_3559_cast_fp16")]; + tensor var_3561_equation_0 = const()[name = tensor("op_3561_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3561_cast_fp16 = einsum(equation = var_3561_equation_0, values = (var_3447_cast_fp16_16, var_3524_cast_fp16))[name = tensor("op_3561_cast_fp16")]; + tensor var_3563_equation_0 = const()[name = tensor("op_3563_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3563_cast_fp16 = einsum(equation = var_3563_equation_0, values = (var_3447_cast_fp16_17, var_3525_cast_fp16))[name = tensor("op_3563_cast_fp16")]; + tensor var_3565_equation_0 = const()[name = tensor("op_3565_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3565_cast_fp16 = einsum(equation = var_3565_equation_0, values = (var_3447_cast_fp16_18, var_3526_cast_fp16))[name = tensor("op_3565_cast_fp16")]; + tensor var_3567_equation_0 = const()[name = tensor("op_3567_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3567_cast_fp16 = einsum(equation = var_3567_equation_0, values = (var_3447_cast_fp16_19, var_3527_cast_fp16))[name = tensor("op_3567_cast_fp16")]; + tensor input_125_interleave_0 = const()[name = tensor("input_125_interleave_0"), val = tensor(false)]; + tensor input_125_cast_fp16 = concat(axis = var_3352, interleave = input_125_interleave_0, values = (var_3529_cast_fp16, var_3531_cast_fp16, var_3533_cast_fp16, var_3535_cast_fp16, var_3537_cast_fp16, var_3539_cast_fp16, var_3541_cast_fp16, var_3543_cast_fp16, var_3545_cast_fp16, var_3547_cast_fp16, var_3549_cast_fp16, var_3551_cast_fp16, var_3553_cast_fp16, var_3555_cast_fp16, var_3557_cast_fp16, var_3559_cast_fp16, var_3561_cast_fp16, var_3563_cast_fp16, var_3565_cast_fp16, var_3567_cast_fp16))[name = tensor("input_125_cast_fp16")]; + tensor var_3576_pad_type_0 = const()[name = tensor("op_3576_pad_type_0"), val = tensor("valid")]; + tensor var_3576_strides_0 = const()[name = tensor("op_3576_strides_0"), val = tensor([1, 1])]; + tensor var_3576_pad_0 = const()[name = tensor("op_3576_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3576_dilations_0 = const()[name = tensor("op_3576_dilations_0"), val = tensor([1, 1])]; + tensor var_3576_groups_0 = const()[name = tensor("op_3576_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_out_weight_to_fp16 = const()[name = tensor("blocks_12_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496370752)))]; + tensor blocks_12_attn_out_bias_to_fp16 = const()[name = tensor("blocks_12_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499647616)))]; + tensor var_3576_cast_fp16 = conv(bias = blocks_12_attn_out_bias_to_fp16, dilations = var_3576_dilations_0, groups = var_3576_groups_0, pad = var_3576_pad_0, pad_type = var_3576_pad_type_0, strides = var_3576_strides_0, weight = blocks_12_attn_out_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = var_3576_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([1])]; + tensor input_127_gamma_0_to_fp16 = const()[name = tensor("input_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499650240)))]; + tensor input_127_beta_0_to_fp16 = const()[name = tensor("input_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499652864)))]; + tensor var_3586_to_fp16 = const()[name = tensor("op_3586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = input_127_beta_0_to_fp16, epsilon = var_3586_to_fp16, gamma = input_127_gamma_0_to_fp16, x = inputs_51_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("valid")]; + tensor input_129_strides_0 = const()[name = tensor("input_129_strides_0"), val = tensor([1, 1])]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_129_dilations_0 = const()[name = tensor("input_129_dilations_0"), val = tensor([1, 1])]; + tensor input_129_groups_0 = const()[name = tensor("input_129_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499655488)))]; + tensor blocks_12_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512762752)))]; + tensor input_129_cast_fp16 = conv(bias = blocks_12_mlp_0_bias_to_fp16, dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = blocks_12_mlp_0_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3612_pad_type_0 = const()[name = tensor("op_3612_pad_type_0"), val = tensor("valid")]; + tensor var_3612_strides_0 = const()[name = tensor("op_3612_strides_0"), val = tensor([1, 1])]; + tensor var_3612_pad_0 = const()[name = tensor("op_3612_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3612_dilations_0 = const()[name = tensor("op_3612_dilations_0"), val = tensor([1, 1])]; + tensor var_3612_groups_0 = const()[name = tensor("op_3612_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512773056)))]; + tensor blocks_12_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525880320)))]; + tensor var_3612_cast_fp16 = conv(bias = blocks_12_mlp_2_bias_to_fp16, dilations = var_3612_dilations_0, groups = var_3612_groups_0, pad = var_3612_pad_0, pad_type = var_3612_pad_type_0, strides = var_3612_strides_0, weight = blocks_12_mlp_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_3612_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3621 = const()[name = tensor("op_3621"), val = tensor(1)]; + tensor input_133_axes_0 = const()[name = tensor("input_133_axes_0"), val = tensor([1])]; + tensor input_133_gamma_0_to_fp16 = const()[name = tensor("input_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525882944)))]; + tensor input_133_beta_0_to_fp16 = const()[name = tensor("input_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525885568)))]; + tensor var_3637_to_fp16 = const()[name = tensor("op_3637_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_133_cast_fp16 = layer_norm(axes = input_133_axes_0, beta = input_133_beta_0_to_fp16, epsilon = var_3637_to_fp16, gamma = input_133_gamma_0_to_fp16, x = inputs_53_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor q_27_pad_type_0 = const()[name = tensor("q_27_pad_type_0"), val = tensor("valid")]; + tensor q_27_strides_0 = const()[name = tensor("q_27_strides_0"), val = tensor([1, 1])]; + tensor q_27_pad_0 = const()[name = tensor("q_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_27_dilations_0 = const()[name = tensor("q_27_dilations_0"), val = tensor([1, 1])]; + tensor q_27_groups_0 = const()[name = tensor("q_27_groups_0"), val = tensor(1)]; + tensor var_3672_weight_0_to_fp16 = const()[name = tensor("op_3672_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525888192)))]; + tensor var_3672_bias_0_to_fp16 = const()[name = tensor("op_3672_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529165056)))]; + tensor var_3672_cast_fp16 = conv(bias = var_3672_bias_0_to_fp16, dilations = q_27_dilations_0, groups = q_27_groups_0, pad = q_27_pad_0, pad_type = q_27_pad_type_0, strides = q_27_strides_0, weight = var_3672_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3672_cast_fp16")]; + tensor k_27_pad_type_0 = const()[name = tensor("k_27_pad_type_0"), val = tensor("valid")]; + tensor k_27_strides_0 = const()[name = tensor("k_27_strides_0"), val = tensor([1, 1])]; + tensor k_27_pad_0 = const()[name = tensor("k_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_27_dilations_0 = const()[name = tensor("k_27_dilations_0"), val = tensor([1, 1])]; + tensor k_27_groups_0 = const()[name = tensor("k_27_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_key_weight_to_fp16 = const()[name = tensor("blocks_13_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529167680)))]; + tensor k_27_cast_fp16 = conv(dilations = k_27_dilations_0, groups = k_27_groups_0, pad = k_27_pad_0, pad_type = k_27_pad_type_0, strides = k_27_strides_0, weight = blocks_13_attn_key_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor var_3670_pad_type_0 = const()[name = tensor("op_3670_pad_type_0"), val = tensor("valid")]; + tensor var_3670_strides_0 = const()[name = tensor("op_3670_strides_0"), val = tensor([1, 1])]; + tensor var_3670_pad_0 = const()[name = tensor("op_3670_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3670_dilations_0 = const()[name = tensor("op_3670_dilations_0"), val = tensor([1, 1])]; + tensor var_3670_groups_0 = const()[name = tensor("op_3670_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_value_weight_to_fp16 = const()[name = tensor("blocks_13_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532444544)))]; + tensor blocks_13_attn_value_bias_to_fp16 = const()[name = tensor("blocks_13_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535721408)))]; + tensor var_3670_cast_fp16 = conv(bias = blocks_13_attn_value_bias_to_fp16, dilations = var_3670_dilations_0, groups = var_3670_groups_0, pad = var_3670_pad_0, pad_type = var_3670_pad_type_0, strides = var_3670_strides_0, weight = blocks_13_attn_value_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3670_cast_fp16")]; + tensor tile_39 = const()[name = tensor("tile_39"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3673_axis_0 = const()[name = tensor("op_3673_axis_0"), val = tensor(1)]; + tensor var_3673_cast_fp16_0, tensor var_3673_cast_fp16_1, tensor var_3673_cast_fp16_2, tensor var_3673_cast_fp16_3, tensor var_3673_cast_fp16_4, tensor var_3673_cast_fp16_5, tensor var_3673_cast_fp16_6, tensor var_3673_cast_fp16_7, tensor var_3673_cast_fp16_8, tensor var_3673_cast_fp16_9, tensor var_3673_cast_fp16_10, tensor var_3673_cast_fp16_11, tensor var_3673_cast_fp16_12, tensor var_3673_cast_fp16_13, tensor var_3673_cast_fp16_14, tensor var_3673_cast_fp16_15, tensor var_3673_cast_fp16_16, tensor var_3673_cast_fp16_17, tensor var_3673_cast_fp16_18, tensor var_3673_cast_fp16_19 = split(axis = var_3673_axis_0, split_sizes = tile_39, x = var_3672_cast_fp16)[name = tensor("op_3673_cast_fp16")]; + tensor var_3694_perm_0 = const()[name = tensor("op_3694_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_40 = const()[name = tensor("tile_40"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3695_axis_0 = const()[name = tensor("op_3695_axis_0"), val = tensor(3)]; + tensor var_3694_cast_fp16 = transpose(perm = var_3694_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_19")]; + tensor var_3695_cast_fp16_0, tensor var_3695_cast_fp16_1, tensor var_3695_cast_fp16_2, tensor var_3695_cast_fp16_3, tensor var_3695_cast_fp16_4, tensor var_3695_cast_fp16_5, tensor var_3695_cast_fp16_6, tensor var_3695_cast_fp16_7, tensor var_3695_cast_fp16_8, tensor var_3695_cast_fp16_9, tensor var_3695_cast_fp16_10, tensor var_3695_cast_fp16_11, tensor var_3695_cast_fp16_12, tensor var_3695_cast_fp16_13, tensor var_3695_cast_fp16_14, tensor var_3695_cast_fp16_15, tensor var_3695_cast_fp16_16, tensor var_3695_cast_fp16_17, tensor var_3695_cast_fp16_18, tensor var_3695_cast_fp16_19 = split(axis = var_3695_axis_0, split_sizes = tile_40, x = var_3694_cast_fp16)[name = tensor("op_3695_cast_fp16")]; + tensor tile_41 = const()[name = tensor("tile_41"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3716_axis_0 = const()[name = tensor("op_3716_axis_0"), val = tensor(1)]; + tensor var_3716_cast_fp16_0, tensor var_3716_cast_fp16_1, tensor var_3716_cast_fp16_2, tensor var_3716_cast_fp16_3, tensor var_3716_cast_fp16_4, tensor var_3716_cast_fp16_5, tensor var_3716_cast_fp16_6, tensor var_3716_cast_fp16_7, tensor var_3716_cast_fp16_8, tensor var_3716_cast_fp16_9, tensor var_3716_cast_fp16_10, tensor var_3716_cast_fp16_11, tensor var_3716_cast_fp16_12, tensor var_3716_cast_fp16_13, tensor var_3716_cast_fp16_14, tensor var_3716_cast_fp16_15, tensor var_3716_cast_fp16_16, tensor var_3716_cast_fp16_17, tensor var_3716_cast_fp16_18, tensor var_3716_cast_fp16_19 = split(axis = var_3716_axis_0, split_sizes = tile_41, x = var_3670_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor aw_521_equation_0 = const()[name = tensor("aw_521_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_521_cast_fp16 = einsum(equation = aw_521_equation_0, values = (var_3695_cast_fp16_0, var_3673_cast_fp16_0))[name = tensor("aw_521_cast_fp16")]; + tensor aw_523_equation_0 = const()[name = tensor("aw_523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_523_cast_fp16 = einsum(equation = aw_523_equation_0, values = (var_3695_cast_fp16_1, var_3673_cast_fp16_1))[name = tensor("aw_523_cast_fp16")]; + tensor aw_525_equation_0 = const()[name = tensor("aw_525_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_525_cast_fp16 = einsum(equation = aw_525_equation_0, values = (var_3695_cast_fp16_2, var_3673_cast_fp16_2))[name = tensor("aw_525_cast_fp16")]; + tensor aw_527_equation_0 = const()[name = tensor("aw_527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_527_cast_fp16 = einsum(equation = aw_527_equation_0, values = (var_3695_cast_fp16_3, var_3673_cast_fp16_3))[name = tensor("aw_527_cast_fp16")]; + tensor aw_529_equation_0 = const()[name = tensor("aw_529_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_529_cast_fp16 = einsum(equation = aw_529_equation_0, values = (var_3695_cast_fp16_4, var_3673_cast_fp16_4))[name = tensor("aw_529_cast_fp16")]; + tensor aw_531_equation_0 = const()[name = tensor("aw_531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_531_cast_fp16 = einsum(equation = aw_531_equation_0, values = (var_3695_cast_fp16_5, var_3673_cast_fp16_5))[name = tensor("aw_531_cast_fp16")]; + tensor aw_533_equation_0 = const()[name = tensor("aw_533_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_533_cast_fp16 = einsum(equation = aw_533_equation_0, values = (var_3695_cast_fp16_6, var_3673_cast_fp16_6))[name = tensor("aw_533_cast_fp16")]; + tensor aw_535_equation_0 = const()[name = tensor("aw_535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_535_cast_fp16 = einsum(equation = aw_535_equation_0, values = (var_3695_cast_fp16_7, var_3673_cast_fp16_7))[name = tensor("aw_535_cast_fp16")]; + tensor aw_537_equation_0 = const()[name = tensor("aw_537_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_537_cast_fp16 = einsum(equation = aw_537_equation_0, values = (var_3695_cast_fp16_8, var_3673_cast_fp16_8))[name = tensor("aw_537_cast_fp16")]; + tensor aw_539_equation_0 = const()[name = tensor("aw_539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_539_cast_fp16 = einsum(equation = aw_539_equation_0, values = (var_3695_cast_fp16_9, var_3673_cast_fp16_9))[name = tensor("aw_539_cast_fp16")]; + tensor aw_541_equation_0 = const()[name = tensor("aw_541_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_541_cast_fp16 = einsum(equation = aw_541_equation_0, values = (var_3695_cast_fp16_10, var_3673_cast_fp16_10))[name = tensor("aw_541_cast_fp16")]; + tensor aw_543_equation_0 = const()[name = tensor("aw_543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_543_cast_fp16 = einsum(equation = aw_543_equation_0, values = (var_3695_cast_fp16_11, var_3673_cast_fp16_11))[name = tensor("aw_543_cast_fp16")]; + tensor aw_545_equation_0 = const()[name = tensor("aw_545_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_545_cast_fp16 = einsum(equation = aw_545_equation_0, values = (var_3695_cast_fp16_12, var_3673_cast_fp16_12))[name = tensor("aw_545_cast_fp16")]; + tensor aw_547_equation_0 = const()[name = tensor("aw_547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_547_cast_fp16 = einsum(equation = aw_547_equation_0, values = (var_3695_cast_fp16_13, var_3673_cast_fp16_13))[name = tensor("aw_547_cast_fp16")]; + tensor aw_549_equation_0 = const()[name = tensor("aw_549_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_549_cast_fp16 = einsum(equation = aw_549_equation_0, values = (var_3695_cast_fp16_14, var_3673_cast_fp16_14))[name = tensor("aw_549_cast_fp16")]; + tensor aw_551_equation_0 = const()[name = tensor("aw_551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_551_cast_fp16 = einsum(equation = aw_551_equation_0, values = (var_3695_cast_fp16_15, var_3673_cast_fp16_15))[name = tensor("aw_551_cast_fp16")]; + tensor aw_553_equation_0 = const()[name = tensor("aw_553_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_553_cast_fp16 = einsum(equation = aw_553_equation_0, values = (var_3695_cast_fp16_16, var_3673_cast_fp16_16))[name = tensor("aw_553_cast_fp16")]; + tensor aw_555_equation_0 = const()[name = tensor("aw_555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_555_cast_fp16 = einsum(equation = aw_555_equation_0, values = (var_3695_cast_fp16_17, var_3673_cast_fp16_17))[name = tensor("aw_555_cast_fp16")]; + tensor aw_557_equation_0 = const()[name = tensor("aw_557_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_557_cast_fp16 = einsum(equation = aw_557_equation_0, values = (var_3695_cast_fp16_18, var_3673_cast_fp16_18))[name = tensor("aw_557_cast_fp16")]; + tensor aw_559_equation_0 = const()[name = tensor("aw_559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_559_cast_fp16 = einsum(equation = aw_559_equation_0, values = (var_3695_cast_fp16_19, var_3673_cast_fp16_19))[name = tensor("aw_559_cast_fp16")]; + tensor var_3777_cast_fp16 = softmax(axis = var_3621, x = aw_521_cast_fp16)[name = tensor("op_3777_cast_fp16")]; + tensor var_3778_cast_fp16 = softmax(axis = var_3621, x = aw_523_cast_fp16)[name = tensor("op_3778_cast_fp16")]; + tensor var_3779_cast_fp16 = softmax(axis = var_3621, x = aw_525_cast_fp16)[name = tensor("op_3779_cast_fp16")]; + tensor var_3780_cast_fp16 = softmax(axis = var_3621, x = aw_527_cast_fp16)[name = tensor("op_3780_cast_fp16")]; + tensor var_3781_cast_fp16 = softmax(axis = var_3621, x = aw_529_cast_fp16)[name = tensor("op_3781_cast_fp16")]; + tensor var_3782_cast_fp16 = softmax(axis = var_3621, x = aw_531_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = softmax(axis = var_3621, x = aw_533_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor var_3784_cast_fp16 = softmax(axis = var_3621, x = aw_535_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3785_cast_fp16 = softmax(axis = var_3621, x = aw_537_cast_fp16)[name = tensor("op_3785_cast_fp16")]; + tensor var_3786_cast_fp16 = softmax(axis = var_3621, x = aw_539_cast_fp16)[name = tensor("op_3786_cast_fp16")]; + tensor var_3787_cast_fp16 = softmax(axis = var_3621, x = aw_541_cast_fp16)[name = tensor("op_3787_cast_fp16")]; + tensor var_3788_cast_fp16 = softmax(axis = var_3621, x = aw_543_cast_fp16)[name = tensor("op_3788_cast_fp16")]; + tensor var_3789_cast_fp16 = softmax(axis = var_3621, x = aw_545_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3790_cast_fp16 = softmax(axis = var_3621, x = aw_547_cast_fp16)[name = tensor("op_3790_cast_fp16")]; + tensor var_3791_cast_fp16 = softmax(axis = var_3621, x = aw_549_cast_fp16)[name = tensor("op_3791_cast_fp16")]; + tensor var_3792_cast_fp16 = softmax(axis = var_3621, x = aw_551_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3793_cast_fp16 = softmax(axis = var_3621, x = aw_553_cast_fp16)[name = tensor("op_3793_cast_fp16")]; + tensor var_3794_cast_fp16 = softmax(axis = var_3621, x = aw_555_cast_fp16)[name = tensor("op_3794_cast_fp16")]; + tensor var_3795_cast_fp16 = softmax(axis = var_3621, x = aw_557_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = softmax(axis = var_3621, x = aw_559_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor var_3798_equation_0 = const()[name = tensor("op_3798_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3798_cast_fp16 = einsum(equation = var_3798_equation_0, values = (var_3716_cast_fp16_0, var_3777_cast_fp16))[name = tensor("op_3798_cast_fp16")]; + tensor var_3800_equation_0 = const()[name = tensor("op_3800_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3800_cast_fp16 = einsum(equation = var_3800_equation_0, values = (var_3716_cast_fp16_1, var_3778_cast_fp16))[name = tensor("op_3800_cast_fp16")]; + tensor var_3802_equation_0 = const()[name = tensor("op_3802_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3802_cast_fp16 = einsum(equation = var_3802_equation_0, values = (var_3716_cast_fp16_2, var_3779_cast_fp16))[name = tensor("op_3802_cast_fp16")]; + tensor var_3804_equation_0 = const()[name = tensor("op_3804_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3804_cast_fp16 = einsum(equation = var_3804_equation_0, values = (var_3716_cast_fp16_3, var_3780_cast_fp16))[name = tensor("op_3804_cast_fp16")]; + tensor var_3806_equation_0 = const()[name = tensor("op_3806_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3806_cast_fp16 = einsum(equation = var_3806_equation_0, values = (var_3716_cast_fp16_4, var_3781_cast_fp16))[name = tensor("op_3806_cast_fp16")]; + tensor var_3808_equation_0 = const()[name = tensor("op_3808_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3808_cast_fp16 = einsum(equation = var_3808_equation_0, values = (var_3716_cast_fp16_5, var_3782_cast_fp16))[name = tensor("op_3808_cast_fp16")]; + tensor var_3810_equation_0 = const()[name = tensor("op_3810_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3810_cast_fp16 = einsum(equation = var_3810_equation_0, values = (var_3716_cast_fp16_6, var_3783_cast_fp16))[name = tensor("op_3810_cast_fp16")]; + tensor var_3812_equation_0 = const()[name = tensor("op_3812_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3812_cast_fp16 = einsum(equation = var_3812_equation_0, values = (var_3716_cast_fp16_7, var_3784_cast_fp16))[name = tensor("op_3812_cast_fp16")]; + tensor var_3814_equation_0 = const()[name = tensor("op_3814_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3814_cast_fp16 = einsum(equation = var_3814_equation_0, values = (var_3716_cast_fp16_8, var_3785_cast_fp16))[name = tensor("op_3814_cast_fp16")]; + tensor var_3816_equation_0 = const()[name = tensor("op_3816_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3816_cast_fp16 = einsum(equation = var_3816_equation_0, values = (var_3716_cast_fp16_9, var_3786_cast_fp16))[name = tensor("op_3816_cast_fp16")]; + tensor var_3818_equation_0 = const()[name = tensor("op_3818_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3818_cast_fp16 = einsum(equation = var_3818_equation_0, values = (var_3716_cast_fp16_10, var_3787_cast_fp16))[name = tensor("op_3818_cast_fp16")]; + tensor var_3820_equation_0 = const()[name = tensor("op_3820_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3820_cast_fp16 = einsum(equation = var_3820_equation_0, values = (var_3716_cast_fp16_11, var_3788_cast_fp16))[name = tensor("op_3820_cast_fp16")]; + tensor var_3822_equation_0 = const()[name = tensor("op_3822_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3822_cast_fp16 = einsum(equation = var_3822_equation_0, values = (var_3716_cast_fp16_12, var_3789_cast_fp16))[name = tensor("op_3822_cast_fp16")]; + tensor var_3824_equation_0 = const()[name = tensor("op_3824_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3824_cast_fp16 = einsum(equation = var_3824_equation_0, values = (var_3716_cast_fp16_13, var_3790_cast_fp16))[name = tensor("op_3824_cast_fp16")]; + tensor var_3826_equation_0 = const()[name = tensor("op_3826_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3826_cast_fp16 = einsum(equation = var_3826_equation_0, values = (var_3716_cast_fp16_14, var_3791_cast_fp16))[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_equation_0 = const()[name = tensor("op_3828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3828_cast_fp16 = einsum(equation = var_3828_equation_0, values = (var_3716_cast_fp16_15, var_3792_cast_fp16))[name = tensor("op_3828_cast_fp16")]; + tensor var_3830_equation_0 = const()[name = tensor("op_3830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3830_cast_fp16 = einsum(equation = var_3830_equation_0, values = (var_3716_cast_fp16_16, var_3793_cast_fp16))[name = tensor("op_3830_cast_fp16")]; + tensor var_3832_equation_0 = const()[name = tensor("op_3832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3832_cast_fp16 = einsum(equation = var_3832_equation_0, values = (var_3716_cast_fp16_17, var_3794_cast_fp16))[name = tensor("op_3832_cast_fp16")]; + tensor var_3834_equation_0 = const()[name = tensor("op_3834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3834_cast_fp16 = einsum(equation = var_3834_equation_0, values = (var_3716_cast_fp16_18, var_3795_cast_fp16))[name = tensor("op_3834_cast_fp16")]; + tensor var_3836_equation_0 = const()[name = tensor("op_3836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3836_cast_fp16 = einsum(equation = var_3836_equation_0, values = (var_3716_cast_fp16_19, var_3796_cast_fp16))[name = tensor("op_3836_cast_fp16")]; + tensor input_135_interleave_0 = const()[name = tensor("input_135_interleave_0"), val = tensor(false)]; + tensor input_135_cast_fp16 = concat(axis = var_3621, interleave = input_135_interleave_0, values = (var_3798_cast_fp16, var_3800_cast_fp16, var_3802_cast_fp16, var_3804_cast_fp16, var_3806_cast_fp16, var_3808_cast_fp16, var_3810_cast_fp16, var_3812_cast_fp16, var_3814_cast_fp16, var_3816_cast_fp16, var_3818_cast_fp16, var_3820_cast_fp16, var_3822_cast_fp16, var_3824_cast_fp16, var_3826_cast_fp16, var_3828_cast_fp16, var_3830_cast_fp16, var_3832_cast_fp16, var_3834_cast_fp16, var_3836_cast_fp16))[name = tensor("input_135_cast_fp16")]; + tensor var_3845_pad_type_0 = const()[name = tensor("op_3845_pad_type_0"), val = tensor("valid")]; + tensor var_3845_strides_0 = const()[name = tensor("op_3845_strides_0"), val = tensor([1, 1])]; + tensor var_3845_pad_0 = const()[name = tensor("op_3845_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3845_dilations_0 = const()[name = tensor("op_3845_dilations_0"), val = tensor([1, 1])]; + tensor var_3845_groups_0 = const()[name = tensor("op_3845_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_out_weight_to_fp16 = const()[name = tensor("blocks_13_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535724032)))]; + tensor blocks_13_attn_out_bias_to_fp16 = const()[name = tensor("blocks_13_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539000896)))]; + tensor var_3845_cast_fp16 = conv(bias = blocks_13_attn_out_bias_to_fp16, dilations = var_3845_dilations_0, groups = var_3845_groups_0, pad = var_3845_pad_0, pad_type = var_3845_pad_type_0, strides = var_3845_strides_0, weight = blocks_13_attn_out_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("op_3845_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = var_3845_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([1])]; + tensor input_137_gamma_0_to_fp16 = const()[name = tensor("input_137_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539003520)))]; + tensor input_137_beta_0_to_fp16 = const()[name = tensor("input_137_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539006144)))]; + tensor var_3855_to_fp16 = const()[name = tensor("op_3855_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = input_137_beta_0_to_fp16, epsilon = var_3855_to_fp16, gamma = input_137_gamma_0_to_fp16, x = inputs_55_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("valid")]; + tensor input_139_strides_0 = const()[name = tensor("input_139_strides_0"), val = tensor([1, 1])]; + tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_139_dilations_0 = const()[name = tensor("input_139_dilations_0"), val = tensor([1, 1])]; + tensor input_139_groups_0 = const()[name = tensor("input_139_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539008768)))]; + tensor blocks_13_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552116032)))]; + tensor input_139_cast_fp16 = conv(bias = blocks_13_mlp_0_bias_to_fp16, dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = blocks_13_mlp_0_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3881_pad_type_0 = const()[name = tensor("op_3881_pad_type_0"), val = tensor("valid")]; + tensor var_3881_strides_0 = const()[name = tensor("op_3881_strides_0"), val = tensor([1, 1])]; + tensor var_3881_pad_0 = const()[name = tensor("op_3881_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3881_dilations_0 = const()[name = tensor("op_3881_dilations_0"), val = tensor([1, 1])]; + tensor var_3881_groups_0 = const()[name = tensor("op_3881_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552126336)))]; + tensor blocks_13_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565233600)))]; + tensor var_3881_cast_fp16 = conv(bias = blocks_13_mlp_2_bias_to_fp16, dilations = var_3881_dilations_0, groups = var_3881_groups_0, pad = var_3881_pad_0, pad_type = var_3881_pad_type_0, strides = var_3881_strides_0, weight = blocks_13_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = var_3881_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3890 = const()[name = tensor("op_3890"), val = tensor(1)]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([1])]; + tensor input_143_gamma_0_to_fp16 = const()[name = tensor("input_143_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565236224)))]; + tensor input_143_beta_0_to_fp16 = const()[name = tensor("input_143_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565238848)))]; + tensor var_3906_to_fp16 = const()[name = tensor("op_3906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = input_143_beta_0_to_fp16, epsilon = var_3906_to_fp16, gamma = input_143_gamma_0_to_fp16, x = inputs_57_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor q_29_pad_type_0 = const()[name = tensor("q_29_pad_type_0"), val = tensor("valid")]; + tensor q_29_strides_0 = const()[name = tensor("q_29_strides_0"), val = tensor([1, 1])]; + tensor q_29_pad_0 = const()[name = tensor("q_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_29_dilations_0 = const()[name = tensor("q_29_dilations_0"), val = tensor([1, 1])]; + tensor q_29_groups_0 = const()[name = tensor("q_29_groups_0"), val = tensor(1)]; + tensor var_3941_weight_0_to_fp16 = const()[name = tensor("op_3941_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565241472)))]; + tensor var_3941_bias_0_to_fp16 = const()[name = tensor("op_3941_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568518336)))]; + tensor var_3941_cast_fp16 = conv(bias = var_3941_bias_0_to_fp16, dilations = q_29_dilations_0, groups = q_29_groups_0, pad = q_29_pad_0, pad_type = q_29_pad_type_0, strides = q_29_strides_0, weight = var_3941_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3941_cast_fp16")]; + tensor k_29_pad_type_0 = const()[name = tensor("k_29_pad_type_0"), val = tensor("valid")]; + tensor k_29_strides_0 = const()[name = tensor("k_29_strides_0"), val = tensor([1, 1])]; + tensor k_29_pad_0 = const()[name = tensor("k_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_29_dilations_0 = const()[name = tensor("k_29_dilations_0"), val = tensor([1, 1])]; + tensor k_29_groups_0 = const()[name = tensor("k_29_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_key_weight_to_fp16 = const()[name = tensor("blocks_14_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568520960)))]; + tensor k_29_cast_fp16 = conv(dilations = k_29_dilations_0, groups = k_29_groups_0, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = k_29_strides_0, weight = blocks_14_attn_key_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor var_3939_pad_type_0 = const()[name = tensor("op_3939_pad_type_0"), val = tensor("valid")]; + tensor var_3939_strides_0 = const()[name = tensor("op_3939_strides_0"), val = tensor([1, 1])]; + tensor var_3939_pad_0 = const()[name = tensor("op_3939_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3939_dilations_0 = const()[name = tensor("op_3939_dilations_0"), val = tensor([1, 1])]; + tensor var_3939_groups_0 = const()[name = tensor("op_3939_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_value_weight_to_fp16 = const()[name = tensor("blocks_14_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571797824)))]; + tensor blocks_14_attn_value_bias_to_fp16 = const()[name = tensor("blocks_14_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575074688)))]; + tensor var_3939_cast_fp16 = conv(bias = blocks_14_attn_value_bias_to_fp16, dilations = var_3939_dilations_0, groups = var_3939_groups_0, pad = var_3939_pad_0, pad_type = var_3939_pad_type_0, strides = var_3939_strides_0, weight = blocks_14_attn_value_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor tile_42 = const()[name = tensor("tile_42"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3942_axis_0 = const()[name = tensor("op_3942_axis_0"), val = tensor(1)]; + tensor var_3942_cast_fp16_0, tensor var_3942_cast_fp16_1, tensor var_3942_cast_fp16_2, tensor var_3942_cast_fp16_3, tensor var_3942_cast_fp16_4, tensor var_3942_cast_fp16_5, tensor var_3942_cast_fp16_6, tensor var_3942_cast_fp16_7, tensor var_3942_cast_fp16_8, tensor var_3942_cast_fp16_9, tensor var_3942_cast_fp16_10, tensor var_3942_cast_fp16_11, tensor var_3942_cast_fp16_12, tensor var_3942_cast_fp16_13, tensor var_3942_cast_fp16_14, tensor var_3942_cast_fp16_15, tensor var_3942_cast_fp16_16, tensor var_3942_cast_fp16_17, tensor var_3942_cast_fp16_18, tensor var_3942_cast_fp16_19 = split(axis = var_3942_axis_0, split_sizes = tile_42, x = var_3941_cast_fp16)[name = tensor("op_3942_cast_fp16")]; + tensor var_3963_perm_0 = const()[name = tensor("op_3963_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_43 = const()[name = tensor("tile_43"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3964_axis_0 = const()[name = tensor("op_3964_axis_0"), val = tensor(3)]; + tensor var_3963_cast_fp16 = transpose(perm = var_3963_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_18")]; + tensor var_3964_cast_fp16_0, tensor var_3964_cast_fp16_1, tensor var_3964_cast_fp16_2, tensor var_3964_cast_fp16_3, tensor var_3964_cast_fp16_4, tensor var_3964_cast_fp16_5, tensor var_3964_cast_fp16_6, tensor var_3964_cast_fp16_7, tensor var_3964_cast_fp16_8, tensor var_3964_cast_fp16_9, tensor var_3964_cast_fp16_10, tensor var_3964_cast_fp16_11, tensor var_3964_cast_fp16_12, tensor var_3964_cast_fp16_13, tensor var_3964_cast_fp16_14, tensor var_3964_cast_fp16_15, tensor var_3964_cast_fp16_16, tensor var_3964_cast_fp16_17, tensor var_3964_cast_fp16_18, tensor var_3964_cast_fp16_19 = split(axis = var_3964_axis_0, split_sizes = tile_43, x = var_3963_cast_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor tile_44 = const()[name = tensor("tile_44"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3985_axis_0 = const()[name = tensor("op_3985_axis_0"), val = tensor(1)]; + tensor var_3985_cast_fp16_0, tensor var_3985_cast_fp16_1, tensor var_3985_cast_fp16_2, tensor var_3985_cast_fp16_3, tensor var_3985_cast_fp16_4, tensor var_3985_cast_fp16_5, tensor var_3985_cast_fp16_6, tensor var_3985_cast_fp16_7, tensor var_3985_cast_fp16_8, tensor var_3985_cast_fp16_9, tensor var_3985_cast_fp16_10, tensor var_3985_cast_fp16_11, tensor var_3985_cast_fp16_12, tensor var_3985_cast_fp16_13, tensor var_3985_cast_fp16_14, tensor var_3985_cast_fp16_15, tensor var_3985_cast_fp16_16, tensor var_3985_cast_fp16_17, tensor var_3985_cast_fp16_18, tensor var_3985_cast_fp16_19 = split(axis = var_3985_axis_0, split_sizes = tile_44, x = var_3939_cast_fp16)[name = tensor("op_3985_cast_fp16")]; + tensor aw_561_equation_0 = const()[name = tensor("aw_561_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_561_cast_fp16 = einsum(equation = aw_561_equation_0, values = (var_3964_cast_fp16_0, var_3942_cast_fp16_0))[name = tensor("aw_561_cast_fp16")]; + tensor aw_563_equation_0 = const()[name = tensor("aw_563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_563_cast_fp16 = einsum(equation = aw_563_equation_0, values = (var_3964_cast_fp16_1, var_3942_cast_fp16_1))[name = tensor("aw_563_cast_fp16")]; + tensor aw_565_equation_0 = const()[name = tensor("aw_565_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_565_cast_fp16 = einsum(equation = aw_565_equation_0, values = (var_3964_cast_fp16_2, var_3942_cast_fp16_2))[name = tensor("aw_565_cast_fp16")]; + tensor aw_567_equation_0 = const()[name = tensor("aw_567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_567_cast_fp16 = einsum(equation = aw_567_equation_0, values = (var_3964_cast_fp16_3, var_3942_cast_fp16_3))[name = tensor("aw_567_cast_fp16")]; + tensor aw_569_equation_0 = const()[name = tensor("aw_569_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_569_cast_fp16 = einsum(equation = aw_569_equation_0, values = (var_3964_cast_fp16_4, var_3942_cast_fp16_4))[name = tensor("aw_569_cast_fp16")]; + tensor aw_571_equation_0 = const()[name = tensor("aw_571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_571_cast_fp16 = einsum(equation = aw_571_equation_0, values = (var_3964_cast_fp16_5, var_3942_cast_fp16_5))[name = tensor("aw_571_cast_fp16")]; + tensor aw_573_equation_0 = const()[name = tensor("aw_573_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_573_cast_fp16 = einsum(equation = aw_573_equation_0, values = (var_3964_cast_fp16_6, var_3942_cast_fp16_6))[name = tensor("aw_573_cast_fp16")]; + tensor aw_575_equation_0 = const()[name = tensor("aw_575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_575_cast_fp16 = einsum(equation = aw_575_equation_0, values = (var_3964_cast_fp16_7, var_3942_cast_fp16_7))[name = tensor("aw_575_cast_fp16")]; + tensor aw_577_equation_0 = const()[name = tensor("aw_577_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_577_cast_fp16 = einsum(equation = aw_577_equation_0, values = (var_3964_cast_fp16_8, var_3942_cast_fp16_8))[name = tensor("aw_577_cast_fp16")]; + tensor aw_579_equation_0 = const()[name = tensor("aw_579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_579_cast_fp16 = einsum(equation = aw_579_equation_0, values = (var_3964_cast_fp16_9, var_3942_cast_fp16_9))[name = tensor("aw_579_cast_fp16")]; + tensor aw_581_equation_0 = const()[name = tensor("aw_581_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_581_cast_fp16 = einsum(equation = aw_581_equation_0, values = (var_3964_cast_fp16_10, var_3942_cast_fp16_10))[name = tensor("aw_581_cast_fp16")]; + tensor aw_583_equation_0 = const()[name = tensor("aw_583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_583_cast_fp16 = einsum(equation = aw_583_equation_0, values = (var_3964_cast_fp16_11, var_3942_cast_fp16_11))[name = tensor("aw_583_cast_fp16")]; + tensor aw_585_equation_0 = const()[name = tensor("aw_585_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_585_cast_fp16 = einsum(equation = aw_585_equation_0, values = (var_3964_cast_fp16_12, var_3942_cast_fp16_12))[name = tensor("aw_585_cast_fp16")]; + tensor aw_587_equation_0 = const()[name = tensor("aw_587_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_587_cast_fp16 = einsum(equation = aw_587_equation_0, values = (var_3964_cast_fp16_13, var_3942_cast_fp16_13))[name = tensor("aw_587_cast_fp16")]; + tensor aw_589_equation_0 = const()[name = tensor("aw_589_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_589_cast_fp16 = einsum(equation = aw_589_equation_0, values = (var_3964_cast_fp16_14, var_3942_cast_fp16_14))[name = tensor("aw_589_cast_fp16")]; + tensor aw_591_equation_0 = const()[name = tensor("aw_591_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_591_cast_fp16 = einsum(equation = aw_591_equation_0, values = (var_3964_cast_fp16_15, var_3942_cast_fp16_15))[name = tensor("aw_591_cast_fp16")]; + tensor aw_593_equation_0 = const()[name = tensor("aw_593_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_593_cast_fp16 = einsum(equation = aw_593_equation_0, values = (var_3964_cast_fp16_16, var_3942_cast_fp16_16))[name = tensor("aw_593_cast_fp16")]; + tensor aw_595_equation_0 = const()[name = tensor("aw_595_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_595_cast_fp16 = einsum(equation = aw_595_equation_0, values = (var_3964_cast_fp16_17, var_3942_cast_fp16_17))[name = tensor("aw_595_cast_fp16")]; + tensor aw_597_equation_0 = const()[name = tensor("aw_597_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_597_cast_fp16 = einsum(equation = aw_597_equation_0, values = (var_3964_cast_fp16_18, var_3942_cast_fp16_18))[name = tensor("aw_597_cast_fp16")]; + tensor aw_599_equation_0 = const()[name = tensor("aw_599_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_599_cast_fp16 = einsum(equation = aw_599_equation_0, values = (var_3964_cast_fp16_19, var_3942_cast_fp16_19))[name = tensor("aw_599_cast_fp16")]; + tensor var_4046_cast_fp16 = softmax(axis = var_3890, x = aw_561_cast_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor var_4047_cast_fp16 = softmax(axis = var_3890, x = aw_563_cast_fp16)[name = tensor("op_4047_cast_fp16")]; + tensor var_4048_cast_fp16 = softmax(axis = var_3890, x = aw_565_cast_fp16)[name = tensor("op_4048_cast_fp16")]; + tensor var_4049_cast_fp16 = softmax(axis = var_3890, x = aw_567_cast_fp16)[name = tensor("op_4049_cast_fp16")]; + tensor var_4050_cast_fp16 = softmax(axis = var_3890, x = aw_569_cast_fp16)[name = tensor("op_4050_cast_fp16")]; + tensor var_4051_cast_fp16 = softmax(axis = var_3890, x = aw_571_cast_fp16)[name = tensor("op_4051_cast_fp16")]; + tensor var_4052_cast_fp16 = softmax(axis = var_3890, x = aw_573_cast_fp16)[name = tensor("op_4052_cast_fp16")]; + tensor var_4053_cast_fp16 = softmax(axis = var_3890, x = aw_575_cast_fp16)[name = tensor("op_4053_cast_fp16")]; + tensor var_4054_cast_fp16 = softmax(axis = var_3890, x = aw_577_cast_fp16)[name = tensor("op_4054_cast_fp16")]; + tensor var_4055_cast_fp16 = softmax(axis = var_3890, x = aw_579_cast_fp16)[name = tensor("op_4055_cast_fp16")]; + tensor var_4056_cast_fp16 = softmax(axis = var_3890, x = aw_581_cast_fp16)[name = tensor("op_4056_cast_fp16")]; + tensor var_4057_cast_fp16 = softmax(axis = var_3890, x = aw_583_cast_fp16)[name = tensor("op_4057_cast_fp16")]; + tensor var_4058_cast_fp16 = softmax(axis = var_3890, x = aw_585_cast_fp16)[name = tensor("op_4058_cast_fp16")]; + tensor var_4059_cast_fp16 = softmax(axis = var_3890, x = aw_587_cast_fp16)[name = tensor("op_4059_cast_fp16")]; + tensor var_4060_cast_fp16 = softmax(axis = var_3890, x = aw_589_cast_fp16)[name = tensor("op_4060_cast_fp16")]; + tensor var_4061_cast_fp16 = softmax(axis = var_3890, x = aw_591_cast_fp16)[name = tensor("op_4061_cast_fp16")]; + tensor var_4062_cast_fp16 = softmax(axis = var_3890, x = aw_593_cast_fp16)[name = tensor("op_4062_cast_fp16")]; + tensor var_4063_cast_fp16 = softmax(axis = var_3890, x = aw_595_cast_fp16)[name = tensor("op_4063_cast_fp16")]; + tensor var_4064_cast_fp16 = softmax(axis = var_3890, x = aw_597_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_cast_fp16 = softmax(axis = var_3890, x = aw_599_cast_fp16)[name = tensor("op_4065_cast_fp16")]; + tensor var_4067_equation_0 = const()[name = tensor("op_4067_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4067_cast_fp16 = einsum(equation = var_4067_equation_0, values = (var_3985_cast_fp16_0, var_4046_cast_fp16))[name = tensor("op_4067_cast_fp16")]; + tensor var_4069_equation_0 = const()[name = tensor("op_4069_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4069_cast_fp16 = einsum(equation = var_4069_equation_0, values = (var_3985_cast_fp16_1, var_4047_cast_fp16))[name = tensor("op_4069_cast_fp16")]; + tensor var_4071_equation_0 = const()[name = tensor("op_4071_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4071_cast_fp16 = einsum(equation = var_4071_equation_0, values = (var_3985_cast_fp16_2, var_4048_cast_fp16))[name = tensor("op_4071_cast_fp16")]; + tensor var_4073_equation_0 = const()[name = tensor("op_4073_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4073_cast_fp16 = einsum(equation = var_4073_equation_0, values = (var_3985_cast_fp16_3, var_4049_cast_fp16))[name = tensor("op_4073_cast_fp16")]; + tensor var_4075_equation_0 = const()[name = tensor("op_4075_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4075_cast_fp16 = einsum(equation = var_4075_equation_0, values = (var_3985_cast_fp16_4, var_4050_cast_fp16))[name = tensor("op_4075_cast_fp16")]; + tensor var_4077_equation_0 = const()[name = tensor("op_4077_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4077_cast_fp16 = einsum(equation = var_4077_equation_0, values = (var_3985_cast_fp16_5, var_4051_cast_fp16))[name = tensor("op_4077_cast_fp16")]; + tensor var_4079_equation_0 = const()[name = tensor("op_4079_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4079_cast_fp16 = einsum(equation = var_4079_equation_0, values = (var_3985_cast_fp16_6, var_4052_cast_fp16))[name = tensor("op_4079_cast_fp16")]; + tensor var_4081_equation_0 = const()[name = tensor("op_4081_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4081_cast_fp16 = einsum(equation = var_4081_equation_0, values = (var_3985_cast_fp16_7, var_4053_cast_fp16))[name = tensor("op_4081_cast_fp16")]; + tensor var_4083_equation_0 = const()[name = tensor("op_4083_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4083_cast_fp16 = einsum(equation = var_4083_equation_0, values = (var_3985_cast_fp16_8, var_4054_cast_fp16))[name = tensor("op_4083_cast_fp16")]; + tensor var_4085_equation_0 = const()[name = tensor("op_4085_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4085_cast_fp16 = einsum(equation = var_4085_equation_0, values = (var_3985_cast_fp16_9, var_4055_cast_fp16))[name = tensor("op_4085_cast_fp16")]; + tensor var_4087_equation_0 = const()[name = tensor("op_4087_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4087_cast_fp16 = einsum(equation = var_4087_equation_0, values = (var_3985_cast_fp16_10, var_4056_cast_fp16))[name = tensor("op_4087_cast_fp16")]; + tensor var_4089_equation_0 = const()[name = tensor("op_4089_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4089_cast_fp16 = einsum(equation = var_4089_equation_0, values = (var_3985_cast_fp16_11, var_4057_cast_fp16))[name = tensor("op_4089_cast_fp16")]; + tensor var_4091_equation_0 = const()[name = tensor("op_4091_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4091_cast_fp16 = einsum(equation = var_4091_equation_0, values = (var_3985_cast_fp16_12, var_4058_cast_fp16))[name = tensor("op_4091_cast_fp16")]; + tensor var_4093_equation_0 = const()[name = tensor("op_4093_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4093_cast_fp16 = einsum(equation = var_4093_equation_0, values = (var_3985_cast_fp16_13, var_4059_cast_fp16))[name = tensor("op_4093_cast_fp16")]; + tensor var_4095_equation_0 = const()[name = tensor("op_4095_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4095_cast_fp16 = einsum(equation = var_4095_equation_0, values = (var_3985_cast_fp16_14, var_4060_cast_fp16))[name = tensor("op_4095_cast_fp16")]; + tensor var_4097_equation_0 = const()[name = tensor("op_4097_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4097_cast_fp16 = einsum(equation = var_4097_equation_0, values = (var_3985_cast_fp16_15, var_4061_cast_fp16))[name = tensor("op_4097_cast_fp16")]; + tensor var_4099_equation_0 = const()[name = tensor("op_4099_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4099_cast_fp16 = einsum(equation = var_4099_equation_0, values = (var_3985_cast_fp16_16, var_4062_cast_fp16))[name = tensor("op_4099_cast_fp16")]; + tensor var_4101_equation_0 = const()[name = tensor("op_4101_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4101_cast_fp16 = einsum(equation = var_4101_equation_0, values = (var_3985_cast_fp16_17, var_4063_cast_fp16))[name = tensor("op_4101_cast_fp16")]; + tensor var_4103_equation_0 = const()[name = tensor("op_4103_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4103_cast_fp16 = einsum(equation = var_4103_equation_0, values = (var_3985_cast_fp16_18, var_4064_cast_fp16))[name = tensor("op_4103_cast_fp16")]; + tensor var_4105_equation_0 = const()[name = tensor("op_4105_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4105_cast_fp16 = einsum(equation = var_4105_equation_0, values = (var_3985_cast_fp16_19, var_4065_cast_fp16))[name = tensor("op_4105_cast_fp16")]; + tensor input_145_interleave_0 = const()[name = tensor("input_145_interleave_0"), val = tensor(false)]; + tensor input_145_cast_fp16 = concat(axis = var_3890, interleave = input_145_interleave_0, values = (var_4067_cast_fp16, var_4069_cast_fp16, var_4071_cast_fp16, var_4073_cast_fp16, var_4075_cast_fp16, var_4077_cast_fp16, var_4079_cast_fp16, var_4081_cast_fp16, var_4083_cast_fp16, var_4085_cast_fp16, var_4087_cast_fp16, var_4089_cast_fp16, var_4091_cast_fp16, var_4093_cast_fp16, var_4095_cast_fp16, var_4097_cast_fp16, var_4099_cast_fp16, var_4101_cast_fp16, var_4103_cast_fp16, var_4105_cast_fp16))[name = tensor("input_145_cast_fp16")]; + tensor var_4114_pad_type_0 = const()[name = tensor("op_4114_pad_type_0"), val = tensor("valid")]; + tensor var_4114_strides_0 = const()[name = tensor("op_4114_strides_0"), val = tensor([1, 1])]; + tensor var_4114_pad_0 = const()[name = tensor("op_4114_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4114_dilations_0 = const()[name = tensor("op_4114_dilations_0"), val = tensor([1, 1])]; + tensor var_4114_groups_0 = const()[name = tensor("op_4114_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_out_weight_to_fp16 = const()[name = tensor("blocks_14_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575077312)))]; + tensor blocks_14_attn_out_bias_to_fp16 = const()[name = tensor("blocks_14_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578354176)))]; + tensor var_4114_cast_fp16 = conv(bias = blocks_14_attn_out_bias_to_fp16, dilations = var_4114_dilations_0, groups = var_4114_groups_0, pad = var_4114_pad_0, pad_type = var_4114_pad_type_0, strides = var_4114_strides_0, weight = blocks_14_attn_out_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("op_4114_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_4114_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([1])]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578356800)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578359424)))]; + tensor var_4124_to_fp16 = const()[name = tensor("op_4124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = input_147_beta_0_to_fp16, epsilon = var_4124_to_fp16, gamma = input_147_gamma_0_to_fp16, x = inputs_59_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578362048)))]; + tensor blocks_14_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591469312)))]; + tensor input_149_cast_fp16 = conv(bias = blocks_14_mlp_0_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = blocks_14_mlp_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_4150_pad_type_0 = const()[name = tensor("op_4150_pad_type_0"), val = tensor("valid")]; + tensor var_4150_strides_0 = const()[name = tensor("op_4150_strides_0"), val = tensor([1, 1])]; + tensor var_4150_pad_0 = const()[name = tensor("op_4150_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4150_dilations_0 = const()[name = tensor("op_4150_dilations_0"), val = tensor([1, 1])]; + tensor var_4150_groups_0 = const()[name = tensor("op_4150_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591479616)))]; + tensor blocks_14_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604586880)))]; + tensor var_4150_cast_fp16 = conv(bias = blocks_14_mlp_2_bias_to_fp16, dilations = var_4150_dilations_0, groups = var_4150_groups_0, pad = var_4150_pad_0, pad_type = var_4150_pad_type_0, strides = var_4150_strides_0, weight = blocks_14_mlp_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = var_4150_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_4159 = const()[name = tensor("op_4159"), val = tensor(1)]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([1])]; + tensor input_153_gamma_0_to_fp16 = const()[name = tensor("input_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604589504)))]; + tensor input_153_beta_0_to_fp16 = const()[name = tensor("input_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604592128)))]; + tensor var_4175_to_fp16 = const()[name = tensor("op_4175_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = input_153_beta_0_to_fp16, epsilon = var_4175_to_fp16, gamma = input_153_gamma_0_to_fp16, x = inputs_61_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor q_31_pad_type_0 = const()[name = tensor("q_31_pad_type_0"), val = tensor("valid")]; + tensor q_31_strides_0 = const()[name = tensor("q_31_strides_0"), val = tensor([1, 1])]; + tensor q_31_pad_0 = const()[name = tensor("q_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_31_dilations_0 = const()[name = tensor("q_31_dilations_0"), val = tensor([1, 1])]; + tensor q_31_groups_0 = const()[name = tensor("q_31_groups_0"), val = tensor(1)]; + tensor var_4210_weight_0_to_fp16 = const()[name = tensor("op_4210_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604594752)))]; + tensor var_4210_bias_0_to_fp16 = const()[name = tensor("op_4210_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607871616)))]; + tensor var_4210_cast_fp16 = conv(bias = var_4210_bias_0_to_fp16, dilations = q_31_dilations_0, groups = q_31_groups_0, pad = q_31_pad_0, pad_type = q_31_pad_type_0, strides = q_31_strides_0, weight = var_4210_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("op_4210_cast_fp16")]; + tensor k_31_pad_type_0 = const()[name = tensor("k_31_pad_type_0"), val = tensor("valid")]; + tensor k_31_strides_0 = const()[name = tensor("k_31_strides_0"), val = tensor([1, 1])]; + tensor k_31_pad_0 = const()[name = tensor("k_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_31_dilations_0 = const()[name = tensor("k_31_dilations_0"), val = tensor([1, 1])]; + tensor k_31_groups_0 = const()[name = tensor("k_31_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_key_weight_to_fp16 = const()[name = tensor("blocks_15_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607874240)))]; + tensor k_31_cast_fp16 = conv(dilations = k_31_dilations_0, groups = k_31_groups_0, pad = k_31_pad_0, pad_type = k_31_pad_type_0, strides = k_31_strides_0, weight = blocks_15_attn_key_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor var_4208_pad_type_0 = const()[name = tensor("op_4208_pad_type_0"), val = tensor("valid")]; + tensor var_4208_strides_0 = const()[name = tensor("op_4208_strides_0"), val = tensor([1, 1])]; + tensor var_4208_pad_0 = const()[name = tensor("op_4208_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4208_dilations_0 = const()[name = tensor("op_4208_dilations_0"), val = tensor([1, 1])]; + tensor var_4208_groups_0 = const()[name = tensor("op_4208_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_value_weight_to_fp16 = const()[name = tensor("blocks_15_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611151104)))]; + tensor blocks_15_attn_value_bias_to_fp16 = const()[name = tensor("blocks_15_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614427968)))]; + tensor var_4208_cast_fp16 = conv(bias = blocks_15_attn_value_bias_to_fp16, dilations = var_4208_dilations_0, groups = var_4208_groups_0, pad = var_4208_pad_0, pad_type = var_4208_pad_type_0, strides = var_4208_strides_0, weight = blocks_15_attn_value_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor tile_45 = const()[name = tensor("tile_45"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4211_axis_0 = const()[name = tensor("op_4211_axis_0"), val = tensor(1)]; + tensor var_4211_cast_fp16_0, tensor var_4211_cast_fp16_1, tensor var_4211_cast_fp16_2, tensor var_4211_cast_fp16_3, tensor var_4211_cast_fp16_4, tensor var_4211_cast_fp16_5, tensor var_4211_cast_fp16_6, tensor var_4211_cast_fp16_7, tensor var_4211_cast_fp16_8, tensor var_4211_cast_fp16_9, tensor var_4211_cast_fp16_10, tensor var_4211_cast_fp16_11, tensor var_4211_cast_fp16_12, tensor var_4211_cast_fp16_13, tensor var_4211_cast_fp16_14, tensor var_4211_cast_fp16_15, tensor var_4211_cast_fp16_16, tensor var_4211_cast_fp16_17, tensor var_4211_cast_fp16_18, tensor var_4211_cast_fp16_19 = split(axis = var_4211_axis_0, split_sizes = tile_45, x = var_4210_cast_fp16)[name = tensor("op_4211_cast_fp16")]; + tensor var_4232_perm_0 = const()[name = tensor("op_4232_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_46 = const()[name = tensor("tile_46"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4233_axis_0 = const()[name = tensor("op_4233_axis_0"), val = tensor(3)]; + tensor var_4232_cast_fp16 = transpose(perm = var_4232_perm_0, x = k_31_cast_fp16)[name = tensor("transpose_17")]; + tensor var_4233_cast_fp16_0, tensor var_4233_cast_fp16_1, tensor var_4233_cast_fp16_2, tensor var_4233_cast_fp16_3, tensor var_4233_cast_fp16_4, tensor var_4233_cast_fp16_5, tensor var_4233_cast_fp16_6, tensor var_4233_cast_fp16_7, tensor var_4233_cast_fp16_8, tensor var_4233_cast_fp16_9, tensor var_4233_cast_fp16_10, tensor var_4233_cast_fp16_11, tensor var_4233_cast_fp16_12, tensor var_4233_cast_fp16_13, tensor var_4233_cast_fp16_14, tensor var_4233_cast_fp16_15, tensor var_4233_cast_fp16_16, tensor var_4233_cast_fp16_17, tensor var_4233_cast_fp16_18, tensor var_4233_cast_fp16_19 = split(axis = var_4233_axis_0, split_sizes = tile_46, x = var_4232_cast_fp16)[name = tensor("op_4233_cast_fp16")]; + tensor tile_47 = const()[name = tensor("tile_47"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4254_axis_0 = const()[name = tensor("op_4254_axis_0"), val = tensor(1)]; + tensor var_4254_cast_fp16_0, tensor var_4254_cast_fp16_1, tensor var_4254_cast_fp16_2, tensor var_4254_cast_fp16_3, tensor var_4254_cast_fp16_4, tensor var_4254_cast_fp16_5, tensor var_4254_cast_fp16_6, tensor var_4254_cast_fp16_7, tensor var_4254_cast_fp16_8, tensor var_4254_cast_fp16_9, tensor var_4254_cast_fp16_10, tensor var_4254_cast_fp16_11, tensor var_4254_cast_fp16_12, tensor var_4254_cast_fp16_13, tensor var_4254_cast_fp16_14, tensor var_4254_cast_fp16_15, tensor var_4254_cast_fp16_16, tensor var_4254_cast_fp16_17, tensor var_4254_cast_fp16_18, tensor var_4254_cast_fp16_19 = split(axis = var_4254_axis_0, split_sizes = tile_47, x = var_4208_cast_fp16)[name = tensor("op_4254_cast_fp16")]; + tensor aw_601_equation_0 = const()[name = tensor("aw_601_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_601_cast_fp16 = einsum(equation = aw_601_equation_0, values = (var_4233_cast_fp16_0, var_4211_cast_fp16_0))[name = tensor("aw_601_cast_fp16")]; + tensor aw_603_equation_0 = const()[name = tensor("aw_603_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_603_cast_fp16 = einsum(equation = aw_603_equation_0, values = (var_4233_cast_fp16_1, var_4211_cast_fp16_1))[name = tensor("aw_603_cast_fp16")]; + tensor aw_605_equation_0 = const()[name = tensor("aw_605_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_605_cast_fp16 = einsum(equation = aw_605_equation_0, values = (var_4233_cast_fp16_2, var_4211_cast_fp16_2))[name = tensor("aw_605_cast_fp16")]; + tensor aw_607_equation_0 = const()[name = tensor("aw_607_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_607_cast_fp16 = einsum(equation = aw_607_equation_0, values = (var_4233_cast_fp16_3, var_4211_cast_fp16_3))[name = tensor("aw_607_cast_fp16")]; + tensor aw_609_equation_0 = const()[name = tensor("aw_609_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_609_cast_fp16 = einsum(equation = aw_609_equation_0, values = (var_4233_cast_fp16_4, var_4211_cast_fp16_4))[name = tensor("aw_609_cast_fp16")]; + tensor aw_611_equation_0 = const()[name = tensor("aw_611_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_611_cast_fp16 = einsum(equation = aw_611_equation_0, values = (var_4233_cast_fp16_5, var_4211_cast_fp16_5))[name = tensor("aw_611_cast_fp16")]; + tensor aw_613_equation_0 = const()[name = tensor("aw_613_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_613_cast_fp16 = einsum(equation = aw_613_equation_0, values = (var_4233_cast_fp16_6, var_4211_cast_fp16_6))[name = tensor("aw_613_cast_fp16")]; + tensor aw_615_equation_0 = const()[name = tensor("aw_615_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_615_cast_fp16 = einsum(equation = aw_615_equation_0, values = (var_4233_cast_fp16_7, var_4211_cast_fp16_7))[name = tensor("aw_615_cast_fp16")]; + tensor aw_617_equation_0 = const()[name = tensor("aw_617_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_617_cast_fp16 = einsum(equation = aw_617_equation_0, values = (var_4233_cast_fp16_8, var_4211_cast_fp16_8))[name = tensor("aw_617_cast_fp16")]; + tensor aw_619_equation_0 = const()[name = tensor("aw_619_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_619_cast_fp16 = einsum(equation = aw_619_equation_0, values = (var_4233_cast_fp16_9, var_4211_cast_fp16_9))[name = tensor("aw_619_cast_fp16")]; + tensor aw_621_equation_0 = const()[name = tensor("aw_621_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_621_cast_fp16 = einsum(equation = aw_621_equation_0, values = (var_4233_cast_fp16_10, var_4211_cast_fp16_10))[name = tensor("aw_621_cast_fp16")]; + tensor aw_623_equation_0 = const()[name = tensor("aw_623_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_623_cast_fp16 = einsum(equation = aw_623_equation_0, values = (var_4233_cast_fp16_11, var_4211_cast_fp16_11))[name = tensor("aw_623_cast_fp16")]; + tensor aw_625_equation_0 = const()[name = tensor("aw_625_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_625_cast_fp16 = einsum(equation = aw_625_equation_0, values = (var_4233_cast_fp16_12, var_4211_cast_fp16_12))[name = tensor("aw_625_cast_fp16")]; + tensor aw_627_equation_0 = const()[name = tensor("aw_627_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_627_cast_fp16 = einsum(equation = aw_627_equation_0, values = (var_4233_cast_fp16_13, var_4211_cast_fp16_13))[name = tensor("aw_627_cast_fp16")]; + tensor aw_629_equation_0 = const()[name = tensor("aw_629_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_629_cast_fp16 = einsum(equation = aw_629_equation_0, values = (var_4233_cast_fp16_14, var_4211_cast_fp16_14))[name = tensor("aw_629_cast_fp16")]; + tensor aw_631_equation_0 = const()[name = tensor("aw_631_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_631_cast_fp16 = einsum(equation = aw_631_equation_0, values = (var_4233_cast_fp16_15, var_4211_cast_fp16_15))[name = tensor("aw_631_cast_fp16")]; + tensor aw_633_equation_0 = const()[name = tensor("aw_633_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_633_cast_fp16 = einsum(equation = aw_633_equation_0, values = (var_4233_cast_fp16_16, var_4211_cast_fp16_16))[name = tensor("aw_633_cast_fp16")]; + tensor aw_635_equation_0 = const()[name = tensor("aw_635_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_635_cast_fp16 = einsum(equation = aw_635_equation_0, values = (var_4233_cast_fp16_17, var_4211_cast_fp16_17))[name = tensor("aw_635_cast_fp16")]; + tensor aw_637_equation_0 = const()[name = tensor("aw_637_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_637_cast_fp16 = einsum(equation = aw_637_equation_0, values = (var_4233_cast_fp16_18, var_4211_cast_fp16_18))[name = tensor("aw_637_cast_fp16")]; + tensor aw_639_equation_0 = const()[name = tensor("aw_639_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_639_cast_fp16 = einsum(equation = aw_639_equation_0, values = (var_4233_cast_fp16_19, var_4211_cast_fp16_19))[name = tensor("aw_639_cast_fp16")]; + tensor var_4315_cast_fp16 = softmax(axis = var_4159, x = aw_601_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4316_cast_fp16 = softmax(axis = var_4159, x = aw_603_cast_fp16)[name = tensor("op_4316_cast_fp16")]; + tensor var_4317_cast_fp16 = softmax(axis = var_4159, x = aw_605_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor var_4318_cast_fp16 = softmax(axis = var_4159, x = aw_607_cast_fp16)[name = tensor("op_4318_cast_fp16")]; + tensor var_4319_cast_fp16 = softmax(axis = var_4159, x = aw_609_cast_fp16)[name = tensor("op_4319_cast_fp16")]; + tensor var_4320_cast_fp16 = softmax(axis = var_4159, x = aw_611_cast_fp16)[name = tensor("op_4320_cast_fp16")]; + tensor var_4321_cast_fp16 = softmax(axis = var_4159, x = aw_613_cast_fp16)[name = tensor("op_4321_cast_fp16")]; + tensor var_4322_cast_fp16 = softmax(axis = var_4159, x = aw_615_cast_fp16)[name = tensor("op_4322_cast_fp16")]; + tensor var_4323_cast_fp16 = softmax(axis = var_4159, x = aw_617_cast_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4324_cast_fp16 = softmax(axis = var_4159, x = aw_619_cast_fp16)[name = tensor("op_4324_cast_fp16")]; + tensor var_4325_cast_fp16 = softmax(axis = var_4159, x = aw_621_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor var_4326_cast_fp16 = softmax(axis = var_4159, x = aw_623_cast_fp16)[name = tensor("op_4326_cast_fp16")]; + tensor var_4327_cast_fp16 = softmax(axis = var_4159, x = aw_625_cast_fp16)[name = tensor("op_4327_cast_fp16")]; + tensor var_4328_cast_fp16 = softmax(axis = var_4159, x = aw_627_cast_fp16)[name = tensor("op_4328_cast_fp16")]; + tensor var_4329_cast_fp16 = softmax(axis = var_4159, x = aw_629_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_cast_fp16 = softmax(axis = var_4159, x = aw_631_cast_fp16)[name = tensor("op_4330_cast_fp16")]; + tensor var_4331_cast_fp16 = softmax(axis = var_4159, x = aw_633_cast_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor var_4332_cast_fp16 = softmax(axis = var_4159, x = aw_635_cast_fp16)[name = tensor("op_4332_cast_fp16")]; + tensor var_4333_cast_fp16 = softmax(axis = var_4159, x = aw_637_cast_fp16)[name = tensor("op_4333_cast_fp16")]; + tensor var_4334_cast_fp16 = softmax(axis = var_4159, x = aw_639_cast_fp16)[name = tensor("op_4334_cast_fp16")]; + tensor var_4336_equation_0 = const()[name = tensor("op_4336_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4336_cast_fp16 = einsum(equation = var_4336_equation_0, values = (var_4254_cast_fp16_0, var_4315_cast_fp16))[name = tensor("op_4336_cast_fp16")]; + tensor var_4338_equation_0 = const()[name = tensor("op_4338_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4338_cast_fp16 = einsum(equation = var_4338_equation_0, values = (var_4254_cast_fp16_1, var_4316_cast_fp16))[name = tensor("op_4338_cast_fp16")]; + tensor var_4340_equation_0 = const()[name = tensor("op_4340_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4340_cast_fp16 = einsum(equation = var_4340_equation_0, values = (var_4254_cast_fp16_2, var_4317_cast_fp16))[name = tensor("op_4340_cast_fp16")]; + tensor var_4342_equation_0 = const()[name = tensor("op_4342_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4342_cast_fp16 = einsum(equation = var_4342_equation_0, values = (var_4254_cast_fp16_3, var_4318_cast_fp16))[name = tensor("op_4342_cast_fp16")]; + tensor var_4344_equation_0 = const()[name = tensor("op_4344_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4344_cast_fp16 = einsum(equation = var_4344_equation_0, values = (var_4254_cast_fp16_4, var_4319_cast_fp16))[name = tensor("op_4344_cast_fp16")]; + tensor var_4346_equation_0 = const()[name = tensor("op_4346_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4346_cast_fp16 = einsum(equation = var_4346_equation_0, values = (var_4254_cast_fp16_5, var_4320_cast_fp16))[name = tensor("op_4346_cast_fp16")]; + tensor var_4348_equation_0 = const()[name = tensor("op_4348_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4348_cast_fp16 = einsum(equation = var_4348_equation_0, values = (var_4254_cast_fp16_6, var_4321_cast_fp16))[name = tensor("op_4348_cast_fp16")]; + tensor var_4350_equation_0 = const()[name = tensor("op_4350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4350_cast_fp16 = einsum(equation = var_4350_equation_0, values = (var_4254_cast_fp16_7, var_4322_cast_fp16))[name = tensor("op_4350_cast_fp16")]; + tensor var_4352_equation_0 = const()[name = tensor("op_4352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4352_cast_fp16 = einsum(equation = var_4352_equation_0, values = (var_4254_cast_fp16_8, var_4323_cast_fp16))[name = tensor("op_4352_cast_fp16")]; + tensor var_4354_equation_0 = const()[name = tensor("op_4354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4354_cast_fp16 = einsum(equation = var_4354_equation_0, values = (var_4254_cast_fp16_9, var_4324_cast_fp16))[name = tensor("op_4354_cast_fp16")]; + tensor var_4356_equation_0 = const()[name = tensor("op_4356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4356_cast_fp16 = einsum(equation = var_4356_equation_0, values = (var_4254_cast_fp16_10, var_4325_cast_fp16))[name = tensor("op_4356_cast_fp16")]; + tensor var_4358_equation_0 = const()[name = tensor("op_4358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4358_cast_fp16 = einsum(equation = var_4358_equation_0, values = (var_4254_cast_fp16_11, var_4326_cast_fp16))[name = tensor("op_4358_cast_fp16")]; + tensor var_4360_equation_0 = const()[name = tensor("op_4360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4360_cast_fp16 = einsum(equation = var_4360_equation_0, values = (var_4254_cast_fp16_12, var_4327_cast_fp16))[name = tensor("op_4360_cast_fp16")]; + tensor var_4362_equation_0 = const()[name = tensor("op_4362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4362_cast_fp16 = einsum(equation = var_4362_equation_0, values = (var_4254_cast_fp16_13, var_4328_cast_fp16))[name = tensor("op_4362_cast_fp16")]; + tensor var_4364_equation_0 = const()[name = tensor("op_4364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4364_cast_fp16 = einsum(equation = var_4364_equation_0, values = (var_4254_cast_fp16_14, var_4329_cast_fp16))[name = tensor("op_4364_cast_fp16")]; + tensor var_4366_equation_0 = const()[name = tensor("op_4366_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4366_cast_fp16 = einsum(equation = var_4366_equation_0, values = (var_4254_cast_fp16_15, var_4330_cast_fp16))[name = tensor("op_4366_cast_fp16")]; + tensor var_4368_equation_0 = const()[name = tensor("op_4368_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4368_cast_fp16 = einsum(equation = var_4368_equation_0, values = (var_4254_cast_fp16_16, var_4331_cast_fp16))[name = tensor("op_4368_cast_fp16")]; + tensor var_4370_equation_0 = const()[name = tensor("op_4370_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4370_cast_fp16 = einsum(equation = var_4370_equation_0, values = (var_4254_cast_fp16_17, var_4332_cast_fp16))[name = tensor("op_4370_cast_fp16")]; + tensor var_4372_equation_0 = const()[name = tensor("op_4372_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4372_cast_fp16 = einsum(equation = var_4372_equation_0, values = (var_4254_cast_fp16_18, var_4333_cast_fp16))[name = tensor("op_4372_cast_fp16")]; + tensor var_4374_equation_0 = const()[name = tensor("op_4374_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4374_cast_fp16 = einsum(equation = var_4374_equation_0, values = (var_4254_cast_fp16_19, var_4334_cast_fp16))[name = tensor("op_4374_cast_fp16")]; + tensor input_155_interleave_0 = const()[name = tensor("input_155_interleave_0"), val = tensor(false)]; + tensor input_155_cast_fp16 = concat(axis = var_4159, interleave = input_155_interleave_0, values = (var_4336_cast_fp16, var_4338_cast_fp16, var_4340_cast_fp16, var_4342_cast_fp16, var_4344_cast_fp16, var_4346_cast_fp16, var_4348_cast_fp16, var_4350_cast_fp16, var_4352_cast_fp16, var_4354_cast_fp16, var_4356_cast_fp16, var_4358_cast_fp16, var_4360_cast_fp16, var_4362_cast_fp16, var_4364_cast_fp16, var_4366_cast_fp16, var_4368_cast_fp16, var_4370_cast_fp16, var_4372_cast_fp16, var_4374_cast_fp16))[name = tensor("input_155_cast_fp16")]; + tensor var_4383_pad_type_0 = const()[name = tensor("op_4383_pad_type_0"), val = tensor("valid")]; + tensor var_4383_strides_0 = const()[name = tensor("op_4383_strides_0"), val = tensor([1, 1])]; + tensor var_4383_pad_0 = const()[name = tensor("op_4383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4383_dilations_0 = const()[name = tensor("op_4383_dilations_0"), val = tensor([1, 1])]; + tensor var_4383_groups_0 = const()[name = tensor("op_4383_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_out_weight_to_fp16 = const()[name = tensor("blocks_15_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614430592)))]; + tensor blocks_15_attn_out_bias_to_fp16 = const()[name = tensor("blocks_15_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617707456)))]; + tensor var_4383_cast_fp16 = conv(bias = blocks_15_attn_out_bias_to_fp16, dilations = var_4383_dilations_0, groups = var_4383_groups_0, pad = var_4383_pad_0, pad_type = var_4383_pad_type_0, strides = var_4383_strides_0, weight = blocks_15_attn_out_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("op_4383_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_4383_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor input_157_axes_0 = const()[name = tensor("input_157_axes_0"), val = tensor([1])]; + tensor input_157_gamma_0_to_fp16 = const()[name = tensor("input_157_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617710080)))]; + tensor input_157_beta_0_to_fp16 = const()[name = tensor("input_157_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617712704)))]; + tensor var_4393_to_fp16 = const()[name = tensor("op_4393_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = input_157_beta_0_to_fp16, epsilon = var_4393_to_fp16, gamma = input_157_gamma_0_to_fp16, x = inputs_63_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("valid")]; + tensor input_159_strides_0 = const()[name = tensor("input_159_strides_0"), val = tensor([1, 1])]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_159_dilations_0 = const()[name = tensor("input_159_dilations_0"), val = tensor([1, 1])]; + tensor input_159_groups_0 = const()[name = tensor("input_159_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617715328)))]; + tensor blocks_15_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630822592)))]; + tensor input_159_cast_fp16 = conv(bias = blocks_15_mlp_0_bias_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = blocks_15_mlp_0_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_4419_pad_type_0 = const()[name = tensor("op_4419_pad_type_0"), val = tensor("valid")]; + tensor var_4419_strides_0 = const()[name = tensor("op_4419_strides_0"), val = tensor([1, 1])]; + tensor var_4419_pad_0 = const()[name = tensor("op_4419_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4419_dilations_0 = const()[name = tensor("op_4419_dilations_0"), val = tensor([1, 1])]; + tensor var_4419_groups_0 = const()[name = tensor("op_4419_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630832896)))]; + tensor blocks_15_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643940160)))]; + tensor var_4419_cast_fp16 = conv(bias = blocks_15_mlp_2_bias_to_fp16, dilations = var_4419_dilations_0, groups = var_4419_groups_0, pad = var_4419_pad_0, pad_type = var_4419_pad_type_0, strides = var_4419_strides_0, weight = blocks_15_mlp_2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("op_4419_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = var_4419_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_4428 = const()[name = tensor("op_4428"), val = tensor(1)]; + tensor input_163_axes_0 = const()[name = tensor("input_163_axes_0"), val = tensor([1])]; + tensor input_163_gamma_0_to_fp16 = const()[name = tensor("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643942784)))]; + tensor input_163_beta_0_to_fp16 = const()[name = tensor("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643945408)))]; + tensor var_4444_to_fp16 = const()[name = tensor("op_4444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = input_163_beta_0_to_fp16, epsilon = var_4444_to_fp16, gamma = input_163_gamma_0_to_fp16, x = inputs_65_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor q_33_pad_type_0 = const()[name = tensor("q_33_pad_type_0"), val = tensor("valid")]; + tensor q_33_strides_0 = const()[name = tensor("q_33_strides_0"), val = tensor([1, 1])]; + tensor q_33_pad_0 = const()[name = tensor("q_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_33_dilations_0 = const()[name = tensor("q_33_dilations_0"), val = tensor([1, 1])]; + tensor q_33_groups_0 = const()[name = tensor("q_33_groups_0"), val = tensor(1)]; + tensor var_4479_weight_0_to_fp16 = const()[name = tensor("op_4479_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643948032)))]; + tensor var_4479_bias_0_to_fp16 = const()[name = tensor("op_4479_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647224896)))]; + tensor var_4479_cast_fp16 = conv(bias = var_4479_bias_0_to_fp16, dilations = q_33_dilations_0, groups = q_33_groups_0, pad = q_33_pad_0, pad_type = q_33_pad_type_0, strides = q_33_strides_0, weight = var_4479_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("op_4479_cast_fp16")]; + tensor k_33_pad_type_0 = const()[name = tensor("k_33_pad_type_0"), val = tensor("valid")]; + tensor k_33_strides_0 = const()[name = tensor("k_33_strides_0"), val = tensor([1, 1])]; + tensor k_33_pad_0 = const()[name = tensor("k_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_33_dilations_0 = const()[name = tensor("k_33_dilations_0"), val = tensor([1, 1])]; + tensor k_33_groups_0 = const()[name = tensor("k_33_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_key_weight_to_fp16 = const()[name = tensor("blocks_16_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647227520)))]; + tensor k_33_cast_fp16 = conv(dilations = k_33_dilations_0, groups = k_33_groups_0, pad = k_33_pad_0, pad_type = k_33_pad_type_0, strides = k_33_strides_0, weight = blocks_16_attn_key_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor var_4477_pad_type_0 = const()[name = tensor("op_4477_pad_type_0"), val = tensor("valid")]; + tensor var_4477_strides_0 = const()[name = tensor("op_4477_strides_0"), val = tensor([1, 1])]; + tensor var_4477_pad_0 = const()[name = tensor("op_4477_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4477_dilations_0 = const()[name = tensor("op_4477_dilations_0"), val = tensor([1, 1])]; + tensor var_4477_groups_0 = const()[name = tensor("op_4477_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_value_weight_to_fp16 = const()[name = tensor("blocks_16_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650504384)))]; + tensor blocks_16_attn_value_bias_to_fp16 = const()[name = tensor("blocks_16_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653781248)))]; + tensor var_4477_cast_fp16 = conv(bias = blocks_16_attn_value_bias_to_fp16, dilations = var_4477_dilations_0, groups = var_4477_groups_0, pad = var_4477_pad_0, pad_type = var_4477_pad_type_0, strides = var_4477_strides_0, weight = blocks_16_attn_value_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("op_4477_cast_fp16")]; + tensor tile_48 = const()[name = tensor("tile_48"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4480_axis_0 = const()[name = tensor("op_4480_axis_0"), val = tensor(1)]; + tensor var_4480_cast_fp16_0, tensor var_4480_cast_fp16_1, tensor var_4480_cast_fp16_2, tensor var_4480_cast_fp16_3, tensor var_4480_cast_fp16_4, tensor var_4480_cast_fp16_5, tensor var_4480_cast_fp16_6, tensor var_4480_cast_fp16_7, tensor var_4480_cast_fp16_8, tensor var_4480_cast_fp16_9, tensor var_4480_cast_fp16_10, tensor var_4480_cast_fp16_11, tensor var_4480_cast_fp16_12, tensor var_4480_cast_fp16_13, tensor var_4480_cast_fp16_14, tensor var_4480_cast_fp16_15, tensor var_4480_cast_fp16_16, tensor var_4480_cast_fp16_17, tensor var_4480_cast_fp16_18, tensor var_4480_cast_fp16_19 = split(axis = var_4480_axis_0, split_sizes = tile_48, x = var_4479_cast_fp16)[name = tensor("op_4480_cast_fp16")]; + tensor var_4501_perm_0 = const()[name = tensor("op_4501_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_49 = const()[name = tensor("tile_49"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4502_axis_0 = const()[name = tensor("op_4502_axis_0"), val = tensor(3)]; + tensor var_4501_cast_fp16 = transpose(perm = var_4501_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_16")]; + tensor var_4502_cast_fp16_0, tensor var_4502_cast_fp16_1, tensor var_4502_cast_fp16_2, tensor var_4502_cast_fp16_3, tensor var_4502_cast_fp16_4, tensor var_4502_cast_fp16_5, tensor var_4502_cast_fp16_6, tensor var_4502_cast_fp16_7, tensor var_4502_cast_fp16_8, tensor var_4502_cast_fp16_9, tensor var_4502_cast_fp16_10, tensor var_4502_cast_fp16_11, tensor var_4502_cast_fp16_12, tensor var_4502_cast_fp16_13, tensor var_4502_cast_fp16_14, tensor var_4502_cast_fp16_15, tensor var_4502_cast_fp16_16, tensor var_4502_cast_fp16_17, tensor var_4502_cast_fp16_18, tensor var_4502_cast_fp16_19 = split(axis = var_4502_axis_0, split_sizes = tile_49, x = var_4501_cast_fp16)[name = tensor("op_4502_cast_fp16")]; + tensor tile_50 = const()[name = tensor("tile_50"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4523_axis_0 = const()[name = tensor("op_4523_axis_0"), val = tensor(1)]; + tensor var_4523_cast_fp16_0, tensor var_4523_cast_fp16_1, tensor var_4523_cast_fp16_2, tensor var_4523_cast_fp16_3, tensor var_4523_cast_fp16_4, tensor var_4523_cast_fp16_5, tensor var_4523_cast_fp16_6, tensor var_4523_cast_fp16_7, tensor var_4523_cast_fp16_8, tensor var_4523_cast_fp16_9, tensor var_4523_cast_fp16_10, tensor var_4523_cast_fp16_11, tensor var_4523_cast_fp16_12, tensor var_4523_cast_fp16_13, tensor var_4523_cast_fp16_14, tensor var_4523_cast_fp16_15, tensor var_4523_cast_fp16_16, tensor var_4523_cast_fp16_17, tensor var_4523_cast_fp16_18, tensor var_4523_cast_fp16_19 = split(axis = var_4523_axis_0, split_sizes = tile_50, x = var_4477_cast_fp16)[name = tensor("op_4523_cast_fp16")]; + tensor aw_641_equation_0 = const()[name = tensor("aw_641_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_641_cast_fp16 = einsum(equation = aw_641_equation_0, values = (var_4502_cast_fp16_0, var_4480_cast_fp16_0))[name = tensor("aw_641_cast_fp16")]; + tensor aw_643_equation_0 = const()[name = tensor("aw_643_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_643_cast_fp16 = einsum(equation = aw_643_equation_0, values = (var_4502_cast_fp16_1, var_4480_cast_fp16_1))[name = tensor("aw_643_cast_fp16")]; + tensor aw_645_equation_0 = const()[name = tensor("aw_645_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_645_cast_fp16 = einsum(equation = aw_645_equation_0, values = (var_4502_cast_fp16_2, var_4480_cast_fp16_2))[name = tensor("aw_645_cast_fp16")]; + tensor aw_647_equation_0 = const()[name = tensor("aw_647_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_647_cast_fp16 = einsum(equation = aw_647_equation_0, values = (var_4502_cast_fp16_3, var_4480_cast_fp16_3))[name = tensor("aw_647_cast_fp16")]; + tensor aw_649_equation_0 = const()[name = tensor("aw_649_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_649_cast_fp16 = einsum(equation = aw_649_equation_0, values = (var_4502_cast_fp16_4, var_4480_cast_fp16_4))[name = tensor("aw_649_cast_fp16")]; + tensor aw_651_equation_0 = const()[name = tensor("aw_651_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_651_cast_fp16 = einsum(equation = aw_651_equation_0, values = (var_4502_cast_fp16_5, var_4480_cast_fp16_5))[name = tensor("aw_651_cast_fp16")]; + tensor aw_653_equation_0 = const()[name = tensor("aw_653_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_653_cast_fp16 = einsum(equation = aw_653_equation_0, values = (var_4502_cast_fp16_6, var_4480_cast_fp16_6))[name = tensor("aw_653_cast_fp16")]; + tensor aw_655_equation_0 = const()[name = tensor("aw_655_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_655_cast_fp16 = einsum(equation = aw_655_equation_0, values = (var_4502_cast_fp16_7, var_4480_cast_fp16_7))[name = tensor("aw_655_cast_fp16")]; + tensor aw_657_equation_0 = const()[name = tensor("aw_657_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_657_cast_fp16 = einsum(equation = aw_657_equation_0, values = (var_4502_cast_fp16_8, var_4480_cast_fp16_8))[name = tensor("aw_657_cast_fp16")]; + tensor aw_659_equation_0 = const()[name = tensor("aw_659_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_659_cast_fp16 = einsum(equation = aw_659_equation_0, values = (var_4502_cast_fp16_9, var_4480_cast_fp16_9))[name = tensor("aw_659_cast_fp16")]; + tensor aw_661_equation_0 = const()[name = tensor("aw_661_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_661_cast_fp16 = einsum(equation = aw_661_equation_0, values = (var_4502_cast_fp16_10, var_4480_cast_fp16_10))[name = tensor("aw_661_cast_fp16")]; + tensor aw_663_equation_0 = const()[name = tensor("aw_663_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_663_cast_fp16 = einsum(equation = aw_663_equation_0, values = (var_4502_cast_fp16_11, var_4480_cast_fp16_11))[name = tensor("aw_663_cast_fp16")]; + tensor aw_665_equation_0 = const()[name = tensor("aw_665_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_665_cast_fp16 = einsum(equation = aw_665_equation_0, values = (var_4502_cast_fp16_12, var_4480_cast_fp16_12))[name = tensor("aw_665_cast_fp16")]; + tensor aw_667_equation_0 = const()[name = tensor("aw_667_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_667_cast_fp16 = einsum(equation = aw_667_equation_0, values = (var_4502_cast_fp16_13, var_4480_cast_fp16_13))[name = tensor("aw_667_cast_fp16")]; + tensor aw_669_equation_0 = const()[name = tensor("aw_669_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_669_cast_fp16 = einsum(equation = aw_669_equation_0, values = (var_4502_cast_fp16_14, var_4480_cast_fp16_14))[name = tensor("aw_669_cast_fp16")]; + tensor aw_671_equation_0 = const()[name = tensor("aw_671_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_671_cast_fp16 = einsum(equation = aw_671_equation_0, values = (var_4502_cast_fp16_15, var_4480_cast_fp16_15))[name = tensor("aw_671_cast_fp16")]; + tensor aw_673_equation_0 = const()[name = tensor("aw_673_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_673_cast_fp16 = einsum(equation = aw_673_equation_0, values = (var_4502_cast_fp16_16, var_4480_cast_fp16_16))[name = tensor("aw_673_cast_fp16")]; + tensor aw_675_equation_0 = const()[name = tensor("aw_675_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_675_cast_fp16 = einsum(equation = aw_675_equation_0, values = (var_4502_cast_fp16_17, var_4480_cast_fp16_17))[name = tensor("aw_675_cast_fp16")]; + tensor aw_677_equation_0 = const()[name = tensor("aw_677_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_677_cast_fp16 = einsum(equation = aw_677_equation_0, values = (var_4502_cast_fp16_18, var_4480_cast_fp16_18))[name = tensor("aw_677_cast_fp16")]; + tensor aw_679_equation_0 = const()[name = tensor("aw_679_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_679_cast_fp16 = einsum(equation = aw_679_equation_0, values = (var_4502_cast_fp16_19, var_4480_cast_fp16_19))[name = tensor("aw_679_cast_fp16")]; + tensor var_4584_cast_fp16 = softmax(axis = var_4428, x = aw_641_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = softmax(axis = var_4428, x = aw_643_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor var_4586_cast_fp16 = softmax(axis = var_4428, x = aw_645_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = softmax(axis = var_4428, x = aw_647_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor var_4588_cast_fp16 = softmax(axis = var_4428, x = aw_649_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = softmax(axis = var_4428, x = aw_651_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor var_4590_cast_fp16 = softmax(axis = var_4428, x = aw_653_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_4428, x = aw_655_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor var_4592_cast_fp16 = softmax(axis = var_4428, x = aw_657_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = softmax(axis = var_4428, x = aw_659_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor var_4594_cast_fp16 = softmax(axis = var_4428, x = aw_661_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = softmax(axis = var_4428, x = aw_663_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor var_4596_cast_fp16 = softmax(axis = var_4428, x = aw_665_cast_fp16)[name = tensor("op_4596_cast_fp16")]; + tensor var_4597_cast_fp16 = softmax(axis = var_4428, x = aw_667_cast_fp16)[name = tensor("op_4597_cast_fp16")]; + tensor var_4598_cast_fp16 = softmax(axis = var_4428, x = aw_669_cast_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4599_cast_fp16 = softmax(axis = var_4428, x = aw_671_cast_fp16)[name = tensor("op_4599_cast_fp16")]; + tensor var_4600_cast_fp16 = softmax(axis = var_4428, x = aw_673_cast_fp16)[name = tensor("op_4600_cast_fp16")]; + tensor var_4601_cast_fp16 = softmax(axis = var_4428, x = aw_675_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor var_4602_cast_fp16 = softmax(axis = var_4428, x = aw_677_cast_fp16)[name = tensor("op_4602_cast_fp16")]; + tensor var_4603_cast_fp16 = softmax(axis = var_4428, x = aw_679_cast_fp16)[name = tensor("op_4603_cast_fp16")]; + tensor var_4605_equation_0 = const()[name = tensor("op_4605_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4605_cast_fp16 = einsum(equation = var_4605_equation_0, values = (var_4523_cast_fp16_0, var_4584_cast_fp16))[name = tensor("op_4605_cast_fp16")]; + tensor var_4607_equation_0 = const()[name = tensor("op_4607_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4607_cast_fp16 = einsum(equation = var_4607_equation_0, values = (var_4523_cast_fp16_1, var_4585_cast_fp16))[name = tensor("op_4607_cast_fp16")]; + tensor var_4609_equation_0 = const()[name = tensor("op_4609_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4609_cast_fp16 = einsum(equation = var_4609_equation_0, values = (var_4523_cast_fp16_2, var_4586_cast_fp16))[name = tensor("op_4609_cast_fp16")]; + tensor var_4611_equation_0 = const()[name = tensor("op_4611_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4611_cast_fp16 = einsum(equation = var_4611_equation_0, values = (var_4523_cast_fp16_3, var_4587_cast_fp16))[name = tensor("op_4611_cast_fp16")]; + tensor var_4613_equation_0 = const()[name = tensor("op_4613_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4613_cast_fp16 = einsum(equation = var_4613_equation_0, values = (var_4523_cast_fp16_4, var_4588_cast_fp16))[name = tensor("op_4613_cast_fp16")]; + tensor var_4615_equation_0 = const()[name = tensor("op_4615_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4615_cast_fp16 = einsum(equation = var_4615_equation_0, values = (var_4523_cast_fp16_5, var_4589_cast_fp16))[name = tensor("op_4615_cast_fp16")]; + tensor var_4617_equation_0 = const()[name = tensor("op_4617_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4617_cast_fp16 = einsum(equation = var_4617_equation_0, values = (var_4523_cast_fp16_6, var_4590_cast_fp16))[name = tensor("op_4617_cast_fp16")]; + tensor var_4619_equation_0 = const()[name = tensor("op_4619_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4619_cast_fp16 = einsum(equation = var_4619_equation_0, values = (var_4523_cast_fp16_7, var_4591_cast_fp16))[name = tensor("op_4619_cast_fp16")]; + tensor var_4621_equation_0 = const()[name = tensor("op_4621_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4621_cast_fp16 = einsum(equation = var_4621_equation_0, values = (var_4523_cast_fp16_8, var_4592_cast_fp16))[name = tensor("op_4621_cast_fp16")]; + tensor var_4623_equation_0 = const()[name = tensor("op_4623_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4623_cast_fp16 = einsum(equation = var_4623_equation_0, values = (var_4523_cast_fp16_9, var_4593_cast_fp16))[name = tensor("op_4623_cast_fp16")]; + tensor var_4625_equation_0 = const()[name = tensor("op_4625_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4625_cast_fp16 = einsum(equation = var_4625_equation_0, values = (var_4523_cast_fp16_10, var_4594_cast_fp16))[name = tensor("op_4625_cast_fp16")]; + tensor var_4627_equation_0 = const()[name = tensor("op_4627_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4627_cast_fp16 = einsum(equation = var_4627_equation_0, values = (var_4523_cast_fp16_11, var_4595_cast_fp16))[name = tensor("op_4627_cast_fp16")]; + tensor var_4629_equation_0 = const()[name = tensor("op_4629_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4629_cast_fp16 = einsum(equation = var_4629_equation_0, values = (var_4523_cast_fp16_12, var_4596_cast_fp16))[name = tensor("op_4629_cast_fp16")]; + tensor var_4631_equation_0 = const()[name = tensor("op_4631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4631_cast_fp16 = einsum(equation = var_4631_equation_0, values = (var_4523_cast_fp16_13, var_4597_cast_fp16))[name = tensor("op_4631_cast_fp16")]; + tensor var_4633_equation_0 = const()[name = tensor("op_4633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4633_cast_fp16 = einsum(equation = var_4633_equation_0, values = (var_4523_cast_fp16_14, var_4598_cast_fp16))[name = tensor("op_4633_cast_fp16")]; + tensor var_4635_equation_0 = const()[name = tensor("op_4635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4635_cast_fp16 = einsum(equation = var_4635_equation_0, values = (var_4523_cast_fp16_15, var_4599_cast_fp16))[name = tensor("op_4635_cast_fp16")]; + tensor var_4637_equation_0 = const()[name = tensor("op_4637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4637_cast_fp16 = einsum(equation = var_4637_equation_0, values = (var_4523_cast_fp16_16, var_4600_cast_fp16))[name = tensor("op_4637_cast_fp16")]; + tensor var_4639_equation_0 = const()[name = tensor("op_4639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4639_cast_fp16 = einsum(equation = var_4639_equation_0, values = (var_4523_cast_fp16_17, var_4601_cast_fp16))[name = tensor("op_4639_cast_fp16")]; + tensor var_4641_equation_0 = const()[name = tensor("op_4641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4641_cast_fp16 = einsum(equation = var_4641_equation_0, values = (var_4523_cast_fp16_18, var_4602_cast_fp16))[name = tensor("op_4641_cast_fp16")]; + tensor var_4643_equation_0 = const()[name = tensor("op_4643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4643_cast_fp16 = einsum(equation = var_4643_equation_0, values = (var_4523_cast_fp16_19, var_4603_cast_fp16))[name = tensor("op_4643_cast_fp16")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165_cast_fp16 = concat(axis = var_4428, interleave = input_165_interleave_0, values = (var_4605_cast_fp16, var_4607_cast_fp16, var_4609_cast_fp16, var_4611_cast_fp16, var_4613_cast_fp16, var_4615_cast_fp16, var_4617_cast_fp16, var_4619_cast_fp16, var_4621_cast_fp16, var_4623_cast_fp16, var_4625_cast_fp16, var_4627_cast_fp16, var_4629_cast_fp16, var_4631_cast_fp16, var_4633_cast_fp16, var_4635_cast_fp16, var_4637_cast_fp16, var_4639_cast_fp16, var_4641_cast_fp16, var_4643_cast_fp16))[name = tensor("input_165_cast_fp16")]; + tensor var_4652_pad_type_0 = const()[name = tensor("op_4652_pad_type_0"), val = tensor("valid")]; + tensor var_4652_strides_0 = const()[name = tensor("op_4652_strides_0"), val = tensor([1, 1])]; + tensor var_4652_pad_0 = const()[name = tensor("op_4652_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4652_dilations_0 = const()[name = tensor("op_4652_dilations_0"), val = tensor([1, 1])]; + tensor var_4652_groups_0 = const()[name = tensor("op_4652_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_out_weight_to_fp16 = const()[name = tensor("blocks_16_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653783872)))]; + tensor blocks_16_attn_out_bias_to_fp16 = const()[name = tensor("blocks_16_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657060736)))]; + tensor var_4652_cast_fp16 = conv(bias = blocks_16_attn_out_bias_to_fp16, dilations = var_4652_dilations_0, groups = var_4652_groups_0, pad = var_4652_pad_0, pad_type = var_4652_pad_type_0, strides = var_4652_strides_0, weight = blocks_16_attn_out_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("op_4652_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = var_4652_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([1])]; + tensor input_167_gamma_0_to_fp16 = const()[name = tensor("input_167_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657063360)))]; + tensor input_167_beta_0_to_fp16 = const()[name = tensor("input_167_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657065984)))]; + tensor var_4662_to_fp16 = const()[name = tensor("op_4662_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = input_167_beta_0_to_fp16, epsilon = var_4662_to_fp16, gamma = input_167_gamma_0_to_fp16, x = inputs_67_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("valid")]; + tensor input_169_strides_0 = const()[name = tensor("input_169_strides_0"), val = tensor([1, 1])]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_169_dilations_0 = const()[name = tensor("input_169_dilations_0"), val = tensor([1, 1])]; + tensor input_169_groups_0 = const()[name = tensor("input_169_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657068608)))]; + tensor blocks_16_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670175872)))]; + tensor input_169_cast_fp16 = conv(bias = blocks_16_mlp_0_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = blocks_16_mlp_0_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_4688_pad_type_0 = const()[name = tensor("op_4688_pad_type_0"), val = tensor("valid")]; + tensor var_4688_strides_0 = const()[name = tensor("op_4688_strides_0"), val = tensor([1, 1])]; + tensor var_4688_pad_0 = const()[name = tensor("op_4688_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4688_dilations_0 = const()[name = tensor("op_4688_dilations_0"), val = tensor([1, 1])]; + tensor var_4688_groups_0 = const()[name = tensor("op_4688_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670186176)))]; + tensor blocks_16_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683293440)))]; + tensor var_4688_cast_fp16 = conv(bias = blocks_16_mlp_2_bias_to_fp16, dilations = var_4688_dilations_0, groups = var_4688_groups_0, pad = var_4688_pad_0, pad_type = var_4688_pad_type_0, strides = var_4688_strides_0, weight = blocks_16_mlp_2_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_4688_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4697 = const()[name = tensor("op_4697"), val = tensor(1)]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([1])]; + tensor input_173_gamma_0_to_fp16 = const()[name = tensor("input_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683296064)))]; + tensor input_173_beta_0_to_fp16 = const()[name = tensor("input_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683298688)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = input_173_beta_0_to_fp16, epsilon = var_4713_to_fp16, gamma = input_173_gamma_0_to_fp16, x = inputs_69_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor q_35_pad_type_0 = const()[name = tensor("q_35_pad_type_0"), val = tensor("valid")]; + tensor q_35_strides_0 = const()[name = tensor("q_35_strides_0"), val = tensor([1, 1])]; + tensor q_35_pad_0 = const()[name = tensor("q_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_35_dilations_0 = const()[name = tensor("q_35_dilations_0"), val = tensor([1, 1])]; + tensor q_35_groups_0 = const()[name = tensor("q_35_groups_0"), val = tensor(1)]; + tensor var_4748_weight_0_to_fp16 = const()[name = tensor("op_4748_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683301312)))]; + tensor var_4748_bias_0_to_fp16 = const()[name = tensor("op_4748_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686578176)))]; + tensor var_4748_cast_fp16 = conv(bias = var_4748_bias_0_to_fp16, dilations = q_35_dilations_0, groups = q_35_groups_0, pad = q_35_pad_0, pad_type = q_35_pad_type_0, strides = q_35_strides_0, weight = var_4748_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4748_cast_fp16")]; + tensor k_35_pad_type_0 = const()[name = tensor("k_35_pad_type_0"), val = tensor("valid")]; + tensor k_35_strides_0 = const()[name = tensor("k_35_strides_0"), val = tensor([1, 1])]; + tensor k_35_pad_0 = const()[name = tensor("k_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_35_dilations_0 = const()[name = tensor("k_35_dilations_0"), val = tensor([1, 1])]; + tensor k_35_groups_0 = const()[name = tensor("k_35_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_key_weight_to_fp16 = const()[name = tensor("blocks_17_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686580800)))]; + tensor k_35_cast_fp16 = conv(dilations = k_35_dilations_0, groups = k_35_groups_0, pad = k_35_pad_0, pad_type = k_35_pad_type_0, strides = k_35_strides_0, weight = blocks_17_attn_key_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_4746_pad_type_0 = const()[name = tensor("op_4746_pad_type_0"), val = tensor("valid")]; + tensor var_4746_strides_0 = const()[name = tensor("op_4746_strides_0"), val = tensor([1, 1])]; + tensor var_4746_pad_0 = const()[name = tensor("op_4746_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4746_dilations_0 = const()[name = tensor("op_4746_dilations_0"), val = tensor([1, 1])]; + tensor var_4746_groups_0 = const()[name = tensor("op_4746_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_value_weight_to_fp16 = const()[name = tensor("blocks_17_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689857664)))]; + tensor blocks_17_attn_value_bias_to_fp16 = const()[name = tensor("blocks_17_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693134528)))]; + tensor var_4746_cast_fp16 = conv(bias = blocks_17_attn_value_bias_to_fp16, dilations = var_4746_dilations_0, groups = var_4746_groups_0, pad = var_4746_pad_0, pad_type = var_4746_pad_type_0, strides = var_4746_strides_0, weight = blocks_17_attn_value_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4746_cast_fp16")]; + tensor tile_51 = const()[name = tensor("tile_51"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4749_axis_0 = const()[name = tensor("op_4749_axis_0"), val = tensor(1)]; + tensor var_4749_cast_fp16_0, tensor var_4749_cast_fp16_1, tensor var_4749_cast_fp16_2, tensor var_4749_cast_fp16_3, tensor var_4749_cast_fp16_4, tensor var_4749_cast_fp16_5, tensor var_4749_cast_fp16_6, tensor var_4749_cast_fp16_7, tensor var_4749_cast_fp16_8, tensor var_4749_cast_fp16_9, tensor var_4749_cast_fp16_10, tensor var_4749_cast_fp16_11, tensor var_4749_cast_fp16_12, tensor var_4749_cast_fp16_13, tensor var_4749_cast_fp16_14, tensor var_4749_cast_fp16_15, tensor var_4749_cast_fp16_16, tensor var_4749_cast_fp16_17, tensor var_4749_cast_fp16_18, tensor var_4749_cast_fp16_19 = split(axis = var_4749_axis_0, split_sizes = tile_51, x = var_4748_cast_fp16)[name = tensor("op_4749_cast_fp16")]; + tensor var_4770_perm_0 = const()[name = tensor("op_4770_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_52 = const()[name = tensor("tile_52"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4771_axis_0 = const()[name = tensor("op_4771_axis_0"), val = tensor(3)]; + tensor var_4770_cast_fp16 = transpose(perm = var_4770_perm_0, x = k_35_cast_fp16)[name = tensor("transpose_15")]; + tensor var_4771_cast_fp16_0, tensor var_4771_cast_fp16_1, tensor var_4771_cast_fp16_2, tensor var_4771_cast_fp16_3, tensor var_4771_cast_fp16_4, tensor var_4771_cast_fp16_5, tensor var_4771_cast_fp16_6, tensor var_4771_cast_fp16_7, tensor var_4771_cast_fp16_8, tensor var_4771_cast_fp16_9, tensor var_4771_cast_fp16_10, tensor var_4771_cast_fp16_11, tensor var_4771_cast_fp16_12, tensor var_4771_cast_fp16_13, tensor var_4771_cast_fp16_14, tensor var_4771_cast_fp16_15, tensor var_4771_cast_fp16_16, tensor var_4771_cast_fp16_17, tensor var_4771_cast_fp16_18, tensor var_4771_cast_fp16_19 = split(axis = var_4771_axis_0, split_sizes = tile_52, x = var_4770_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor tile_53 = const()[name = tensor("tile_53"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4792_axis_0 = const()[name = tensor("op_4792_axis_0"), val = tensor(1)]; + tensor var_4792_cast_fp16_0, tensor var_4792_cast_fp16_1, tensor var_4792_cast_fp16_2, tensor var_4792_cast_fp16_3, tensor var_4792_cast_fp16_4, tensor var_4792_cast_fp16_5, tensor var_4792_cast_fp16_6, tensor var_4792_cast_fp16_7, tensor var_4792_cast_fp16_8, tensor var_4792_cast_fp16_9, tensor var_4792_cast_fp16_10, tensor var_4792_cast_fp16_11, tensor var_4792_cast_fp16_12, tensor var_4792_cast_fp16_13, tensor var_4792_cast_fp16_14, tensor var_4792_cast_fp16_15, tensor var_4792_cast_fp16_16, tensor var_4792_cast_fp16_17, tensor var_4792_cast_fp16_18, tensor var_4792_cast_fp16_19 = split(axis = var_4792_axis_0, split_sizes = tile_53, x = var_4746_cast_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor aw_681_equation_0 = const()[name = tensor("aw_681_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_681_cast_fp16 = einsum(equation = aw_681_equation_0, values = (var_4771_cast_fp16_0, var_4749_cast_fp16_0))[name = tensor("aw_681_cast_fp16")]; + tensor aw_683_equation_0 = const()[name = tensor("aw_683_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_683_cast_fp16 = einsum(equation = aw_683_equation_0, values = (var_4771_cast_fp16_1, var_4749_cast_fp16_1))[name = tensor("aw_683_cast_fp16")]; + tensor aw_685_equation_0 = const()[name = tensor("aw_685_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_685_cast_fp16 = einsum(equation = aw_685_equation_0, values = (var_4771_cast_fp16_2, var_4749_cast_fp16_2))[name = tensor("aw_685_cast_fp16")]; + tensor aw_687_equation_0 = const()[name = tensor("aw_687_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_687_cast_fp16 = einsum(equation = aw_687_equation_0, values = (var_4771_cast_fp16_3, var_4749_cast_fp16_3))[name = tensor("aw_687_cast_fp16")]; + tensor aw_689_equation_0 = const()[name = tensor("aw_689_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_689_cast_fp16 = einsum(equation = aw_689_equation_0, values = (var_4771_cast_fp16_4, var_4749_cast_fp16_4))[name = tensor("aw_689_cast_fp16")]; + tensor aw_691_equation_0 = const()[name = tensor("aw_691_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_691_cast_fp16 = einsum(equation = aw_691_equation_0, values = (var_4771_cast_fp16_5, var_4749_cast_fp16_5))[name = tensor("aw_691_cast_fp16")]; + tensor aw_693_equation_0 = const()[name = tensor("aw_693_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_693_cast_fp16 = einsum(equation = aw_693_equation_0, values = (var_4771_cast_fp16_6, var_4749_cast_fp16_6))[name = tensor("aw_693_cast_fp16")]; + tensor aw_695_equation_0 = const()[name = tensor("aw_695_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_695_cast_fp16 = einsum(equation = aw_695_equation_0, values = (var_4771_cast_fp16_7, var_4749_cast_fp16_7))[name = tensor("aw_695_cast_fp16")]; + tensor aw_697_equation_0 = const()[name = tensor("aw_697_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_697_cast_fp16 = einsum(equation = aw_697_equation_0, values = (var_4771_cast_fp16_8, var_4749_cast_fp16_8))[name = tensor("aw_697_cast_fp16")]; + tensor aw_699_equation_0 = const()[name = tensor("aw_699_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_699_cast_fp16 = einsum(equation = aw_699_equation_0, values = (var_4771_cast_fp16_9, var_4749_cast_fp16_9))[name = tensor("aw_699_cast_fp16")]; + tensor aw_701_equation_0 = const()[name = tensor("aw_701_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_701_cast_fp16 = einsum(equation = aw_701_equation_0, values = (var_4771_cast_fp16_10, var_4749_cast_fp16_10))[name = tensor("aw_701_cast_fp16")]; + tensor aw_703_equation_0 = const()[name = tensor("aw_703_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_703_cast_fp16 = einsum(equation = aw_703_equation_0, values = (var_4771_cast_fp16_11, var_4749_cast_fp16_11))[name = tensor("aw_703_cast_fp16")]; + tensor aw_705_equation_0 = const()[name = tensor("aw_705_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_705_cast_fp16 = einsum(equation = aw_705_equation_0, values = (var_4771_cast_fp16_12, var_4749_cast_fp16_12))[name = tensor("aw_705_cast_fp16")]; + tensor aw_707_equation_0 = const()[name = tensor("aw_707_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_707_cast_fp16 = einsum(equation = aw_707_equation_0, values = (var_4771_cast_fp16_13, var_4749_cast_fp16_13))[name = tensor("aw_707_cast_fp16")]; + tensor aw_709_equation_0 = const()[name = tensor("aw_709_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_709_cast_fp16 = einsum(equation = aw_709_equation_0, values = (var_4771_cast_fp16_14, var_4749_cast_fp16_14))[name = tensor("aw_709_cast_fp16")]; + tensor aw_711_equation_0 = const()[name = tensor("aw_711_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_711_cast_fp16 = einsum(equation = aw_711_equation_0, values = (var_4771_cast_fp16_15, var_4749_cast_fp16_15))[name = tensor("aw_711_cast_fp16")]; + tensor aw_713_equation_0 = const()[name = tensor("aw_713_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_713_cast_fp16 = einsum(equation = aw_713_equation_0, values = (var_4771_cast_fp16_16, var_4749_cast_fp16_16))[name = tensor("aw_713_cast_fp16")]; + tensor aw_715_equation_0 = const()[name = tensor("aw_715_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_715_cast_fp16 = einsum(equation = aw_715_equation_0, values = (var_4771_cast_fp16_17, var_4749_cast_fp16_17))[name = tensor("aw_715_cast_fp16")]; + tensor aw_717_equation_0 = const()[name = tensor("aw_717_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_717_cast_fp16 = einsum(equation = aw_717_equation_0, values = (var_4771_cast_fp16_18, var_4749_cast_fp16_18))[name = tensor("aw_717_cast_fp16")]; + tensor aw_719_equation_0 = const()[name = tensor("aw_719_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_719_cast_fp16 = einsum(equation = aw_719_equation_0, values = (var_4771_cast_fp16_19, var_4749_cast_fp16_19))[name = tensor("aw_719_cast_fp16")]; + tensor var_4853_cast_fp16 = softmax(axis = var_4697, x = aw_681_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = softmax(axis = var_4697, x = aw_683_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor var_4855_cast_fp16 = softmax(axis = var_4697, x = aw_685_cast_fp16)[name = tensor("op_4855_cast_fp16")]; + tensor var_4856_cast_fp16 = softmax(axis = var_4697, x = aw_687_cast_fp16)[name = tensor("op_4856_cast_fp16")]; + tensor var_4857_cast_fp16 = softmax(axis = var_4697, x = aw_689_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = softmax(axis = var_4697, x = aw_691_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor var_4859_cast_fp16 = softmax(axis = var_4697, x = aw_693_cast_fp16)[name = tensor("op_4859_cast_fp16")]; + tensor var_4860_cast_fp16 = softmax(axis = var_4697, x = aw_695_cast_fp16)[name = tensor("op_4860_cast_fp16")]; + tensor var_4861_cast_fp16 = softmax(axis = var_4697, x = aw_697_cast_fp16)[name = tensor("op_4861_cast_fp16")]; + tensor var_4862_cast_fp16 = softmax(axis = var_4697, x = aw_699_cast_fp16)[name = tensor("op_4862_cast_fp16")]; + tensor var_4863_cast_fp16 = softmax(axis = var_4697, x = aw_701_cast_fp16)[name = tensor("op_4863_cast_fp16")]; + tensor var_4864_cast_fp16 = softmax(axis = var_4697, x = aw_703_cast_fp16)[name = tensor("op_4864_cast_fp16")]; + tensor var_4865_cast_fp16 = softmax(axis = var_4697, x = aw_705_cast_fp16)[name = tensor("op_4865_cast_fp16")]; + tensor var_4866_cast_fp16 = softmax(axis = var_4697, x = aw_707_cast_fp16)[name = tensor("op_4866_cast_fp16")]; + tensor var_4867_cast_fp16 = softmax(axis = var_4697, x = aw_709_cast_fp16)[name = tensor("op_4867_cast_fp16")]; + tensor var_4868_cast_fp16 = softmax(axis = var_4697, x = aw_711_cast_fp16)[name = tensor("op_4868_cast_fp16")]; + tensor var_4869_cast_fp16 = softmax(axis = var_4697, x = aw_713_cast_fp16)[name = tensor("op_4869_cast_fp16")]; + tensor var_4870_cast_fp16 = softmax(axis = var_4697, x = aw_715_cast_fp16)[name = tensor("op_4870_cast_fp16")]; + tensor var_4871_cast_fp16 = softmax(axis = var_4697, x = aw_717_cast_fp16)[name = tensor("op_4871_cast_fp16")]; + tensor var_4872_cast_fp16 = softmax(axis = var_4697, x = aw_719_cast_fp16)[name = tensor("op_4872_cast_fp16")]; + tensor var_4874_equation_0 = const()[name = tensor("op_4874_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4874_cast_fp16 = einsum(equation = var_4874_equation_0, values = (var_4792_cast_fp16_0, var_4853_cast_fp16))[name = tensor("op_4874_cast_fp16")]; + tensor var_4876_equation_0 = const()[name = tensor("op_4876_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4876_cast_fp16 = einsum(equation = var_4876_equation_0, values = (var_4792_cast_fp16_1, var_4854_cast_fp16))[name = tensor("op_4876_cast_fp16")]; + tensor var_4878_equation_0 = const()[name = tensor("op_4878_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4878_cast_fp16 = einsum(equation = var_4878_equation_0, values = (var_4792_cast_fp16_2, var_4855_cast_fp16))[name = tensor("op_4878_cast_fp16")]; + tensor var_4880_equation_0 = const()[name = tensor("op_4880_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4880_cast_fp16 = einsum(equation = var_4880_equation_0, values = (var_4792_cast_fp16_3, var_4856_cast_fp16))[name = tensor("op_4880_cast_fp16")]; + tensor var_4882_equation_0 = const()[name = tensor("op_4882_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4882_cast_fp16 = einsum(equation = var_4882_equation_0, values = (var_4792_cast_fp16_4, var_4857_cast_fp16))[name = tensor("op_4882_cast_fp16")]; + tensor var_4884_equation_0 = const()[name = tensor("op_4884_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4884_cast_fp16 = einsum(equation = var_4884_equation_0, values = (var_4792_cast_fp16_5, var_4858_cast_fp16))[name = tensor("op_4884_cast_fp16")]; + tensor var_4886_equation_0 = const()[name = tensor("op_4886_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4886_cast_fp16 = einsum(equation = var_4886_equation_0, values = (var_4792_cast_fp16_6, var_4859_cast_fp16))[name = tensor("op_4886_cast_fp16")]; + tensor var_4888_equation_0 = const()[name = tensor("op_4888_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4888_cast_fp16 = einsum(equation = var_4888_equation_0, values = (var_4792_cast_fp16_7, var_4860_cast_fp16))[name = tensor("op_4888_cast_fp16")]; + tensor var_4890_equation_0 = const()[name = tensor("op_4890_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4890_cast_fp16 = einsum(equation = var_4890_equation_0, values = (var_4792_cast_fp16_8, var_4861_cast_fp16))[name = tensor("op_4890_cast_fp16")]; + tensor var_4892_equation_0 = const()[name = tensor("op_4892_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4892_cast_fp16 = einsum(equation = var_4892_equation_0, values = (var_4792_cast_fp16_9, var_4862_cast_fp16))[name = tensor("op_4892_cast_fp16")]; + tensor var_4894_equation_0 = const()[name = tensor("op_4894_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4894_cast_fp16 = einsum(equation = var_4894_equation_0, values = (var_4792_cast_fp16_10, var_4863_cast_fp16))[name = tensor("op_4894_cast_fp16")]; + tensor var_4896_equation_0 = const()[name = tensor("op_4896_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4896_cast_fp16 = einsum(equation = var_4896_equation_0, values = (var_4792_cast_fp16_11, var_4864_cast_fp16))[name = tensor("op_4896_cast_fp16")]; + tensor var_4898_equation_0 = const()[name = tensor("op_4898_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4898_cast_fp16 = einsum(equation = var_4898_equation_0, values = (var_4792_cast_fp16_12, var_4865_cast_fp16))[name = tensor("op_4898_cast_fp16")]; + tensor var_4900_equation_0 = const()[name = tensor("op_4900_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4900_cast_fp16 = einsum(equation = var_4900_equation_0, values = (var_4792_cast_fp16_13, var_4866_cast_fp16))[name = tensor("op_4900_cast_fp16")]; + tensor var_4902_equation_0 = const()[name = tensor("op_4902_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4902_cast_fp16 = einsum(equation = var_4902_equation_0, values = (var_4792_cast_fp16_14, var_4867_cast_fp16))[name = tensor("op_4902_cast_fp16")]; + tensor var_4904_equation_0 = const()[name = tensor("op_4904_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4904_cast_fp16 = einsum(equation = var_4904_equation_0, values = (var_4792_cast_fp16_15, var_4868_cast_fp16))[name = tensor("op_4904_cast_fp16")]; + tensor var_4906_equation_0 = const()[name = tensor("op_4906_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4906_cast_fp16 = einsum(equation = var_4906_equation_0, values = (var_4792_cast_fp16_16, var_4869_cast_fp16))[name = tensor("op_4906_cast_fp16")]; + tensor var_4908_equation_0 = const()[name = tensor("op_4908_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4908_cast_fp16 = einsum(equation = var_4908_equation_0, values = (var_4792_cast_fp16_17, var_4870_cast_fp16))[name = tensor("op_4908_cast_fp16")]; + tensor var_4910_equation_0 = const()[name = tensor("op_4910_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4910_cast_fp16 = einsum(equation = var_4910_equation_0, values = (var_4792_cast_fp16_18, var_4871_cast_fp16))[name = tensor("op_4910_cast_fp16")]; + tensor var_4912_equation_0 = const()[name = tensor("op_4912_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4912_cast_fp16 = einsum(equation = var_4912_equation_0, values = (var_4792_cast_fp16_19, var_4872_cast_fp16))[name = tensor("op_4912_cast_fp16")]; + tensor input_175_interleave_0 = const()[name = tensor("input_175_interleave_0"), val = tensor(false)]; + tensor input_175_cast_fp16 = concat(axis = var_4697, interleave = input_175_interleave_0, values = (var_4874_cast_fp16, var_4876_cast_fp16, var_4878_cast_fp16, var_4880_cast_fp16, var_4882_cast_fp16, var_4884_cast_fp16, var_4886_cast_fp16, var_4888_cast_fp16, var_4890_cast_fp16, var_4892_cast_fp16, var_4894_cast_fp16, var_4896_cast_fp16, var_4898_cast_fp16, var_4900_cast_fp16, var_4902_cast_fp16, var_4904_cast_fp16, var_4906_cast_fp16, var_4908_cast_fp16, var_4910_cast_fp16, var_4912_cast_fp16))[name = tensor("input_175_cast_fp16")]; + tensor var_4921_pad_type_0 = const()[name = tensor("op_4921_pad_type_0"), val = tensor("valid")]; + tensor var_4921_strides_0 = const()[name = tensor("op_4921_strides_0"), val = tensor([1, 1])]; + tensor var_4921_pad_0 = const()[name = tensor("op_4921_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4921_dilations_0 = const()[name = tensor("op_4921_dilations_0"), val = tensor([1, 1])]; + tensor var_4921_groups_0 = const()[name = tensor("op_4921_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_out_weight_to_fp16 = const()[name = tensor("blocks_17_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693137152)))]; + tensor blocks_17_attn_out_bias_to_fp16 = const()[name = tensor("blocks_17_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696414016)))]; + tensor var_4921_cast_fp16 = conv(bias = blocks_17_attn_out_bias_to_fp16, dilations = var_4921_dilations_0, groups = var_4921_groups_0, pad = var_4921_pad_0, pad_type = var_4921_pad_type_0, strides = var_4921_strides_0, weight = blocks_17_attn_out_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = var_4921_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([1])]; + tensor input_177_gamma_0_to_fp16 = const()[name = tensor("input_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696416640)))]; + tensor input_177_beta_0_to_fp16 = const()[name = tensor("input_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696419264)))]; + tensor var_4931_to_fp16 = const()[name = tensor("op_4931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = input_177_beta_0_to_fp16, epsilon = var_4931_to_fp16, gamma = input_177_gamma_0_to_fp16, x = inputs_71_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("valid")]; + tensor input_179_strides_0 = const()[name = tensor("input_179_strides_0"), val = tensor([1, 1])]; + tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_179_dilations_0 = const()[name = tensor("input_179_dilations_0"), val = tensor([1, 1])]; + tensor input_179_groups_0 = const()[name = tensor("input_179_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696421888)))]; + tensor blocks_17_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709529152)))]; + tensor input_179_cast_fp16 = conv(bias = blocks_17_mlp_0_bias_to_fp16, dilations = input_179_dilations_0, groups = input_179_groups_0, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = input_179_strides_0, weight = blocks_17_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4957_pad_type_0 = const()[name = tensor("op_4957_pad_type_0"), val = tensor("valid")]; + tensor var_4957_strides_0 = const()[name = tensor("op_4957_strides_0"), val = tensor([1, 1])]; + tensor var_4957_pad_0 = const()[name = tensor("op_4957_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4957_dilations_0 = const()[name = tensor("op_4957_dilations_0"), val = tensor([1, 1])]; + tensor var_4957_groups_0 = const()[name = tensor("op_4957_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709539456)))]; + tensor blocks_17_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722646720)))]; + tensor var_4957_cast_fp16 = conv(bias = blocks_17_mlp_2_bias_to_fp16, dilations = var_4957_dilations_0, groups = var_4957_groups_0, pad = var_4957_pad_0, pad_type = var_4957_pad_type_0, strides = var_4957_strides_0, weight = blocks_17_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("op_4957_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_4957_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4966 = const()[name = tensor("op_4966"), val = tensor(1)]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([1])]; + tensor input_183_gamma_0_to_fp16 = const()[name = tensor("input_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722649344)))]; + tensor input_183_beta_0_to_fp16 = const()[name = tensor("input_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722651968)))]; + tensor var_4982_to_fp16 = const()[name = tensor("op_4982_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = input_183_beta_0_to_fp16, epsilon = var_4982_to_fp16, gamma = input_183_gamma_0_to_fp16, x = inputs_73_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor q_37_pad_type_0 = const()[name = tensor("q_37_pad_type_0"), val = tensor("valid")]; + tensor q_37_strides_0 = const()[name = tensor("q_37_strides_0"), val = tensor([1, 1])]; + tensor q_37_pad_0 = const()[name = tensor("q_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_37_dilations_0 = const()[name = tensor("q_37_dilations_0"), val = tensor([1, 1])]; + tensor q_37_groups_0 = const()[name = tensor("q_37_groups_0"), val = tensor(1)]; + tensor var_5017_weight_0_to_fp16 = const()[name = tensor("op_5017_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722654592)))]; + tensor var_5017_bias_0_to_fp16 = const()[name = tensor("op_5017_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725931456)))]; + tensor var_5017_cast_fp16 = conv(bias = var_5017_bias_0_to_fp16, dilations = q_37_dilations_0, groups = q_37_groups_0, pad = q_37_pad_0, pad_type = q_37_pad_type_0, strides = q_37_strides_0, weight = var_5017_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("op_5017_cast_fp16")]; + tensor k_37_pad_type_0 = const()[name = tensor("k_37_pad_type_0"), val = tensor("valid")]; + tensor k_37_strides_0 = const()[name = tensor("k_37_strides_0"), val = tensor([1, 1])]; + tensor k_37_pad_0 = const()[name = tensor("k_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_37_dilations_0 = const()[name = tensor("k_37_dilations_0"), val = tensor([1, 1])]; + tensor k_37_groups_0 = const()[name = tensor("k_37_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_key_weight_to_fp16 = const()[name = tensor("blocks_18_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725934080)))]; + tensor k_37_cast_fp16 = conv(dilations = k_37_dilations_0, groups = k_37_groups_0, pad = k_37_pad_0, pad_type = k_37_pad_type_0, strides = k_37_strides_0, weight = blocks_18_attn_key_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor var_5015_pad_type_0 = const()[name = tensor("op_5015_pad_type_0"), val = tensor("valid")]; + tensor var_5015_strides_0 = const()[name = tensor("op_5015_strides_0"), val = tensor([1, 1])]; + tensor var_5015_pad_0 = const()[name = tensor("op_5015_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5015_dilations_0 = const()[name = tensor("op_5015_dilations_0"), val = tensor([1, 1])]; + tensor var_5015_groups_0 = const()[name = tensor("op_5015_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_value_weight_to_fp16 = const()[name = tensor("blocks_18_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729210944)))]; + tensor blocks_18_attn_value_bias_to_fp16 = const()[name = tensor("blocks_18_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732487808)))]; + tensor var_5015_cast_fp16 = conv(bias = blocks_18_attn_value_bias_to_fp16, dilations = var_5015_dilations_0, groups = var_5015_groups_0, pad = var_5015_pad_0, pad_type = var_5015_pad_type_0, strides = var_5015_strides_0, weight = blocks_18_attn_value_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor tile_54 = const()[name = tensor("tile_54"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5018_axis_0 = const()[name = tensor("op_5018_axis_0"), val = tensor(1)]; + tensor var_5018_cast_fp16_0, tensor var_5018_cast_fp16_1, tensor var_5018_cast_fp16_2, tensor var_5018_cast_fp16_3, tensor var_5018_cast_fp16_4, tensor var_5018_cast_fp16_5, tensor var_5018_cast_fp16_6, tensor var_5018_cast_fp16_7, tensor var_5018_cast_fp16_8, tensor var_5018_cast_fp16_9, tensor var_5018_cast_fp16_10, tensor var_5018_cast_fp16_11, tensor var_5018_cast_fp16_12, tensor var_5018_cast_fp16_13, tensor var_5018_cast_fp16_14, tensor var_5018_cast_fp16_15, tensor var_5018_cast_fp16_16, tensor var_5018_cast_fp16_17, tensor var_5018_cast_fp16_18, tensor var_5018_cast_fp16_19 = split(axis = var_5018_axis_0, split_sizes = tile_54, x = var_5017_cast_fp16)[name = tensor("op_5018_cast_fp16")]; + tensor var_5039_perm_0 = const()[name = tensor("op_5039_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_55 = const()[name = tensor("tile_55"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5040_axis_0 = const()[name = tensor("op_5040_axis_0"), val = tensor(3)]; + tensor var_5039_cast_fp16 = transpose(perm = var_5039_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_14")]; + tensor var_5040_cast_fp16_0, tensor var_5040_cast_fp16_1, tensor var_5040_cast_fp16_2, tensor var_5040_cast_fp16_3, tensor var_5040_cast_fp16_4, tensor var_5040_cast_fp16_5, tensor var_5040_cast_fp16_6, tensor var_5040_cast_fp16_7, tensor var_5040_cast_fp16_8, tensor var_5040_cast_fp16_9, tensor var_5040_cast_fp16_10, tensor var_5040_cast_fp16_11, tensor var_5040_cast_fp16_12, tensor var_5040_cast_fp16_13, tensor var_5040_cast_fp16_14, tensor var_5040_cast_fp16_15, tensor var_5040_cast_fp16_16, tensor var_5040_cast_fp16_17, tensor var_5040_cast_fp16_18, tensor var_5040_cast_fp16_19 = split(axis = var_5040_axis_0, split_sizes = tile_55, x = var_5039_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor tile_56 = const()[name = tensor("tile_56"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5061_axis_0 = const()[name = tensor("op_5061_axis_0"), val = tensor(1)]; + tensor var_5061_cast_fp16_0, tensor var_5061_cast_fp16_1, tensor var_5061_cast_fp16_2, tensor var_5061_cast_fp16_3, tensor var_5061_cast_fp16_4, tensor var_5061_cast_fp16_5, tensor var_5061_cast_fp16_6, tensor var_5061_cast_fp16_7, tensor var_5061_cast_fp16_8, tensor var_5061_cast_fp16_9, tensor var_5061_cast_fp16_10, tensor var_5061_cast_fp16_11, tensor var_5061_cast_fp16_12, tensor var_5061_cast_fp16_13, tensor var_5061_cast_fp16_14, tensor var_5061_cast_fp16_15, tensor var_5061_cast_fp16_16, tensor var_5061_cast_fp16_17, tensor var_5061_cast_fp16_18, tensor var_5061_cast_fp16_19 = split(axis = var_5061_axis_0, split_sizes = tile_56, x = var_5015_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor aw_721_equation_0 = const()[name = tensor("aw_721_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_721_cast_fp16 = einsum(equation = aw_721_equation_0, values = (var_5040_cast_fp16_0, var_5018_cast_fp16_0))[name = tensor("aw_721_cast_fp16")]; + tensor aw_723_equation_0 = const()[name = tensor("aw_723_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_723_cast_fp16 = einsum(equation = aw_723_equation_0, values = (var_5040_cast_fp16_1, var_5018_cast_fp16_1))[name = tensor("aw_723_cast_fp16")]; + tensor aw_725_equation_0 = const()[name = tensor("aw_725_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_725_cast_fp16 = einsum(equation = aw_725_equation_0, values = (var_5040_cast_fp16_2, var_5018_cast_fp16_2))[name = tensor("aw_725_cast_fp16")]; + tensor aw_727_equation_0 = const()[name = tensor("aw_727_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_727_cast_fp16 = einsum(equation = aw_727_equation_0, values = (var_5040_cast_fp16_3, var_5018_cast_fp16_3))[name = tensor("aw_727_cast_fp16")]; + tensor aw_729_equation_0 = const()[name = tensor("aw_729_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_729_cast_fp16 = einsum(equation = aw_729_equation_0, values = (var_5040_cast_fp16_4, var_5018_cast_fp16_4))[name = tensor("aw_729_cast_fp16")]; + tensor aw_731_equation_0 = const()[name = tensor("aw_731_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_731_cast_fp16 = einsum(equation = aw_731_equation_0, values = (var_5040_cast_fp16_5, var_5018_cast_fp16_5))[name = tensor("aw_731_cast_fp16")]; + tensor aw_733_equation_0 = const()[name = tensor("aw_733_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_733_cast_fp16 = einsum(equation = aw_733_equation_0, values = (var_5040_cast_fp16_6, var_5018_cast_fp16_6))[name = tensor("aw_733_cast_fp16")]; + tensor aw_735_equation_0 = const()[name = tensor("aw_735_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_735_cast_fp16 = einsum(equation = aw_735_equation_0, values = (var_5040_cast_fp16_7, var_5018_cast_fp16_7))[name = tensor("aw_735_cast_fp16")]; + tensor aw_737_equation_0 = const()[name = tensor("aw_737_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_737_cast_fp16 = einsum(equation = aw_737_equation_0, values = (var_5040_cast_fp16_8, var_5018_cast_fp16_8))[name = tensor("aw_737_cast_fp16")]; + tensor aw_739_equation_0 = const()[name = tensor("aw_739_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_739_cast_fp16 = einsum(equation = aw_739_equation_0, values = (var_5040_cast_fp16_9, var_5018_cast_fp16_9))[name = tensor("aw_739_cast_fp16")]; + tensor aw_741_equation_0 = const()[name = tensor("aw_741_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_741_cast_fp16 = einsum(equation = aw_741_equation_0, values = (var_5040_cast_fp16_10, var_5018_cast_fp16_10))[name = tensor("aw_741_cast_fp16")]; + tensor aw_743_equation_0 = const()[name = tensor("aw_743_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_743_cast_fp16 = einsum(equation = aw_743_equation_0, values = (var_5040_cast_fp16_11, var_5018_cast_fp16_11))[name = tensor("aw_743_cast_fp16")]; + tensor aw_745_equation_0 = const()[name = tensor("aw_745_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_745_cast_fp16 = einsum(equation = aw_745_equation_0, values = (var_5040_cast_fp16_12, var_5018_cast_fp16_12))[name = tensor("aw_745_cast_fp16")]; + tensor aw_747_equation_0 = const()[name = tensor("aw_747_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_747_cast_fp16 = einsum(equation = aw_747_equation_0, values = (var_5040_cast_fp16_13, var_5018_cast_fp16_13))[name = tensor("aw_747_cast_fp16")]; + tensor aw_749_equation_0 = const()[name = tensor("aw_749_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_749_cast_fp16 = einsum(equation = aw_749_equation_0, values = (var_5040_cast_fp16_14, var_5018_cast_fp16_14))[name = tensor("aw_749_cast_fp16")]; + tensor aw_751_equation_0 = const()[name = tensor("aw_751_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_751_cast_fp16 = einsum(equation = aw_751_equation_0, values = (var_5040_cast_fp16_15, var_5018_cast_fp16_15))[name = tensor("aw_751_cast_fp16")]; + tensor aw_753_equation_0 = const()[name = tensor("aw_753_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_753_cast_fp16 = einsum(equation = aw_753_equation_0, values = (var_5040_cast_fp16_16, var_5018_cast_fp16_16))[name = tensor("aw_753_cast_fp16")]; + tensor aw_755_equation_0 = const()[name = tensor("aw_755_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_755_cast_fp16 = einsum(equation = aw_755_equation_0, values = (var_5040_cast_fp16_17, var_5018_cast_fp16_17))[name = tensor("aw_755_cast_fp16")]; + tensor aw_757_equation_0 = const()[name = tensor("aw_757_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_757_cast_fp16 = einsum(equation = aw_757_equation_0, values = (var_5040_cast_fp16_18, var_5018_cast_fp16_18))[name = tensor("aw_757_cast_fp16")]; + tensor aw_759_equation_0 = const()[name = tensor("aw_759_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_759_cast_fp16 = einsum(equation = aw_759_equation_0, values = (var_5040_cast_fp16_19, var_5018_cast_fp16_19))[name = tensor("aw_759_cast_fp16")]; + tensor var_5122_cast_fp16 = softmax(axis = var_4966, x = aw_721_cast_fp16)[name = tensor("op_5122_cast_fp16")]; + tensor var_5123_cast_fp16 = softmax(axis = var_4966, x = aw_723_cast_fp16)[name = tensor("op_5123_cast_fp16")]; + tensor var_5124_cast_fp16 = softmax(axis = var_4966, x = aw_725_cast_fp16)[name = tensor("op_5124_cast_fp16")]; + tensor var_5125_cast_fp16 = softmax(axis = var_4966, x = aw_727_cast_fp16)[name = tensor("op_5125_cast_fp16")]; + tensor var_5126_cast_fp16 = softmax(axis = var_4966, x = aw_729_cast_fp16)[name = tensor("op_5126_cast_fp16")]; + tensor var_5127_cast_fp16 = softmax(axis = var_4966, x = aw_731_cast_fp16)[name = tensor("op_5127_cast_fp16")]; + tensor var_5128_cast_fp16 = softmax(axis = var_4966, x = aw_733_cast_fp16)[name = tensor("op_5128_cast_fp16")]; + tensor var_5129_cast_fp16 = softmax(axis = var_4966, x = aw_735_cast_fp16)[name = tensor("op_5129_cast_fp16")]; + tensor var_5130_cast_fp16 = softmax(axis = var_4966, x = aw_737_cast_fp16)[name = tensor("op_5130_cast_fp16")]; + tensor var_5131_cast_fp16 = softmax(axis = var_4966, x = aw_739_cast_fp16)[name = tensor("op_5131_cast_fp16")]; + tensor var_5132_cast_fp16 = softmax(axis = var_4966, x = aw_741_cast_fp16)[name = tensor("op_5132_cast_fp16")]; + tensor var_5133_cast_fp16 = softmax(axis = var_4966, x = aw_743_cast_fp16)[name = tensor("op_5133_cast_fp16")]; + tensor var_5134_cast_fp16 = softmax(axis = var_4966, x = aw_745_cast_fp16)[name = tensor("op_5134_cast_fp16")]; + tensor var_5135_cast_fp16 = softmax(axis = var_4966, x = aw_747_cast_fp16)[name = tensor("op_5135_cast_fp16")]; + tensor var_5136_cast_fp16 = softmax(axis = var_4966, x = aw_749_cast_fp16)[name = tensor("op_5136_cast_fp16")]; + tensor var_5137_cast_fp16 = softmax(axis = var_4966, x = aw_751_cast_fp16)[name = tensor("op_5137_cast_fp16")]; + tensor var_5138_cast_fp16 = softmax(axis = var_4966, x = aw_753_cast_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5139_cast_fp16 = softmax(axis = var_4966, x = aw_755_cast_fp16)[name = tensor("op_5139_cast_fp16")]; + tensor var_5140_cast_fp16 = softmax(axis = var_4966, x = aw_757_cast_fp16)[name = tensor("op_5140_cast_fp16")]; + tensor var_5141_cast_fp16 = softmax(axis = var_4966, x = aw_759_cast_fp16)[name = tensor("op_5141_cast_fp16")]; + tensor var_5143_equation_0 = const()[name = tensor("op_5143_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5143_cast_fp16 = einsum(equation = var_5143_equation_0, values = (var_5061_cast_fp16_0, var_5122_cast_fp16))[name = tensor("op_5143_cast_fp16")]; + tensor var_5145_equation_0 = const()[name = tensor("op_5145_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5145_cast_fp16 = einsum(equation = var_5145_equation_0, values = (var_5061_cast_fp16_1, var_5123_cast_fp16))[name = tensor("op_5145_cast_fp16")]; + tensor var_5147_equation_0 = const()[name = tensor("op_5147_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5147_cast_fp16 = einsum(equation = var_5147_equation_0, values = (var_5061_cast_fp16_2, var_5124_cast_fp16))[name = tensor("op_5147_cast_fp16")]; + tensor var_5149_equation_0 = const()[name = tensor("op_5149_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5149_cast_fp16 = einsum(equation = var_5149_equation_0, values = (var_5061_cast_fp16_3, var_5125_cast_fp16))[name = tensor("op_5149_cast_fp16")]; + tensor var_5151_equation_0 = const()[name = tensor("op_5151_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5151_cast_fp16 = einsum(equation = var_5151_equation_0, values = (var_5061_cast_fp16_4, var_5126_cast_fp16))[name = tensor("op_5151_cast_fp16")]; + tensor var_5153_equation_0 = const()[name = tensor("op_5153_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5153_cast_fp16 = einsum(equation = var_5153_equation_0, values = (var_5061_cast_fp16_5, var_5127_cast_fp16))[name = tensor("op_5153_cast_fp16")]; + tensor var_5155_equation_0 = const()[name = tensor("op_5155_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5155_cast_fp16 = einsum(equation = var_5155_equation_0, values = (var_5061_cast_fp16_6, var_5128_cast_fp16))[name = tensor("op_5155_cast_fp16")]; + tensor var_5157_equation_0 = const()[name = tensor("op_5157_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5157_cast_fp16 = einsum(equation = var_5157_equation_0, values = (var_5061_cast_fp16_7, var_5129_cast_fp16))[name = tensor("op_5157_cast_fp16")]; + tensor var_5159_equation_0 = const()[name = tensor("op_5159_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5159_cast_fp16 = einsum(equation = var_5159_equation_0, values = (var_5061_cast_fp16_8, var_5130_cast_fp16))[name = tensor("op_5159_cast_fp16")]; + tensor var_5161_equation_0 = const()[name = tensor("op_5161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5161_cast_fp16 = einsum(equation = var_5161_equation_0, values = (var_5061_cast_fp16_9, var_5131_cast_fp16))[name = tensor("op_5161_cast_fp16")]; + tensor var_5163_equation_0 = const()[name = tensor("op_5163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5163_cast_fp16 = einsum(equation = var_5163_equation_0, values = (var_5061_cast_fp16_10, var_5132_cast_fp16))[name = tensor("op_5163_cast_fp16")]; + tensor var_5165_equation_0 = const()[name = tensor("op_5165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5165_cast_fp16 = einsum(equation = var_5165_equation_0, values = (var_5061_cast_fp16_11, var_5133_cast_fp16))[name = tensor("op_5165_cast_fp16")]; + tensor var_5167_equation_0 = const()[name = tensor("op_5167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5167_cast_fp16 = einsum(equation = var_5167_equation_0, values = (var_5061_cast_fp16_12, var_5134_cast_fp16))[name = tensor("op_5167_cast_fp16")]; + tensor var_5169_equation_0 = const()[name = tensor("op_5169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5169_cast_fp16 = einsum(equation = var_5169_equation_0, values = (var_5061_cast_fp16_13, var_5135_cast_fp16))[name = tensor("op_5169_cast_fp16")]; + tensor var_5171_equation_0 = const()[name = tensor("op_5171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5171_cast_fp16 = einsum(equation = var_5171_equation_0, values = (var_5061_cast_fp16_14, var_5136_cast_fp16))[name = tensor("op_5171_cast_fp16")]; + tensor var_5173_equation_0 = const()[name = tensor("op_5173_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5173_cast_fp16 = einsum(equation = var_5173_equation_0, values = (var_5061_cast_fp16_15, var_5137_cast_fp16))[name = tensor("op_5173_cast_fp16")]; + tensor var_5175_equation_0 = const()[name = tensor("op_5175_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5175_cast_fp16 = einsum(equation = var_5175_equation_0, values = (var_5061_cast_fp16_16, var_5138_cast_fp16))[name = tensor("op_5175_cast_fp16")]; + tensor var_5177_equation_0 = const()[name = tensor("op_5177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5177_cast_fp16 = einsum(equation = var_5177_equation_0, values = (var_5061_cast_fp16_17, var_5139_cast_fp16))[name = tensor("op_5177_cast_fp16")]; + tensor var_5179_equation_0 = const()[name = tensor("op_5179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5179_cast_fp16 = einsum(equation = var_5179_equation_0, values = (var_5061_cast_fp16_18, var_5140_cast_fp16))[name = tensor("op_5179_cast_fp16")]; + tensor var_5181_equation_0 = const()[name = tensor("op_5181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5181_cast_fp16 = einsum(equation = var_5181_equation_0, values = (var_5061_cast_fp16_19, var_5141_cast_fp16))[name = tensor("op_5181_cast_fp16")]; + tensor input_185_interleave_0 = const()[name = tensor("input_185_interleave_0"), val = tensor(false)]; + tensor input_185_cast_fp16 = concat(axis = var_4966, interleave = input_185_interleave_0, values = (var_5143_cast_fp16, var_5145_cast_fp16, var_5147_cast_fp16, var_5149_cast_fp16, var_5151_cast_fp16, var_5153_cast_fp16, var_5155_cast_fp16, var_5157_cast_fp16, var_5159_cast_fp16, var_5161_cast_fp16, var_5163_cast_fp16, var_5165_cast_fp16, var_5167_cast_fp16, var_5169_cast_fp16, var_5171_cast_fp16, var_5173_cast_fp16, var_5175_cast_fp16, var_5177_cast_fp16, var_5179_cast_fp16, var_5181_cast_fp16))[name = tensor("input_185_cast_fp16")]; + tensor var_5190_pad_type_0 = const()[name = tensor("op_5190_pad_type_0"), val = tensor("valid")]; + tensor var_5190_strides_0 = const()[name = tensor("op_5190_strides_0"), val = tensor([1, 1])]; + tensor var_5190_pad_0 = const()[name = tensor("op_5190_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5190_dilations_0 = const()[name = tensor("op_5190_dilations_0"), val = tensor([1, 1])]; + tensor var_5190_groups_0 = const()[name = tensor("op_5190_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_out_weight_to_fp16 = const()[name = tensor("blocks_18_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732490432)))]; + tensor blocks_18_attn_out_bias_to_fp16 = const()[name = tensor("blocks_18_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735767296)))]; + tensor var_5190_cast_fp16 = conv(bias = blocks_18_attn_out_bias_to_fp16, dilations = var_5190_dilations_0, groups = var_5190_groups_0, pad = var_5190_pad_0, pad_type = var_5190_pad_type_0, strides = var_5190_strides_0, weight = blocks_18_attn_out_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("op_5190_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = var_5190_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor input_187_axes_0 = const()[name = tensor("input_187_axes_0"), val = tensor([1])]; + tensor input_187_gamma_0_to_fp16 = const()[name = tensor("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735769920)))]; + tensor input_187_beta_0_to_fp16 = const()[name = tensor("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735772544)))]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = input_187_beta_0_to_fp16, epsilon = var_5200_to_fp16, gamma = input_187_gamma_0_to_fp16, x = inputs_75_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("valid")]; + tensor input_189_strides_0 = const()[name = tensor("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = tensor("input_189_dilations_0"), val = tensor([1, 1])]; + tensor input_189_groups_0 = const()[name = tensor("input_189_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735775168)))]; + tensor blocks_18_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748882432)))]; + tensor input_189_cast_fp16 = conv(bias = blocks_18_mlp_0_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = blocks_18_mlp_0_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_5226_pad_type_0 = const()[name = tensor("op_5226_pad_type_0"), val = tensor("valid")]; + tensor var_5226_strides_0 = const()[name = tensor("op_5226_strides_0"), val = tensor([1, 1])]; + tensor var_5226_pad_0 = const()[name = tensor("op_5226_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5226_dilations_0 = const()[name = tensor("op_5226_dilations_0"), val = tensor([1, 1])]; + tensor var_5226_groups_0 = const()[name = tensor("op_5226_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748892736)))]; + tensor blocks_18_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762000000)))]; + tensor var_5226_cast_fp16 = conv(bias = blocks_18_mlp_2_bias_to_fp16, dilations = var_5226_dilations_0, groups = var_5226_groups_0, pad = var_5226_pad_0, pad_type = var_5226_pad_type_0, strides = var_5226_strides_0, weight = blocks_18_mlp_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("op_5226_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = var_5226_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor(1)]; + tensor input_193_axes_0 = const()[name = tensor("input_193_axes_0"), val = tensor([1])]; + tensor input_193_gamma_0_to_fp16 = const()[name = tensor("input_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762002624)))]; + tensor input_193_beta_0_to_fp16 = const()[name = tensor("input_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762005248)))]; + tensor var_5251_to_fp16 = const()[name = tensor("op_5251_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = input_193_beta_0_to_fp16, epsilon = var_5251_to_fp16, gamma = input_193_gamma_0_to_fp16, x = inputs_77_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor q_39_pad_type_0 = const()[name = tensor("q_39_pad_type_0"), val = tensor("valid")]; + tensor q_39_strides_0 = const()[name = tensor("q_39_strides_0"), val = tensor([1, 1])]; + tensor q_39_pad_0 = const()[name = tensor("q_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_39_dilations_0 = const()[name = tensor("q_39_dilations_0"), val = tensor([1, 1])]; + tensor q_39_groups_0 = const()[name = tensor("q_39_groups_0"), val = tensor(1)]; + tensor var_5286_weight_0_to_fp16 = const()[name = tensor("op_5286_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762007872)))]; + tensor var_5286_bias_0_to_fp16 = const()[name = tensor("op_5286_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765284736)))]; + tensor var_5286_cast_fp16 = conv(bias = var_5286_bias_0_to_fp16, dilations = q_39_dilations_0, groups = q_39_groups_0, pad = q_39_pad_0, pad_type = q_39_pad_type_0, strides = q_39_strides_0, weight = var_5286_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor k_39_pad_type_0 = const()[name = tensor("k_39_pad_type_0"), val = tensor("valid")]; + tensor k_39_strides_0 = const()[name = tensor("k_39_strides_0"), val = tensor([1, 1])]; + tensor k_39_pad_0 = const()[name = tensor("k_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_39_dilations_0 = const()[name = tensor("k_39_dilations_0"), val = tensor([1, 1])]; + tensor k_39_groups_0 = const()[name = tensor("k_39_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_key_weight_to_fp16 = const()[name = tensor("blocks_19_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765287360)))]; + tensor k_39_cast_fp16 = conv(dilations = k_39_dilations_0, groups = k_39_groups_0, pad = k_39_pad_0, pad_type = k_39_pad_type_0, strides = k_39_strides_0, weight = blocks_19_attn_key_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor var_5284_pad_type_0 = const()[name = tensor("op_5284_pad_type_0"), val = tensor("valid")]; + tensor var_5284_strides_0 = const()[name = tensor("op_5284_strides_0"), val = tensor([1, 1])]; + tensor var_5284_pad_0 = const()[name = tensor("op_5284_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5284_dilations_0 = const()[name = tensor("op_5284_dilations_0"), val = tensor([1, 1])]; + tensor var_5284_groups_0 = const()[name = tensor("op_5284_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_value_weight_to_fp16 = const()[name = tensor("blocks_19_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768564224)))]; + tensor blocks_19_attn_value_bias_to_fp16 = const()[name = tensor("blocks_19_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771841088)))]; + tensor var_5284_cast_fp16 = conv(bias = blocks_19_attn_value_bias_to_fp16, dilations = var_5284_dilations_0, groups = var_5284_groups_0, pad = var_5284_pad_0, pad_type = var_5284_pad_type_0, strides = var_5284_strides_0, weight = blocks_19_attn_value_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor tile_57 = const()[name = tensor("tile_57"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5287_axis_0 = const()[name = tensor("op_5287_axis_0"), val = tensor(1)]; + tensor var_5287_cast_fp16_0, tensor var_5287_cast_fp16_1, tensor var_5287_cast_fp16_2, tensor var_5287_cast_fp16_3, tensor var_5287_cast_fp16_4, tensor var_5287_cast_fp16_5, tensor var_5287_cast_fp16_6, tensor var_5287_cast_fp16_7, tensor var_5287_cast_fp16_8, tensor var_5287_cast_fp16_9, tensor var_5287_cast_fp16_10, tensor var_5287_cast_fp16_11, tensor var_5287_cast_fp16_12, tensor var_5287_cast_fp16_13, tensor var_5287_cast_fp16_14, tensor var_5287_cast_fp16_15, tensor var_5287_cast_fp16_16, tensor var_5287_cast_fp16_17, tensor var_5287_cast_fp16_18, tensor var_5287_cast_fp16_19 = split(axis = var_5287_axis_0, split_sizes = tile_57, x = var_5286_cast_fp16)[name = tensor("op_5287_cast_fp16")]; + tensor var_5308_perm_0 = const()[name = tensor("op_5308_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_58 = const()[name = tensor("tile_58"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5309_axis_0 = const()[name = tensor("op_5309_axis_0"), val = tensor(3)]; + tensor var_5308_cast_fp16 = transpose(perm = var_5308_perm_0, x = k_39_cast_fp16)[name = tensor("transpose_13")]; + tensor var_5309_cast_fp16_0, tensor var_5309_cast_fp16_1, tensor var_5309_cast_fp16_2, tensor var_5309_cast_fp16_3, tensor var_5309_cast_fp16_4, tensor var_5309_cast_fp16_5, tensor var_5309_cast_fp16_6, tensor var_5309_cast_fp16_7, tensor var_5309_cast_fp16_8, tensor var_5309_cast_fp16_9, tensor var_5309_cast_fp16_10, tensor var_5309_cast_fp16_11, tensor var_5309_cast_fp16_12, tensor var_5309_cast_fp16_13, tensor var_5309_cast_fp16_14, tensor var_5309_cast_fp16_15, tensor var_5309_cast_fp16_16, tensor var_5309_cast_fp16_17, tensor var_5309_cast_fp16_18, tensor var_5309_cast_fp16_19 = split(axis = var_5309_axis_0, split_sizes = tile_58, x = var_5308_cast_fp16)[name = tensor("op_5309_cast_fp16")]; + tensor tile_59 = const()[name = tensor("tile_59"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5330_axis_0 = const()[name = tensor("op_5330_axis_0"), val = tensor(1)]; + tensor var_5330_cast_fp16_0, tensor var_5330_cast_fp16_1, tensor var_5330_cast_fp16_2, tensor var_5330_cast_fp16_3, tensor var_5330_cast_fp16_4, tensor var_5330_cast_fp16_5, tensor var_5330_cast_fp16_6, tensor var_5330_cast_fp16_7, tensor var_5330_cast_fp16_8, tensor var_5330_cast_fp16_9, tensor var_5330_cast_fp16_10, tensor var_5330_cast_fp16_11, tensor var_5330_cast_fp16_12, tensor var_5330_cast_fp16_13, tensor var_5330_cast_fp16_14, tensor var_5330_cast_fp16_15, tensor var_5330_cast_fp16_16, tensor var_5330_cast_fp16_17, tensor var_5330_cast_fp16_18, tensor var_5330_cast_fp16_19 = split(axis = var_5330_axis_0, split_sizes = tile_59, x = var_5284_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor aw_761_equation_0 = const()[name = tensor("aw_761_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_761_cast_fp16 = einsum(equation = aw_761_equation_0, values = (var_5309_cast_fp16_0, var_5287_cast_fp16_0))[name = tensor("aw_761_cast_fp16")]; + tensor aw_763_equation_0 = const()[name = tensor("aw_763_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_763_cast_fp16 = einsum(equation = aw_763_equation_0, values = (var_5309_cast_fp16_1, var_5287_cast_fp16_1))[name = tensor("aw_763_cast_fp16")]; + tensor aw_765_equation_0 = const()[name = tensor("aw_765_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_765_cast_fp16 = einsum(equation = aw_765_equation_0, values = (var_5309_cast_fp16_2, var_5287_cast_fp16_2))[name = tensor("aw_765_cast_fp16")]; + tensor aw_767_equation_0 = const()[name = tensor("aw_767_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_767_cast_fp16 = einsum(equation = aw_767_equation_0, values = (var_5309_cast_fp16_3, var_5287_cast_fp16_3))[name = tensor("aw_767_cast_fp16")]; + tensor aw_769_equation_0 = const()[name = tensor("aw_769_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_769_cast_fp16 = einsum(equation = aw_769_equation_0, values = (var_5309_cast_fp16_4, var_5287_cast_fp16_4))[name = tensor("aw_769_cast_fp16")]; + tensor aw_771_equation_0 = const()[name = tensor("aw_771_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_771_cast_fp16 = einsum(equation = aw_771_equation_0, values = (var_5309_cast_fp16_5, var_5287_cast_fp16_5))[name = tensor("aw_771_cast_fp16")]; + tensor aw_773_equation_0 = const()[name = tensor("aw_773_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_773_cast_fp16 = einsum(equation = aw_773_equation_0, values = (var_5309_cast_fp16_6, var_5287_cast_fp16_6))[name = tensor("aw_773_cast_fp16")]; + tensor aw_775_equation_0 = const()[name = tensor("aw_775_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_775_cast_fp16 = einsum(equation = aw_775_equation_0, values = (var_5309_cast_fp16_7, var_5287_cast_fp16_7))[name = tensor("aw_775_cast_fp16")]; + tensor aw_777_equation_0 = const()[name = tensor("aw_777_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_777_cast_fp16 = einsum(equation = aw_777_equation_0, values = (var_5309_cast_fp16_8, var_5287_cast_fp16_8))[name = tensor("aw_777_cast_fp16")]; + tensor aw_779_equation_0 = const()[name = tensor("aw_779_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_779_cast_fp16 = einsum(equation = aw_779_equation_0, values = (var_5309_cast_fp16_9, var_5287_cast_fp16_9))[name = tensor("aw_779_cast_fp16")]; + tensor aw_781_equation_0 = const()[name = tensor("aw_781_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_781_cast_fp16 = einsum(equation = aw_781_equation_0, values = (var_5309_cast_fp16_10, var_5287_cast_fp16_10))[name = tensor("aw_781_cast_fp16")]; + tensor aw_783_equation_0 = const()[name = tensor("aw_783_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_783_cast_fp16 = einsum(equation = aw_783_equation_0, values = (var_5309_cast_fp16_11, var_5287_cast_fp16_11))[name = tensor("aw_783_cast_fp16")]; + tensor aw_785_equation_0 = const()[name = tensor("aw_785_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_785_cast_fp16 = einsum(equation = aw_785_equation_0, values = (var_5309_cast_fp16_12, var_5287_cast_fp16_12))[name = tensor("aw_785_cast_fp16")]; + tensor aw_787_equation_0 = const()[name = tensor("aw_787_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_787_cast_fp16 = einsum(equation = aw_787_equation_0, values = (var_5309_cast_fp16_13, var_5287_cast_fp16_13))[name = tensor("aw_787_cast_fp16")]; + tensor aw_789_equation_0 = const()[name = tensor("aw_789_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_789_cast_fp16 = einsum(equation = aw_789_equation_0, values = (var_5309_cast_fp16_14, var_5287_cast_fp16_14))[name = tensor("aw_789_cast_fp16")]; + tensor aw_791_equation_0 = const()[name = tensor("aw_791_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_791_cast_fp16 = einsum(equation = aw_791_equation_0, values = (var_5309_cast_fp16_15, var_5287_cast_fp16_15))[name = tensor("aw_791_cast_fp16")]; + tensor aw_793_equation_0 = const()[name = tensor("aw_793_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_793_cast_fp16 = einsum(equation = aw_793_equation_0, values = (var_5309_cast_fp16_16, var_5287_cast_fp16_16))[name = tensor("aw_793_cast_fp16")]; + tensor aw_795_equation_0 = const()[name = tensor("aw_795_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_795_cast_fp16 = einsum(equation = aw_795_equation_0, values = (var_5309_cast_fp16_17, var_5287_cast_fp16_17))[name = tensor("aw_795_cast_fp16")]; + tensor aw_797_equation_0 = const()[name = tensor("aw_797_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_797_cast_fp16 = einsum(equation = aw_797_equation_0, values = (var_5309_cast_fp16_18, var_5287_cast_fp16_18))[name = tensor("aw_797_cast_fp16")]; + tensor aw_799_equation_0 = const()[name = tensor("aw_799_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_799_cast_fp16 = einsum(equation = aw_799_equation_0, values = (var_5309_cast_fp16_19, var_5287_cast_fp16_19))[name = tensor("aw_799_cast_fp16")]; + tensor var_5391_cast_fp16 = softmax(axis = var_5235, x = aw_761_cast_fp16)[name = tensor("op_5391_cast_fp16")]; + tensor var_5392_cast_fp16 = softmax(axis = var_5235, x = aw_763_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5393_cast_fp16 = softmax(axis = var_5235, x = aw_765_cast_fp16)[name = tensor("op_5393_cast_fp16")]; + tensor var_5394_cast_fp16 = softmax(axis = var_5235, x = aw_767_cast_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor var_5395_cast_fp16 = softmax(axis = var_5235, x = aw_769_cast_fp16)[name = tensor("op_5395_cast_fp16")]; + tensor var_5396_cast_fp16 = softmax(axis = var_5235, x = aw_771_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor var_5397_cast_fp16 = softmax(axis = var_5235, x = aw_773_cast_fp16)[name = tensor("op_5397_cast_fp16")]; + tensor var_5398_cast_fp16 = softmax(axis = var_5235, x = aw_775_cast_fp16)[name = tensor("op_5398_cast_fp16")]; + tensor var_5399_cast_fp16 = softmax(axis = var_5235, x = aw_777_cast_fp16)[name = tensor("op_5399_cast_fp16")]; + tensor var_5400_cast_fp16 = softmax(axis = var_5235, x = aw_779_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5401_cast_fp16 = softmax(axis = var_5235, x = aw_781_cast_fp16)[name = tensor("op_5401_cast_fp16")]; + tensor var_5402_cast_fp16 = softmax(axis = var_5235, x = aw_783_cast_fp16)[name = tensor("op_5402_cast_fp16")]; + tensor var_5403_cast_fp16 = softmax(axis = var_5235, x = aw_785_cast_fp16)[name = tensor("op_5403_cast_fp16")]; + tensor var_5404_cast_fp16 = softmax(axis = var_5235, x = aw_787_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor var_5405_cast_fp16 = softmax(axis = var_5235, x = aw_789_cast_fp16)[name = tensor("op_5405_cast_fp16")]; + tensor var_5406_cast_fp16 = softmax(axis = var_5235, x = aw_791_cast_fp16)[name = tensor("op_5406_cast_fp16")]; + tensor var_5407_cast_fp16 = softmax(axis = var_5235, x = aw_793_cast_fp16)[name = tensor("op_5407_cast_fp16")]; + tensor var_5408_cast_fp16 = softmax(axis = var_5235, x = aw_795_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor var_5409_cast_fp16 = softmax(axis = var_5235, x = aw_797_cast_fp16)[name = tensor("op_5409_cast_fp16")]; + tensor var_5410_cast_fp16 = softmax(axis = var_5235, x = aw_799_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5412_equation_0 = const()[name = tensor("op_5412_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5412_cast_fp16 = einsum(equation = var_5412_equation_0, values = (var_5330_cast_fp16_0, var_5391_cast_fp16))[name = tensor("op_5412_cast_fp16")]; + tensor var_5414_equation_0 = const()[name = tensor("op_5414_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5414_cast_fp16 = einsum(equation = var_5414_equation_0, values = (var_5330_cast_fp16_1, var_5392_cast_fp16))[name = tensor("op_5414_cast_fp16")]; + tensor var_5416_equation_0 = const()[name = tensor("op_5416_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5416_cast_fp16 = einsum(equation = var_5416_equation_0, values = (var_5330_cast_fp16_2, var_5393_cast_fp16))[name = tensor("op_5416_cast_fp16")]; + tensor var_5418_equation_0 = const()[name = tensor("op_5418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5418_cast_fp16 = einsum(equation = var_5418_equation_0, values = (var_5330_cast_fp16_3, var_5394_cast_fp16))[name = tensor("op_5418_cast_fp16")]; + tensor var_5420_equation_0 = const()[name = tensor("op_5420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5420_cast_fp16 = einsum(equation = var_5420_equation_0, values = (var_5330_cast_fp16_4, var_5395_cast_fp16))[name = tensor("op_5420_cast_fp16")]; + tensor var_5422_equation_0 = const()[name = tensor("op_5422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5422_cast_fp16 = einsum(equation = var_5422_equation_0, values = (var_5330_cast_fp16_5, var_5396_cast_fp16))[name = tensor("op_5422_cast_fp16")]; + tensor var_5424_equation_0 = const()[name = tensor("op_5424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5424_cast_fp16 = einsum(equation = var_5424_equation_0, values = (var_5330_cast_fp16_6, var_5397_cast_fp16))[name = tensor("op_5424_cast_fp16")]; + tensor var_5426_equation_0 = const()[name = tensor("op_5426_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5426_cast_fp16 = einsum(equation = var_5426_equation_0, values = (var_5330_cast_fp16_7, var_5398_cast_fp16))[name = tensor("op_5426_cast_fp16")]; + tensor var_5428_equation_0 = const()[name = tensor("op_5428_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5428_cast_fp16 = einsum(equation = var_5428_equation_0, values = (var_5330_cast_fp16_8, var_5399_cast_fp16))[name = tensor("op_5428_cast_fp16")]; + tensor var_5430_equation_0 = const()[name = tensor("op_5430_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5430_cast_fp16 = einsum(equation = var_5430_equation_0, values = (var_5330_cast_fp16_9, var_5400_cast_fp16))[name = tensor("op_5430_cast_fp16")]; + tensor var_5432_equation_0 = const()[name = tensor("op_5432_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5432_cast_fp16 = einsum(equation = var_5432_equation_0, values = (var_5330_cast_fp16_10, var_5401_cast_fp16))[name = tensor("op_5432_cast_fp16")]; + tensor var_5434_equation_0 = const()[name = tensor("op_5434_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5434_cast_fp16 = einsum(equation = var_5434_equation_0, values = (var_5330_cast_fp16_11, var_5402_cast_fp16))[name = tensor("op_5434_cast_fp16")]; + tensor var_5436_equation_0 = const()[name = tensor("op_5436_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5436_cast_fp16 = einsum(equation = var_5436_equation_0, values = (var_5330_cast_fp16_12, var_5403_cast_fp16))[name = tensor("op_5436_cast_fp16")]; + tensor var_5438_equation_0 = const()[name = tensor("op_5438_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5438_cast_fp16 = einsum(equation = var_5438_equation_0, values = (var_5330_cast_fp16_13, var_5404_cast_fp16))[name = tensor("op_5438_cast_fp16")]; + tensor var_5440_equation_0 = const()[name = tensor("op_5440_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5440_cast_fp16 = einsum(equation = var_5440_equation_0, values = (var_5330_cast_fp16_14, var_5405_cast_fp16))[name = tensor("op_5440_cast_fp16")]; + tensor var_5442_equation_0 = const()[name = tensor("op_5442_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5442_cast_fp16 = einsum(equation = var_5442_equation_0, values = (var_5330_cast_fp16_15, var_5406_cast_fp16))[name = tensor("op_5442_cast_fp16")]; + tensor var_5444_equation_0 = const()[name = tensor("op_5444_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5444_cast_fp16 = einsum(equation = var_5444_equation_0, values = (var_5330_cast_fp16_16, var_5407_cast_fp16))[name = tensor("op_5444_cast_fp16")]; + tensor var_5446_equation_0 = const()[name = tensor("op_5446_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5446_cast_fp16 = einsum(equation = var_5446_equation_0, values = (var_5330_cast_fp16_17, var_5408_cast_fp16))[name = tensor("op_5446_cast_fp16")]; + tensor var_5448_equation_0 = const()[name = tensor("op_5448_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5448_cast_fp16 = einsum(equation = var_5448_equation_0, values = (var_5330_cast_fp16_18, var_5409_cast_fp16))[name = tensor("op_5448_cast_fp16")]; + tensor var_5450_equation_0 = const()[name = tensor("op_5450_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5450_cast_fp16 = einsum(equation = var_5450_equation_0, values = (var_5330_cast_fp16_19, var_5410_cast_fp16))[name = tensor("op_5450_cast_fp16")]; + tensor input_195_interleave_0 = const()[name = tensor("input_195_interleave_0"), val = tensor(false)]; + tensor input_195_cast_fp16 = concat(axis = var_5235, interleave = input_195_interleave_0, values = (var_5412_cast_fp16, var_5414_cast_fp16, var_5416_cast_fp16, var_5418_cast_fp16, var_5420_cast_fp16, var_5422_cast_fp16, var_5424_cast_fp16, var_5426_cast_fp16, var_5428_cast_fp16, var_5430_cast_fp16, var_5432_cast_fp16, var_5434_cast_fp16, var_5436_cast_fp16, var_5438_cast_fp16, var_5440_cast_fp16, var_5442_cast_fp16, var_5444_cast_fp16, var_5446_cast_fp16, var_5448_cast_fp16, var_5450_cast_fp16))[name = tensor("input_195_cast_fp16")]; + tensor var_5459_pad_type_0 = const()[name = tensor("op_5459_pad_type_0"), val = tensor("valid")]; + tensor var_5459_strides_0 = const()[name = tensor("op_5459_strides_0"), val = tensor([1, 1])]; + tensor var_5459_pad_0 = const()[name = tensor("op_5459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5459_dilations_0 = const()[name = tensor("op_5459_dilations_0"), val = tensor([1, 1])]; + tensor var_5459_groups_0 = const()[name = tensor("op_5459_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_out_weight_to_fp16 = const()[name = tensor("blocks_19_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771843712)))]; + tensor blocks_19_attn_out_bias_to_fp16 = const()[name = tensor("blocks_19_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775120576)))]; + tensor var_5459_cast_fp16 = conv(bias = blocks_19_attn_out_bias_to_fp16, dilations = var_5459_dilations_0, groups = var_5459_groups_0, pad = var_5459_pad_0, pad_type = var_5459_pad_type_0, strides = var_5459_strides_0, weight = blocks_19_attn_out_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("op_5459_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_5459_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([1])]; + tensor input_197_gamma_0_to_fp16 = const()[name = tensor("input_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775123200)))]; + tensor input_197_beta_0_to_fp16 = const()[name = tensor("input_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775125824)))]; + tensor var_5469_to_fp16 = const()[name = tensor("op_5469_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = input_197_beta_0_to_fp16, epsilon = var_5469_to_fp16, gamma = input_197_gamma_0_to_fp16, x = inputs_79_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("valid")]; + tensor input_199_strides_0 = const()[name = tensor("input_199_strides_0"), val = tensor([1, 1])]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_199_dilations_0 = const()[name = tensor("input_199_dilations_0"), val = tensor([1, 1])]; + tensor input_199_groups_0 = const()[name = tensor("input_199_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775128448)))]; + tensor blocks_19_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788235712)))]; + tensor input_199_cast_fp16 = conv(bias = blocks_19_mlp_0_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = blocks_19_mlp_0_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_5495_pad_type_0 = const()[name = tensor("op_5495_pad_type_0"), val = tensor("valid")]; + tensor var_5495_strides_0 = const()[name = tensor("op_5495_strides_0"), val = tensor([1, 1])]; + tensor var_5495_pad_0 = const()[name = tensor("op_5495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5495_dilations_0 = const()[name = tensor("op_5495_dilations_0"), val = tensor([1, 1])]; + tensor var_5495_groups_0 = const()[name = tensor("op_5495_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788246016)))]; + tensor blocks_19_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801353280)))]; + tensor var_5495_cast_fp16 = conv(bias = blocks_19_mlp_2_bias_to_fp16, dilations = var_5495_dilations_0, groups = var_5495_groups_0, pad = var_5495_pad_0, pad_type = var_5495_pad_type_0, strides = var_5495_strides_0, weight = blocks_19_mlp_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = var_5495_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_5504 = const()[name = tensor("op_5504"), val = tensor(1)]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([1])]; + tensor input_203_gamma_0_to_fp16 = const()[name = tensor("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801355904)))]; + tensor input_203_beta_0_to_fp16 = const()[name = tensor("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801358528)))]; + tensor var_5520_to_fp16 = const()[name = tensor("op_5520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = input_203_beta_0_to_fp16, epsilon = var_5520_to_fp16, gamma = input_203_gamma_0_to_fp16, x = inputs_81_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor q_41_pad_type_0 = const()[name = tensor("q_41_pad_type_0"), val = tensor("valid")]; + tensor q_41_strides_0 = const()[name = tensor("q_41_strides_0"), val = tensor([1, 1])]; + tensor q_41_pad_0 = const()[name = tensor("q_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_41_dilations_0 = const()[name = tensor("q_41_dilations_0"), val = tensor([1, 1])]; + tensor q_41_groups_0 = const()[name = tensor("q_41_groups_0"), val = tensor(1)]; + tensor var_5555_weight_0_to_fp16 = const()[name = tensor("op_5555_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801361152)))]; + tensor var_5555_bias_0_to_fp16 = const()[name = tensor("op_5555_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804638016)))]; + tensor var_5555_cast_fp16 = conv(bias = var_5555_bias_0_to_fp16, dilations = q_41_dilations_0, groups = q_41_groups_0, pad = q_41_pad_0, pad_type = q_41_pad_type_0, strides = q_41_strides_0, weight = var_5555_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("op_5555_cast_fp16")]; + tensor k_41_pad_type_0 = const()[name = tensor("k_41_pad_type_0"), val = tensor("valid")]; + tensor k_41_strides_0 = const()[name = tensor("k_41_strides_0"), val = tensor([1, 1])]; + tensor k_41_pad_0 = const()[name = tensor("k_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_41_dilations_0 = const()[name = tensor("k_41_dilations_0"), val = tensor([1, 1])]; + tensor k_41_groups_0 = const()[name = tensor("k_41_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_key_weight_to_fp16 = const()[name = tensor("blocks_20_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804640640)))]; + tensor k_41_cast_fp16 = conv(dilations = k_41_dilations_0, groups = k_41_groups_0, pad = k_41_pad_0, pad_type = k_41_pad_type_0, strides = k_41_strides_0, weight = blocks_20_attn_key_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor var_5553_pad_type_0 = const()[name = tensor("op_5553_pad_type_0"), val = tensor("valid")]; + tensor var_5553_strides_0 = const()[name = tensor("op_5553_strides_0"), val = tensor([1, 1])]; + tensor var_5553_pad_0 = const()[name = tensor("op_5553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5553_dilations_0 = const()[name = tensor("op_5553_dilations_0"), val = tensor([1, 1])]; + tensor var_5553_groups_0 = const()[name = tensor("op_5553_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_value_weight_to_fp16 = const()[name = tensor("blocks_20_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807917504)))]; + tensor blocks_20_attn_value_bias_to_fp16 = const()[name = tensor("blocks_20_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811194368)))]; + tensor var_5553_cast_fp16 = conv(bias = blocks_20_attn_value_bias_to_fp16, dilations = var_5553_dilations_0, groups = var_5553_groups_0, pad = var_5553_pad_0, pad_type = var_5553_pad_type_0, strides = var_5553_strides_0, weight = blocks_20_attn_value_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("op_5553_cast_fp16")]; + tensor tile_60 = const()[name = tensor("tile_60"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5556_axis_0 = const()[name = tensor("op_5556_axis_0"), val = tensor(1)]; + tensor var_5556_cast_fp16_0, tensor var_5556_cast_fp16_1, tensor var_5556_cast_fp16_2, tensor var_5556_cast_fp16_3, tensor var_5556_cast_fp16_4, tensor var_5556_cast_fp16_5, tensor var_5556_cast_fp16_6, tensor var_5556_cast_fp16_7, tensor var_5556_cast_fp16_8, tensor var_5556_cast_fp16_9, tensor var_5556_cast_fp16_10, tensor var_5556_cast_fp16_11, tensor var_5556_cast_fp16_12, tensor var_5556_cast_fp16_13, tensor var_5556_cast_fp16_14, tensor var_5556_cast_fp16_15, tensor var_5556_cast_fp16_16, tensor var_5556_cast_fp16_17, tensor var_5556_cast_fp16_18, tensor var_5556_cast_fp16_19 = split(axis = var_5556_axis_0, split_sizes = tile_60, x = var_5555_cast_fp16)[name = tensor("op_5556_cast_fp16")]; + tensor var_5577_perm_0 = const()[name = tensor("op_5577_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_61 = const()[name = tensor("tile_61"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5578_axis_0 = const()[name = tensor("op_5578_axis_0"), val = tensor(3)]; + tensor var_5577_cast_fp16 = transpose(perm = var_5577_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_12")]; + tensor var_5578_cast_fp16_0, tensor var_5578_cast_fp16_1, tensor var_5578_cast_fp16_2, tensor var_5578_cast_fp16_3, tensor var_5578_cast_fp16_4, tensor var_5578_cast_fp16_5, tensor var_5578_cast_fp16_6, tensor var_5578_cast_fp16_7, tensor var_5578_cast_fp16_8, tensor var_5578_cast_fp16_9, tensor var_5578_cast_fp16_10, tensor var_5578_cast_fp16_11, tensor var_5578_cast_fp16_12, tensor var_5578_cast_fp16_13, tensor var_5578_cast_fp16_14, tensor var_5578_cast_fp16_15, tensor var_5578_cast_fp16_16, tensor var_5578_cast_fp16_17, tensor var_5578_cast_fp16_18, tensor var_5578_cast_fp16_19 = split(axis = var_5578_axis_0, split_sizes = tile_61, x = var_5577_cast_fp16)[name = tensor("op_5578_cast_fp16")]; + tensor tile_62 = const()[name = tensor("tile_62"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5599_axis_0 = const()[name = tensor("op_5599_axis_0"), val = tensor(1)]; + tensor var_5599_cast_fp16_0, tensor var_5599_cast_fp16_1, tensor var_5599_cast_fp16_2, tensor var_5599_cast_fp16_3, tensor var_5599_cast_fp16_4, tensor var_5599_cast_fp16_5, tensor var_5599_cast_fp16_6, tensor var_5599_cast_fp16_7, tensor var_5599_cast_fp16_8, tensor var_5599_cast_fp16_9, tensor var_5599_cast_fp16_10, tensor var_5599_cast_fp16_11, tensor var_5599_cast_fp16_12, tensor var_5599_cast_fp16_13, tensor var_5599_cast_fp16_14, tensor var_5599_cast_fp16_15, tensor var_5599_cast_fp16_16, tensor var_5599_cast_fp16_17, tensor var_5599_cast_fp16_18, tensor var_5599_cast_fp16_19 = split(axis = var_5599_axis_0, split_sizes = tile_62, x = var_5553_cast_fp16)[name = tensor("op_5599_cast_fp16")]; + tensor aw_801_equation_0 = const()[name = tensor("aw_801_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_801_cast_fp16 = einsum(equation = aw_801_equation_0, values = (var_5578_cast_fp16_0, var_5556_cast_fp16_0))[name = tensor("aw_801_cast_fp16")]; + tensor aw_803_equation_0 = const()[name = tensor("aw_803_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_803_cast_fp16 = einsum(equation = aw_803_equation_0, values = (var_5578_cast_fp16_1, var_5556_cast_fp16_1))[name = tensor("aw_803_cast_fp16")]; + tensor aw_805_equation_0 = const()[name = tensor("aw_805_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_805_cast_fp16 = einsum(equation = aw_805_equation_0, values = (var_5578_cast_fp16_2, var_5556_cast_fp16_2))[name = tensor("aw_805_cast_fp16")]; + tensor aw_807_equation_0 = const()[name = tensor("aw_807_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_807_cast_fp16 = einsum(equation = aw_807_equation_0, values = (var_5578_cast_fp16_3, var_5556_cast_fp16_3))[name = tensor("aw_807_cast_fp16")]; + tensor aw_809_equation_0 = const()[name = tensor("aw_809_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_809_cast_fp16 = einsum(equation = aw_809_equation_0, values = (var_5578_cast_fp16_4, var_5556_cast_fp16_4))[name = tensor("aw_809_cast_fp16")]; + tensor aw_811_equation_0 = const()[name = tensor("aw_811_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_811_cast_fp16 = einsum(equation = aw_811_equation_0, values = (var_5578_cast_fp16_5, var_5556_cast_fp16_5))[name = tensor("aw_811_cast_fp16")]; + tensor aw_813_equation_0 = const()[name = tensor("aw_813_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_813_cast_fp16 = einsum(equation = aw_813_equation_0, values = (var_5578_cast_fp16_6, var_5556_cast_fp16_6))[name = tensor("aw_813_cast_fp16")]; + tensor aw_815_equation_0 = const()[name = tensor("aw_815_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_815_cast_fp16 = einsum(equation = aw_815_equation_0, values = (var_5578_cast_fp16_7, var_5556_cast_fp16_7))[name = tensor("aw_815_cast_fp16")]; + tensor aw_817_equation_0 = const()[name = tensor("aw_817_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_817_cast_fp16 = einsum(equation = aw_817_equation_0, values = (var_5578_cast_fp16_8, var_5556_cast_fp16_8))[name = tensor("aw_817_cast_fp16")]; + tensor aw_819_equation_0 = const()[name = tensor("aw_819_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_819_cast_fp16 = einsum(equation = aw_819_equation_0, values = (var_5578_cast_fp16_9, var_5556_cast_fp16_9))[name = tensor("aw_819_cast_fp16")]; + tensor aw_821_equation_0 = const()[name = tensor("aw_821_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_821_cast_fp16 = einsum(equation = aw_821_equation_0, values = (var_5578_cast_fp16_10, var_5556_cast_fp16_10))[name = tensor("aw_821_cast_fp16")]; + tensor aw_823_equation_0 = const()[name = tensor("aw_823_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_823_cast_fp16 = einsum(equation = aw_823_equation_0, values = (var_5578_cast_fp16_11, var_5556_cast_fp16_11))[name = tensor("aw_823_cast_fp16")]; + tensor aw_825_equation_0 = const()[name = tensor("aw_825_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_825_cast_fp16 = einsum(equation = aw_825_equation_0, values = (var_5578_cast_fp16_12, var_5556_cast_fp16_12))[name = tensor("aw_825_cast_fp16")]; + tensor aw_827_equation_0 = const()[name = tensor("aw_827_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_827_cast_fp16 = einsum(equation = aw_827_equation_0, values = (var_5578_cast_fp16_13, var_5556_cast_fp16_13))[name = tensor("aw_827_cast_fp16")]; + tensor aw_829_equation_0 = const()[name = tensor("aw_829_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_829_cast_fp16 = einsum(equation = aw_829_equation_0, values = (var_5578_cast_fp16_14, var_5556_cast_fp16_14))[name = tensor("aw_829_cast_fp16")]; + tensor aw_831_equation_0 = const()[name = tensor("aw_831_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_831_cast_fp16 = einsum(equation = aw_831_equation_0, values = (var_5578_cast_fp16_15, var_5556_cast_fp16_15))[name = tensor("aw_831_cast_fp16")]; + tensor aw_833_equation_0 = const()[name = tensor("aw_833_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_833_cast_fp16 = einsum(equation = aw_833_equation_0, values = (var_5578_cast_fp16_16, var_5556_cast_fp16_16))[name = tensor("aw_833_cast_fp16")]; + tensor aw_835_equation_0 = const()[name = tensor("aw_835_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_835_cast_fp16 = einsum(equation = aw_835_equation_0, values = (var_5578_cast_fp16_17, var_5556_cast_fp16_17))[name = tensor("aw_835_cast_fp16")]; + tensor aw_837_equation_0 = const()[name = tensor("aw_837_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_837_cast_fp16 = einsum(equation = aw_837_equation_0, values = (var_5578_cast_fp16_18, var_5556_cast_fp16_18))[name = tensor("aw_837_cast_fp16")]; + tensor aw_839_equation_0 = const()[name = tensor("aw_839_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_839_cast_fp16 = einsum(equation = aw_839_equation_0, values = (var_5578_cast_fp16_19, var_5556_cast_fp16_19))[name = tensor("aw_839_cast_fp16")]; + tensor var_5660_cast_fp16 = softmax(axis = var_5504, x = aw_801_cast_fp16)[name = tensor("op_5660_cast_fp16")]; + tensor var_5661_cast_fp16 = softmax(axis = var_5504, x = aw_803_cast_fp16)[name = tensor("op_5661_cast_fp16")]; + tensor var_5662_cast_fp16 = softmax(axis = var_5504, x = aw_805_cast_fp16)[name = tensor("op_5662_cast_fp16")]; + tensor var_5663_cast_fp16 = softmax(axis = var_5504, x = aw_807_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor var_5664_cast_fp16 = softmax(axis = var_5504, x = aw_809_cast_fp16)[name = tensor("op_5664_cast_fp16")]; + tensor var_5665_cast_fp16 = softmax(axis = var_5504, x = aw_811_cast_fp16)[name = tensor("op_5665_cast_fp16")]; + tensor var_5666_cast_fp16 = softmax(axis = var_5504, x = aw_813_cast_fp16)[name = tensor("op_5666_cast_fp16")]; + tensor var_5667_cast_fp16 = softmax(axis = var_5504, x = aw_815_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = softmax(axis = var_5504, x = aw_817_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor var_5669_cast_fp16 = softmax(axis = var_5504, x = aw_819_cast_fp16)[name = tensor("op_5669_cast_fp16")]; + tensor var_5670_cast_fp16 = softmax(axis = var_5504, x = aw_821_cast_fp16)[name = tensor("op_5670_cast_fp16")]; + tensor var_5671_cast_fp16 = softmax(axis = var_5504, x = aw_823_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = softmax(axis = var_5504, x = aw_825_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor var_5673_cast_fp16 = softmax(axis = var_5504, x = aw_827_cast_fp16)[name = tensor("op_5673_cast_fp16")]; + tensor var_5674_cast_fp16 = softmax(axis = var_5504, x = aw_829_cast_fp16)[name = tensor("op_5674_cast_fp16")]; + tensor var_5675_cast_fp16 = softmax(axis = var_5504, x = aw_831_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = softmax(axis = var_5504, x = aw_833_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor var_5677_cast_fp16 = softmax(axis = var_5504, x = aw_835_cast_fp16)[name = tensor("op_5677_cast_fp16")]; + tensor var_5678_cast_fp16 = softmax(axis = var_5504, x = aw_837_cast_fp16)[name = tensor("op_5678_cast_fp16")]; + tensor var_5679_cast_fp16 = softmax(axis = var_5504, x = aw_839_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5681_equation_0 = const()[name = tensor("op_5681_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5681_cast_fp16 = einsum(equation = var_5681_equation_0, values = (var_5599_cast_fp16_0, var_5660_cast_fp16))[name = tensor("op_5681_cast_fp16")]; + tensor var_5683_equation_0 = const()[name = tensor("op_5683_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5683_cast_fp16 = einsum(equation = var_5683_equation_0, values = (var_5599_cast_fp16_1, var_5661_cast_fp16))[name = tensor("op_5683_cast_fp16")]; + tensor var_5685_equation_0 = const()[name = tensor("op_5685_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5685_cast_fp16 = einsum(equation = var_5685_equation_0, values = (var_5599_cast_fp16_2, var_5662_cast_fp16))[name = tensor("op_5685_cast_fp16")]; + tensor var_5687_equation_0 = const()[name = tensor("op_5687_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5687_cast_fp16 = einsum(equation = var_5687_equation_0, values = (var_5599_cast_fp16_3, var_5663_cast_fp16))[name = tensor("op_5687_cast_fp16")]; + tensor var_5689_equation_0 = const()[name = tensor("op_5689_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5689_cast_fp16 = einsum(equation = var_5689_equation_0, values = (var_5599_cast_fp16_4, var_5664_cast_fp16))[name = tensor("op_5689_cast_fp16")]; + tensor var_5691_equation_0 = const()[name = tensor("op_5691_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5691_cast_fp16 = einsum(equation = var_5691_equation_0, values = (var_5599_cast_fp16_5, var_5665_cast_fp16))[name = tensor("op_5691_cast_fp16")]; + tensor var_5693_equation_0 = const()[name = tensor("op_5693_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5693_cast_fp16 = einsum(equation = var_5693_equation_0, values = (var_5599_cast_fp16_6, var_5666_cast_fp16))[name = tensor("op_5693_cast_fp16")]; + tensor var_5695_equation_0 = const()[name = tensor("op_5695_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5695_cast_fp16 = einsum(equation = var_5695_equation_0, values = (var_5599_cast_fp16_7, var_5667_cast_fp16))[name = tensor("op_5695_cast_fp16")]; + tensor var_5697_equation_0 = const()[name = tensor("op_5697_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5697_cast_fp16 = einsum(equation = var_5697_equation_0, values = (var_5599_cast_fp16_8, var_5668_cast_fp16))[name = tensor("op_5697_cast_fp16")]; + tensor var_5699_equation_0 = const()[name = tensor("op_5699_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5699_cast_fp16 = einsum(equation = var_5699_equation_0, values = (var_5599_cast_fp16_9, var_5669_cast_fp16))[name = tensor("op_5699_cast_fp16")]; + tensor var_5701_equation_0 = const()[name = tensor("op_5701_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5701_cast_fp16 = einsum(equation = var_5701_equation_0, values = (var_5599_cast_fp16_10, var_5670_cast_fp16))[name = tensor("op_5701_cast_fp16")]; + tensor var_5703_equation_0 = const()[name = tensor("op_5703_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5703_cast_fp16 = einsum(equation = var_5703_equation_0, values = (var_5599_cast_fp16_11, var_5671_cast_fp16))[name = tensor("op_5703_cast_fp16")]; + tensor var_5705_equation_0 = const()[name = tensor("op_5705_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5705_cast_fp16 = einsum(equation = var_5705_equation_0, values = (var_5599_cast_fp16_12, var_5672_cast_fp16))[name = tensor("op_5705_cast_fp16")]; + tensor var_5707_equation_0 = const()[name = tensor("op_5707_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5707_cast_fp16 = einsum(equation = var_5707_equation_0, values = (var_5599_cast_fp16_13, var_5673_cast_fp16))[name = tensor("op_5707_cast_fp16")]; + tensor var_5709_equation_0 = const()[name = tensor("op_5709_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5709_cast_fp16 = einsum(equation = var_5709_equation_0, values = (var_5599_cast_fp16_14, var_5674_cast_fp16))[name = tensor("op_5709_cast_fp16")]; + tensor var_5711_equation_0 = const()[name = tensor("op_5711_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5711_cast_fp16 = einsum(equation = var_5711_equation_0, values = (var_5599_cast_fp16_15, var_5675_cast_fp16))[name = tensor("op_5711_cast_fp16")]; + tensor var_5713_equation_0 = const()[name = tensor("op_5713_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5713_cast_fp16 = einsum(equation = var_5713_equation_0, values = (var_5599_cast_fp16_16, var_5676_cast_fp16))[name = tensor("op_5713_cast_fp16")]; + tensor var_5715_equation_0 = const()[name = tensor("op_5715_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5715_cast_fp16 = einsum(equation = var_5715_equation_0, values = (var_5599_cast_fp16_17, var_5677_cast_fp16))[name = tensor("op_5715_cast_fp16")]; + tensor var_5717_equation_0 = const()[name = tensor("op_5717_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5717_cast_fp16 = einsum(equation = var_5717_equation_0, values = (var_5599_cast_fp16_18, var_5678_cast_fp16))[name = tensor("op_5717_cast_fp16")]; + tensor var_5719_equation_0 = const()[name = tensor("op_5719_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5719_cast_fp16 = einsum(equation = var_5719_equation_0, values = (var_5599_cast_fp16_19, var_5679_cast_fp16))[name = tensor("op_5719_cast_fp16")]; + tensor input_205_interleave_0 = const()[name = tensor("input_205_interleave_0"), val = tensor(false)]; + tensor input_205_cast_fp16 = concat(axis = var_5504, interleave = input_205_interleave_0, values = (var_5681_cast_fp16, var_5683_cast_fp16, var_5685_cast_fp16, var_5687_cast_fp16, var_5689_cast_fp16, var_5691_cast_fp16, var_5693_cast_fp16, var_5695_cast_fp16, var_5697_cast_fp16, var_5699_cast_fp16, var_5701_cast_fp16, var_5703_cast_fp16, var_5705_cast_fp16, var_5707_cast_fp16, var_5709_cast_fp16, var_5711_cast_fp16, var_5713_cast_fp16, var_5715_cast_fp16, var_5717_cast_fp16, var_5719_cast_fp16))[name = tensor("input_205_cast_fp16")]; + tensor var_5728_pad_type_0 = const()[name = tensor("op_5728_pad_type_0"), val = tensor("valid")]; + tensor var_5728_strides_0 = const()[name = tensor("op_5728_strides_0"), val = tensor([1, 1])]; + tensor var_5728_pad_0 = const()[name = tensor("op_5728_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5728_dilations_0 = const()[name = tensor("op_5728_dilations_0"), val = tensor([1, 1])]; + tensor var_5728_groups_0 = const()[name = tensor("op_5728_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_out_weight_to_fp16 = const()[name = tensor("blocks_20_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811196992)))]; + tensor blocks_20_attn_out_bias_to_fp16 = const()[name = tensor("blocks_20_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814473856)))]; + tensor var_5728_cast_fp16 = conv(bias = blocks_20_attn_out_bias_to_fp16, dilations = var_5728_dilations_0, groups = var_5728_groups_0, pad = var_5728_pad_0, pad_type = var_5728_pad_type_0, strides = var_5728_strides_0, weight = blocks_20_attn_out_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_5728_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([1])]; + tensor input_207_gamma_0_to_fp16 = const()[name = tensor("input_207_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814476480)))]; + tensor input_207_beta_0_to_fp16 = const()[name = tensor("input_207_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814479104)))]; + tensor var_5738_to_fp16 = const()[name = tensor("op_5738_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = input_207_beta_0_to_fp16, epsilon = var_5738_to_fp16, gamma = input_207_gamma_0_to_fp16, x = inputs_83_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("valid")]; + tensor input_209_strides_0 = const()[name = tensor("input_209_strides_0"), val = tensor([1, 1])]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_209_dilations_0 = const()[name = tensor("input_209_dilations_0"), val = tensor([1, 1])]; + tensor input_209_groups_0 = const()[name = tensor("input_209_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814481728)))]; + tensor blocks_20_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827588992)))]; + tensor input_209_cast_fp16 = conv(bias = blocks_20_mlp_0_bias_to_fp16, dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = blocks_20_mlp_0_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = input_209_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_5764_pad_type_0 = const()[name = tensor("op_5764_pad_type_0"), val = tensor("valid")]; + tensor var_5764_strides_0 = const()[name = tensor("op_5764_strides_0"), val = tensor([1, 1])]; + tensor var_5764_pad_0 = const()[name = tensor("op_5764_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5764_dilations_0 = const()[name = tensor("op_5764_dilations_0"), val = tensor([1, 1])]; + tensor var_5764_groups_0 = const()[name = tensor("op_5764_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827599296)))]; + tensor blocks_20_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840706560)))]; + tensor var_5764_cast_fp16 = conv(bias = blocks_20_mlp_2_bias_to_fp16, dilations = var_5764_dilations_0, groups = var_5764_groups_0, pad = var_5764_pad_0, pad_type = var_5764_pad_type_0, strides = var_5764_strides_0, weight = blocks_20_mlp_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("op_5764_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = var_5764_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_5773 = const()[name = tensor("op_5773"), val = tensor(1)]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([1])]; + tensor input_213_gamma_0_to_fp16 = const()[name = tensor("input_213_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840709184)))]; + tensor input_213_beta_0_to_fp16 = const()[name = tensor("input_213_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840711808)))]; + tensor var_5789_to_fp16 = const()[name = tensor("op_5789_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = input_213_beta_0_to_fp16, epsilon = var_5789_to_fp16, gamma = input_213_gamma_0_to_fp16, x = inputs_85_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor q_43_pad_type_0 = const()[name = tensor("q_43_pad_type_0"), val = tensor("valid")]; + tensor q_43_strides_0 = const()[name = tensor("q_43_strides_0"), val = tensor([1, 1])]; + tensor q_43_pad_0 = const()[name = tensor("q_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_43_dilations_0 = const()[name = tensor("q_43_dilations_0"), val = tensor([1, 1])]; + tensor q_43_groups_0 = const()[name = tensor("q_43_groups_0"), val = tensor(1)]; + tensor var_5824_weight_0_to_fp16 = const()[name = tensor("op_5824_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840714432)))]; + tensor var_5824_bias_0_to_fp16 = const()[name = tensor("op_5824_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(843991296)))]; + tensor var_5824_cast_fp16 = conv(bias = var_5824_bias_0_to_fp16, dilations = q_43_dilations_0, groups = q_43_groups_0, pad = q_43_pad_0, pad_type = q_43_pad_type_0, strides = q_43_strides_0, weight = var_5824_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5824_cast_fp16")]; + tensor k_43_pad_type_0 = const()[name = tensor("k_43_pad_type_0"), val = tensor("valid")]; + tensor k_43_strides_0 = const()[name = tensor("k_43_strides_0"), val = tensor([1, 1])]; + tensor k_43_pad_0 = const()[name = tensor("k_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_43_dilations_0 = const()[name = tensor("k_43_dilations_0"), val = tensor([1, 1])]; + tensor k_43_groups_0 = const()[name = tensor("k_43_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_key_weight_to_fp16 = const()[name = tensor("blocks_21_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(843993920)))]; + tensor k_43_cast_fp16 = conv(dilations = k_43_dilations_0, groups = k_43_groups_0, pad = k_43_pad_0, pad_type = k_43_pad_type_0, strides = k_43_strides_0, weight = blocks_21_attn_key_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor var_5822_pad_type_0 = const()[name = tensor("op_5822_pad_type_0"), val = tensor("valid")]; + tensor var_5822_strides_0 = const()[name = tensor("op_5822_strides_0"), val = tensor([1, 1])]; + tensor var_5822_pad_0 = const()[name = tensor("op_5822_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5822_dilations_0 = const()[name = tensor("op_5822_dilations_0"), val = tensor([1, 1])]; + tensor var_5822_groups_0 = const()[name = tensor("op_5822_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_value_weight_to_fp16 = const()[name = tensor("blocks_21_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847270784)))]; + tensor blocks_21_attn_value_bias_to_fp16 = const()[name = tensor("blocks_21_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850547648)))]; + tensor var_5822_cast_fp16 = conv(bias = blocks_21_attn_value_bias_to_fp16, dilations = var_5822_dilations_0, groups = var_5822_groups_0, pad = var_5822_pad_0, pad_type = var_5822_pad_type_0, strides = var_5822_strides_0, weight = blocks_21_attn_value_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5822_cast_fp16")]; + tensor tile_63 = const()[name = tensor("tile_63"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5825_axis_0 = const()[name = tensor("op_5825_axis_0"), val = tensor(1)]; + tensor var_5825_cast_fp16_0, tensor var_5825_cast_fp16_1, tensor var_5825_cast_fp16_2, tensor var_5825_cast_fp16_3, tensor var_5825_cast_fp16_4, tensor var_5825_cast_fp16_5, tensor var_5825_cast_fp16_6, tensor var_5825_cast_fp16_7, tensor var_5825_cast_fp16_8, tensor var_5825_cast_fp16_9, tensor var_5825_cast_fp16_10, tensor var_5825_cast_fp16_11, tensor var_5825_cast_fp16_12, tensor var_5825_cast_fp16_13, tensor var_5825_cast_fp16_14, tensor var_5825_cast_fp16_15, tensor var_5825_cast_fp16_16, tensor var_5825_cast_fp16_17, tensor var_5825_cast_fp16_18, tensor var_5825_cast_fp16_19 = split(axis = var_5825_axis_0, split_sizes = tile_63, x = var_5824_cast_fp16)[name = tensor("op_5825_cast_fp16")]; + tensor var_5846_perm_0 = const()[name = tensor("op_5846_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_64 = const()[name = tensor("tile_64"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5847_axis_0 = const()[name = tensor("op_5847_axis_0"), val = tensor(3)]; + tensor var_5846_cast_fp16 = transpose(perm = var_5846_perm_0, x = k_43_cast_fp16)[name = tensor("transpose_11")]; + tensor var_5847_cast_fp16_0, tensor var_5847_cast_fp16_1, tensor var_5847_cast_fp16_2, tensor var_5847_cast_fp16_3, tensor var_5847_cast_fp16_4, tensor var_5847_cast_fp16_5, tensor var_5847_cast_fp16_6, tensor var_5847_cast_fp16_7, tensor var_5847_cast_fp16_8, tensor var_5847_cast_fp16_9, tensor var_5847_cast_fp16_10, tensor var_5847_cast_fp16_11, tensor var_5847_cast_fp16_12, tensor var_5847_cast_fp16_13, tensor var_5847_cast_fp16_14, tensor var_5847_cast_fp16_15, tensor var_5847_cast_fp16_16, tensor var_5847_cast_fp16_17, tensor var_5847_cast_fp16_18, tensor var_5847_cast_fp16_19 = split(axis = var_5847_axis_0, split_sizes = tile_64, x = var_5846_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor tile_65 = const()[name = tensor("tile_65"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5868_axis_0 = const()[name = tensor("op_5868_axis_0"), val = tensor(1)]; + tensor var_5868_cast_fp16_0, tensor var_5868_cast_fp16_1, tensor var_5868_cast_fp16_2, tensor var_5868_cast_fp16_3, tensor var_5868_cast_fp16_4, tensor var_5868_cast_fp16_5, tensor var_5868_cast_fp16_6, tensor var_5868_cast_fp16_7, tensor var_5868_cast_fp16_8, tensor var_5868_cast_fp16_9, tensor var_5868_cast_fp16_10, tensor var_5868_cast_fp16_11, tensor var_5868_cast_fp16_12, tensor var_5868_cast_fp16_13, tensor var_5868_cast_fp16_14, tensor var_5868_cast_fp16_15, tensor var_5868_cast_fp16_16, tensor var_5868_cast_fp16_17, tensor var_5868_cast_fp16_18, tensor var_5868_cast_fp16_19 = split(axis = var_5868_axis_0, split_sizes = tile_65, x = var_5822_cast_fp16)[name = tensor("op_5868_cast_fp16")]; + tensor aw_841_equation_0 = const()[name = tensor("aw_841_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_841_cast_fp16 = einsum(equation = aw_841_equation_0, values = (var_5847_cast_fp16_0, var_5825_cast_fp16_0))[name = tensor("aw_841_cast_fp16")]; + tensor aw_843_equation_0 = const()[name = tensor("aw_843_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_843_cast_fp16 = einsum(equation = aw_843_equation_0, values = (var_5847_cast_fp16_1, var_5825_cast_fp16_1))[name = tensor("aw_843_cast_fp16")]; + tensor aw_845_equation_0 = const()[name = tensor("aw_845_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_845_cast_fp16 = einsum(equation = aw_845_equation_0, values = (var_5847_cast_fp16_2, var_5825_cast_fp16_2))[name = tensor("aw_845_cast_fp16")]; + tensor aw_847_equation_0 = const()[name = tensor("aw_847_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_847_cast_fp16 = einsum(equation = aw_847_equation_0, values = (var_5847_cast_fp16_3, var_5825_cast_fp16_3))[name = tensor("aw_847_cast_fp16")]; + tensor aw_849_equation_0 = const()[name = tensor("aw_849_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_849_cast_fp16 = einsum(equation = aw_849_equation_0, values = (var_5847_cast_fp16_4, var_5825_cast_fp16_4))[name = tensor("aw_849_cast_fp16")]; + tensor aw_851_equation_0 = const()[name = tensor("aw_851_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_851_cast_fp16 = einsum(equation = aw_851_equation_0, values = (var_5847_cast_fp16_5, var_5825_cast_fp16_5))[name = tensor("aw_851_cast_fp16")]; + tensor aw_853_equation_0 = const()[name = tensor("aw_853_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_853_cast_fp16 = einsum(equation = aw_853_equation_0, values = (var_5847_cast_fp16_6, var_5825_cast_fp16_6))[name = tensor("aw_853_cast_fp16")]; + tensor aw_855_equation_0 = const()[name = tensor("aw_855_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_855_cast_fp16 = einsum(equation = aw_855_equation_0, values = (var_5847_cast_fp16_7, var_5825_cast_fp16_7))[name = tensor("aw_855_cast_fp16")]; + tensor aw_857_equation_0 = const()[name = tensor("aw_857_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_857_cast_fp16 = einsum(equation = aw_857_equation_0, values = (var_5847_cast_fp16_8, var_5825_cast_fp16_8))[name = tensor("aw_857_cast_fp16")]; + tensor aw_859_equation_0 = const()[name = tensor("aw_859_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_859_cast_fp16 = einsum(equation = aw_859_equation_0, values = (var_5847_cast_fp16_9, var_5825_cast_fp16_9))[name = tensor("aw_859_cast_fp16")]; + tensor aw_861_equation_0 = const()[name = tensor("aw_861_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_861_cast_fp16 = einsum(equation = aw_861_equation_0, values = (var_5847_cast_fp16_10, var_5825_cast_fp16_10))[name = tensor("aw_861_cast_fp16")]; + tensor aw_863_equation_0 = const()[name = tensor("aw_863_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_863_cast_fp16 = einsum(equation = aw_863_equation_0, values = (var_5847_cast_fp16_11, var_5825_cast_fp16_11))[name = tensor("aw_863_cast_fp16")]; + tensor aw_865_equation_0 = const()[name = tensor("aw_865_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_865_cast_fp16 = einsum(equation = aw_865_equation_0, values = (var_5847_cast_fp16_12, var_5825_cast_fp16_12))[name = tensor("aw_865_cast_fp16")]; + tensor aw_867_equation_0 = const()[name = tensor("aw_867_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_867_cast_fp16 = einsum(equation = aw_867_equation_0, values = (var_5847_cast_fp16_13, var_5825_cast_fp16_13))[name = tensor("aw_867_cast_fp16")]; + tensor aw_869_equation_0 = const()[name = tensor("aw_869_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_869_cast_fp16 = einsum(equation = aw_869_equation_0, values = (var_5847_cast_fp16_14, var_5825_cast_fp16_14))[name = tensor("aw_869_cast_fp16")]; + tensor aw_871_equation_0 = const()[name = tensor("aw_871_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_871_cast_fp16 = einsum(equation = aw_871_equation_0, values = (var_5847_cast_fp16_15, var_5825_cast_fp16_15))[name = tensor("aw_871_cast_fp16")]; + tensor aw_873_equation_0 = const()[name = tensor("aw_873_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_873_cast_fp16 = einsum(equation = aw_873_equation_0, values = (var_5847_cast_fp16_16, var_5825_cast_fp16_16))[name = tensor("aw_873_cast_fp16")]; + tensor aw_875_equation_0 = const()[name = tensor("aw_875_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_875_cast_fp16 = einsum(equation = aw_875_equation_0, values = (var_5847_cast_fp16_17, var_5825_cast_fp16_17))[name = tensor("aw_875_cast_fp16")]; + tensor aw_877_equation_0 = const()[name = tensor("aw_877_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_877_cast_fp16 = einsum(equation = aw_877_equation_0, values = (var_5847_cast_fp16_18, var_5825_cast_fp16_18))[name = tensor("aw_877_cast_fp16")]; + tensor aw_879_equation_0 = const()[name = tensor("aw_879_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_879_cast_fp16 = einsum(equation = aw_879_equation_0, values = (var_5847_cast_fp16_19, var_5825_cast_fp16_19))[name = tensor("aw_879_cast_fp16")]; + tensor var_5929_cast_fp16 = softmax(axis = var_5773, x = aw_841_cast_fp16)[name = tensor("op_5929_cast_fp16")]; + tensor var_5930_cast_fp16 = softmax(axis = var_5773, x = aw_843_cast_fp16)[name = tensor("op_5930_cast_fp16")]; + tensor var_5931_cast_fp16 = softmax(axis = var_5773, x = aw_845_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5932_cast_fp16 = softmax(axis = var_5773, x = aw_847_cast_fp16)[name = tensor("op_5932_cast_fp16")]; + tensor var_5933_cast_fp16 = softmax(axis = var_5773, x = aw_849_cast_fp16)[name = tensor("op_5933_cast_fp16")]; + tensor var_5934_cast_fp16 = softmax(axis = var_5773, x = aw_851_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = softmax(axis = var_5773, x = aw_853_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor var_5936_cast_fp16 = softmax(axis = var_5773, x = aw_855_cast_fp16)[name = tensor("op_5936_cast_fp16")]; + tensor var_5937_cast_fp16 = softmax(axis = var_5773, x = aw_857_cast_fp16)[name = tensor("op_5937_cast_fp16")]; + tensor var_5938_cast_fp16 = softmax(axis = var_5773, x = aw_859_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = softmax(axis = var_5773, x = aw_861_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor var_5940_cast_fp16 = softmax(axis = var_5773, x = aw_863_cast_fp16)[name = tensor("op_5940_cast_fp16")]; + tensor var_5941_cast_fp16 = softmax(axis = var_5773, x = aw_865_cast_fp16)[name = tensor("op_5941_cast_fp16")]; + tensor var_5942_cast_fp16 = softmax(axis = var_5773, x = aw_867_cast_fp16)[name = tensor("op_5942_cast_fp16")]; + tensor var_5943_cast_fp16 = softmax(axis = var_5773, x = aw_869_cast_fp16)[name = tensor("op_5943_cast_fp16")]; + tensor var_5944_cast_fp16 = softmax(axis = var_5773, x = aw_871_cast_fp16)[name = tensor("op_5944_cast_fp16")]; + tensor var_5945_cast_fp16 = softmax(axis = var_5773, x = aw_873_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor var_5946_cast_fp16 = softmax(axis = var_5773, x = aw_875_cast_fp16)[name = tensor("op_5946_cast_fp16")]; + tensor var_5947_cast_fp16 = softmax(axis = var_5773, x = aw_877_cast_fp16)[name = tensor("op_5947_cast_fp16")]; + tensor var_5948_cast_fp16 = softmax(axis = var_5773, x = aw_879_cast_fp16)[name = tensor("op_5948_cast_fp16")]; + tensor var_5950_equation_0 = const()[name = tensor("op_5950_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5950_cast_fp16 = einsum(equation = var_5950_equation_0, values = (var_5868_cast_fp16_0, var_5929_cast_fp16))[name = tensor("op_5950_cast_fp16")]; + tensor var_5952_equation_0 = const()[name = tensor("op_5952_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5952_cast_fp16 = einsum(equation = var_5952_equation_0, values = (var_5868_cast_fp16_1, var_5930_cast_fp16))[name = tensor("op_5952_cast_fp16")]; + tensor var_5954_equation_0 = const()[name = tensor("op_5954_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5954_cast_fp16 = einsum(equation = var_5954_equation_0, values = (var_5868_cast_fp16_2, var_5931_cast_fp16))[name = tensor("op_5954_cast_fp16")]; + tensor var_5956_equation_0 = const()[name = tensor("op_5956_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5956_cast_fp16 = einsum(equation = var_5956_equation_0, values = (var_5868_cast_fp16_3, var_5932_cast_fp16))[name = tensor("op_5956_cast_fp16")]; + tensor var_5958_equation_0 = const()[name = tensor("op_5958_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5958_cast_fp16 = einsum(equation = var_5958_equation_0, values = (var_5868_cast_fp16_4, var_5933_cast_fp16))[name = tensor("op_5958_cast_fp16")]; + tensor var_5960_equation_0 = const()[name = tensor("op_5960_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5960_cast_fp16 = einsum(equation = var_5960_equation_0, values = (var_5868_cast_fp16_5, var_5934_cast_fp16))[name = tensor("op_5960_cast_fp16")]; + tensor var_5962_equation_0 = const()[name = tensor("op_5962_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5962_cast_fp16 = einsum(equation = var_5962_equation_0, values = (var_5868_cast_fp16_6, var_5935_cast_fp16))[name = tensor("op_5962_cast_fp16")]; + tensor var_5964_equation_0 = const()[name = tensor("op_5964_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5964_cast_fp16 = einsum(equation = var_5964_equation_0, values = (var_5868_cast_fp16_7, var_5936_cast_fp16))[name = tensor("op_5964_cast_fp16")]; + tensor var_5966_equation_0 = const()[name = tensor("op_5966_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5966_cast_fp16 = einsum(equation = var_5966_equation_0, values = (var_5868_cast_fp16_8, var_5937_cast_fp16))[name = tensor("op_5966_cast_fp16")]; + tensor var_5968_equation_0 = const()[name = tensor("op_5968_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5968_cast_fp16 = einsum(equation = var_5968_equation_0, values = (var_5868_cast_fp16_9, var_5938_cast_fp16))[name = tensor("op_5968_cast_fp16")]; + tensor var_5970_equation_0 = const()[name = tensor("op_5970_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5970_cast_fp16 = einsum(equation = var_5970_equation_0, values = (var_5868_cast_fp16_10, var_5939_cast_fp16))[name = tensor("op_5970_cast_fp16")]; + tensor var_5972_equation_0 = const()[name = tensor("op_5972_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5972_cast_fp16 = einsum(equation = var_5972_equation_0, values = (var_5868_cast_fp16_11, var_5940_cast_fp16))[name = tensor("op_5972_cast_fp16")]; + tensor var_5974_equation_0 = const()[name = tensor("op_5974_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5974_cast_fp16 = einsum(equation = var_5974_equation_0, values = (var_5868_cast_fp16_12, var_5941_cast_fp16))[name = tensor("op_5974_cast_fp16")]; + tensor var_5976_equation_0 = const()[name = tensor("op_5976_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5976_cast_fp16 = einsum(equation = var_5976_equation_0, values = (var_5868_cast_fp16_13, var_5942_cast_fp16))[name = tensor("op_5976_cast_fp16")]; + tensor var_5978_equation_0 = const()[name = tensor("op_5978_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5978_cast_fp16 = einsum(equation = var_5978_equation_0, values = (var_5868_cast_fp16_14, var_5943_cast_fp16))[name = tensor("op_5978_cast_fp16")]; + tensor var_5980_equation_0 = const()[name = tensor("op_5980_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5980_cast_fp16 = einsum(equation = var_5980_equation_0, values = (var_5868_cast_fp16_15, var_5944_cast_fp16))[name = tensor("op_5980_cast_fp16")]; + tensor var_5982_equation_0 = const()[name = tensor("op_5982_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5982_cast_fp16 = einsum(equation = var_5982_equation_0, values = (var_5868_cast_fp16_16, var_5945_cast_fp16))[name = tensor("op_5982_cast_fp16")]; + tensor var_5984_equation_0 = const()[name = tensor("op_5984_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5984_cast_fp16 = einsum(equation = var_5984_equation_0, values = (var_5868_cast_fp16_17, var_5946_cast_fp16))[name = tensor("op_5984_cast_fp16")]; + tensor var_5986_equation_0 = const()[name = tensor("op_5986_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5986_cast_fp16 = einsum(equation = var_5986_equation_0, values = (var_5868_cast_fp16_18, var_5947_cast_fp16))[name = tensor("op_5986_cast_fp16")]; + tensor var_5988_equation_0 = const()[name = tensor("op_5988_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5988_cast_fp16 = einsum(equation = var_5988_equation_0, values = (var_5868_cast_fp16_19, var_5948_cast_fp16))[name = tensor("op_5988_cast_fp16")]; + tensor input_215_interleave_0 = const()[name = tensor("input_215_interleave_0"), val = tensor(false)]; + tensor input_215_cast_fp16 = concat(axis = var_5773, interleave = input_215_interleave_0, values = (var_5950_cast_fp16, var_5952_cast_fp16, var_5954_cast_fp16, var_5956_cast_fp16, var_5958_cast_fp16, var_5960_cast_fp16, var_5962_cast_fp16, var_5964_cast_fp16, var_5966_cast_fp16, var_5968_cast_fp16, var_5970_cast_fp16, var_5972_cast_fp16, var_5974_cast_fp16, var_5976_cast_fp16, var_5978_cast_fp16, var_5980_cast_fp16, var_5982_cast_fp16, var_5984_cast_fp16, var_5986_cast_fp16, var_5988_cast_fp16))[name = tensor("input_215_cast_fp16")]; + tensor var_5997_pad_type_0 = const()[name = tensor("op_5997_pad_type_0"), val = tensor("valid")]; + tensor var_5997_strides_0 = const()[name = tensor("op_5997_strides_0"), val = tensor([1, 1])]; + tensor var_5997_pad_0 = const()[name = tensor("op_5997_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5997_dilations_0 = const()[name = tensor("op_5997_dilations_0"), val = tensor([1, 1])]; + tensor var_5997_groups_0 = const()[name = tensor("op_5997_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_out_weight_to_fp16 = const()[name = tensor("blocks_21_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850550272)))]; + tensor blocks_21_attn_out_bias_to_fp16 = const()[name = tensor("blocks_21_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853827136)))]; + tensor var_5997_cast_fp16 = conv(bias = blocks_21_attn_out_bias_to_fp16, dilations = var_5997_dilations_0, groups = var_5997_groups_0, pad = var_5997_pad_0, pad_type = var_5997_pad_type_0, strides = var_5997_strides_0, weight = blocks_21_attn_out_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("op_5997_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = var_5997_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor input_217_axes_0 = const()[name = tensor("input_217_axes_0"), val = tensor([1])]; + tensor input_217_gamma_0_to_fp16 = const()[name = tensor("input_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853829760)))]; + tensor input_217_beta_0_to_fp16 = const()[name = tensor("input_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853832384)))]; + tensor var_6007_to_fp16 = const()[name = tensor("op_6007_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_217_cast_fp16 = layer_norm(axes = input_217_axes_0, beta = input_217_beta_0_to_fp16, epsilon = var_6007_to_fp16, gamma = input_217_gamma_0_to_fp16, x = inputs_87_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("valid")]; + tensor input_219_strides_0 = const()[name = tensor("input_219_strides_0"), val = tensor([1, 1])]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_219_dilations_0 = const()[name = tensor("input_219_dilations_0"), val = tensor([1, 1])]; + tensor input_219_groups_0 = const()[name = tensor("input_219_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853835008)))]; + tensor blocks_21_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(866942272)))]; + tensor input_219_cast_fp16 = conv(bias = blocks_21_mlp_0_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = blocks_21_mlp_0_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_6033_pad_type_0 = const()[name = tensor("op_6033_pad_type_0"), val = tensor("valid")]; + tensor var_6033_strides_0 = const()[name = tensor("op_6033_strides_0"), val = tensor([1, 1])]; + tensor var_6033_pad_0 = const()[name = tensor("op_6033_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6033_dilations_0 = const()[name = tensor("op_6033_dilations_0"), val = tensor([1, 1])]; + tensor var_6033_groups_0 = const()[name = tensor("op_6033_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(866952576)))]; + tensor blocks_21_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880059840)))]; + tensor var_6033_cast_fp16 = conv(bias = blocks_21_mlp_2_bias_to_fp16, dilations = var_6033_dilations_0, groups = var_6033_groups_0, pad = var_6033_pad_0, pad_type = var_6033_pad_type_0, strides = var_6033_strides_0, weight = blocks_21_mlp_2_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("op_6033_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_6033_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(1)]; + tensor input_223_axes_0 = const()[name = tensor("input_223_axes_0"), val = tensor([1])]; + tensor input_223_gamma_0_to_fp16 = const()[name = tensor("input_223_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880062464)))]; + tensor input_223_beta_0_to_fp16 = const()[name = tensor("input_223_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880065088)))]; + tensor var_6058_to_fp16 = const()[name = tensor("op_6058_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_223_cast_fp16 = layer_norm(axes = input_223_axes_0, beta = input_223_beta_0_to_fp16, epsilon = var_6058_to_fp16, gamma = input_223_gamma_0_to_fp16, x = inputs_89_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor q_45_pad_type_0 = const()[name = tensor("q_45_pad_type_0"), val = tensor("valid")]; + tensor q_45_strides_0 = const()[name = tensor("q_45_strides_0"), val = tensor([1, 1])]; + tensor q_45_pad_0 = const()[name = tensor("q_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_45_dilations_0 = const()[name = tensor("q_45_dilations_0"), val = tensor([1, 1])]; + tensor q_45_groups_0 = const()[name = tensor("q_45_groups_0"), val = tensor(1)]; + tensor var_6093_weight_0_to_fp16 = const()[name = tensor("op_6093_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880067712)))]; + tensor var_6093_bias_0_to_fp16 = const()[name = tensor("op_6093_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883344576)))]; + tensor var_6093_cast_fp16 = conv(bias = var_6093_bias_0_to_fp16, dilations = q_45_dilations_0, groups = q_45_groups_0, pad = q_45_pad_0, pad_type = q_45_pad_type_0, strides = q_45_strides_0, weight = var_6093_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("op_6093_cast_fp16")]; + tensor k_45_pad_type_0 = const()[name = tensor("k_45_pad_type_0"), val = tensor("valid")]; + tensor k_45_strides_0 = const()[name = tensor("k_45_strides_0"), val = tensor([1, 1])]; + tensor k_45_pad_0 = const()[name = tensor("k_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_45_dilations_0 = const()[name = tensor("k_45_dilations_0"), val = tensor([1, 1])]; + tensor k_45_groups_0 = const()[name = tensor("k_45_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_key_weight_to_fp16 = const()[name = tensor("blocks_22_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883347200)))]; + tensor k_45_cast_fp16 = conv(dilations = k_45_dilations_0, groups = k_45_groups_0, pad = k_45_pad_0, pad_type = k_45_pad_type_0, strides = k_45_strides_0, weight = blocks_22_attn_key_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor var_6091_pad_type_0 = const()[name = tensor("op_6091_pad_type_0"), val = tensor("valid")]; + tensor var_6091_strides_0 = const()[name = tensor("op_6091_strides_0"), val = tensor([1, 1])]; + tensor var_6091_pad_0 = const()[name = tensor("op_6091_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6091_dilations_0 = const()[name = tensor("op_6091_dilations_0"), val = tensor([1, 1])]; + tensor var_6091_groups_0 = const()[name = tensor("op_6091_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_value_weight_to_fp16 = const()[name = tensor("blocks_22_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886624064)))]; + tensor blocks_22_attn_value_bias_to_fp16 = const()[name = tensor("blocks_22_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889900928)))]; + tensor var_6091_cast_fp16 = conv(bias = blocks_22_attn_value_bias_to_fp16, dilations = var_6091_dilations_0, groups = var_6091_groups_0, pad = var_6091_pad_0, pad_type = var_6091_pad_type_0, strides = var_6091_strides_0, weight = blocks_22_attn_value_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor tile_66 = const()[name = tensor("tile_66"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6094_axis_0 = const()[name = tensor("op_6094_axis_0"), val = tensor(1)]; + tensor var_6094_cast_fp16_0, tensor var_6094_cast_fp16_1, tensor var_6094_cast_fp16_2, tensor var_6094_cast_fp16_3, tensor var_6094_cast_fp16_4, tensor var_6094_cast_fp16_5, tensor var_6094_cast_fp16_6, tensor var_6094_cast_fp16_7, tensor var_6094_cast_fp16_8, tensor var_6094_cast_fp16_9, tensor var_6094_cast_fp16_10, tensor var_6094_cast_fp16_11, tensor var_6094_cast_fp16_12, tensor var_6094_cast_fp16_13, tensor var_6094_cast_fp16_14, tensor var_6094_cast_fp16_15, tensor var_6094_cast_fp16_16, tensor var_6094_cast_fp16_17, tensor var_6094_cast_fp16_18, tensor var_6094_cast_fp16_19 = split(axis = var_6094_axis_0, split_sizes = tile_66, x = var_6093_cast_fp16)[name = tensor("op_6094_cast_fp16")]; + tensor var_6115_perm_0 = const()[name = tensor("op_6115_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_67 = const()[name = tensor("tile_67"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6116_axis_0 = const()[name = tensor("op_6116_axis_0"), val = tensor(3)]; + tensor var_6115_cast_fp16 = transpose(perm = var_6115_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_10")]; + tensor var_6116_cast_fp16_0, tensor var_6116_cast_fp16_1, tensor var_6116_cast_fp16_2, tensor var_6116_cast_fp16_3, tensor var_6116_cast_fp16_4, tensor var_6116_cast_fp16_5, tensor var_6116_cast_fp16_6, tensor var_6116_cast_fp16_7, tensor var_6116_cast_fp16_8, tensor var_6116_cast_fp16_9, tensor var_6116_cast_fp16_10, tensor var_6116_cast_fp16_11, tensor var_6116_cast_fp16_12, tensor var_6116_cast_fp16_13, tensor var_6116_cast_fp16_14, tensor var_6116_cast_fp16_15, tensor var_6116_cast_fp16_16, tensor var_6116_cast_fp16_17, tensor var_6116_cast_fp16_18, tensor var_6116_cast_fp16_19 = split(axis = var_6116_axis_0, split_sizes = tile_67, x = var_6115_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor tile_68 = const()[name = tensor("tile_68"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6137_axis_0 = const()[name = tensor("op_6137_axis_0"), val = tensor(1)]; + tensor var_6137_cast_fp16_0, tensor var_6137_cast_fp16_1, tensor var_6137_cast_fp16_2, tensor var_6137_cast_fp16_3, tensor var_6137_cast_fp16_4, tensor var_6137_cast_fp16_5, tensor var_6137_cast_fp16_6, tensor var_6137_cast_fp16_7, tensor var_6137_cast_fp16_8, tensor var_6137_cast_fp16_9, tensor var_6137_cast_fp16_10, tensor var_6137_cast_fp16_11, tensor var_6137_cast_fp16_12, tensor var_6137_cast_fp16_13, tensor var_6137_cast_fp16_14, tensor var_6137_cast_fp16_15, tensor var_6137_cast_fp16_16, tensor var_6137_cast_fp16_17, tensor var_6137_cast_fp16_18, tensor var_6137_cast_fp16_19 = split(axis = var_6137_axis_0, split_sizes = tile_68, x = var_6091_cast_fp16)[name = tensor("op_6137_cast_fp16")]; + tensor aw_881_equation_0 = const()[name = tensor("aw_881_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_881_cast_fp16 = einsum(equation = aw_881_equation_0, values = (var_6116_cast_fp16_0, var_6094_cast_fp16_0))[name = tensor("aw_881_cast_fp16")]; + tensor aw_883_equation_0 = const()[name = tensor("aw_883_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_883_cast_fp16 = einsum(equation = aw_883_equation_0, values = (var_6116_cast_fp16_1, var_6094_cast_fp16_1))[name = tensor("aw_883_cast_fp16")]; + tensor aw_885_equation_0 = const()[name = tensor("aw_885_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_885_cast_fp16 = einsum(equation = aw_885_equation_0, values = (var_6116_cast_fp16_2, var_6094_cast_fp16_2))[name = tensor("aw_885_cast_fp16")]; + tensor aw_887_equation_0 = const()[name = tensor("aw_887_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_887_cast_fp16 = einsum(equation = aw_887_equation_0, values = (var_6116_cast_fp16_3, var_6094_cast_fp16_3))[name = tensor("aw_887_cast_fp16")]; + tensor aw_889_equation_0 = const()[name = tensor("aw_889_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_889_cast_fp16 = einsum(equation = aw_889_equation_0, values = (var_6116_cast_fp16_4, var_6094_cast_fp16_4))[name = tensor("aw_889_cast_fp16")]; + tensor aw_891_equation_0 = const()[name = tensor("aw_891_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_891_cast_fp16 = einsum(equation = aw_891_equation_0, values = (var_6116_cast_fp16_5, var_6094_cast_fp16_5))[name = tensor("aw_891_cast_fp16")]; + tensor aw_893_equation_0 = const()[name = tensor("aw_893_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_893_cast_fp16 = einsum(equation = aw_893_equation_0, values = (var_6116_cast_fp16_6, var_6094_cast_fp16_6))[name = tensor("aw_893_cast_fp16")]; + tensor aw_895_equation_0 = const()[name = tensor("aw_895_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_895_cast_fp16 = einsum(equation = aw_895_equation_0, values = (var_6116_cast_fp16_7, var_6094_cast_fp16_7))[name = tensor("aw_895_cast_fp16")]; + tensor aw_897_equation_0 = const()[name = tensor("aw_897_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_897_cast_fp16 = einsum(equation = aw_897_equation_0, values = (var_6116_cast_fp16_8, var_6094_cast_fp16_8))[name = tensor("aw_897_cast_fp16")]; + tensor aw_899_equation_0 = const()[name = tensor("aw_899_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_899_cast_fp16 = einsum(equation = aw_899_equation_0, values = (var_6116_cast_fp16_9, var_6094_cast_fp16_9))[name = tensor("aw_899_cast_fp16")]; + tensor aw_901_equation_0 = const()[name = tensor("aw_901_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_901_cast_fp16 = einsum(equation = aw_901_equation_0, values = (var_6116_cast_fp16_10, var_6094_cast_fp16_10))[name = tensor("aw_901_cast_fp16")]; + tensor aw_903_equation_0 = const()[name = tensor("aw_903_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_903_cast_fp16 = einsum(equation = aw_903_equation_0, values = (var_6116_cast_fp16_11, var_6094_cast_fp16_11))[name = tensor("aw_903_cast_fp16")]; + tensor aw_905_equation_0 = const()[name = tensor("aw_905_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_905_cast_fp16 = einsum(equation = aw_905_equation_0, values = (var_6116_cast_fp16_12, var_6094_cast_fp16_12))[name = tensor("aw_905_cast_fp16")]; + tensor aw_907_equation_0 = const()[name = tensor("aw_907_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_907_cast_fp16 = einsum(equation = aw_907_equation_0, values = (var_6116_cast_fp16_13, var_6094_cast_fp16_13))[name = tensor("aw_907_cast_fp16")]; + tensor aw_909_equation_0 = const()[name = tensor("aw_909_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_909_cast_fp16 = einsum(equation = aw_909_equation_0, values = (var_6116_cast_fp16_14, var_6094_cast_fp16_14))[name = tensor("aw_909_cast_fp16")]; + tensor aw_911_equation_0 = const()[name = tensor("aw_911_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_911_cast_fp16 = einsum(equation = aw_911_equation_0, values = (var_6116_cast_fp16_15, var_6094_cast_fp16_15))[name = tensor("aw_911_cast_fp16")]; + tensor aw_913_equation_0 = const()[name = tensor("aw_913_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_913_cast_fp16 = einsum(equation = aw_913_equation_0, values = (var_6116_cast_fp16_16, var_6094_cast_fp16_16))[name = tensor("aw_913_cast_fp16")]; + tensor aw_915_equation_0 = const()[name = tensor("aw_915_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_915_cast_fp16 = einsum(equation = aw_915_equation_0, values = (var_6116_cast_fp16_17, var_6094_cast_fp16_17))[name = tensor("aw_915_cast_fp16")]; + tensor aw_917_equation_0 = const()[name = tensor("aw_917_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_917_cast_fp16 = einsum(equation = aw_917_equation_0, values = (var_6116_cast_fp16_18, var_6094_cast_fp16_18))[name = tensor("aw_917_cast_fp16")]; + tensor aw_919_equation_0 = const()[name = tensor("aw_919_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_919_cast_fp16 = einsum(equation = aw_919_equation_0, values = (var_6116_cast_fp16_19, var_6094_cast_fp16_19))[name = tensor("aw_919_cast_fp16")]; + tensor var_6198_cast_fp16 = softmax(axis = var_6042, x = aw_881_cast_fp16)[name = tensor("op_6198_cast_fp16")]; + tensor var_6199_cast_fp16 = softmax(axis = var_6042, x = aw_883_cast_fp16)[name = tensor("op_6199_cast_fp16")]; + tensor var_6200_cast_fp16 = softmax(axis = var_6042, x = aw_885_cast_fp16)[name = tensor("op_6200_cast_fp16")]; + tensor var_6201_cast_fp16 = softmax(axis = var_6042, x = aw_887_cast_fp16)[name = tensor("op_6201_cast_fp16")]; + tensor var_6202_cast_fp16 = softmax(axis = var_6042, x = aw_889_cast_fp16)[name = tensor("op_6202_cast_fp16")]; + tensor var_6203_cast_fp16 = softmax(axis = var_6042, x = aw_891_cast_fp16)[name = tensor("op_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = softmax(axis = var_6042, x = aw_893_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_cast_fp16 = softmax(axis = var_6042, x = aw_895_cast_fp16)[name = tensor("op_6205_cast_fp16")]; + tensor var_6206_cast_fp16 = softmax(axis = var_6042, x = aw_897_cast_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor var_6207_cast_fp16 = softmax(axis = var_6042, x = aw_899_cast_fp16)[name = tensor("op_6207_cast_fp16")]; + tensor var_6208_cast_fp16 = softmax(axis = var_6042, x = aw_901_cast_fp16)[name = tensor("op_6208_cast_fp16")]; + tensor var_6209_cast_fp16 = softmax(axis = var_6042, x = aw_903_cast_fp16)[name = tensor("op_6209_cast_fp16")]; + tensor var_6210_cast_fp16 = softmax(axis = var_6042, x = aw_905_cast_fp16)[name = tensor("op_6210_cast_fp16")]; + tensor var_6211_cast_fp16 = softmax(axis = var_6042, x = aw_907_cast_fp16)[name = tensor("op_6211_cast_fp16")]; + tensor var_6212_cast_fp16 = softmax(axis = var_6042, x = aw_909_cast_fp16)[name = tensor("op_6212_cast_fp16")]; + tensor var_6213_cast_fp16 = softmax(axis = var_6042, x = aw_911_cast_fp16)[name = tensor("op_6213_cast_fp16")]; + tensor var_6214_cast_fp16 = softmax(axis = var_6042, x = aw_913_cast_fp16)[name = tensor("op_6214_cast_fp16")]; + tensor var_6215_cast_fp16 = softmax(axis = var_6042, x = aw_915_cast_fp16)[name = tensor("op_6215_cast_fp16")]; + tensor var_6216_cast_fp16 = softmax(axis = var_6042, x = aw_917_cast_fp16)[name = tensor("op_6216_cast_fp16")]; + tensor var_6217_cast_fp16 = softmax(axis = var_6042, x = aw_919_cast_fp16)[name = tensor("op_6217_cast_fp16")]; + tensor var_6219_equation_0 = const()[name = tensor("op_6219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6219_cast_fp16 = einsum(equation = var_6219_equation_0, values = (var_6137_cast_fp16_0, var_6198_cast_fp16))[name = tensor("op_6219_cast_fp16")]; + tensor var_6221_equation_0 = const()[name = tensor("op_6221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6221_cast_fp16 = einsum(equation = var_6221_equation_0, values = (var_6137_cast_fp16_1, var_6199_cast_fp16))[name = tensor("op_6221_cast_fp16")]; + tensor var_6223_equation_0 = const()[name = tensor("op_6223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6223_cast_fp16 = einsum(equation = var_6223_equation_0, values = (var_6137_cast_fp16_2, var_6200_cast_fp16))[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_equation_0 = const()[name = tensor("op_6225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6225_cast_fp16 = einsum(equation = var_6225_equation_0, values = (var_6137_cast_fp16_3, var_6201_cast_fp16))[name = tensor("op_6225_cast_fp16")]; + tensor var_6227_equation_0 = const()[name = tensor("op_6227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6227_cast_fp16 = einsum(equation = var_6227_equation_0, values = (var_6137_cast_fp16_4, var_6202_cast_fp16))[name = tensor("op_6227_cast_fp16")]; + tensor var_6229_equation_0 = const()[name = tensor("op_6229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6229_cast_fp16 = einsum(equation = var_6229_equation_0, values = (var_6137_cast_fp16_5, var_6203_cast_fp16))[name = tensor("op_6229_cast_fp16")]; + tensor var_6231_equation_0 = const()[name = tensor("op_6231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6231_cast_fp16 = einsum(equation = var_6231_equation_0, values = (var_6137_cast_fp16_6, var_6204_cast_fp16))[name = tensor("op_6231_cast_fp16")]; + tensor var_6233_equation_0 = const()[name = tensor("op_6233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6233_cast_fp16 = einsum(equation = var_6233_equation_0, values = (var_6137_cast_fp16_7, var_6205_cast_fp16))[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_equation_0 = const()[name = tensor("op_6235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6235_cast_fp16 = einsum(equation = var_6235_equation_0, values = (var_6137_cast_fp16_8, var_6206_cast_fp16))[name = tensor("op_6235_cast_fp16")]; + tensor var_6237_equation_0 = const()[name = tensor("op_6237_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6237_cast_fp16 = einsum(equation = var_6237_equation_0, values = (var_6137_cast_fp16_9, var_6207_cast_fp16))[name = tensor("op_6237_cast_fp16")]; + tensor var_6239_equation_0 = const()[name = tensor("op_6239_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6239_cast_fp16 = einsum(equation = var_6239_equation_0, values = (var_6137_cast_fp16_10, var_6208_cast_fp16))[name = tensor("op_6239_cast_fp16")]; + tensor var_6241_equation_0 = const()[name = tensor("op_6241_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6241_cast_fp16 = einsum(equation = var_6241_equation_0, values = (var_6137_cast_fp16_11, var_6209_cast_fp16))[name = tensor("op_6241_cast_fp16")]; + tensor var_6243_equation_0 = const()[name = tensor("op_6243_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6243_cast_fp16 = einsum(equation = var_6243_equation_0, values = (var_6137_cast_fp16_12, var_6210_cast_fp16))[name = tensor("op_6243_cast_fp16")]; + tensor var_6245_equation_0 = const()[name = tensor("op_6245_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6245_cast_fp16 = einsum(equation = var_6245_equation_0, values = (var_6137_cast_fp16_13, var_6211_cast_fp16))[name = tensor("op_6245_cast_fp16")]; + tensor var_6247_equation_0 = const()[name = tensor("op_6247_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6247_cast_fp16 = einsum(equation = var_6247_equation_0, values = (var_6137_cast_fp16_14, var_6212_cast_fp16))[name = tensor("op_6247_cast_fp16")]; + tensor var_6249_equation_0 = const()[name = tensor("op_6249_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6249_cast_fp16 = einsum(equation = var_6249_equation_0, values = (var_6137_cast_fp16_15, var_6213_cast_fp16))[name = tensor("op_6249_cast_fp16")]; + tensor var_6251_equation_0 = const()[name = tensor("op_6251_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6251_cast_fp16 = einsum(equation = var_6251_equation_0, values = (var_6137_cast_fp16_16, var_6214_cast_fp16))[name = tensor("op_6251_cast_fp16")]; + tensor var_6253_equation_0 = const()[name = tensor("op_6253_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6253_cast_fp16 = einsum(equation = var_6253_equation_0, values = (var_6137_cast_fp16_17, var_6215_cast_fp16))[name = tensor("op_6253_cast_fp16")]; + tensor var_6255_equation_0 = const()[name = tensor("op_6255_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6255_cast_fp16 = einsum(equation = var_6255_equation_0, values = (var_6137_cast_fp16_18, var_6216_cast_fp16))[name = tensor("op_6255_cast_fp16")]; + tensor var_6257_equation_0 = const()[name = tensor("op_6257_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6257_cast_fp16 = einsum(equation = var_6257_equation_0, values = (var_6137_cast_fp16_19, var_6217_cast_fp16))[name = tensor("op_6257_cast_fp16")]; + tensor input_225_interleave_0 = const()[name = tensor("input_225_interleave_0"), val = tensor(false)]; + tensor input_225_cast_fp16 = concat(axis = var_6042, interleave = input_225_interleave_0, values = (var_6219_cast_fp16, var_6221_cast_fp16, var_6223_cast_fp16, var_6225_cast_fp16, var_6227_cast_fp16, var_6229_cast_fp16, var_6231_cast_fp16, var_6233_cast_fp16, var_6235_cast_fp16, var_6237_cast_fp16, var_6239_cast_fp16, var_6241_cast_fp16, var_6243_cast_fp16, var_6245_cast_fp16, var_6247_cast_fp16, var_6249_cast_fp16, var_6251_cast_fp16, var_6253_cast_fp16, var_6255_cast_fp16, var_6257_cast_fp16))[name = tensor("input_225_cast_fp16")]; + tensor var_6266_pad_type_0 = const()[name = tensor("op_6266_pad_type_0"), val = tensor("valid")]; + tensor var_6266_strides_0 = const()[name = tensor("op_6266_strides_0"), val = tensor([1, 1])]; + tensor var_6266_pad_0 = const()[name = tensor("op_6266_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6266_dilations_0 = const()[name = tensor("op_6266_dilations_0"), val = tensor([1, 1])]; + tensor var_6266_groups_0 = const()[name = tensor("op_6266_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_out_weight_to_fp16 = const()[name = tensor("blocks_22_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889903552)))]; + tensor blocks_22_attn_out_bias_to_fp16 = const()[name = tensor("blocks_22_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893180416)))]; + tensor var_6266_cast_fp16 = conv(bias = blocks_22_attn_out_bias_to_fp16, dilations = var_6266_dilations_0, groups = var_6266_groups_0, pad = var_6266_pad_0, pad_type = var_6266_pad_type_0, strides = var_6266_strides_0, weight = blocks_22_attn_out_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("op_6266_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = var_6266_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([1])]; + tensor input_227_gamma_0_to_fp16 = const()[name = tensor("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893183040)))]; + tensor input_227_beta_0_to_fp16 = const()[name = tensor("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893185664)))]; + tensor var_6276_to_fp16 = const()[name = tensor("op_6276_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = input_227_beta_0_to_fp16, epsilon = var_6276_to_fp16, gamma = input_227_gamma_0_to_fp16, x = inputs_91_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("valid")]; + tensor input_229_strides_0 = const()[name = tensor("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = tensor("input_229_dilations_0"), val = tensor([1, 1])]; + tensor input_229_groups_0 = const()[name = tensor("input_229_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893188288)))]; + tensor blocks_22_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906295552)))]; + tensor input_229_cast_fp16 = conv(bias = blocks_22_mlp_0_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = blocks_22_mlp_0_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_6302_pad_type_0 = const()[name = tensor("op_6302_pad_type_0"), val = tensor("valid")]; + tensor var_6302_strides_0 = const()[name = tensor("op_6302_strides_0"), val = tensor([1, 1])]; + tensor var_6302_pad_0 = const()[name = tensor("op_6302_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6302_dilations_0 = const()[name = tensor("op_6302_dilations_0"), val = tensor([1, 1])]; + tensor var_6302_groups_0 = const()[name = tensor("op_6302_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906305856)))]; + tensor blocks_22_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919413120)))]; + tensor var_6302_cast_fp16 = conv(bias = blocks_22_mlp_2_bias_to_fp16, dilations = var_6302_dilations_0, groups = var_6302_groups_0, pad = var_6302_pad_0, pad_type = var_6302_pad_type_0, strides = var_6302_strides_0, weight = blocks_22_mlp_2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("op_6302_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_6302_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_6311 = const()[name = tensor("op_6311"), val = tensor(1)]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([1])]; + tensor input_233_gamma_0_to_fp16 = const()[name = tensor("input_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919415744)))]; + tensor input_233_beta_0_to_fp16 = const()[name = tensor("input_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919418368)))]; + tensor var_6327_to_fp16 = const()[name = tensor("op_6327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = input_233_beta_0_to_fp16, epsilon = var_6327_to_fp16, gamma = input_233_gamma_0_to_fp16, x = inputs_93_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor q_47_pad_type_0 = const()[name = tensor("q_47_pad_type_0"), val = tensor("valid")]; + tensor q_47_strides_0 = const()[name = tensor("q_47_strides_0"), val = tensor([1, 1])]; + tensor q_47_pad_0 = const()[name = tensor("q_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_47_dilations_0 = const()[name = tensor("q_47_dilations_0"), val = tensor([1, 1])]; + tensor q_47_groups_0 = const()[name = tensor("q_47_groups_0"), val = tensor(1)]; + tensor var_6362_weight_0_to_fp16 = const()[name = tensor("op_6362_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919420992)))]; + tensor var_6362_bias_0_to_fp16 = const()[name = tensor("op_6362_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922697856)))]; + tensor var_6362_cast_fp16 = conv(bias = var_6362_bias_0_to_fp16, dilations = q_47_dilations_0, groups = q_47_groups_0, pad = q_47_pad_0, pad_type = q_47_pad_type_0, strides = q_47_strides_0, weight = var_6362_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("op_6362_cast_fp16")]; + tensor k_47_pad_type_0 = const()[name = tensor("k_47_pad_type_0"), val = tensor("valid")]; + tensor k_47_strides_0 = const()[name = tensor("k_47_strides_0"), val = tensor([1, 1])]; + tensor k_47_pad_0 = const()[name = tensor("k_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_47_dilations_0 = const()[name = tensor("k_47_dilations_0"), val = tensor([1, 1])]; + tensor k_47_groups_0 = const()[name = tensor("k_47_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_key_weight_to_fp16 = const()[name = tensor("blocks_23_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922700480)))]; + tensor k_47_cast_fp16 = conv(dilations = k_47_dilations_0, groups = k_47_groups_0, pad = k_47_pad_0, pad_type = k_47_pad_type_0, strides = k_47_strides_0, weight = blocks_23_attn_key_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor var_6360_pad_type_0 = const()[name = tensor("op_6360_pad_type_0"), val = tensor("valid")]; + tensor var_6360_strides_0 = const()[name = tensor("op_6360_strides_0"), val = tensor([1, 1])]; + tensor var_6360_pad_0 = const()[name = tensor("op_6360_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6360_dilations_0 = const()[name = tensor("op_6360_dilations_0"), val = tensor([1, 1])]; + tensor var_6360_groups_0 = const()[name = tensor("op_6360_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_value_weight_to_fp16 = const()[name = tensor("blocks_23_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(925977344)))]; + tensor blocks_23_attn_value_bias_to_fp16 = const()[name = tensor("blocks_23_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929254208)))]; + tensor var_6360_cast_fp16 = conv(bias = blocks_23_attn_value_bias_to_fp16, dilations = var_6360_dilations_0, groups = var_6360_groups_0, pad = var_6360_pad_0, pad_type = var_6360_pad_type_0, strides = var_6360_strides_0, weight = blocks_23_attn_value_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("op_6360_cast_fp16")]; + tensor tile_69 = const()[name = tensor("tile_69"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6363_axis_0 = const()[name = tensor("op_6363_axis_0"), val = tensor(1)]; + tensor var_6363_cast_fp16_0, tensor var_6363_cast_fp16_1, tensor var_6363_cast_fp16_2, tensor var_6363_cast_fp16_3, tensor var_6363_cast_fp16_4, tensor var_6363_cast_fp16_5, tensor var_6363_cast_fp16_6, tensor var_6363_cast_fp16_7, tensor var_6363_cast_fp16_8, tensor var_6363_cast_fp16_9, tensor var_6363_cast_fp16_10, tensor var_6363_cast_fp16_11, tensor var_6363_cast_fp16_12, tensor var_6363_cast_fp16_13, tensor var_6363_cast_fp16_14, tensor var_6363_cast_fp16_15, tensor var_6363_cast_fp16_16, tensor var_6363_cast_fp16_17, tensor var_6363_cast_fp16_18, tensor var_6363_cast_fp16_19 = split(axis = var_6363_axis_0, split_sizes = tile_69, x = var_6362_cast_fp16)[name = tensor("op_6363_cast_fp16")]; + tensor var_6384_perm_0 = const()[name = tensor("op_6384_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_70 = const()[name = tensor("tile_70"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6385_axis_0 = const()[name = tensor("op_6385_axis_0"), val = tensor(3)]; + tensor var_6384_cast_fp16 = transpose(perm = var_6384_perm_0, x = k_47_cast_fp16)[name = tensor("transpose_9")]; + tensor var_6385_cast_fp16_0, tensor var_6385_cast_fp16_1, tensor var_6385_cast_fp16_2, tensor var_6385_cast_fp16_3, tensor var_6385_cast_fp16_4, tensor var_6385_cast_fp16_5, tensor var_6385_cast_fp16_6, tensor var_6385_cast_fp16_7, tensor var_6385_cast_fp16_8, tensor var_6385_cast_fp16_9, tensor var_6385_cast_fp16_10, tensor var_6385_cast_fp16_11, tensor var_6385_cast_fp16_12, tensor var_6385_cast_fp16_13, tensor var_6385_cast_fp16_14, tensor var_6385_cast_fp16_15, tensor var_6385_cast_fp16_16, tensor var_6385_cast_fp16_17, tensor var_6385_cast_fp16_18, tensor var_6385_cast_fp16_19 = split(axis = var_6385_axis_0, split_sizes = tile_70, x = var_6384_cast_fp16)[name = tensor("op_6385_cast_fp16")]; + tensor tile_71 = const()[name = tensor("tile_71"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6406_axis_0 = const()[name = tensor("op_6406_axis_0"), val = tensor(1)]; + tensor var_6406_cast_fp16_0, tensor var_6406_cast_fp16_1, tensor var_6406_cast_fp16_2, tensor var_6406_cast_fp16_3, tensor var_6406_cast_fp16_4, tensor var_6406_cast_fp16_5, tensor var_6406_cast_fp16_6, tensor var_6406_cast_fp16_7, tensor var_6406_cast_fp16_8, tensor var_6406_cast_fp16_9, tensor var_6406_cast_fp16_10, tensor var_6406_cast_fp16_11, tensor var_6406_cast_fp16_12, tensor var_6406_cast_fp16_13, tensor var_6406_cast_fp16_14, tensor var_6406_cast_fp16_15, tensor var_6406_cast_fp16_16, tensor var_6406_cast_fp16_17, tensor var_6406_cast_fp16_18, tensor var_6406_cast_fp16_19 = split(axis = var_6406_axis_0, split_sizes = tile_71, x = var_6360_cast_fp16)[name = tensor("op_6406_cast_fp16")]; + tensor aw_921_equation_0 = const()[name = tensor("aw_921_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_921_cast_fp16 = einsum(equation = aw_921_equation_0, values = (var_6385_cast_fp16_0, var_6363_cast_fp16_0))[name = tensor("aw_921_cast_fp16")]; + tensor aw_923_equation_0 = const()[name = tensor("aw_923_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_923_cast_fp16 = einsum(equation = aw_923_equation_0, values = (var_6385_cast_fp16_1, var_6363_cast_fp16_1))[name = tensor("aw_923_cast_fp16")]; + tensor aw_925_equation_0 = const()[name = tensor("aw_925_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_925_cast_fp16 = einsum(equation = aw_925_equation_0, values = (var_6385_cast_fp16_2, var_6363_cast_fp16_2))[name = tensor("aw_925_cast_fp16")]; + tensor aw_927_equation_0 = const()[name = tensor("aw_927_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_927_cast_fp16 = einsum(equation = aw_927_equation_0, values = (var_6385_cast_fp16_3, var_6363_cast_fp16_3))[name = tensor("aw_927_cast_fp16")]; + tensor aw_929_equation_0 = const()[name = tensor("aw_929_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_929_cast_fp16 = einsum(equation = aw_929_equation_0, values = (var_6385_cast_fp16_4, var_6363_cast_fp16_4))[name = tensor("aw_929_cast_fp16")]; + tensor aw_931_equation_0 = const()[name = tensor("aw_931_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_931_cast_fp16 = einsum(equation = aw_931_equation_0, values = (var_6385_cast_fp16_5, var_6363_cast_fp16_5))[name = tensor("aw_931_cast_fp16")]; + tensor aw_933_equation_0 = const()[name = tensor("aw_933_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_933_cast_fp16 = einsum(equation = aw_933_equation_0, values = (var_6385_cast_fp16_6, var_6363_cast_fp16_6))[name = tensor("aw_933_cast_fp16")]; + tensor aw_935_equation_0 = const()[name = tensor("aw_935_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_935_cast_fp16 = einsum(equation = aw_935_equation_0, values = (var_6385_cast_fp16_7, var_6363_cast_fp16_7))[name = tensor("aw_935_cast_fp16")]; + tensor aw_937_equation_0 = const()[name = tensor("aw_937_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_937_cast_fp16 = einsum(equation = aw_937_equation_0, values = (var_6385_cast_fp16_8, var_6363_cast_fp16_8))[name = tensor("aw_937_cast_fp16")]; + tensor aw_939_equation_0 = const()[name = tensor("aw_939_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_939_cast_fp16 = einsum(equation = aw_939_equation_0, values = (var_6385_cast_fp16_9, var_6363_cast_fp16_9))[name = tensor("aw_939_cast_fp16")]; + tensor aw_941_equation_0 = const()[name = tensor("aw_941_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_941_cast_fp16 = einsum(equation = aw_941_equation_0, values = (var_6385_cast_fp16_10, var_6363_cast_fp16_10))[name = tensor("aw_941_cast_fp16")]; + tensor aw_943_equation_0 = const()[name = tensor("aw_943_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_943_cast_fp16 = einsum(equation = aw_943_equation_0, values = (var_6385_cast_fp16_11, var_6363_cast_fp16_11))[name = tensor("aw_943_cast_fp16")]; + tensor aw_945_equation_0 = const()[name = tensor("aw_945_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_945_cast_fp16 = einsum(equation = aw_945_equation_0, values = (var_6385_cast_fp16_12, var_6363_cast_fp16_12))[name = tensor("aw_945_cast_fp16")]; + tensor aw_947_equation_0 = const()[name = tensor("aw_947_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_947_cast_fp16 = einsum(equation = aw_947_equation_0, values = (var_6385_cast_fp16_13, var_6363_cast_fp16_13))[name = tensor("aw_947_cast_fp16")]; + tensor aw_949_equation_0 = const()[name = tensor("aw_949_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_949_cast_fp16 = einsum(equation = aw_949_equation_0, values = (var_6385_cast_fp16_14, var_6363_cast_fp16_14))[name = tensor("aw_949_cast_fp16")]; + tensor aw_951_equation_0 = const()[name = tensor("aw_951_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_951_cast_fp16 = einsum(equation = aw_951_equation_0, values = (var_6385_cast_fp16_15, var_6363_cast_fp16_15))[name = tensor("aw_951_cast_fp16")]; + tensor aw_953_equation_0 = const()[name = tensor("aw_953_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_953_cast_fp16 = einsum(equation = aw_953_equation_0, values = (var_6385_cast_fp16_16, var_6363_cast_fp16_16))[name = tensor("aw_953_cast_fp16")]; + tensor aw_955_equation_0 = const()[name = tensor("aw_955_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_955_cast_fp16 = einsum(equation = aw_955_equation_0, values = (var_6385_cast_fp16_17, var_6363_cast_fp16_17))[name = tensor("aw_955_cast_fp16")]; + tensor aw_957_equation_0 = const()[name = tensor("aw_957_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_957_cast_fp16 = einsum(equation = aw_957_equation_0, values = (var_6385_cast_fp16_18, var_6363_cast_fp16_18))[name = tensor("aw_957_cast_fp16")]; + tensor aw_959_equation_0 = const()[name = tensor("aw_959_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_959_cast_fp16 = einsum(equation = aw_959_equation_0, values = (var_6385_cast_fp16_19, var_6363_cast_fp16_19))[name = tensor("aw_959_cast_fp16")]; + tensor var_6467_cast_fp16 = softmax(axis = var_6311, x = aw_921_cast_fp16)[name = tensor("op_6467_cast_fp16")]; + tensor var_6468_cast_fp16 = softmax(axis = var_6311, x = aw_923_cast_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6469_cast_fp16 = softmax(axis = var_6311, x = aw_925_cast_fp16)[name = tensor("op_6469_cast_fp16")]; + tensor var_6470_cast_fp16 = softmax(axis = var_6311, x = aw_927_cast_fp16)[name = tensor("op_6470_cast_fp16")]; + tensor var_6471_cast_fp16 = softmax(axis = var_6311, x = aw_929_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6472_cast_fp16 = softmax(axis = var_6311, x = aw_931_cast_fp16)[name = tensor("op_6472_cast_fp16")]; + tensor var_6473_cast_fp16 = softmax(axis = var_6311, x = aw_933_cast_fp16)[name = tensor("op_6473_cast_fp16")]; + tensor var_6474_cast_fp16 = softmax(axis = var_6311, x = aw_935_cast_fp16)[name = tensor("op_6474_cast_fp16")]; + tensor var_6475_cast_fp16 = softmax(axis = var_6311, x = aw_937_cast_fp16)[name = tensor("op_6475_cast_fp16")]; + tensor var_6476_cast_fp16 = softmax(axis = var_6311, x = aw_939_cast_fp16)[name = tensor("op_6476_cast_fp16")]; + tensor var_6477_cast_fp16 = softmax(axis = var_6311, x = aw_941_cast_fp16)[name = tensor("op_6477_cast_fp16")]; + tensor var_6478_cast_fp16 = softmax(axis = var_6311, x = aw_943_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6479_cast_fp16 = softmax(axis = var_6311, x = aw_945_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor var_6480_cast_fp16 = softmax(axis = var_6311, x = aw_947_cast_fp16)[name = tensor("op_6480_cast_fp16")]; + tensor var_6481_cast_fp16 = softmax(axis = var_6311, x = aw_949_cast_fp16)[name = tensor("op_6481_cast_fp16")]; + tensor var_6482_cast_fp16 = softmax(axis = var_6311, x = aw_951_cast_fp16)[name = tensor("op_6482_cast_fp16")]; + tensor var_6483_cast_fp16 = softmax(axis = var_6311, x = aw_953_cast_fp16)[name = tensor("op_6483_cast_fp16")]; + tensor var_6484_cast_fp16 = softmax(axis = var_6311, x = aw_955_cast_fp16)[name = tensor("op_6484_cast_fp16")]; + tensor var_6485_cast_fp16 = softmax(axis = var_6311, x = aw_957_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor var_6486_cast_fp16 = softmax(axis = var_6311, x = aw_959_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6488_equation_0 = const()[name = tensor("op_6488_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6488_cast_fp16 = einsum(equation = var_6488_equation_0, values = (var_6406_cast_fp16_0, var_6467_cast_fp16))[name = tensor("op_6488_cast_fp16")]; + tensor var_6490_equation_0 = const()[name = tensor("op_6490_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6490_cast_fp16 = einsum(equation = var_6490_equation_0, values = (var_6406_cast_fp16_1, var_6468_cast_fp16))[name = tensor("op_6490_cast_fp16")]; + tensor var_6492_equation_0 = const()[name = tensor("op_6492_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6492_cast_fp16 = einsum(equation = var_6492_equation_0, values = (var_6406_cast_fp16_2, var_6469_cast_fp16))[name = tensor("op_6492_cast_fp16")]; + tensor var_6494_equation_0 = const()[name = tensor("op_6494_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6494_cast_fp16 = einsum(equation = var_6494_equation_0, values = (var_6406_cast_fp16_3, var_6470_cast_fp16))[name = tensor("op_6494_cast_fp16")]; + tensor var_6496_equation_0 = const()[name = tensor("op_6496_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6496_cast_fp16 = einsum(equation = var_6496_equation_0, values = (var_6406_cast_fp16_4, var_6471_cast_fp16))[name = tensor("op_6496_cast_fp16")]; + tensor var_6498_equation_0 = const()[name = tensor("op_6498_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6498_cast_fp16 = einsum(equation = var_6498_equation_0, values = (var_6406_cast_fp16_5, var_6472_cast_fp16))[name = tensor("op_6498_cast_fp16")]; + tensor var_6500_equation_0 = const()[name = tensor("op_6500_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6500_cast_fp16 = einsum(equation = var_6500_equation_0, values = (var_6406_cast_fp16_6, var_6473_cast_fp16))[name = tensor("op_6500_cast_fp16")]; + tensor var_6502_equation_0 = const()[name = tensor("op_6502_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6502_cast_fp16 = einsum(equation = var_6502_equation_0, values = (var_6406_cast_fp16_7, var_6474_cast_fp16))[name = tensor("op_6502_cast_fp16")]; + tensor var_6504_equation_0 = const()[name = tensor("op_6504_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6504_cast_fp16 = einsum(equation = var_6504_equation_0, values = (var_6406_cast_fp16_8, var_6475_cast_fp16))[name = tensor("op_6504_cast_fp16")]; + tensor var_6506_equation_0 = const()[name = tensor("op_6506_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6506_cast_fp16 = einsum(equation = var_6506_equation_0, values = (var_6406_cast_fp16_9, var_6476_cast_fp16))[name = tensor("op_6506_cast_fp16")]; + tensor var_6508_equation_0 = const()[name = tensor("op_6508_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6508_cast_fp16 = einsum(equation = var_6508_equation_0, values = (var_6406_cast_fp16_10, var_6477_cast_fp16))[name = tensor("op_6508_cast_fp16")]; + tensor var_6510_equation_0 = const()[name = tensor("op_6510_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6510_cast_fp16 = einsum(equation = var_6510_equation_0, values = (var_6406_cast_fp16_11, var_6478_cast_fp16))[name = tensor("op_6510_cast_fp16")]; + tensor var_6512_equation_0 = const()[name = tensor("op_6512_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6512_cast_fp16 = einsum(equation = var_6512_equation_0, values = (var_6406_cast_fp16_12, var_6479_cast_fp16))[name = tensor("op_6512_cast_fp16")]; + tensor var_6514_equation_0 = const()[name = tensor("op_6514_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6514_cast_fp16 = einsum(equation = var_6514_equation_0, values = (var_6406_cast_fp16_13, var_6480_cast_fp16))[name = tensor("op_6514_cast_fp16")]; + tensor var_6516_equation_0 = const()[name = tensor("op_6516_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6516_cast_fp16 = einsum(equation = var_6516_equation_0, values = (var_6406_cast_fp16_14, var_6481_cast_fp16))[name = tensor("op_6516_cast_fp16")]; + tensor var_6518_equation_0 = const()[name = tensor("op_6518_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6518_cast_fp16 = einsum(equation = var_6518_equation_0, values = (var_6406_cast_fp16_15, var_6482_cast_fp16))[name = tensor("op_6518_cast_fp16")]; + tensor var_6520_equation_0 = const()[name = tensor("op_6520_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6520_cast_fp16 = einsum(equation = var_6520_equation_0, values = (var_6406_cast_fp16_16, var_6483_cast_fp16))[name = tensor("op_6520_cast_fp16")]; + tensor var_6522_equation_0 = const()[name = tensor("op_6522_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6522_cast_fp16 = einsum(equation = var_6522_equation_0, values = (var_6406_cast_fp16_17, var_6484_cast_fp16))[name = tensor("op_6522_cast_fp16")]; + tensor var_6524_equation_0 = const()[name = tensor("op_6524_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6524_cast_fp16 = einsum(equation = var_6524_equation_0, values = (var_6406_cast_fp16_18, var_6485_cast_fp16))[name = tensor("op_6524_cast_fp16")]; + tensor var_6526_equation_0 = const()[name = tensor("op_6526_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6526_cast_fp16 = einsum(equation = var_6526_equation_0, values = (var_6406_cast_fp16_19, var_6486_cast_fp16))[name = tensor("op_6526_cast_fp16")]; + tensor input_235_interleave_0 = const()[name = tensor("input_235_interleave_0"), val = tensor(false)]; + tensor input_235_cast_fp16 = concat(axis = var_6311, interleave = input_235_interleave_0, values = (var_6488_cast_fp16, var_6490_cast_fp16, var_6492_cast_fp16, var_6494_cast_fp16, var_6496_cast_fp16, var_6498_cast_fp16, var_6500_cast_fp16, var_6502_cast_fp16, var_6504_cast_fp16, var_6506_cast_fp16, var_6508_cast_fp16, var_6510_cast_fp16, var_6512_cast_fp16, var_6514_cast_fp16, var_6516_cast_fp16, var_6518_cast_fp16, var_6520_cast_fp16, var_6522_cast_fp16, var_6524_cast_fp16, var_6526_cast_fp16))[name = tensor("input_235_cast_fp16")]; + tensor var_6535_pad_type_0 = const()[name = tensor("op_6535_pad_type_0"), val = tensor("valid")]; + tensor var_6535_strides_0 = const()[name = tensor("op_6535_strides_0"), val = tensor([1, 1])]; + tensor var_6535_pad_0 = const()[name = tensor("op_6535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6535_dilations_0 = const()[name = tensor("op_6535_dilations_0"), val = tensor([1, 1])]; + tensor var_6535_groups_0 = const()[name = tensor("op_6535_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_out_weight_to_fp16 = const()[name = tensor("blocks_23_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929256832)))]; + tensor blocks_23_attn_out_bias_to_fp16 = const()[name = tensor("blocks_23_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932533696)))]; + tensor var_6535_cast_fp16 = conv(bias = blocks_23_attn_out_bias_to_fp16, dilations = var_6535_dilations_0, groups = var_6535_groups_0, pad = var_6535_pad_0, pad_type = var_6535_pad_type_0, strides = var_6535_strides_0, weight = blocks_23_attn_out_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("op_6535_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = var_6535_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([1])]; + tensor input_237_gamma_0_to_fp16 = const()[name = tensor("input_237_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932536320)))]; + tensor input_237_beta_0_to_fp16 = const()[name = tensor("input_237_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932538944)))]; + tensor var_6545_to_fp16 = const()[name = tensor("op_6545_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = input_237_beta_0_to_fp16, epsilon = var_6545_to_fp16, gamma = input_237_gamma_0_to_fp16, x = inputs_95_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("valid")]; + tensor input_239_strides_0 = const()[name = tensor("input_239_strides_0"), val = tensor([1, 1])]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_239_dilations_0 = const()[name = tensor("input_239_dilations_0"), val = tensor([1, 1])]; + tensor input_239_groups_0 = const()[name = tensor("input_239_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932541568)))]; + tensor blocks_23_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945648832)))]; + tensor input_239_cast_fp16 = conv(bias = blocks_23_mlp_0_bias_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = blocks_23_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = input_239_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_6571_pad_type_0 = const()[name = tensor("op_6571_pad_type_0"), val = tensor("valid")]; + tensor var_6571_strides_0 = const()[name = tensor("op_6571_strides_0"), val = tensor([1, 1])]; + tensor var_6571_pad_0 = const()[name = tensor("op_6571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6571_dilations_0 = const()[name = tensor("op_6571_dilations_0"), val = tensor([1, 1])]; + tensor var_6571_groups_0 = const()[name = tensor("op_6571_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945659136)))]; + tensor blocks_23_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958766400)))]; + tensor var_6571_cast_fp16 = conv(bias = blocks_23_mlp_2_bias_to_fp16, dilations = var_6571_dilations_0, groups = var_6571_groups_0, pad = var_6571_pad_0, pad_type = var_6571_pad_type_0, strides = var_6571_strides_0, weight = blocks_23_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("op_6571_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = var_6571_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_6580 = const()[name = tensor("op_6580"), val = tensor(1)]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([1])]; + tensor input_243_gamma_0_to_fp16 = const()[name = tensor("input_243_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958769024)))]; + tensor input_243_beta_0_to_fp16 = const()[name = tensor("input_243_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958771648)))]; + tensor var_6596_to_fp16 = const()[name = tensor("op_6596_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = input_243_beta_0_to_fp16, epsilon = var_6596_to_fp16, gamma = input_243_gamma_0_to_fp16, x = inputs_97_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor q_49_pad_type_0 = const()[name = tensor("q_49_pad_type_0"), val = tensor("valid")]; + tensor q_49_strides_0 = const()[name = tensor("q_49_strides_0"), val = tensor([1, 1])]; + tensor q_49_pad_0 = const()[name = tensor("q_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_49_dilations_0 = const()[name = tensor("q_49_dilations_0"), val = tensor([1, 1])]; + tensor q_49_groups_0 = const()[name = tensor("q_49_groups_0"), val = tensor(1)]; + tensor var_6631_weight_0_to_fp16 = const()[name = tensor("op_6631_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958774272)))]; + tensor var_6631_bias_0_to_fp16 = const()[name = tensor("op_6631_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962051136)))]; + tensor var_6631_cast_fp16 = conv(bias = var_6631_bias_0_to_fp16, dilations = q_49_dilations_0, groups = q_49_groups_0, pad = q_49_pad_0, pad_type = q_49_pad_type_0, strides = q_49_strides_0, weight = var_6631_weight_0_to_fp16, x = input_243_cast_fp16)[name = tensor("op_6631_cast_fp16")]; + tensor k_49_pad_type_0 = const()[name = tensor("k_49_pad_type_0"), val = tensor("valid")]; + tensor k_49_strides_0 = const()[name = tensor("k_49_strides_0"), val = tensor([1, 1])]; + tensor k_49_pad_0 = const()[name = tensor("k_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_49_dilations_0 = const()[name = tensor("k_49_dilations_0"), val = tensor([1, 1])]; + tensor k_49_groups_0 = const()[name = tensor("k_49_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_key_weight_to_fp16 = const()[name = tensor("blocks_24_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962053760)))]; + tensor k_49_cast_fp16 = conv(dilations = k_49_dilations_0, groups = k_49_groups_0, pad = k_49_pad_0, pad_type = k_49_pad_type_0, strides = k_49_strides_0, weight = blocks_24_attn_key_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor var_6629_pad_type_0 = const()[name = tensor("op_6629_pad_type_0"), val = tensor("valid")]; + tensor var_6629_strides_0 = const()[name = tensor("op_6629_strides_0"), val = tensor([1, 1])]; + tensor var_6629_pad_0 = const()[name = tensor("op_6629_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6629_dilations_0 = const()[name = tensor("op_6629_dilations_0"), val = tensor([1, 1])]; + tensor var_6629_groups_0 = const()[name = tensor("op_6629_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_value_weight_to_fp16 = const()[name = tensor("blocks_24_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965330624)))]; + tensor blocks_24_attn_value_bias_to_fp16 = const()[name = tensor("blocks_24_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968607488)))]; + tensor var_6629_cast_fp16 = conv(bias = blocks_24_attn_value_bias_to_fp16, dilations = var_6629_dilations_0, groups = var_6629_groups_0, pad = var_6629_pad_0, pad_type = var_6629_pad_type_0, strides = var_6629_strides_0, weight = blocks_24_attn_value_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor tile_72 = const()[name = tensor("tile_72"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6632_axis_0 = const()[name = tensor("op_6632_axis_0"), val = tensor(1)]; + tensor var_6632_cast_fp16_0, tensor var_6632_cast_fp16_1, tensor var_6632_cast_fp16_2, tensor var_6632_cast_fp16_3, tensor var_6632_cast_fp16_4, tensor var_6632_cast_fp16_5, tensor var_6632_cast_fp16_6, tensor var_6632_cast_fp16_7, tensor var_6632_cast_fp16_8, tensor var_6632_cast_fp16_9, tensor var_6632_cast_fp16_10, tensor var_6632_cast_fp16_11, tensor var_6632_cast_fp16_12, tensor var_6632_cast_fp16_13, tensor var_6632_cast_fp16_14, tensor var_6632_cast_fp16_15, tensor var_6632_cast_fp16_16, tensor var_6632_cast_fp16_17, tensor var_6632_cast_fp16_18, tensor var_6632_cast_fp16_19 = split(axis = var_6632_axis_0, split_sizes = tile_72, x = var_6631_cast_fp16)[name = tensor("op_6632_cast_fp16")]; + tensor var_6653_perm_0 = const()[name = tensor("op_6653_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_73 = const()[name = tensor("tile_73"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6654_axis_0 = const()[name = tensor("op_6654_axis_0"), val = tensor(3)]; + tensor var_6653_cast_fp16 = transpose(perm = var_6653_perm_0, x = k_49_cast_fp16)[name = tensor("transpose_8")]; + tensor var_6654_cast_fp16_0, tensor var_6654_cast_fp16_1, tensor var_6654_cast_fp16_2, tensor var_6654_cast_fp16_3, tensor var_6654_cast_fp16_4, tensor var_6654_cast_fp16_5, tensor var_6654_cast_fp16_6, tensor var_6654_cast_fp16_7, tensor var_6654_cast_fp16_8, tensor var_6654_cast_fp16_9, tensor var_6654_cast_fp16_10, tensor var_6654_cast_fp16_11, tensor var_6654_cast_fp16_12, tensor var_6654_cast_fp16_13, tensor var_6654_cast_fp16_14, tensor var_6654_cast_fp16_15, tensor var_6654_cast_fp16_16, tensor var_6654_cast_fp16_17, tensor var_6654_cast_fp16_18, tensor var_6654_cast_fp16_19 = split(axis = var_6654_axis_0, split_sizes = tile_73, x = var_6653_cast_fp16)[name = tensor("op_6654_cast_fp16")]; + tensor tile_74 = const()[name = tensor("tile_74"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6675_axis_0 = const()[name = tensor("op_6675_axis_0"), val = tensor(1)]; + tensor var_6675_cast_fp16_0, tensor var_6675_cast_fp16_1, tensor var_6675_cast_fp16_2, tensor var_6675_cast_fp16_3, tensor var_6675_cast_fp16_4, tensor var_6675_cast_fp16_5, tensor var_6675_cast_fp16_6, tensor var_6675_cast_fp16_7, tensor var_6675_cast_fp16_8, tensor var_6675_cast_fp16_9, tensor var_6675_cast_fp16_10, tensor var_6675_cast_fp16_11, tensor var_6675_cast_fp16_12, tensor var_6675_cast_fp16_13, tensor var_6675_cast_fp16_14, tensor var_6675_cast_fp16_15, tensor var_6675_cast_fp16_16, tensor var_6675_cast_fp16_17, tensor var_6675_cast_fp16_18, tensor var_6675_cast_fp16_19 = split(axis = var_6675_axis_0, split_sizes = tile_74, x = var_6629_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor aw_961_equation_0 = const()[name = tensor("aw_961_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_961_cast_fp16 = einsum(equation = aw_961_equation_0, values = (var_6654_cast_fp16_0, var_6632_cast_fp16_0))[name = tensor("aw_961_cast_fp16")]; + tensor aw_963_equation_0 = const()[name = tensor("aw_963_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_963_cast_fp16 = einsum(equation = aw_963_equation_0, values = (var_6654_cast_fp16_1, var_6632_cast_fp16_1))[name = tensor("aw_963_cast_fp16")]; + tensor aw_965_equation_0 = const()[name = tensor("aw_965_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_965_cast_fp16 = einsum(equation = aw_965_equation_0, values = (var_6654_cast_fp16_2, var_6632_cast_fp16_2))[name = tensor("aw_965_cast_fp16")]; + tensor aw_967_equation_0 = const()[name = tensor("aw_967_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_967_cast_fp16 = einsum(equation = aw_967_equation_0, values = (var_6654_cast_fp16_3, var_6632_cast_fp16_3))[name = tensor("aw_967_cast_fp16")]; + tensor aw_969_equation_0 = const()[name = tensor("aw_969_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_969_cast_fp16 = einsum(equation = aw_969_equation_0, values = (var_6654_cast_fp16_4, var_6632_cast_fp16_4))[name = tensor("aw_969_cast_fp16")]; + tensor aw_971_equation_0 = const()[name = tensor("aw_971_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_971_cast_fp16 = einsum(equation = aw_971_equation_0, values = (var_6654_cast_fp16_5, var_6632_cast_fp16_5))[name = tensor("aw_971_cast_fp16")]; + tensor aw_973_equation_0 = const()[name = tensor("aw_973_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_973_cast_fp16 = einsum(equation = aw_973_equation_0, values = (var_6654_cast_fp16_6, var_6632_cast_fp16_6))[name = tensor("aw_973_cast_fp16")]; + tensor aw_975_equation_0 = const()[name = tensor("aw_975_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_975_cast_fp16 = einsum(equation = aw_975_equation_0, values = (var_6654_cast_fp16_7, var_6632_cast_fp16_7))[name = tensor("aw_975_cast_fp16")]; + tensor aw_977_equation_0 = const()[name = tensor("aw_977_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_977_cast_fp16 = einsum(equation = aw_977_equation_0, values = (var_6654_cast_fp16_8, var_6632_cast_fp16_8))[name = tensor("aw_977_cast_fp16")]; + tensor aw_979_equation_0 = const()[name = tensor("aw_979_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_979_cast_fp16 = einsum(equation = aw_979_equation_0, values = (var_6654_cast_fp16_9, var_6632_cast_fp16_9))[name = tensor("aw_979_cast_fp16")]; + tensor aw_981_equation_0 = const()[name = tensor("aw_981_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_981_cast_fp16 = einsum(equation = aw_981_equation_0, values = (var_6654_cast_fp16_10, var_6632_cast_fp16_10))[name = tensor("aw_981_cast_fp16")]; + tensor aw_983_equation_0 = const()[name = tensor("aw_983_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_983_cast_fp16 = einsum(equation = aw_983_equation_0, values = (var_6654_cast_fp16_11, var_6632_cast_fp16_11))[name = tensor("aw_983_cast_fp16")]; + tensor aw_985_equation_0 = const()[name = tensor("aw_985_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_985_cast_fp16 = einsum(equation = aw_985_equation_0, values = (var_6654_cast_fp16_12, var_6632_cast_fp16_12))[name = tensor("aw_985_cast_fp16")]; + tensor aw_987_equation_0 = const()[name = tensor("aw_987_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_987_cast_fp16 = einsum(equation = aw_987_equation_0, values = (var_6654_cast_fp16_13, var_6632_cast_fp16_13))[name = tensor("aw_987_cast_fp16")]; + tensor aw_989_equation_0 = const()[name = tensor("aw_989_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_989_cast_fp16 = einsum(equation = aw_989_equation_0, values = (var_6654_cast_fp16_14, var_6632_cast_fp16_14))[name = tensor("aw_989_cast_fp16")]; + tensor aw_991_equation_0 = const()[name = tensor("aw_991_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_991_cast_fp16 = einsum(equation = aw_991_equation_0, values = (var_6654_cast_fp16_15, var_6632_cast_fp16_15))[name = tensor("aw_991_cast_fp16")]; + tensor aw_993_equation_0 = const()[name = tensor("aw_993_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_993_cast_fp16 = einsum(equation = aw_993_equation_0, values = (var_6654_cast_fp16_16, var_6632_cast_fp16_16))[name = tensor("aw_993_cast_fp16")]; + tensor aw_995_equation_0 = const()[name = tensor("aw_995_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_995_cast_fp16 = einsum(equation = aw_995_equation_0, values = (var_6654_cast_fp16_17, var_6632_cast_fp16_17))[name = tensor("aw_995_cast_fp16")]; + tensor aw_997_equation_0 = const()[name = tensor("aw_997_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_997_cast_fp16 = einsum(equation = aw_997_equation_0, values = (var_6654_cast_fp16_18, var_6632_cast_fp16_18))[name = tensor("aw_997_cast_fp16")]; + tensor aw_999_equation_0 = const()[name = tensor("aw_999_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_999_cast_fp16 = einsum(equation = aw_999_equation_0, values = (var_6654_cast_fp16_19, var_6632_cast_fp16_19))[name = tensor("aw_999_cast_fp16")]; + tensor var_6736_cast_fp16 = softmax(axis = var_6580, x = aw_961_cast_fp16)[name = tensor("op_6736_cast_fp16")]; + tensor var_6737_cast_fp16 = softmax(axis = var_6580, x = aw_963_cast_fp16)[name = tensor("op_6737_cast_fp16")]; + tensor var_6738_cast_fp16 = softmax(axis = var_6580, x = aw_965_cast_fp16)[name = tensor("op_6738_cast_fp16")]; + tensor var_6739_cast_fp16 = softmax(axis = var_6580, x = aw_967_cast_fp16)[name = tensor("op_6739_cast_fp16")]; + tensor var_6740_cast_fp16 = softmax(axis = var_6580, x = aw_969_cast_fp16)[name = tensor("op_6740_cast_fp16")]; + tensor var_6741_cast_fp16 = softmax(axis = var_6580, x = aw_971_cast_fp16)[name = tensor("op_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = softmax(axis = var_6580, x = aw_973_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_cast_fp16 = softmax(axis = var_6580, x = aw_975_cast_fp16)[name = tensor("op_6743_cast_fp16")]; + tensor var_6744_cast_fp16 = softmax(axis = var_6580, x = aw_977_cast_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor var_6745_cast_fp16 = softmax(axis = var_6580, x = aw_979_cast_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6746_cast_fp16 = softmax(axis = var_6580, x = aw_981_cast_fp16)[name = tensor("op_6746_cast_fp16")]; + tensor var_6747_cast_fp16 = softmax(axis = var_6580, x = aw_983_cast_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6748_cast_fp16 = softmax(axis = var_6580, x = aw_985_cast_fp16)[name = tensor("op_6748_cast_fp16")]; + tensor var_6749_cast_fp16 = softmax(axis = var_6580, x = aw_987_cast_fp16)[name = tensor("op_6749_cast_fp16")]; + tensor var_6750_cast_fp16 = softmax(axis = var_6580, x = aw_989_cast_fp16)[name = tensor("op_6750_cast_fp16")]; + tensor var_6751_cast_fp16 = softmax(axis = var_6580, x = aw_991_cast_fp16)[name = tensor("op_6751_cast_fp16")]; + tensor var_6752_cast_fp16 = softmax(axis = var_6580, x = aw_993_cast_fp16)[name = tensor("op_6752_cast_fp16")]; + tensor var_6753_cast_fp16 = softmax(axis = var_6580, x = aw_995_cast_fp16)[name = tensor("op_6753_cast_fp16")]; + tensor var_6754_cast_fp16 = softmax(axis = var_6580, x = aw_997_cast_fp16)[name = tensor("op_6754_cast_fp16")]; + tensor var_6755_cast_fp16 = softmax(axis = var_6580, x = aw_999_cast_fp16)[name = tensor("op_6755_cast_fp16")]; + tensor var_6757_equation_0 = const()[name = tensor("op_6757_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6757_cast_fp16 = einsum(equation = var_6757_equation_0, values = (var_6675_cast_fp16_0, var_6736_cast_fp16))[name = tensor("op_6757_cast_fp16")]; + tensor var_6759_equation_0 = const()[name = tensor("op_6759_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6759_cast_fp16 = einsum(equation = var_6759_equation_0, values = (var_6675_cast_fp16_1, var_6737_cast_fp16))[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_equation_0 = const()[name = tensor("op_6761_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6761_cast_fp16 = einsum(equation = var_6761_equation_0, values = (var_6675_cast_fp16_2, var_6738_cast_fp16))[name = tensor("op_6761_cast_fp16")]; + tensor var_6763_equation_0 = const()[name = tensor("op_6763_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6763_cast_fp16 = einsum(equation = var_6763_equation_0, values = (var_6675_cast_fp16_3, var_6739_cast_fp16))[name = tensor("op_6763_cast_fp16")]; + tensor var_6765_equation_0 = const()[name = tensor("op_6765_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6765_cast_fp16 = einsum(equation = var_6765_equation_0, values = (var_6675_cast_fp16_4, var_6740_cast_fp16))[name = tensor("op_6765_cast_fp16")]; + tensor var_6767_equation_0 = const()[name = tensor("op_6767_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6767_cast_fp16 = einsum(equation = var_6767_equation_0, values = (var_6675_cast_fp16_5, var_6741_cast_fp16))[name = tensor("op_6767_cast_fp16")]; + tensor var_6769_equation_0 = const()[name = tensor("op_6769_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6769_cast_fp16 = einsum(equation = var_6769_equation_0, values = (var_6675_cast_fp16_6, var_6742_cast_fp16))[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_equation_0 = const()[name = tensor("op_6771_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6771_cast_fp16 = einsum(equation = var_6771_equation_0, values = (var_6675_cast_fp16_7, var_6743_cast_fp16))[name = tensor("op_6771_cast_fp16")]; + tensor var_6773_equation_0 = const()[name = tensor("op_6773_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6773_cast_fp16 = einsum(equation = var_6773_equation_0, values = (var_6675_cast_fp16_8, var_6744_cast_fp16))[name = tensor("op_6773_cast_fp16")]; + tensor var_6775_equation_0 = const()[name = tensor("op_6775_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6775_cast_fp16 = einsum(equation = var_6775_equation_0, values = (var_6675_cast_fp16_9, var_6745_cast_fp16))[name = tensor("op_6775_cast_fp16")]; + tensor var_6777_equation_0 = const()[name = tensor("op_6777_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6777_cast_fp16 = einsum(equation = var_6777_equation_0, values = (var_6675_cast_fp16_10, var_6746_cast_fp16))[name = tensor("op_6777_cast_fp16")]; + tensor var_6779_equation_0 = const()[name = tensor("op_6779_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6779_cast_fp16 = einsum(equation = var_6779_equation_0, values = (var_6675_cast_fp16_11, var_6747_cast_fp16))[name = tensor("op_6779_cast_fp16")]; + tensor var_6781_equation_0 = const()[name = tensor("op_6781_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6781_cast_fp16 = einsum(equation = var_6781_equation_0, values = (var_6675_cast_fp16_12, var_6748_cast_fp16))[name = tensor("op_6781_cast_fp16")]; + tensor var_6783_equation_0 = const()[name = tensor("op_6783_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6783_cast_fp16 = einsum(equation = var_6783_equation_0, values = (var_6675_cast_fp16_13, var_6749_cast_fp16))[name = tensor("op_6783_cast_fp16")]; + tensor var_6785_equation_0 = const()[name = tensor("op_6785_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6785_cast_fp16 = einsum(equation = var_6785_equation_0, values = (var_6675_cast_fp16_14, var_6750_cast_fp16))[name = tensor("op_6785_cast_fp16")]; + tensor var_6787_equation_0 = const()[name = tensor("op_6787_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6787_cast_fp16 = einsum(equation = var_6787_equation_0, values = (var_6675_cast_fp16_15, var_6751_cast_fp16))[name = tensor("op_6787_cast_fp16")]; + tensor var_6789_equation_0 = const()[name = tensor("op_6789_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6789_cast_fp16 = einsum(equation = var_6789_equation_0, values = (var_6675_cast_fp16_16, var_6752_cast_fp16))[name = tensor("op_6789_cast_fp16")]; + tensor var_6791_equation_0 = const()[name = tensor("op_6791_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6791_cast_fp16 = einsum(equation = var_6791_equation_0, values = (var_6675_cast_fp16_17, var_6753_cast_fp16))[name = tensor("op_6791_cast_fp16")]; + tensor var_6793_equation_0 = const()[name = tensor("op_6793_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6793_cast_fp16 = einsum(equation = var_6793_equation_0, values = (var_6675_cast_fp16_18, var_6754_cast_fp16))[name = tensor("op_6793_cast_fp16")]; + tensor var_6795_equation_0 = const()[name = tensor("op_6795_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6795_cast_fp16 = einsum(equation = var_6795_equation_0, values = (var_6675_cast_fp16_19, var_6755_cast_fp16))[name = tensor("op_6795_cast_fp16")]; + tensor input_245_interleave_0 = const()[name = tensor("input_245_interleave_0"), val = tensor(false)]; + tensor input_245_cast_fp16 = concat(axis = var_6580, interleave = input_245_interleave_0, values = (var_6757_cast_fp16, var_6759_cast_fp16, var_6761_cast_fp16, var_6763_cast_fp16, var_6765_cast_fp16, var_6767_cast_fp16, var_6769_cast_fp16, var_6771_cast_fp16, var_6773_cast_fp16, var_6775_cast_fp16, var_6777_cast_fp16, var_6779_cast_fp16, var_6781_cast_fp16, var_6783_cast_fp16, var_6785_cast_fp16, var_6787_cast_fp16, var_6789_cast_fp16, var_6791_cast_fp16, var_6793_cast_fp16, var_6795_cast_fp16))[name = tensor("input_245_cast_fp16")]; + tensor var_6804_pad_type_0 = const()[name = tensor("op_6804_pad_type_0"), val = tensor("valid")]; + tensor var_6804_strides_0 = const()[name = tensor("op_6804_strides_0"), val = tensor([1, 1])]; + tensor var_6804_pad_0 = const()[name = tensor("op_6804_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6804_dilations_0 = const()[name = tensor("op_6804_dilations_0"), val = tensor([1, 1])]; + tensor var_6804_groups_0 = const()[name = tensor("op_6804_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_out_weight_to_fp16 = const()[name = tensor("blocks_24_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968610112)))]; + tensor blocks_24_attn_out_bias_to_fp16 = const()[name = tensor("blocks_24_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971886976)))]; + tensor var_6804_cast_fp16 = conv(bias = blocks_24_attn_out_bias_to_fp16, dilations = var_6804_dilations_0, groups = var_6804_groups_0, pad = var_6804_pad_0, pad_type = var_6804_pad_type_0, strides = var_6804_strides_0, weight = blocks_24_attn_out_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("op_6804_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_6804_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor input_247_axes_0 = const()[name = tensor("input_247_axes_0"), val = tensor([1])]; + tensor input_247_gamma_0_to_fp16 = const()[name = tensor("input_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971889600)))]; + tensor input_247_beta_0_to_fp16 = const()[name = tensor("input_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971892224)))]; + tensor var_6814_to_fp16 = const()[name = tensor("op_6814_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, beta = input_247_beta_0_to_fp16, epsilon = var_6814_to_fp16, gamma = input_247_gamma_0_to_fp16, x = inputs_99_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_pad_type_0 = const()[name = tensor("input_249_pad_type_0"), val = tensor("valid")]; + tensor input_249_strides_0 = const()[name = tensor("input_249_strides_0"), val = tensor([1, 1])]; + tensor input_249_pad_0 = const()[name = tensor("input_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_249_dilations_0 = const()[name = tensor("input_249_dilations_0"), val = tensor([1, 1])]; + tensor input_249_groups_0 = const()[name = tensor("input_249_groups_0"), val = tensor(1)]; + tensor blocks_24_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971894848)))]; + tensor blocks_24_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985002112)))]; + tensor input_249_cast_fp16 = conv(bias = blocks_24_mlp_0_bias_to_fp16, dilations = input_249_dilations_0, groups = input_249_groups_0, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = input_249_strides_0, weight = blocks_24_mlp_0_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_6840_pad_type_0 = const()[name = tensor("op_6840_pad_type_0"), val = tensor("valid")]; + tensor var_6840_strides_0 = const()[name = tensor("op_6840_strides_0"), val = tensor([1, 1])]; + tensor var_6840_pad_0 = const()[name = tensor("op_6840_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6840_dilations_0 = const()[name = tensor("op_6840_dilations_0"), val = tensor([1, 1])]; + tensor var_6840_groups_0 = const()[name = tensor("op_6840_groups_0"), val = tensor(1)]; + tensor blocks_24_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985012416)))]; + tensor blocks_24_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998119680)))]; + tensor var_6840_cast_fp16 = conv(bias = blocks_24_mlp_2_bias_to_fp16, dilations = var_6840_dilations_0, groups = var_6840_groups_0, pad = var_6840_pad_0, pad_type = var_6840_pad_type_0, strides = var_6840_strides_0, weight = blocks_24_mlp_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("op_6840_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = var_6840_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_6849 = const()[name = tensor("op_6849"), val = tensor(1)]; + tensor input_253_axes_0 = const()[name = tensor("input_253_axes_0"), val = tensor([1])]; + tensor input_253_gamma_0_to_fp16 = const()[name = tensor("input_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998122304)))]; + tensor input_253_beta_0_to_fp16 = const()[name = tensor("input_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998124928)))]; + tensor var_6865_to_fp16 = const()[name = tensor("op_6865_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_253_cast_fp16 = layer_norm(axes = input_253_axes_0, beta = input_253_beta_0_to_fp16, epsilon = var_6865_to_fp16, gamma = input_253_gamma_0_to_fp16, x = inputs_101_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor q_51_pad_type_0 = const()[name = tensor("q_51_pad_type_0"), val = tensor("valid")]; + tensor q_51_strides_0 = const()[name = tensor("q_51_strides_0"), val = tensor([1, 1])]; + tensor q_51_pad_0 = const()[name = tensor("q_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_51_dilations_0 = const()[name = tensor("q_51_dilations_0"), val = tensor([1, 1])]; + tensor q_51_groups_0 = const()[name = tensor("q_51_groups_0"), val = tensor(1)]; + tensor var_6900_weight_0_to_fp16 = const()[name = tensor("op_6900_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998127552)))]; + tensor var_6900_bias_0_to_fp16 = const()[name = tensor("op_6900_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001404416)))]; + tensor var_6900_cast_fp16 = conv(bias = var_6900_bias_0_to_fp16, dilations = q_51_dilations_0, groups = q_51_groups_0, pad = q_51_pad_0, pad_type = q_51_pad_type_0, strides = q_51_strides_0, weight = var_6900_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor k_51_pad_type_0 = const()[name = tensor("k_51_pad_type_0"), val = tensor("valid")]; + tensor k_51_strides_0 = const()[name = tensor("k_51_strides_0"), val = tensor([1, 1])]; + tensor k_51_pad_0 = const()[name = tensor("k_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_51_dilations_0 = const()[name = tensor("k_51_dilations_0"), val = tensor([1, 1])]; + tensor k_51_groups_0 = const()[name = tensor("k_51_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_key_weight_to_fp16 = const()[name = tensor("blocks_25_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001407040)))]; + tensor k_51_cast_fp16 = conv(dilations = k_51_dilations_0, groups = k_51_groups_0, pad = k_51_pad_0, pad_type = k_51_pad_type_0, strides = k_51_strides_0, weight = blocks_25_attn_key_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor var_6898_pad_type_0 = const()[name = tensor("op_6898_pad_type_0"), val = tensor("valid")]; + tensor var_6898_strides_0 = const()[name = tensor("op_6898_strides_0"), val = tensor([1, 1])]; + tensor var_6898_pad_0 = const()[name = tensor("op_6898_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6898_dilations_0 = const()[name = tensor("op_6898_dilations_0"), val = tensor([1, 1])]; + tensor var_6898_groups_0 = const()[name = tensor("op_6898_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_value_weight_to_fp16 = const()[name = tensor("blocks_25_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1004683904)))]; + tensor blocks_25_attn_value_bias_to_fp16 = const()[name = tensor("blocks_25_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1007960768)))]; + tensor var_6898_cast_fp16 = conv(bias = blocks_25_attn_value_bias_to_fp16, dilations = var_6898_dilations_0, groups = var_6898_groups_0, pad = var_6898_pad_0, pad_type = var_6898_pad_type_0, strides = var_6898_strides_0, weight = blocks_25_attn_value_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor tile_75 = const()[name = tensor("tile_75"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6901_axis_0 = const()[name = tensor("op_6901_axis_0"), val = tensor(1)]; + tensor var_6901_cast_fp16_0, tensor var_6901_cast_fp16_1, tensor var_6901_cast_fp16_2, tensor var_6901_cast_fp16_3, tensor var_6901_cast_fp16_4, tensor var_6901_cast_fp16_5, tensor var_6901_cast_fp16_6, tensor var_6901_cast_fp16_7, tensor var_6901_cast_fp16_8, tensor var_6901_cast_fp16_9, tensor var_6901_cast_fp16_10, tensor var_6901_cast_fp16_11, tensor var_6901_cast_fp16_12, tensor var_6901_cast_fp16_13, tensor var_6901_cast_fp16_14, tensor var_6901_cast_fp16_15, tensor var_6901_cast_fp16_16, tensor var_6901_cast_fp16_17, tensor var_6901_cast_fp16_18, tensor var_6901_cast_fp16_19 = split(axis = var_6901_axis_0, split_sizes = tile_75, x = var_6900_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor var_6922_perm_0 = const()[name = tensor("op_6922_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_76 = const()[name = tensor("tile_76"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6923_axis_0 = const()[name = tensor("op_6923_axis_0"), val = tensor(3)]; + tensor var_6922_cast_fp16 = transpose(perm = var_6922_perm_0, x = k_51_cast_fp16)[name = tensor("transpose_7")]; + tensor var_6923_cast_fp16_0, tensor var_6923_cast_fp16_1, tensor var_6923_cast_fp16_2, tensor var_6923_cast_fp16_3, tensor var_6923_cast_fp16_4, tensor var_6923_cast_fp16_5, tensor var_6923_cast_fp16_6, tensor var_6923_cast_fp16_7, tensor var_6923_cast_fp16_8, tensor var_6923_cast_fp16_9, tensor var_6923_cast_fp16_10, tensor var_6923_cast_fp16_11, tensor var_6923_cast_fp16_12, tensor var_6923_cast_fp16_13, tensor var_6923_cast_fp16_14, tensor var_6923_cast_fp16_15, tensor var_6923_cast_fp16_16, tensor var_6923_cast_fp16_17, tensor var_6923_cast_fp16_18, tensor var_6923_cast_fp16_19 = split(axis = var_6923_axis_0, split_sizes = tile_76, x = var_6922_cast_fp16)[name = tensor("op_6923_cast_fp16")]; + tensor tile_77 = const()[name = tensor("tile_77"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6944_axis_0 = const()[name = tensor("op_6944_axis_0"), val = tensor(1)]; + tensor var_6944_cast_fp16_0, tensor var_6944_cast_fp16_1, tensor var_6944_cast_fp16_2, tensor var_6944_cast_fp16_3, tensor var_6944_cast_fp16_4, tensor var_6944_cast_fp16_5, tensor var_6944_cast_fp16_6, tensor var_6944_cast_fp16_7, tensor var_6944_cast_fp16_8, tensor var_6944_cast_fp16_9, tensor var_6944_cast_fp16_10, tensor var_6944_cast_fp16_11, tensor var_6944_cast_fp16_12, tensor var_6944_cast_fp16_13, tensor var_6944_cast_fp16_14, tensor var_6944_cast_fp16_15, tensor var_6944_cast_fp16_16, tensor var_6944_cast_fp16_17, tensor var_6944_cast_fp16_18, tensor var_6944_cast_fp16_19 = split(axis = var_6944_axis_0, split_sizes = tile_77, x = var_6898_cast_fp16)[name = tensor("op_6944_cast_fp16")]; + tensor aw_1001_equation_0 = const()[name = tensor("aw_1001_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1001_cast_fp16 = einsum(equation = aw_1001_equation_0, values = (var_6923_cast_fp16_0, var_6901_cast_fp16_0))[name = tensor("aw_1001_cast_fp16")]; + tensor aw_1003_equation_0 = const()[name = tensor("aw_1003_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1003_cast_fp16 = einsum(equation = aw_1003_equation_0, values = (var_6923_cast_fp16_1, var_6901_cast_fp16_1))[name = tensor("aw_1003_cast_fp16")]; + tensor aw_1005_equation_0 = const()[name = tensor("aw_1005_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1005_cast_fp16 = einsum(equation = aw_1005_equation_0, values = (var_6923_cast_fp16_2, var_6901_cast_fp16_2))[name = tensor("aw_1005_cast_fp16")]; + tensor aw_1007_equation_0 = const()[name = tensor("aw_1007_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1007_cast_fp16 = einsum(equation = aw_1007_equation_0, values = (var_6923_cast_fp16_3, var_6901_cast_fp16_3))[name = tensor("aw_1007_cast_fp16")]; + tensor aw_1009_equation_0 = const()[name = tensor("aw_1009_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1009_cast_fp16 = einsum(equation = aw_1009_equation_0, values = (var_6923_cast_fp16_4, var_6901_cast_fp16_4))[name = tensor("aw_1009_cast_fp16")]; + tensor aw_1011_equation_0 = const()[name = tensor("aw_1011_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1011_cast_fp16 = einsum(equation = aw_1011_equation_0, values = (var_6923_cast_fp16_5, var_6901_cast_fp16_5))[name = tensor("aw_1011_cast_fp16")]; + tensor aw_1013_equation_0 = const()[name = tensor("aw_1013_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1013_cast_fp16 = einsum(equation = aw_1013_equation_0, values = (var_6923_cast_fp16_6, var_6901_cast_fp16_6))[name = tensor("aw_1013_cast_fp16")]; + tensor aw_1015_equation_0 = const()[name = tensor("aw_1015_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1015_cast_fp16 = einsum(equation = aw_1015_equation_0, values = (var_6923_cast_fp16_7, var_6901_cast_fp16_7))[name = tensor("aw_1015_cast_fp16")]; + tensor aw_1017_equation_0 = const()[name = tensor("aw_1017_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1017_cast_fp16 = einsum(equation = aw_1017_equation_0, values = (var_6923_cast_fp16_8, var_6901_cast_fp16_8))[name = tensor("aw_1017_cast_fp16")]; + tensor aw_1019_equation_0 = const()[name = tensor("aw_1019_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1019_cast_fp16 = einsum(equation = aw_1019_equation_0, values = (var_6923_cast_fp16_9, var_6901_cast_fp16_9))[name = tensor("aw_1019_cast_fp16")]; + tensor aw_1021_equation_0 = const()[name = tensor("aw_1021_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1021_cast_fp16 = einsum(equation = aw_1021_equation_0, values = (var_6923_cast_fp16_10, var_6901_cast_fp16_10))[name = tensor("aw_1021_cast_fp16")]; + tensor aw_1023_equation_0 = const()[name = tensor("aw_1023_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1023_cast_fp16 = einsum(equation = aw_1023_equation_0, values = (var_6923_cast_fp16_11, var_6901_cast_fp16_11))[name = tensor("aw_1023_cast_fp16")]; + tensor aw_1025_equation_0 = const()[name = tensor("aw_1025_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1025_cast_fp16 = einsum(equation = aw_1025_equation_0, values = (var_6923_cast_fp16_12, var_6901_cast_fp16_12))[name = tensor("aw_1025_cast_fp16")]; + tensor aw_1027_equation_0 = const()[name = tensor("aw_1027_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1027_cast_fp16 = einsum(equation = aw_1027_equation_0, values = (var_6923_cast_fp16_13, var_6901_cast_fp16_13))[name = tensor("aw_1027_cast_fp16")]; + tensor aw_1029_equation_0 = const()[name = tensor("aw_1029_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1029_cast_fp16 = einsum(equation = aw_1029_equation_0, values = (var_6923_cast_fp16_14, var_6901_cast_fp16_14))[name = tensor("aw_1029_cast_fp16")]; + tensor aw_1031_equation_0 = const()[name = tensor("aw_1031_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1031_cast_fp16 = einsum(equation = aw_1031_equation_0, values = (var_6923_cast_fp16_15, var_6901_cast_fp16_15))[name = tensor("aw_1031_cast_fp16")]; + tensor aw_1033_equation_0 = const()[name = tensor("aw_1033_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1033_cast_fp16 = einsum(equation = aw_1033_equation_0, values = (var_6923_cast_fp16_16, var_6901_cast_fp16_16))[name = tensor("aw_1033_cast_fp16")]; + tensor aw_1035_equation_0 = const()[name = tensor("aw_1035_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1035_cast_fp16 = einsum(equation = aw_1035_equation_0, values = (var_6923_cast_fp16_17, var_6901_cast_fp16_17))[name = tensor("aw_1035_cast_fp16")]; + tensor aw_1037_equation_0 = const()[name = tensor("aw_1037_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1037_cast_fp16 = einsum(equation = aw_1037_equation_0, values = (var_6923_cast_fp16_18, var_6901_cast_fp16_18))[name = tensor("aw_1037_cast_fp16")]; + tensor aw_1039_equation_0 = const()[name = tensor("aw_1039_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1039_cast_fp16 = einsum(equation = aw_1039_equation_0, values = (var_6923_cast_fp16_19, var_6901_cast_fp16_19))[name = tensor("aw_1039_cast_fp16")]; + tensor var_7005_cast_fp16 = softmax(axis = var_6849, x = aw_1001_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7006_cast_fp16 = softmax(axis = var_6849, x = aw_1003_cast_fp16)[name = tensor("op_7006_cast_fp16")]; + tensor var_7007_cast_fp16 = softmax(axis = var_6849, x = aw_1005_cast_fp16)[name = tensor("op_7007_cast_fp16")]; + tensor var_7008_cast_fp16 = softmax(axis = var_6849, x = aw_1007_cast_fp16)[name = tensor("op_7008_cast_fp16")]; + tensor var_7009_cast_fp16 = softmax(axis = var_6849, x = aw_1009_cast_fp16)[name = tensor("op_7009_cast_fp16")]; + tensor var_7010_cast_fp16 = softmax(axis = var_6849, x = aw_1011_cast_fp16)[name = tensor("op_7010_cast_fp16")]; + tensor var_7011_cast_fp16 = softmax(axis = var_6849, x = aw_1013_cast_fp16)[name = tensor("op_7011_cast_fp16")]; + tensor var_7012_cast_fp16 = softmax(axis = var_6849, x = aw_1015_cast_fp16)[name = tensor("op_7012_cast_fp16")]; + tensor var_7013_cast_fp16 = softmax(axis = var_6849, x = aw_1017_cast_fp16)[name = tensor("op_7013_cast_fp16")]; + tensor var_7014_cast_fp16 = softmax(axis = var_6849, x = aw_1019_cast_fp16)[name = tensor("op_7014_cast_fp16")]; + tensor var_7015_cast_fp16 = softmax(axis = var_6849, x = aw_1021_cast_fp16)[name = tensor("op_7015_cast_fp16")]; + tensor var_7016_cast_fp16 = softmax(axis = var_6849, x = aw_1023_cast_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7017_cast_fp16 = softmax(axis = var_6849, x = aw_1025_cast_fp16)[name = tensor("op_7017_cast_fp16")]; + tensor var_7018_cast_fp16 = softmax(axis = var_6849, x = aw_1027_cast_fp16)[name = tensor("op_7018_cast_fp16")]; + tensor var_7019_cast_fp16 = softmax(axis = var_6849, x = aw_1029_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor var_7020_cast_fp16 = softmax(axis = var_6849, x = aw_1031_cast_fp16)[name = tensor("op_7020_cast_fp16")]; + tensor var_7021_cast_fp16 = softmax(axis = var_6849, x = aw_1033_cast_fp16)[name = tensor("op_7021_cast_fp16")]; + tensor var_7022_cast_fp16 = softmax(axis = var_6849, x = aw_1035_cast_fp16)[name = tensor("op_7022_cast_fp16")]; + tensor var_7023_cast_fp16 = softmax(axis = var_6849, x = aw_1037_cast_fp16)[name = tensor("op_7023_cast_fp16")]; + tensor var_7024_cast_fp16 = softmax(axis = var_6849, x = aw_1039_cast_fp16)[name = tensor("op_7024_cast_fp16")]; + tensor var_7026_equation_0 = const()[name = tensor("op_7026_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7026_cast_fp16 = einsum(equation = var_7026_equation_0, values = (var_6944_cast_fp16_0, var_7005_cast_fp16))[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_equation_0 = const()[name = tensor("op_7028_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7028_cast_fp16 = einsum(equation = var_7028_equation_0, values = (var_6944_cast_fp16_1, var_7006_cast_fp16))[name = tensor("op_7028_cast_fp16")]; + tensor var_7030_equation_0 = const()[name = tensor("op_7030_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7030_cast_fp16 = einsum(equation = var_7030_equation_0, values = (var_6944_cast_fp16_2, var_7007_cast_fp16))[name = tensor("op_7030_cast_fp16")]; + tensor var_7032_equation_0 = const()[name = tensor("op_7032_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7032_cast_fp16 = einsum(equation = var_7032_equation_0, values = (var_6944_cast_fp16_3, var_7008_cast_fp16))[name = tensor("op_7032_cast_fp16")]; + tensor var_7034_equation_0 = const()[name = tensor("op_7034_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7034_cast_fp16 = einsum(equation = var_7034_equation_0, values = (var_6944_cast_fp16_4, var_7009_cast_fp16))[name = tensor("op_7034_cast_fp16")]; + tensor var_7036_equation_0 = const()[name = tensor("op_7036_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7036_cast_fp16 = einsum(equation = var_7036_equation_0, values = (var_6944_cast_fp16_5, var_7010_cast_fp16))[name = tensor("op_7036_cast_fp16")]; + tensor var_7038_equation_0 = const()[name = tensor("op_7038_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7038_cast_fp16 = einsum(equation = var_7038_equation_0, values = (var_6944_cast_fp16_6, var_7011_cast_fp16))[name = tensor("op_7038_cast_fp16")]; + tensor var_7040_equation_0 = const()[name = tensor("op_7040_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7040_cast_fp16 = einsum(equation = var_7040_equation_0, values = (var_6944_cast_fp16_7, var_7012_cast_fp16))[name = tensor("op_7040_cast_fp16")]; + tensor var_7042_equation_0 = const()[name = tensor("op_7042_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7042_cast_fp16 = einsum(equation = var_7042_equation_0, values = (var_6944_cast_fp16_8, var_7013_cast_fp16))[name = tensor("op_7042_cast_fp16")]; + tensor var_7044_equation_0 = const()[name = tensor("op_7044_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7044_cast_fp16 = einsum(equation = var_7044_equation_0, values = (var_6944_cast_fp16_9, var_7014_cast_fp16))[name = tensor("op_7044_cast_fp16")]; + tensor var_7046_equation_0 = const()[name = tensor("op_7046_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7046_cast_fp16 = einsum(equation = var_7046_equation_0, values = (var_6944_cast_fp16_10, var_7015_cast_fp16))[name = tensor("op_7046_cast_fp16")]; + tensor var_7048_equation_0 = const()[name = tensor("op_7048_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7048_cast_fp16 = einsum(equation = var_7048_equation_0, values = (var_6944_cast_fp16_11, var_7016_cast_fp16))[name = tensor("op_7048_cast_fp16")]; + tensor var_7050_equation_0 = const()[name = tensor("op_7050_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7050_cast_fp16 = einsum(equation = var_7050_equation_0, values = (var_6944_cast_fp16_12, var_7017_cast_fp16))[name = tensor("op_7050_cast_fp16")]; + tensor var_7052_equation_0 = const()[name = tensor("op_7052_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7052_cast_fp16 = einsum(equation = var_7052_equation_0, values = (var_6944_cast_fp16_13, var_7018_cast_fp16))[name = tensor("op_7052_cast_fp16")]; + tensor var_7054_equation_0 = const()[name = tensor("op_7054_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7054_cast_fp16 = einsum(equation = var_7054_equation_0, values = (var_6944_cast_fp16_14, var_7019_cast_fp16))[name = tensor("op_7054_cast_fp16")]; + tensor var_7056_equation_0 = const()[name = tensor("op_7056_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7056_cast_fp16 = einsum(equation = var_7056_equation_0, values = (var_6944_cast_fp16_15, var_7020_cast_fp16))[name = tensor("op_7056_cast_fp16")]; + tensor var_7058_equation_0 = const()[name = tensor("op_7058_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7058_cast_fp16 = einsum(equation = var_7058_equation_0, values = (var_6944_cast_fp16_16, var_7021_cast_fp16))[name = tensor("op_7058_cast_fp16")]; + tensor var_7060_equation_0 = const()[name = tensor("op_7060_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7060_cast_fp16 = einsum(equation = var_7060_equation_0, values = (var_6944_cast_fp16_17, var_7022_cast_fp16))[name = tensor("op_7060_cast_fp16")]; + tensor var_7062_equation_0 = const()[name = tensor("op_7062_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7062_cast_fp16 = einsum(equation = var_7062_equation_0, values = (var_6944_cast_fp16_18, var_7023_cast_fp16))[name = tensor("op_7062_cast_fp16")]; + tensor var_7064_equation_0 = const()[name = tensor("op_7064_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7064_cast_fp16 = einsum(equation = var_7064_equation_0, values = (var_6944_cast_fp16_19, var_7024_cast_fp16))[name = tensor("op_7064_cast_fp16")]; + tensor input_255_interleave_0 = const()[name = tensor("input_255_interleave_0"), val = tensor(false)]; + tensor input_255_cast_fp16 = concat(axis = var_6849, interleave = input_255_interleave_0, values = (var_7026_cast_fp16, var_7028_cast_fp16, var_7030_cast_fp16, var_7032_cast_fp16, var_7034_cast_fp16, var_7036_cast_fp16, var_7038_cast_fp16, var_7040_cast_fp16, var_7042_cast_fp16, var_7044_cast_fp16, var_7046_cast_fp16, var_7048_cast_fp16, var_7050_cast_fp16, var_7052_cast_fp16, var_7054_cast_fp16, var_7056_cast_fp16, var_7058_cast_fp16, var_7060_cast_fp16, var_7062_cast_fp16, var_7064_cast_fp16))[name = tensor("input_255_cast_fp16")]; + tensor var_7073_pad_type_0 = const()[name = tensor("op_7073_pad_type_0"), val = tensor("valid")]; + tensor var_7073_strides_0 = const()[name = tensor("op_7073_strides_0"), val = tensor([1, 1])]; + tensor var_7073_pad_0 = const()[name = tensor("op_7073_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7073_dilations_0 = const()[name = tensor("op_7073_dilations_0"), val = tensor([1, 1])]; + tensor var_7073_groups_0 = const()[name = tensor("op_7073_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_out_weight_to_fp16 = const()[name = tensor("blocks_25_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1007963392)))]; + tensor blocks_25_attn_out_bias_to_fp16 = const()[name = tensor("blocks_25_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011240256)))]; + tensor var_7073_cast_fp16 = conv(bias = blocks_25_attn_out_bias_to_fp16, dilations = var_7073_dilations_0, groups = var_7073_groups_0, pad = var_7073_pad_0, pad_type = var_7073_pad_type_0, strides = var_7073_strides_0, weight = blocks_25_attn_out_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("op_7073_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_7073_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([1])]; + tensor input_257_gamma_0_to_fp16 = const()[name = tensor("input_257_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011242880)))]; + tensor input_257_beta_0_to_fp16 = const()[name = tensor("input_257_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011245504)))]; + tensor var_7083_to_fp16 = const()[name = tensor("op_7083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = input_257_beta_0_to_fp16, epsilon = var_7083_to_fp16, gamma = input_257_gamma_0_to_fp16, x = inputs_103_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("valid")]; + tensor input_259_strides_0 = const()[name = tensor("input_259_strides_0"), val = tensor([1, 1])]; + tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_259_dilations_0 = const()[name = tensor("input_259_dilations_0"), val = tensor([1, 1])]; + tensor input_259_groups_0 = const()[name = tensor("input_259_groups_0"), val = tensor(1)]; + tensor blocks_25_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011248128)))]; + tensor blocks_25_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024355392)))]; + tensor input_259_cast_fp16 = conv(bias = blocks_25_mlp_0_bias_to_fp16, dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = blocks_25_mlp_0_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_mode_0 = const()[name = tensor("input_261_mode_0"), val = tensor("EXACT")]; + tensor input_261_cast_fp16 = gelu(mode = input_261_mode_0, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_7109_pad_type_0 = const()[name = tensor("op_7109_pad_type_0"), val = tensor("valid")]; + tensor var_7109_strides_0 = const()[name = tensor("op_7109_strides_0"), val = tensor([1, 1])]; + tensor var_7109_pad_0 = const()[name = tensor("op_7109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7109_dilations_0 = const()[name = tensor("op_7109_dilations_0"), val = tensor([1, 1])]; + tensor var_7109_groups_0 = const()[name = tensor("op_7109_groups_0"), val = tensor(1)]; + tensor blocks_25_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024365696)))]; + tensor blocks_25_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037472960)))]; + tensor var_7109_cast_fp16 = conv(bias = blocks_25_mlp_2_bias_to_fp16, dilations = var_7109_dilations_0, groups = var_7109_groups_0, pad = var_7109_pad_0, pad_type = var_7109_pad_type_0, strides = var_7109_strides_0, weight = blocks_25_mlp_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("op_7109_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = var_7109_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_7118 = const()[name = tensor("op_7118"), val = tensor(1)]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([1])]; + tensor input_263_gamma_0_to_fp16 = const()[name = tensor("input_263_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037475584)))]; + tensor input_263_beta_0_to_fp16 = const()[name = tensor("input_263_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037478208)))]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = input_263_beta_0_to_fp16, epsilon = var_7134_to_fp16, gamma = input_263_gamma_0_to_fp16, x = inputs_105_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor q_53_pad_type_0 = const()[name = tensor("q_53_pad_type_0"), val = tensor("valid")]; + tensor q_53_strides_0 = const()[name = tensor("q_53_strides_0"), val = tensor([1, 1])]; + tensor q_53_pad_0 = const()[name = tensor("q_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_53_dilations_0 = const()[name = tensor("q_53_dilations_0"), val = tensor([1, 1])]; + tensor q_53_groups_0 = const()[name = tensor("q_53_groups_0"), val = tensor(1)]; + tensor var_7169_weight_0_to_fp16 = const()[name = tensor("op_7169_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037480832)))]; + tensor var_7169_bias_0_to_fp16 = const()[name = tensor("op_7169_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040757696)))]; + tensor var_7169_cast_fp16 = conv(bias = var_7169_bias_0_to_fp16, dilations = q_53_dilations_0, groups = q_53_groups_0, pad = q_53_pad_0, pad_type = q_53_pad_type_0, strides = q_53_strides_0, weight = var_7169_weight_0_to_fp16, x = input_263_cast_fp16)[name = tensor("op_7169_cast_fp16")]; + tensor k_53_pad_type_0 = const()[name = tensor("k_53_pad_type_0"), val = tensor("valid")]; + tensor k_53_strides_0 = const()[name = tensor("k_53_strides_0"), val = tensor([1, 1])]; + tensor k_53_pad_0 = const()[name = tensor("k_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_53_dilations_0 = const()[name = tensor("k_53_dilations_0"), val = tensor([1, 1])]; + tensor k_53_groups_0 = const()[name = tensor("k_53_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_key_weight_to_fp16 = const()[name = tensor("blocks_26_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040760320)))]; + tensor k_53_cast_fp16 = conv(dilations = k_53_dilations_0, groups = k_53_groups_0, pad = k_53_pad_0, pad_type = k_53_pad_type_0, strides = k_53_strides_0, weight = blocks_26_attn_key_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor var_7167_pad_type_0 = const()[name = tensor("op_7167_pad_type_0"), val = tensor("valid")]; + tensor var_7167_strides_0 = const()[name = tensor("op_7167_strides_0"), val = tensor([1, 1])]; + tensor var_7167_pad_0 = const()[name = tensor("op_7167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7167_dilations_0 = const()[name = tensor("op_7167_dilations_0"), val = tensor([1, 1])]; + tensor var_7167_groups_0 = const()[name = tensor("op_7167_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_value_weight_to_fp16 = const()[name = tensor("blocks_26_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044037184)))]; + tensor blocks_26_attn_value_bias_to_fp16 = const()[name = tensor("blocks_26_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047314048)))]; + tensor var_7167_cast_fp16 = conv(bias = blocks_26_attn_value_bias_to_fp16, dilations = var_7167_dilations_0, groups = var_7167_groups_0, pad = var_7167_pad_0, pad_type = var_7167_pad_type_0, strides = var_7167_strides_0, weight = blocks_26_attn_value_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("op_7167_cast_fp16")]; + tensor tile_78 = const()[name = tensor("tile_78"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7170_axis_0 = const()[name = tensor("op_7170_axis_0"), val = tensor(1)]; + tensor var_7170_cast_fp16_0, tensor var_7170_cast_fp16_1, tensor var_7170_cast_fp16_2, tensor var_7170_cast_fp16_3, tensor var_7170_cast_fp16_4, tensor var_7170_cast_fp16_5, tensor var_7170_cast_fp16_6, tensor var_7170_cast_fp16_7, tensor var_7170_cast_fp16_8, tensor var_7170_cast_fp16_9, tensor var_7170_cast_fp16_10, tensor var_7170_cast_fp16_11, tensor var_7170_cast_fp16_12, tensor var_7170_cast_fp16_13, tensor var_7170_cast_fp16_14, tensor var_7170_cast_fp16_15, tensor var_7170_cast_fp16_16, tensor var_7170_cast_fp16_17, tensor var_7170_cast_fp16_18, tensor var_7170_cast_fp16_19 = split(axis = var_7170_axis_0, split_sizes = tile_78, x = var_7169_cast_fp16)[name = tensor("op_7170_cast_fp16")]; + tensor var_7191_perm_0 = const()[name = tensor("op_7191_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_79 = const()[name = tensor("tile_79"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7192_axis_0 = const()[name = tensor("op_7192_axis_0"), val = tensor(3)]; + tensor var_7191_cast_fp16 = transpose(perm = var_7191_perm_0, x = k_53_cast_fp16)[name = tensor("transpose_6")]; + tensor var_7192_cast_fp16_0, tensor var_7192_cast_fp16_1, tensor var_7192_cast_fp16_2, tensor var_7192_cast_fp16_3, tensor var_7192_cast_fp16_4, tensor var_7192_cast_fp16_5, tensor var_7192_cast_fp16_6, tensor var_7192_cast_fp16_7, tensor var_7192_cast_fp16_8, tensor var_7192_cast_fp16_9, tensor var_7192_cast_fp16_10, tensor var_7192_cast_fp16_11, tensor var_7192_cast_fp16_12, tensor var_7192_cast_fp16_13, tensor var_7192_cast_fp16_14, tensor var_7192_cast_fp16_15, tensor var_7192_cast_fp16_16, tensor var_7192_cast_fp16_17, tensor var_7192_cast_fp16_18, tensor var_7192_cast_fp16_19 = split(axis = var_7192_axis_0, split_sizes = tile_79, x = var_7191_cast_fp16)[name = tensor("op_7192_cast_fp16")]; + tensor tile_80 = const()[name = tensor("tile_80"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7213_axis_0 = const()[name = tensor("op_7213_axis_0"), val = tensor(1)]; + tensor var_7213_cast_fp16_0, tensor var_7213_cast_fp16_1, tensor var_7213_cast_fp16_2, tensor var_7213_cast_fp16_3, tensor var_7213_cast_fp16_4, tensor var_7213_cast_fp16_5, tensor var_7213_cast_fp16_6, tensor var_7213_cast_fp16_7, tensor var_7213_cast_fp16_8, tensor var_7213_cast_fp16_9, tensor var_7213_cast_fp16_10, tensor var_7213_cast_fp16_11, tensor var_7213_cast_fp16_12, tensor var_7213_cast_fp16_13, tensor var_7213_cast_fp16_14, tensor var_7213_cast_fp16_15, tensor var_7213_cast_fp16_16, tensor var_7213_cast_fp16_17, tensor var_7213_cast_fp16_18, tensor var_7213_cast_fp16_19 = split(axis = var_7213_axis_0, split_sizes = tile_80, x = var_7167_cast_fp16)[name = tensor("op_7213_cast_fp16")]; + tensor aw_1041_equation_0 = const()[name = tensor("aw_1041_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1041_cast_fp16 = einsum(equation = aw_1041_equation_0, values = (var_7192_cast_fp16_0, var_7170_cast_fp16_0))[name = tensor("aw_1041_cast_fp16")]; + tensor aw_1043_equation_0 = const()[name = tensor("aw_1043_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1043_cast_fp16 = einsum(equation = aw_1043_equation_0, values = (var_7192_cast_fp16_1, var_7170_cast_fp16_1))[name = tensor("aw_1043_cast_fp16")]; + tensor aw_1045_equation_0 = const()[name = tensor("aw_1045_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1045_cast_fp16 = einsum(equation = aw_1045_equation_0, values = (var_7192_cast_fp16_2, var_7170_cast_fp16_2))[name = tensor("aw_1045_cast_fp16")]; + tensor aw_1047_equation_0 = const()[name = tensor("aw_1047_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1047_cast_fp16 = einsum(equation = aw_1047_equation_0, values = (var_7192_cast_fp16_3, var_7170_cast_fp16_3))[name = tensor("aw_1047_cast_fp16")]; + tensor aw_1049_equation_0 = const()[name = tensor("aw_1049_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1049_cast_fp16 = einsum(equation = aw_1049_equation_0, values = (var_7192_cast_fp16_4, var_7170_cast_fp16_4))[name = tensor("aw_1049_cast_fp16")]; + tensor aw_1051_equation_0 = const()[name = tensor("aw_1051_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1051_cast_fp16 = einsum(equation = aw_1051_equation_0, values = (var_7192_cast_fp16_5, var_7170_cast_fp16_5))[name = tensor("aw_1051_cast_fp16")]; + tensor aw_1053_equation_0 = const()[name = tensor("aw_1053_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1053_cast_fp16 = einsum(equation = aw_1053_equation_0, values = (var_7192_cast_fp16_6, var_7170_cast_fp16_6))[name = tensor("aw_1053_cast_fp16")]; + tensor aw_1055_equation_0 = const()[name = tensor("aw_1055_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1055_cast_fp16 = einsum(equation = aw_1055_equation_0, values = (var_7192_cast_fp16_7, var_7170_cast_fp16_7))[name = tensor("aw_1055_cast_fp16")]; + tensor aw_1057_equation_0 = const()[name = tensor("aw_1057_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1057_cast_fp16 = einsum(equation = aw_1057_equation_0, values = (var_7192_cast_fp16_8, var_7170_cast_fp16_8))[name = tensor("aw_1057_cast_fp16")]; + tensor aw_1059_equation_0 = const()[name = tensor("aw_1059_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1059_cast_fp16 = einsum(equation = aw_1059_equation_0, values = (var_7192_cast_fp16_9, var_7170_cast_fp16_9))[name = tensor("aw_1059_cast_fp16")]; + tensor aw_1061_equation_0 = const()[name = tensor("aw_1061_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1061_cast_fp16 = einsum(equation = aw_1061_equation_0, values = (var_7192_cast_fp16_10, var_7170_cast_fp16_10))[name = tensor("aw_1061_cast_fp16")]; + tensor aw_1063_equation_0 = const()[name = tensor("aw_1063_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1063_cast_fp16 = einsum(equation = aw_1063_equation_0, values = (var_7192_cast_fp16_11, var_7170_cast_fp16_11))[name = tensor("aw_1063_cast_fp16")]; + tensor aw_1065_equation_0 = const()[name = tensor("aw_1065_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1065_cast_fp16 = einsum(equation = aw_1065_equation_0, values = (var_7192_cast_fp16_12, var_7170_cast_fp16_12))[name = tensor("aw_1065_cast_fp16")]; + tensor aw_1067_equation_0 = const()[name = tensor("aw_1067_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1067_cast_fp16 = einsum(equation = aw_1067_equation_0, values = (var_7192_cast_fp16_13, var_7170_cast_fp16_13))[name = tensor("aw_1067_cast_fp16")]; + tensor aw_1069_equation_0 = const()[name = tensor("aw_1069_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1069_cast_fp16 = einsum(equation = aw_1069_equation_0, values = (var_7192_cast_fp16_14, var_7170_cast_fp16_14))[name = tensor("aw_1069_cast_fp16")]; + tensor aw_1071_equation_0 = const()[name = tensor("aw_1071_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1071_cast_fp16 = einsum(equation = aw_1071_equation_0, values = (var_7192_cast_fp16_15, var_7170_cast_fp16_15))[name = tensor("aw_1071_cast_fp16")]; + tensor aw_1073_equation_0 = const()[name = tensor("aw_1073_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1073_cast_fp16 = einsum(equation = aw_1073_equation_0, values = (var_7192_cast_fp16_16, var_7170_cast_fp16_16))[name = tensor("aw_1073_cast_fp16")]; + tensor aw_1075_equation_0 = const()[name = tensor("aw_1075_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1075_cast_fp16 = einsum(equation = aw_1075_equation_0, values = (var_7192_cast_fp16_17, var_7170_cast_fp16_17))[name = tensor("aw_1075_cast_fp16")]; + tensor aw_1077_equation_0 = const()[name = tensor("aw_1077_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1077_cast_fp16 = einsum(equation = aw_1077_equation_0, values = (var_7192_cast_fp16_18, var_7170_cast_fp16_18))[name = tensor("aw_1077_cast_fp16")]; + tensor aw_1079_equation_0 = const()[name = tensor("aw_1079_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1079_cast_fp16 = einsum(equation = aw_1079_equation_0, values = (var_7192_cast_fp16_19, var_7170_cast_fp16_19))[name = tensor("aw_1079_cast_fp16")]; + tensor var_7274_cast_fp16 = softmax(axis = var_7118, x = aw_1041_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor var_7275_cast_fp16 = softmax(axis = var_7118, x = aw_1043_cast_fp16)[name = tensor("op_7275_cast_fp16")]; + tensor var_7276_cast_fp16 = softmax(axis = var_7118, x = aw_1045_cast_fp16)[name = tensor("op_7276_cast_fp16")]; + tensor var_7277_cast_fp16 = softmax(axis = var_7118, x = aw_1047_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7278_cast_fp16 = softmax(axis = var_7118, x = aw_1049_cast_fp16)[name = tensor("op_7278_cast_fp16")]; + tensor var_7279_cast_fp16 = softmax(axis = var_7118, x = aw_1051_cast_fp16)[name = tensor("op_7279_cast_fp16")]; + tensor var_7280_cast_fp16 = softmax(axis = var_7118, x = aw_1053_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor var_7281_cast_fp16 = softmax(axis = var_7118, x = aw_1055_cast_fp16)[name = tensor("op_7281_cast_fp16")]; + tensor var_7282_cast_fp16 = softmax(axis = var_7118, x = aw_1057_cast_fp16)[name = tensor("op_7282_cast_fp16")]; + tensor var_7283_cast_fp16 = softmax(axis = var_7118, x = aw_1059_cast_fp16)[name = tensor("op_7283_cast_fp16")]; + tensor var_7284_cast_fp16 = softmax(axis = var_7118, x = aw_1061_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = softmax(axis = var_7118, x = aw_1063_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7286_cast_fp16 = softmax(axis = var_7118, x = aw_1065_cast_fp16)[name = tensor("op_7286_cast_fp16")]; + tensor var_7287_cast_fp16 = softmax(axis = var_7118, x = aw_1067_cast_fp16)[name = tensor("op_7287_cast_fp16")]; + tensor var_7288_cast_fp16 = softmax(axis = var_7118, x = aw_1069_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = softmax(axis = var_7118, x = aw_1071_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor var_7290_cast_fp16 = softmax(axis = var_7118, x = aw_1073_cast_fp16)[name = tensor("op_7290_cast_fp16")]; + tensor var_7291_cast_fp16 = softmax(axis = var_7118, x = aw_1075_cast_fp16)[name = tensor("op_7291_cast_fp16")]; + tensor var_7292_cast_fp16 = softmax(axis = var_7118, x = aw_1077_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = softmax(axis = var_7118, x = aw_1079_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7295_equation_0 = const()[name = tensor("op_7295_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7295_cast_fp16 = einsum(equation = var_7295_equation_0, values = (var_7213_cast_fp16_0, var_7274_cast_fp16))[name = tensor("op_7295_cast_fp16")]; + tensor var_7297_equation_0 = const()[name = tensor("op_7297_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7297_cast_fp16 = einsum(equation = var_7297_equation_0, values = (var_7213_cast_fp16_1, var_7275_cast_fp16))[name = tensor("op_7297_cast_fp16")]; + tensor var_7299_equation_0 = const()[name = tensor("op_7299_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7299_cast_fp16 = einsum(equation = var_7299_equation_0, values = (var_7213_cast_fp16_2, var_7276_cast_fp16))[name = tensor("op_7299_cast_fp16")]; + tensor var_7301_equation_0 = const()[name = tensor("op_7301_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7301_cast_fp16 = einsum(equation = var_7301_equation_0, values = (var_7213_cast_fp16_3, var_7277_cast_fp16))[name = tensor("op_7301_cast_fp16")]; + tensor var_7303_equation_0 = const()[name = tensor("op_7303_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7303_cast_fp16 = einsum(equation = var_7303_equation_0, values = (var_7213_cast_fp16_4, var_7278_cast_fp16))[name = tensor("op_7303_cast_fp16")]; + tensor var_7305_equation_0 = const()[name = tensor("op_7305_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7305_cast_fp16 = einsum(equation = var_7305_equation_0, values = (var_7213_cast_fp16_5, var_7279_cast_fp16))[name = tensor("op_7305_cast_fp16")]; + tensor var_7307_equation_0 = const()[name = tensor("op_7307_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7307_cast_fp16 = einsum(equation = var_7307_equation_0, values = (var_7213_cast_fp16_6, var_7280_cast_fp16))[name = tensor("op_7307_cast_fp16")]; + tensor var_7309_equation_0 = const()[name = tensor("op_7309_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7309_cast_fp16 = einsum(equation = var_7309_equation_0, values = (var_7213_cast_fp16_7, var_7281_cast_fp16))[name = tensor("op_7309_cast_fp16")]; + tensor var_7311_equation_0 = const()[name = tensor("op_7311_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7311_cast_fp16 = einsum(equation = var_7311_equation_0, values = (var_7213_cast_fp16_8, var_7282_cast_fp16))[name = tensor("op_7311_cast_fp16")]; + tensor var_7313_equation_0 = const()[name = tensor("op_7313_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7313_cast_fp16 = einsum(equation = var_7313_equation_0, values = (var_7213_cast_fp16_9, var_7283_cast_fp16))[name = tensor("op_7313_cast_fp16")]; + tensor var_7315_equation_0 = const()[name = tensor("op_7315_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7315_cast_fp16 = einsum(equation = var_7315_equation_0, values = (var_7213_cast_fp16_10, var_7284_cast_fp16))[name = tensor("op_7315_cast_fp16")]; + tensor var_7317_equation_0 = const()[name = tensor("op_7317_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7317_cast_fp16 = einsum(equation = var_7317_equation_0, values = (var_7213_cast_fp16_11, var_7285_cast_fp16))[name = tensor("op_7317_cast_fp16")]; + tensor var_7319_equation_0 = const()[name = tensor("op_7319_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7319_cast_fp16 = einsum(equation = var_7319_equation_0, values = (var_7213_cast_fp16_12, var_7286_cast_fp16))[name = tensor("op_7319_cast_fp16")]; + tensor var_7321_equation_0 = const()[name = tensor("op_7321_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7321_cast_fp16 = einsum(equation = var_7321_equation_0, values = (var_7213_cast_fp16_13, var_7287_cast_fp16))[name = tensor("op_7321_cast_fp16")]; + tensor var_7323_equation_0 = const()[name = tensor("op_7323_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7323_cast_fp16 = einsum(equation = var_7323_equation_0, values = (var_7213_cast_fp16_14, var_7288_cast_fp16))[name = tensor("op_7323_cast_fp16")]; + tensor var_7325_equation_0 = const()[name = tensor("op_7325_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7325_cast_fp16 = einsum(equation = var_7325_equation_0, values = (var_7213_cast_fp16_15, var_7289_cast_fp16))[name = tensor("op_7325_cast_fp16")]; + tensor var_7327_equation_0 = const()[name = tensor("op_7327_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7327_cast_fp16 = einsum(equation = var_7327_equation_0, values = (var_7213_cast_fp16_16, var_7290_cast_fp16))[name = tensor("op_7327_cast_fp16")]; + tensor var_7329_equation_0 = const()[name = tensor("op_7329_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7329_cast_fp16 = einsum(equation = var_7329_equation_0, values = (var_7213_cast_fp16_17, var_7291_cast_fp16))[name = tensor("op_7329_cast_fp16")]; + tensor var_7331_equation_0 = const()[name = tensor("op_7331_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7331_cast_fp16 = einsum(equation = var_7331_equation_0, values = (var_7213_cast_fp16_18, var_7292_cast_fp16))[name = tensor("op_7331_cast_fp16")]; + tensor var_7333_equation_0 = const()[name = tensor("op_7333_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7333_cast_fp16 = einsum(equation = var_7333_equation_0, values = (var_7213_cast_fp16_19, var_7293_cast_fp16))[name = tensor("op_7333_cast_fp16")]; + tensor input_265_interleave_0 = const()[name = tensor("input_265_interleave_0"), val = tensor(false)]; + tensor input_265_cast_fp16 = concat(axis = var_7118, interleave = input_265_interleave_0, values = (var_7295_cast_fp16, var_7297_cast_fp16, var_7299_cast_fp16, var_7301_cast_fp16, var_7303_cast_fp16, var_7305_cast_fp16, var_7307_cast_fp16, var_7309_cast_fp16, var_7311_cast_fp16, var_7313_cast_fp16, var_7315_cast_fp16, var_7317_cast_fp16, var_7319_cast_fp16, var_7321_cast_fp16, var_7323_cast_fp16, var_7325_cast_fp16, var_7327_cast_fp16, var_7329_cast_fp16, var_7331_cast_fp16, var_7333_cast_fp16))[name = tensor("input_265_cast_fp16")]; + tensor var_7342_pad_type_0 = const()[name = tensor("op_7342_pad_type_0"), val = tensor("valid")]; + tensor var_7342_strides_0 = const()[name = tensor("op_7342_strides_0"), val = tensor([1, 1])]; + tensor var_7342_pad_0 = const()[name = tensor("op_7342_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7342_dilations_0 = const()[name = tensor("op_7342_dilations_0"), val = tensor([1, 1])]; + tensor var_7342_groups_0 = const()[name = tensor("op_7342_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_out_weight_to_fp16 = const()[name = tensor("blocks_26_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047316672)))]; + tensor blocks_26_attn_out_bias_to_fp16 = const()[name = tensor("blocks_26_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050593536)))]; + tensor var_7342_cast_fp16 = conv(bias = blocks_26_attn_out_bias_to_fp16, dilations = var_7342_dilations_0, groups = var_7342_groups_0, pad = var_7342_pad_0, pad_type = var_7342_pad_type_0, strides = var_7342_strides_0, weight = blocks_26_attn_out_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("op_7342_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = var_7342_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([1])]; + tensor input_267_gamma_0_to_fp16 = const()[name = tensor("input_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050596160)))]; + tensor input_267_beta_0_to_fp16 = const()[name = tensor("input_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050598784)))]; + tensor var_7352_to_fp16 = const()[name = tensor("op_7352_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = input_267_beta_0_to_fp16, epsilon = var_7352_to_fp16, gamma = input_267_gamma_0_to_fp16, x = inputs_107_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_pad_type_0 = const()[name = tensor("input_269_pad_type_0"), val = tensor("valid")]; + tensor input_269_strides_0 = const()[name = tensor("input_269_strides_0"), val = tensor([1, 1])]; + tensor input_269_pad_0 = const()[name = tensor("input_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_269_dilations_0 = const()[name = tensor("input_269_dilations_0"), val = tensor([1, 1])]; + tensor input_269_groups_0 = const()[name = tensor("input_269_groups_0"), val = tensor(1)]; + tensor blocks_26_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050601408)))]; + tensor blocks_26_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063708672)))]; + tensor input_269_cast_fp16 = conv(bias = blocks_26_mlp_0_bias_to_fp16, dilations = input_269_dilations_0, groups = input_269_groups_0, pad = input_269_pad_0, pad_type = input_269_pad_type_0, strides = input_269_strides_0, weight = blocks_26_mlp_0_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor input_271_mode_0 = const()[name = tensor("input_271_mode_0"), val = tensor("EXACT")]; + tensor input_271_cast_fp16 = gelu(mode = input_271_mode_0, x = input_269_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor var_7378_pad_type_0 = const()[name = tensor("op_7378_pad_type_0"), val = tensor("valid")]; + tensor var_7378_strides_0 = const()[name = tensor("op_7378_strides_0"), val = tensor([1, 1])]; + tensor var_7378_pad_0 = const()[name = tensor("op_7378_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7378_dilations_0 = const()[name = tensor("op_7378_dilations_0"), val = tensor([1, 1])]; + tensor var_7378_groups_0 = const()[name = tensor("op_7378_groups_0"), val = tensor(1)]; + tensor blocks_26_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063718976)))]; + tensor blocks_26_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076826240)))]; + tensor var_7378_cast_fp16 = conv(bias = blocks_26_mlp_2_bias_to_fp16, dilations = var_7378_dilations_0, groups = var_7378_groups_0, pad = var_7378_pad_0, pad_type = var_7378_pad_type_0, strides = var_7378_strides_0, weight = blocks_26_mlp_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("op_7378_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_7378_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_7387 = const()[name = tensor("op_7387"), val = tensor(1)]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([1])]; + tensor input_273_gamma_0_to_fp16 = const()[name = tensor("input_273_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076828864)))]; + tensor input_273_beta_0_to_fp16 = const()[name = tensor("input_273_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076831488)))]; + tensor var_7403_to_fp16 = const()[name = tensor("op_7403_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = input_273_beta_0_to_fp16, epsilon = var_7403_to_fp16, gamma = input_273_gamma_0_to_fp16, x = inputs_109_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor q_55_pad_type_0 = const()[name = tensor("q_55_pad_type_0"), val = tensor("valid")]; + tensor q_55_strides_0 = const()[name = tensor("q_55_strides_0"), val = tensor([1, 1])]; + tensor q_55_pad_0 = const()[name = tensor("q_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_55_dilations_0 = const()[name = tensor("q_55_dilations_0"), val = tensor([1, 1])]; + tensor q_55_groups_0 = const()[name = tensor("q_55_groups_0"), val = tensor(1)]; + tensor var_7438_weight_0_to_fp16 = const()[name = tensor("op_7438_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076834112)))]; + tensor var_7438_bias_0_to_fp16 = const()[name = tensor("op_7438_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080110976)))]; + tensor var_7438_cast_fp16 = conv(bias = var_7438_bias_0_to_fp16, dilations = q_55_dilations_0, groups = q_55_groups_0, pad = q_55_pad_0, pad_type = q_55_pad_type_0, strides = q_55_strides_0, weight = var_7438_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("op_7438_cast_fp16")]; + tensor k_55_pad_type_0 = const()[name = tensor("k_55_pad_type_0"), val = tensor("valid")]; + tensor k_55_strides_0 = const()[name = tensor("k_55_strides_0"), val = tensor([1, 1])]; + tensor k_55_pad_0 = const()[name = tensor("k_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_55_dilations_0 = const()[name = tensor("k_55_dilations_0"), val = tensor([1, 1])]; + tensor k_55_groups_0 = const()[name = tensor("k_55_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_key_weight_to_fp16 = const()[name = tensor("blocks_27_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080113600)))]; + tensor k_55_cast_fp16 = conv(dilations = k_55_dilations_0, groups = k_55_groups_0, pad = k_55_pad_0, pad_type = k_55_pad_type_0, strides = k_55_strides_0, weight = blocks_27_attn_key_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor var_7436_pad_type_0 = const()[name = tensor("op_7436_pad_type_0"), val = tensor("valid")]; + tensor var_7436_strides_0 = const()[name = tensor("op_7436_strides_0"), val = tensor([1, 1])]; + tensor var_7436_pad_0 = const()[name = tensor("op_7436_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7436_dilations_0 = const()[name = tensor("op_7436_dilations_0"), val = tensor([1, 1])]; + tensor var_7436_groups_0 = const()[name = tensor("op_7436_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_value_weight_to_fp16 = const()[name = tensor("blocks_27_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083390464)))]; + tensor blocks_27_attn_value_bias_to_fp16 = const()[name = tensor("blocks_27_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086667328)))]; + tensor var_7436_cast_fp16 = conv(bias = blocks_27_attn_value_bias_to_fp16, dilations = var_7436_dilations_0, groups = var_7436_groups_0, pad = var_7436_pad_0, pad_type = var_7436_pad_type_0, strides = var_7436_strides_0, weight = blocks_27_attn_value_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("op_7436_cast_fp16")]; + tensor tile_81 = const()[name = tensor("tile_81"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_cast_fp16_0, tensor var_7439_cast_fp16_1, tensor var_7439_cast_fp16_2, tensor var_7439_cast_fp16_3, tensor var_7439_cast_fp16_4, tensor var_7439_cast_fp16_5, tensor var_7439_cast_fp16_6, tensor var_7439_cast_fp16_7, tensor var_7439_cast_fp16_8, tensor var_7439_cast_fp16_9, tensor var_7439_cast_fp16_10, tensor var_7439_cast_fp16_11, tensor var_7439_cast_fp16_12, tensor var_7439_cast_fp16_13, tensor var_7439_cast_fp16_14, tensor var_7439_cast_fp16_15, tensor var_7439_cast_fp16_16, tensor var_7439_cast_fp16_17, tensor var_7439_cast_fp16_18, tensor var_7439_cast_fp16_19 = split(axis = var_7439_axis_0, split_sizes = tile_81, x = var_7438_cast_fp16)[name = tensor("op_7439_cast_fp16")]; + tensor var_7460_perm_0 = const()[name = tensor("op_7460_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_82 = const()[name = tensor("tile_82"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7461_axis_0 = const()[name = tensor("op_7461_axis_0"), val = tensor(3)]; + tensor var_7460_cast_fp16 = transpose(perm = var_7460_perm_0, x = k_55_cast_fp16)[name = tensor("transpose_5")]; + tensor var_7461_cast_fp16_0, tensor var_7461_cast_fp16_1, tensor var_7461_cast_fp16_2, tensor var_7461_cast_fp16_3, tensor var_7461_cast_fp16_4, tensor var_7461_cast_fp16_5, tensor var_7461_cast_fp16_6, tensor var_7461_cast_fp16_7, tensor var_7461_cast_fp16_8, tensor var_7461_cast_fp16_9, tensor var_7461_cast_fp16_10, tensor var_7461_cast_fp16_11, tensor var_7461_cast_fp16_12, tensor var_7461_cast_fp16_13, tensor var_7461_cast_fp16_14, tensor var_7461_cast_fp16_15, tensor var_7461_cast_fp16_16, tensor var_7461_cast_fp16_17, tensor var_7461_cast_fp16_18, tensor var_7461_cast_fp16_19 = split(axis = var_7461_axis_0, split_sizes = tile_82, x = var_7460_cast_fp16)[name = tensor("op_7461_cast_fp16")]; + tensor tile_83 = const()[name = tensor("tile_83"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7482_axis_0 = const()[name = tensor("op_7482_axis_0"), val = tensor(1)]; + tensor var_7482_cast_fp16_0, tensor var_7482_cast_fp16_1, tensor var_7482_cast_fp16_2, tensor var_7482_cast_fp16_3, tensor var_7482_cast_fp16_4, tensor var_7482_cast_fp16_5, tensor var_7482_cast_fp16_6, tensor var_7482_cast_fp16_7, tensor var_7482_cast_fp16_8, tensor var_7482_cast_fp16_9, tensor var_7482_cast_fp16_10, tensor var_7482_cast_fp16_11, tensor var_7482_cast_fp16_12, tensor var_7482_cast_fp16_13, tensor var_7482_cast_fp16_14, tensor var_7482_cast_fp16_15, tensor var_7482_cast_fp16_16, tensor var_7482_cast_fp16_17, tensor var_7482_cast_fp16_18, tensor var_7482_cast_fp16_19 = split(axis = var_7482_axis_0, split_sizes = tile_83, x = var_7436_cast_fp16)[name = tensor("op_7482_cast_fp16")]; + tensor aw_1081_equation_0 = const()[name = tensor("aw_1081_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1081_cast_fp16 = einsum(equation = aw_1081_equation_0, values = (var_7461_cast_fp16_0, var_7439_cast_fp16_0))[name = tensor("aw_1081_cast_fp16")]; + tensor aw_1083_equation_0 = const()[name = tensor("aw_1083_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1083_cast_fp16 = einsum(equation = aw_1083_equation_0, values = (var_7461_cast_fp16_1, var_7439_cast_fp16_1))[name = tensor("aw_1083_cast_fp16")]; + tensor aw_1085_equation_0 = const()[name = tensor("aw_1085_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1085_cast_fp16 = einsum(equation = aw_1085_equation_0, values = (var_7461_cast_fp16_2, var_7439_cast_fp16_2))[name = tensor("aw_1085_cast_fp16")]; + tensor aw_1087_equation_0 = const()[name = tensor("aw_1087_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1087_cast_fp16 = einsum(equation = aw_1087_equation_0, values = (var_7461_cast_fp16_3, var_7439_cast_fp16_3))[name = tensor("aw_1087_cast_fp16")]; + tensor aw_1089_equation_0 = const()[name = tensor("aw_1089_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1089_cast_fp16 = einsum(equation = aw_1089_equation_0, values = (var_7461_cast_fp16_4, var_7439_cast_fp16_4))[name = tensor("aw_1089_cast_fp16")]; + tensor aw_1091_equation_0 = const()[name = tensor("aw_1091_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1091_cast_fp16 = einsum(equation = aw_1091_equation_0, values = (var_7461_cast_fp16_5, var_7439_cast_fp16_5))[name = tensor("aw_1091_cast_fp16")]; + tensor aw_1093_equation_0 = const()[name = tensor("aw_1093_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1093_cast_fp16 = einsum(equation = aw_1093_equation_0, values = (var_7461_cast_fp16_6, var_7439_cast_fp16_6))[name = tensor("aw_1093_cast_fp16")]; + tensor aw_1095_equation_0 = const()[name = tensor("aw_1095_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1095_cast_fp16 = einsum(equation = aw_1095_equation_0, values = (var_7461_cast_fp16_7, var_7439_cast_fp16_7))[name = tensor("aw_1095_cast_fp16")]; + tensor aw_1097_equation_0 = const()[name = tensor("aw_1097_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1097_cast_fp16 = einsum(equation = aw_1097_equation_0, values = (var_7461_cast_fp16_8, var_7439_cast_fp16_8))[name = tensor("aw_1097_cast_fp16")]; + tensor aw_1099_equation_0 = const()[name = tensor("aw_1099_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1099_cast_fp16 = einsum(equation = aw_1099_equation_0, values = (var_7461_cast_fp16_9, var_7439_cast_fp16_9))[name = tensor("aw_1099_cast_fp16")]; + tensor aw_1101_equation_0 = const()[name = tensor("aw_1101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1101_cast_fp16 = einsum(equation = aw_1101_equation_0, values = (var_7461_cast_fp16_10, var_7439_cast_fp16_10))[name = tensor("aw_1101_cast_fp16")]; + tensor aw_1103_equation_0 = const()[name = tensor("aw_1103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1103_cast_fp16 = einsum(equation = aw_1103_equation_0, values = (var_7461_cast_fp16_11, var_7439_cast_fp16_11))[name = tensor("aw_1103_cast_fp16")]; + tensor aw_1105_equation_0 = const()[name = tensor("aw_1105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1105_cast_fp16 = einsum(equation = aw_1105_equation_0, values = (var_7461_cast_fp16_12, var_7439_cast_fp16_12))[name = tensor("aw_1105_cast_fp16")]; + tensor aw_1107_equation_0 = const()[name = tensor("aw_1107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1107_cast_fp16 = einsum(equation = aw_1107_equation_0, values = (var_7461_cast_fp16_13, var_7439_cast_fp16_13))[name = tensor("aw_1107_cast_fp16")]; + tensor aw_1109_equation_0 = const()[name = tensor("aw_1109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1109_cast_fp16 = einsum(equation = aw_1109_equation_0, values = (var_7461_cast_fp16_14, var_7439_cast_fp16_14))[name = tensor("aw_1109_cast_fp16")]; + tensor aw_1111_equation_0 = const()[name = tensor("aw_1111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1111_cast_fp16 = einsum(equation = aw_1111_equation_0, values = (var_7461_cast_fp16_15, var_7439_cast_fp16_15))[name = tensor("aw_1111_cast_fp16")]; + tensor aw_1113_equation_0 = const()[name = tensor("aw_1113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1113_cast_fp16 = einsum(equation = aw_1113_equation_0, values = (var_7461_cast_fp16_16, var_7439_cast_fp16_16))[name = tensor("aw_1113_cast_fp16")]; + tensor aw_1115_equation_0 = const()[name = tensor("aw_1115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1115_cast_fp16 = einsum(equation = aw_1115_equation_0, values = (var_7461_cast_fp16_17, var_7439_cast_fp16_17))[name = tensor("aw_1115_cast_fp16")]; + tensor aw_1117_equation_0 = const()[name = tensor("aw_1117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1117_cast_fp16 = einsum(equation = aw_1117_equation_0, values = (var_7461_cast_fp16_18, var_7439_cast_fp16_18))[name = tensor("aw_1117_cast_fp16")]; + tensor aw_1119_equation_0 = const()[name = tensor("aw_1119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1119_cast_fp16 = einsum(equation = aw_1119_equation_0, values = (var_7461_cast_fp16_19, var_7439_cast_fp16_19))[name = tensor("aw_1119_cast_fp16")]; + tensor var_7543_cast_fp16 = softmax(axis = var_7387, x = aw_1081_cast_fp16)[name = tensor("op_7543_cast_fp16")]; + tensor var_7544_cast_fp16 = softmax(axis = var_7387, x = aw_1083_cast_fp16)[name = tensor("op_7544_cast_fp16")]; + tensor var_7545_cast_fp16 = softmax(axis = var_7387, x = aw_1085_cast_fp16)[name = tensor("op_7545_cast_fp16")]; + tensor var_7546_cast_fp16 = softmax(axis = var_7387, x = aw_1087_cast_fp16)[name = tensor("op_7546_cast_fp16")]; + tensor var_7547_cast_fp16 = softmax(axis = var_7387, x = aw_1089_cast_fp16)[name = tensor("op_7547_cast_fp16")]; + tensor var_7548_cast_fp16 = softmax(axis = var_7387, x = aw_1091_cast_fp16)[name = tensor("op_7548_cast_fp16")]; + tensor var_7549_cast_fp16 = softmax(axis = var_7387, x = aw_1093_cast_fp16)[name = tensor("op_7549_cast_fp16")]; + tensor var_7550_cast_fp16 = softmax(axis = var_7387, x = aw_1095_cast_fp16)[name = tensor("op_7550_cast_fp16")]; + tensor var_7551_cast_fp16 = softmax(axis = var_7387, x = aw_1097_cast_fp16)[name = tensor("op_7551_cast_fp16")]; + tensor var_7552_cast_fp16 = softmax(axis = var_7387, x = aw_1099_cast_fp16)[name = tensor("op_7552_cast_fp16")]; + tensor var_7553_cast_fp16 = softmax(axis = var_7387, x = aw_1101_cast_fp16)[name = tensor("op_7553_cast_fp16")]; + tensor var_7554_cast_fp16 = softmax(axis = var_7387, x = aw_1103_cast_fp16)[name = tensor("op_7554_cast_fp16")]; + tensor var_7555_cast_fp16 = softmax(axis = var_7387, x = aw_1105_cast_fp16)[name = tensor("op_7555_cast_fp16")]; + tensor var_7556_cast_fp16 = softmax(axis = var_7387, x = aw_1107_cast_fp16)[name = tensor("op_7556_cast_fp16")]; + tensor var_7557_cast_fp16 = softmax(axis = var_7387, x = aw_1109_cast_fp16)[name = tensor("op_7557_cast_fp16")]; + tensor var_7558_cast_fp16 = softmax(axis = var_7387, x = aw_1111_cast_fp16)[name = tensor("op_7558_cast_fp16")]; + tensor var_7559_cast_fp16 = softmax(axis = var_7387, x = aw_1113_cast_fp16)[name = tensor("op_7559_cast_fp16")]; + tensor var_7560_cast_fp16 = softmax(axis = var_7387, x = aw_1115_cast_fp16)[name = tensor("op_7560_cast_fp16")]; + tensor var_7561_cast_fp16 = softmax(axis = var_7387, x = aw_1117_cast_fp16)[name = tensor("op_7561_cast_fp16")]; + tensor var_7562_cast_fp16 = softmax(axis = var_7387, x = aw_1119_cast_fp16)[name = tensor("op_7562_cast_fp16")]; + tensor var_7564_equation_0 = const()[name = tensor("op_7564_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7564_cast_fp16 = einsum(equation = var_7564_equation_0, values = (var_7482_cast_fp16_0, var_7543_cast_fp16))[name = tensor("op_7564_cast_fp16")]; + tensor var_7566_equation_0 = const()[name = tensor("op_7566_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7566_cast_fp16 = einsum(equation = var_7566_equation_0, values = (var_7482_cast_fp16_1, var_7544_cast_fp16))[name = tensor("op_7566_cast_fp16")]; + tensor var_7568_equation_0 = const()[name = tensor("op_7568_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7568_cast_fp16 = einsum(equation = var_7568_equation_0, values = (var_7482_cast_fp16_2, var_7545_cast_fp16))[name = tensor("op_7568_cast_fp16")]; + tensor var_7570_equation_0 = const()[name = tensor("op_7570_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7570_cast_fp16 = einsum(equation = var_7570_equation_0, values = (var_7482_cast_fp16_3, var_7546_cast_fp16))[name = tensor("op_7570_cast_fp16")]; + tensor var_7572_equation_0 = const()[name = tensor("op_7572_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7572_cast_fp16 = einsum(equation = var_7572_equation_0, values = (var_7482_cast_fp16_4, var_7547_cast_fp16))[name = tensor("op_7572_cast_fp16")]; + tensor var_7574_equation_0 = const()[name = tensor("op_7574_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7574_cast_fp16 = einsum(equation = var_7574_equation_0, values = (var_7482_cast_fp16_5, var_7548_cast_fp16))[name = tensor("op_7574_cast_fp16")]; + tensor var_7576_equation_0 = const()[name = tensor("op_7576_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7576_cast_fp16 = einsum(equation = var_7576_equation_0, values = (var_7482_cast_fp16_6, var_7549_cast_fp16))[name = tensor("op_7576_cast_fp16")]; + tensor var_7578_equation_0 = const()[name = tensor("op_7578_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7578_cast_fp16 = einsum(equation = var_7578_equation_0, values = (var_7482_cast_fp16_7, var_7550_cast_fp16))[name = tensor("op_7578_cast_fp16")]; + tensor var_7580_equation_0 = const()[name = tensor("op_7580_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7580_cast_fp16 = einsum(equation = var_7580_equation_0, values = (var_7482_cast_fp16_8, var_7551_cast_fp16))[name = tensor("op_7580_cast_fp16")]; + tensor var_7582_equation_0 = const()[name = tensor("op_7582_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7582_cast_fp16 = einsum(equation = var_7582_equation_0, values = (var_7482_cast_fp16_9, var_7552_cast_fp16))[name = tensor("op_7582_cast_fp16")]; + tensor var_7584_equation_0 = const()[name = tensor("op_7584_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7584_cast_fp16 = einsum(equation = var_7584_equation_0, values = (var_7482_cast_fp16_10, var_7553_cast_fp16))[name = tensor("op_7584_cast_fp16")]; + tensor var_7586_equation_0 = const()[name = tensor("op_7586_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7586_cast_fp16 = einsum(equation = var_7586_equation_0, values = (var_7482_cast_fp16_11, var_7554_cast_fp16))[name = tensor("op_7586_cast_fp16")]; + tensor var_7588_equation_0 = const()[name = tensor("op_7588_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7588_cast_fp16 = einsum(equation = var_7588_equation_0, values = (var_7482_cast_fp16_12, var_7555_cast_fp16))[name = tensor("op_7588_cast_fp16")]; + tensor var_7590_equation_0 = const()[name = tensor("op_7590_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7590_cast_fp16 = einsum(equation = var_7590_equation_0, values = (var_7482_cast_fp16_13, var_7556_cast_fp16))[name = tensor("op_7590_cast_fp16")]; + tensor var_7592_equation_0 = const()[name = tensor("op_7592_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7592_cast_fp16 = einsum(equation = var_7592_equation_0, values = (var_7482_cast_fp16_14, var_7557_cast_fp16))[name = tensor("op_7592_cast_fp16")]; + tensor var_7594_equation_0 = const()[name = tensor("op_7594_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7594_cast_fp16 = einsum(equation = var_7594_equation_0, values = (var_7482_cast_fp16_15, var_7558_cast_fp16))[name = tensor("op_7594_cast_fp16")]; + tensor var_7596_equation_0 = const()[name = tensor("op_7596_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7596_cast_fp16 = einsum(equation = var_7596_equation_0, values = (var_7482_cast_fp16_16, var_7559_cast_fp16))[name = tensor("op_7596_cast_fp16")]; + tensor var_7598_equation_0 = const()[name = tensor("op_7598_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7598_cast_fp16 = einsum(equation = var_7598_equation_0, values = (var_7482_cast_fp16_17, var_7560_cast_fp16))[name = tensor("op_7598_cast_fp16")]; + tensor var_7600_equation_0 = const()[name = tensor("op_7600_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7600_cast_fp16 = einsum(equation = var_7600_equation_0, values = (var_7482_cast_fp16_18, var_7561_cast_fp16))[name = tensor("op_7600_cast_fp16")]; + tensor var_7602_equation_0 = const()[name = tensor("op_7602_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7602_cast_fp16 = einsum(equation = var_7602_equation_0, values = (var_7482_cast_fp16_19, var_7562_cast_fp16))[name = tensor("op_7602_cast_fp16")]; + tensor input_275_interleave_0 = const()[name = tensor("input_275_interleave_0"), val = tensor(false)]; + tensor input_275_cast_fp16 = concat(axis = var_7387, interleave = input_275_interleave_0, values = (var_7564_cast_fp16, var_7566_cast_fp16, var_7568_cast_fp16, var_7570_cast_fp16, var_7572_cast_fp16, var_7574_cast_fp16, var_7576_cast_fp16, var_7578_cast_fp16, var_7580_cast_fp16, var_7582_cast_fp16, var_7584_cast_fp16, var_7586_cast_fp16, var_7588_cast_fp16, var_7590_cast_fp16, var_7592_cast_fp16, var_7594_cast_fp16, var_7596_cast_fp16, var_7598_cast_fp16, var_7600_cast_fp16, var_7602_cast_fp16))[name = tensor("input_275_cast_fp16")]; + tensor var_7611_pad_type_0 = const()[name = tensor("op_7611_pad_type_0"), val = tensor("valid")]; + tensor var_7611_strides_0 = const()[name = tensor("op_7611_strides_0"), val = tensor([1, 1])]; + tensor var_7611_pad_0 = const()[name = tensor("op_7611_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7611_dilations_0 = const()[name = tensor("op_7611_dilations_0"), val = tensor([1, 1])]; + tensor var_7611_groups_0 = const()[name = tensor("op_7611_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_out_weight_to_fp16 = const()[name = tensor("blocks_27_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086669952)))]; + tensor blocks_27_attn_out_bias_to_fp16 = const()[name = tensor("blocks_27_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089946816)))]; + tensor var_7611_cast_fp16 = conv(bias = blocks_27_attn_out_bias_to_fp16, dilations = var_7611_dilations_0, groups = var_7611_groups_0, pad = var_7611_pad_0, pad_type = var_7611_pad_type_0, strides = var_7611_strides_0, weight = blocks_27_attn_out_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("op_7611_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = var_7611_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor input_277_axes_0 = const()[name = tensor("input_277_axes_0"), val = tensor([1])]; + tensor input_277_gamma_0_to_fp16 = const()[name = tensor("input_277_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089949440)))]; + tensor input_277_beta_0_to_fp16 = const()[name = tensor("input_277_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089952064)))]; + tensor var_7621_to_fp16 = const()[name = tensor("op_7621_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_277_cast_fp16 = layer_norm(axes = input_277_axes_0, beta = input_277_beta_0_to_fp16, epsilon = var_7621_to_fp16, gamma = input_277_gamma_0_to_fp16, x = inputs_111_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_pad_type_0 = const()[name = tensor("input_279_pad_type_0"), val = tensor("valid")]; + tensor input_279_strides_0 = const()[name = tensor("input_279_strides_0"), val = tensor([1, 1])]; + tensor input_279_pad_0 = const()[name = tensor("input_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_279_dilations_0 = const()[name = tensor("input_279_dilations_0"), val = tensor([1, 1])]; + tensor input_279_groups_0 = const()[name = tensor("input_279_groups_0"), val = tensor(1)]; + tensor blocks_27_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089954688)))]; + tensor blocks_27_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103061952)))]; + tensor input_279_cast_fp16 = conv(bias = blocks_27_mlp_0_bias_to_fp16, dilations = input_279_dilations_0, groups = input_279_groups_0, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = input_279_strides_0, weight = blocks_27_mlp_0_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_mode_0 = const()[name = tensor("input_281_mode_0"), val = tensor("EXACT")]; + tensor input_281_cast_fp16 = gelu(mode = input_281_mode_0, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_7647_pad_type_0 = const()[name = tensor("op_7647_pad_type_0"), val = tensor("valid")]; + tensor var_7647_strides_0 = const()[name = tensor("op_7647_strides_0"), val = tensor([1, 1])]; + tensor var_7647_pad_0 = const()[name = tensor("op_7647_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7647_dilations_0 = const()[name = tensor("op_7647_dilations_0"), val = tensor([1, 1])]; + tensor var_7647_groups_0 = const()[name = tensor("op_7647_groups_0"), val = tensor(1)]; + tensor blocks_27_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103072256)))]; + tensor blocks_27_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116179520)))]; + tensor var_7647_cast_fp16 = conv(bias = blocks_27_mlp_2_bias_to_fp16, dilations = var_7647_dilations_0, groups = var_7647_groups_0, pad = var_7647_pad_0, pad_type = var_7647_pad_type_0, strides = var_7647_strides_0, weight = blocks_27_mlp_2_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("op_7647_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_7647_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_7656 = const()[name = tensor("op_7656"), val = tensor(1)]; + tensor input_283_axes_0 = const()[name = tensor("input_283_axes_0"), val = tensor([1])]; + tensor input_283_gamma_0_to_fp16 = const()[name = tensor("input_283_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116182144)))]; + tensor input_283_beta_0_to_fp16 = const()[name = tensor("input_283_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116184768)))]; + tensor var_7672_to_fp16 = const()[name = tensor("op_7672_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = input_283_beta_0_to_fp16, epsilon = var_7672_to_fp16, gamma = input_283_gamma_0_to_fp16, x = inputs_113_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor q_57_pad_type_0 = const()[name = tensor("q_57_pad_type_0"), val = tensor("valid")]; + tensor q_57_strides_0 = const()[name = tensor("q_57_strides_0"), val = tensor([1, 1])]; + tensor q_57_pad_0 = const()[name = tensor("q_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_57_dilations_0 = const()[name = tensor("q_57_dilations_0"), val = tensor([1, 1])]; + tensor q_57_groups_0 = const()[name = tensor("q_57_groups_0"), val = tensor(1)]; + tensor var_7707_weight_0_to_fp16 = const()[name = tensor("op_7707_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116187392)))]; + tensor var_7707_bias_0_to_fp16 = const()[name = tensor("op_7707_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119464256)))]; + tensor var_7707_cast_fp16 = conv(bias = var_7707_bias_0_to_fp16, dilations = q_57_dilations_0, groups = q_57_groups_0, pad = q_57_pad_0, pad_type = q_57_pad_type_0, strides = q_57_strides_0, weight = var_7707_weight_0_to_fp16, x = input_283_cast_fp16)[name = tensor("op_7707_cast_fp16")]; + tensor k_57_pad_type_0 = const()[name = tensor("k_57_pad_type_0"), val = tensor("valid")]; + tensor k_57_strides_0 = const()[name = tensor("k_57_strides_0"), val = tensor([1, 1])]; + tensor k_57_pad_0 = const()[name = tensor("k_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_57_dilations_0 = const()[name = tensor("k_57_dilations_0"), val = tensor([1, 1])]; + tensor k_57_groups_0 = const()[name = tensor("k_57_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_key_weight_to_fp16 = const()[name = tensor("blocks_28_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119466880)))]; + tensor k_57_cast_fp16 = conv(dilations = k_57_dilations_0, groups = k_57_groups_0, pad = k_57_pad_0, pad_type = k_57_pad_type_0, strides = k_57_strides_0, weight = blocks_28_attn_key_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor var_7705_pad_type_0 = const()[name = tensor("op_7705_pad_type_0"), val = tensor("valid")]; + tensor var_7705_strides_0 = const()[name = tensor("op_7705_strides_0"), val = tensor([1, 1])]; + tensor var_7705_pad_0 = const()[name = tensor("op_7705_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7705_dilations_0 = const()[name = tensor("op_7705_dilations_0"), val = tensor([1, 1])]; + tensor var_7705_groups_0 = const()[name = tensor("op_7705_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_value_weight_to_fp16 = const()[name = tensor("blocks_28_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1122743744)))]; + tensor blocks_28_attn_value_bias_to_fp16 = const()[name = tensor("blocks_28_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126020608)))]; + tensor var_7705_cast_fp16 = conv(bias = blocks_28_attn_value_bias_to_fp16, dilations = var_7705_dilations_0, groups = var_7705_groups_0, pad = var_7705_pad_0, pad_type = var_7705_pad_type_0, strides = var_7705_strides_0, weight = blocks_28_attn_value_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("op_7705_cast_fp16")]; + tensor tile_84 = const()[name = tensor("tile_84"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7708_axis_0 = const()[name = tensor("op_7708_axis_0"), val = tensor(1)]; + tensor var_7708_cast_fp16_0, tensor var_7708_cast_fp16_1, tensor var_7708_cast_fp16_2, tensor var_7708_cast_fp16_3, tensor var_7708_cast_fp16_4, tensor var_7708_cast_fp16_5, tensor var_7708_cast_fp16_6, tensor var_7708_cast_fp16_7, tensor var_7708_cast_fp16_8, tensor var_7708_cast_fp16_9, tensor var_7708_cast_fp16_10, tensor var_7708_cast_fp16_11, tensor var_7708_cast_fp16_12, tensor var_7708_cast_fp16_13, tensor var_7708_cast_fp16_14, tensor var_7708_cast_fp16_15, tensor var_7708_cast_fp16_16, tensor var_7708_cast_fp16_17, tensor var_7708_cast_fp16_18, tensor var_7708_cast_fp16_19 = split(axis = var_7708_axis_0, split_sizes = tile_84, x = var_7707_cast_fp16)[name = tensor("op_7708_cast_fp16")]; + tensor var_7729_perm_0 = const()[name = tensor("op_7729_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_85 = const()[name = tensor("tile_85"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7730_axis_0 = const()[name = tensor("op_7730_axis_0"), val = tensor(3)]; + tensor var_7729_cast_fp16 = transpose(perm = var_7729_perm_0, x = k_57_cast_fp16)[name = tensor("transpose_4")]; + tensor var_7730_cast_fp16_0, tensor var_7730_cast_fp16_1, tensor var_7730_cast_fp16_2, tensor var_7730_cast_fp16_3, tensor var_7730_cast_fp16_4, tensor var_7730_cast_fp16_5, tensor var_7730_cast_fp16_6, tensor var_7730_cast_fp16_7, tensor var_7730_cast_fp16_8, tensor var_7730_cast_fp16_9, tensor var_7730_cast_fp16_10, tensor var_7730_cast_fp16_11, tensor var_7730_cast_fp16_12, tensor var_7730_cast_fp16_13, tensor var_7730_cast_fp16_14, tensor var_7730_cast_fp16_15, tensor var_7730_cast_fp16_16, tensor var_7730_cast_fp16_17, tensor var_7730_cast_fp16_18, tensor var_7730_cast_fp16_19 = split(axis = var_7730_axis_0, split_sizes = tile_85, x = var_7729_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor tile_86 = const()[name = tensor("tile_86"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7751_axis_0 = const()[name = tensor("op_7751_axis_0"), val = tensor(1)]; + tensor var_7751_cast_fp16_0, tensor var_7751_cast_fp16_1, tensor var_7751_cast_fp16_2, tensor var_7751_cast_fp16_3, tensor var_7751_cast_fp16_4, tensor var_7751_cast_fp16_5, tensor var_7751_cast_fp16_6, tensor var_7751_cast_fp16_7, tensor var_7751_cast_fp16_8, tensor var_7751_cast_fp16_9, tensor var_7751_cast_fp16_10, tensor var_7751_cast_fp16_11, tensor var_7751_cast_fp16_12, tensor var_7751_cast_fp16_13, tensor var_7751_cast_fp16_14, tensor var_7751_cast_fp16_15, tensor var_7751_cast_fp16_16, tensor var_7751_cast_fp16_17, tensor var_7751_cast_fp16_18, tensor var_7751_cast_fp16_19 = split(axis = var_7751_axis_0, split_sizes = tile_86, x = var_7705_cast_fp16)[name = tensor("op_7751_cast_fp16")]; + tensor aw_1121_equation_0 = const()[name = tensor("aw_1121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1121_cast_fp16 = einsum(equation = aw_1121_equation_0, values = (var_7730_cast_fp16_0, var_7708_cast_fp16_0))[name = tensor("aw_1121_cast_fp16")]; + tensor aw_1123_equation_0 = const()[name = tensor("aw_1123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1123_cast_fp16 = einsum(equation = aw_1123_equation_0, values = (var_7730_cast_fp16_1, var_7708_cast_fp16_1))[name = tensor("aw_1123_cast_fp16")]; + tensor aw_1125_equation_0 = const()[name = tensor("aw_1125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1125_cast_fp16 = einsum(equation = aw_1125_equation_0, values = (var_7730_cast_fp16_2, var_7708_cast_fp16_2))[name = tensor("aw_1125_cast_fp16")]; + tensor aw_1127_equation_0 = const()[name = tensor("aw_1127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1127_cast_fp16 = einsum(equation = aw_1127_equation_0, values = (var_7730_cast_fp16_3, var_7708_cast_fp16_3))[name = tensor("aw_1127_cast_fp16")]; + tensor aw_1129_equation_0 = const()[name = tensor("aw_1129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1129_cast_fp16 = einsum(equation = aw_1129_equation_0, values = (var_7730_cast_fp16_4, var_7708_cast_fp16_4))[name = tensor("aw_1129_cast_fp16")]; + tensor aw_1131_equation_0 = const()[name = tensor("aw_1131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1131_cast_fp16 = einsum(equation = aw_1131_equation_0, values = (var_7730_cast_fp16_5, var_7708_cast_fp16_5))[name = tensor("aw_1131_cast_fp16")]; + tensor aw_1133_equation_0 = const()[name = tensor("aw_1133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1133_cast_fp16 = einsum(equation = aw_1133_equation_0, values = (var_7730_cast_fp16_6, var_7708_cast_fp16_6))[name = tensor("aw_1133_cast_fp16")]; + tensor aw_1135_equation_0 = const()[name = tensor("aw_1135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1135_cast_fp16 = einsum(equation = aw_1135_equation_0, values = (var_7730_cast_fp16_7, var_7708_cast_fp16_7))[name = tensor("aw_1135_cast_fp16")]; + tensor aw_1137_equation_0 = const()[name = tensor("aw_1137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1137_cast_fp16 = einsum(equation = aw_1137_equation_0, values = (var_7730_cast_fp16_8, var_7708_cast_fp16_8))[name = tensor("aw_1137_cast_fp16")]; + tensor aw_1139_equation_0 = const()[name = tensor("aw_1139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1139_cast_fp16 = einsum(equation = aw_1139_equation_0, values = (var_7730_cast_fp16_9, var_7708_cast_fp16_9))[name = tensor("aw_1139_cast_fp16")]; + tensor aw_1141_equation_0 = const()[name = tensor("aw_1141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1141_cast_fp16 = einsum(equation = aw_1141_equation_0, values = (var_7730_cast_fp16_10, var_7708_cast_fp16_10))[name = tensor("aw_1141_cast_fp16")]; + tensor aw_1143_equation_0 = const()[name = tensor("aw_1143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1143_cast_fp16 = einsum(equation = aw_1143_equation_0, values = (var_7730_cast_fp16_11, var_7708_cast_fp16_11))[name = tensor("aw_1143_cast_fp16")]; + tensor aw_1145_equation_0 = const()[name = tensor("aw_1145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1145_cast_fp16 = einsum(equation = aw_1145_equation_0, values = (var_7730_cast_fp16_12, var_7708_cast_fp16_12))[name = tensor("aw_1145_cast_fp16")]; + tensor aw_1147_equation_0 = const()[name = tensor("aw_1147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1147_cast_fp16 = einsum(equation = aw_1147_equation_0, values = (var_7730_cast_fp16_13, var_7708_cast_fp16_13))[name = tensor("aw_1147_cast_fp16")]; + tensor aw_1149_equation_0 = const()[name = tensor("aw_1149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1149_cast_fp16 = einsum(equation = aw_1149_equation_0, values = (var_7730_cast_fp16_14, var_7708_cast_fp16_14))[name = tensor("aw_1149_cast_fp16")]; + tensor aw_1151_equation_0 = const()[name = tensor("aw_1151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1151_cast_fp16 = einsum(equation = aw_1151_equation_0, values = (var_7730_cast_fp16_15, var_7708_cast_fp16_15))[name = tensor("aw_1151_cast_fp16")]; + tensor aw_1153_equation_0 = const()[name = tensor("aw_1153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1153_cast_fp16 = einsum(equation = aw_1153_equation_0, values = (var_7730_cast_fp16_16, var_7708_cast_fp16_16))[name = tensor("aw_1153_cast_fp16")]; + tensor aw_1155_equation_0 = const()[name = tensor("aw_1155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1155_cast_fp16 = einsum(equation = aw_1155_equation_0, values = (var_7730_cast_fp16_17, var_7708_cast_fp16_17))[name = tensor("aw_1155_cast_fp16")]; + tensor aw_1157_equation_0 = const()[name = tensor("aw_1157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1157_cast_fp16 = einsum(equation = aw_1157_equation_0, values = (var_7730_cast_fp16_18, var_7708_cast_fp16_18))[name = tensor("aw_1157_cast_fp16")]; + tensor aw_1159_equation_0 = const()[name = tensor("aw_1159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1159_cast_fp16 = einsum(equation = aw_1159_equation_0, values = (var_7730_cast_fp16_19, var_7708_cast_fp16_19))[name = tensor("aw_1159_cast_fp16")]; + tensor var_7812_cast_fp16 = softmax(axis = var_7656, x = aw_1121_cast_fp16)[name = tensor("op_7812_cast_fp16")]; + tensor var_7813_cast_fp16 = softmax(axis = var_7656, x = aw_1123_cast_fp16)[name = tensor("op_7813_cast_fp16")]; + tensor var_7814_cast_fp16 = softmax(axis = var_7656, x = aw_1125_cast_fp16)[name = tensor("op_7814_cast_fp16")]; + tensor var_7815_cast_fp16 = softmax(axis = var_7656, x = aw_1127_cast_fp16)[name = tensor("op_7815_cast_fp16")]; + tensor var_7816_cast_fp16 = softmax(axis = var_7656, x = aw_1129_cast_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor var_7817_cast_fp16 = softmax(axis = var_7656, x = aw_1131_cast_fp16)[name = tensor("op_7817_cast_fp16")]; + tensor var_7818_cast_fp16 = softmax(axis = var_7656, x = aw_1133_cast_fp16)[name = tensor("op_7818_cast_fp16")]; + tensor var_7819_cast_fp16 = softmax(axis = var_7656, x = aw_1135_cast_fp16)[name = tensor("op_7819_cast_fp16")]; + tensor var_7820_cast_fp16 = softmax(axis = var_7656, x = aw_1137_cast_fp16)[name = tensor("op_7820_cast_fp16")]; + tensor var_7821_cast_fp16 = softmax(axis = var_7656, x = aw_1139_cast_fp16)[name = tensor("op_7821_cast_fp16")]; + tensor var_7822_cast_fp16 = softmax(axis = var_7656, x = aw_1141_cast_fp16)[name = tensor("op_7822_cast_fp16")]; + tensor var_7823_cast_fp16 = softmax(axis = var_7656, x = aw_1143_cast_fp16)[name = tensor("op_7823_cast_fp16")]; + tensor var_7824_cast_fp16 = softmax(axis = var_7656, x = aw_1145_cast_fp16)[name = tensor("op_7824_cast_fp16")]; + tensor var_7825_cast_fp16 = softmax(axis = var_7656, x = aw_1147_cast_fp16)[name = tensor("op_7825_cast_fp16")]; + tensor var_7826_cast_fp16 = softmax(axis = var_7656, x = aw_1149_cast_fp16)[name = tensor("op_7826_cast_fp16")]; + tensor var_7827_cast_fp16 = softmax(axis = var_7656, x = aw_1151_cast_fp16)[name = tensor("op_7827_cast_fp16")]; + tensor var_7828_cast_fp16 = softmax(axis = var_7656, x = aw_1153_cast_fp16)[name = tensor("op_7828_cast_fp16")]; + tensor var_7829_cast_fp16 = softmax(axis = var_7656, x = aw_1155_cast_fp16)[name = tensor("op_7829_cast_fp16")]; + tensor var_7830_cast_fp16 = softmax(axis = var_7656, x = aw_1157_cast_fp16)[name = tensor("op_7830_cast_fp16")]; + tensor var_7831_cast_fp16 = softmax(axis = var_7656, x = aw_1159_cast_fp16)[name = tensor("op_7831_cast_fp16")]; + tensor var_7833_equation_0 = const()[name = tensor("op_7833_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7833_cast_fp16 = einsum(equation = var_7833_equation_0, values = (var_7751_cast_fp16_0, var_7812_cast_fp16))[name = tensor("op_7833_cast_fp16")]; + tensor var_7835_equation_0 = const()[name = tensor("op_7835_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7835_cast_fp16 = einsum(equation = var_7835_equation_0, values = (var_7751_cast_fp16_1, var_7813_cast_fp16))[name = tensor("op_7835_cast_fp16")]; + tensor var_7837_equation_0 = const()[name = tensor("op_7837_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7837_cast_fp16 = einsum(equation = var_7837_equation_0, values = (var_7751_cast_fp16_2, var_7814_cast_fp16))[name = tensor("op_7837_cast_fp16")]; + tensor var_7839_equation_0 = const()[name = tensor("op_7839_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7839_cast_fp16 = einsum(equation = var_7839_equation_0, values = (var_7751_cast_fp16_3, var_7815_cast_fp16))[name = tensor("op_7839_cast_fp16")]; + tensor var_7841_equation_0 = const()[name = tensor("op_7841_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7841_cast_fp16 = einsum(equation = var_7841_equation_0, values = (var_7751_cast_fp16_4, var_7816_cast_fp16))[name = tensor("op_7841_cast_fp16")]; + tensor var_7843_equation_0 = const()[name = tensor("op_7843_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7843_cast_fp16 = einsum(equation = var_7843_equation_0, values = (var_7751_cast_fp16_5, var_7817_cast_fp16))[name = tensor("op_7843_cast_fp16")]; + tensor var_7845_equation_0 = const()[name = tensor("op_7845_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7845_cast_fp16 = einsum(equation = var_7845_equation_0, values = (var_7751_cast_fp16_6, var_7818_cast_fp16))[name = tensor("op_7845_cast_fp16")]; + tensor var_7847_equation_0 = const()[name = tensor("op_7847_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7847_cast_fp16 = einsum(equation = var_7847_equation_0, values = (var_7751_cast_fp16_7, var_7819_cast_fp16))[name = tensor("op_7847_cast_fp16")]; + tensor var_7849_equation_0 = const()[name = tensor("op_7849_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7849_cast_fp16 = einsum(equation = var_7849_equation_0, values = (var_7751_cast_fp16_8, var_7820_cast_fp16))[name = tensor("op_7849_cast_fp16")]; + tensor var_7851_equation_0 = const()[name = tensor("op_7851_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7851_cast_fp16 = einsum(equation = var_7851_equation_0, values = (var_7751_cast_fp16_9, var_7821_cast_fp16))[name = tensor("op_7851_cast_fp16")]; + tensor var_7853_equation_0 = const()[name = tensor("op_7853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7853_cast_fp16 = einsum(equation = var_7853_equation_0, values = (var_7751_cast_fp16_10, var_7822_cast_fp16))[name = tensor("op_7853_cast_fp16")]; + tensor var_7855_equation_0 = const()[name = tensor("op_7855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7855_cast_fp16 = einsum(equation = var_7855_equation_0, values = (var_7751_cast_fp16_11, var_7823_cast_fp16))[name = tensor("op_7855_cast_fp16")]; + tensor var_7857_equation_0 = const()[name = tensor("op_7857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7857_cast_fp16 = einsum(equation = var_7857_equation_0, values = (var_7751_cast_fp16_12, var_7824_cast_fp16))[name = tensor("op_7857_cast_fp16")]; + tensor var_7859_equation_0 = const()[name = tensor("op_7859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7859_cast_fp16 = einsum(equation = var_7859_equation_0, values = (var_7751_cast_fp16_13, var_7825_cast_fp16))[name = tensor("op_7859_cast_fp16")]; + tensor var_7861_equation_0 = const()[name = tensor("op_7861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7861_cast_fp16 = einsum(equation = var_7861_equation_0, values = (var_7751_cast_fp16_14, var_7826_cast_fp16))[name = tensor("op_7861_cast_fp16")]; + tensor var_7863_equation_0 = const()[name = tensor("op_7863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7863_cast_fp16 = einsum(equation = var_7863_equation_0, values = (var_7751_cast_fp16_15, var_7827_cast_fp16))[name = tensor("op_7863_cast_fp16")]; + tensor var_7865_equation_0 = const()[name = tensor("op_7865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7865_cast_fp16 = einsum(equation = var_7865_equation_0, values = (var_7751_cast_fp16_16, var_7828_cast_fp16))[name = tensor("op_7865_cast_fp16")]; + tensor var_7867_equation_0 = const()[name = tensor("op_7867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7867_cast_fp16 = einsum(equation = var_7867_equation_0, values = (var_7751_cast_fp16_17, var_7829_cast_fp16))[name = tensor("op_7867_cast_fp16")]; + tensor var_7869_equation_0 = const()[name = tensor("op_7869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7869_cast_fp16 = einsum(equation = var_7869_equation_0, values = (var_7751_cast_fp16_18, var_7830_cast_fp16))[name = tensor("op_7869_cast_fp16")]; + tensor var_7871_equation_0 = const()[name = tensor("op_7871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7871_cast_fp16 = einsum(equation = var_7871_equation_0, values = (var_7751_cast_fp16_19, var_7831_cast_fp16))[name = tensor("op_7871_cast_fp16")]; + tensor input_285_interleave_0 = const()[name = tensor("input_285_interleave_0"), val = tensor(false)]; + tensor input_285_cast_fp16 = concat(axis = var_7656, interleave = input_285_interleave_0, values = (var_7833_cast_fp16, var_7835_cast_fp16, var_7837_cast_fp16, var_7839_cast_fp16, var_7841_cast_fp16, var_7843_cast_fp16, var_7845_cast_fp16, var_7847_cast_fp16, var_7849_cast_fp16, var_7851_cast_fp16, var_7853_cast_fp16, var_7855_cast_fp16, var_7857_cast_fp16, var_7859_cast_fp16, var_7861_cast_fp16, var_7863_cast_fp16, var_7865_cast_fp16, var_7867_cast_fp16, var_7869_cast_fp16, var_7871_cast_fp16))[name = tensor("input_285_cast_fp16")]; + tensor var_7880_pad_type_0 = const()[name = tensor("op_7880_pad_type_0"), val = tensor("valid")]; + tensor var_7880_strides_0 = const()[name = tensor("op_7880_strides_0"), val = tensor([1, 1])]; + tensor var_7880_pad_0 = const()[name = tensor("op_7880_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7880_dilations_0 = const()[name = tensor("op_7880_dilations_0"), val = tensor([1, 1])]; + tensor var_7880_groups_0 = const()[name = tensor("op_7880_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_out_weight_to_fp16 = const()[name = tensor("blocks_28_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126023232)))]; + tensor blocks_28_attn_out_bias_to_fp16 = const()[name = tensor("blocks_28_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129300096)))]; + tensor var_7880_cast_fp16 = conv(bias = blocks_28_attn_out_bias_to_fp16, dilations = var_7880_dilations_0, groups = var_7880_groups_0, pad = var_7880_pad_0, pad_type = var_7880_pad_type_0, strides = var_7880_strides_0, weight = blocks_28_attn_out_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("op_7880_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = var_7880_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([1])]; + tensor input_287_gamma_0_to_fp16 = const()[name = tensor("input_287_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129302720)))]; + tensor input_287_beta_0_to_fp16 = const()[name = tensor("input_287_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129305344)))]; + tensor var_7890_to_fp16 = const()[name = tensor("op_7890_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = input_287_beta_0_to_fp16, epsilon = var_7890_to_fp16, gamma = input_287_gamma_0_to_fp16, x = inputs_115_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_pad_type_0 = const()[name = tensor("input_289_pad_type_0"), val = tensor("valid")]; + tensor input_289_strides_0 = const()[name = tensor("input_289_strides_0"), val = tensor([1, 1])]; + tensor input_289_pad_0 = const()[name = tensor("input_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_289_dilations_0 = const()[name = tensor("input_289_dilations_0"), val = tensor([1, 1])]; + tensor input_289_groups_0 = const()[name = tensor("input_289_groups_0"), val = tensor(1)]; + tensor blocks_28_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129307968)))]; + tensor blocks_28_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142415232)))]; + tensor input_289_cast_fp16 = conv(bias = blocks_28_mlp_0_bias_to_fp16, dilations = input_289_dilations_0, groups = input_289_groups_0, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = input_289_strides_0, weight = blocks_28_mlp_0_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor input_291_mode_0 = const()[name = tensor("input_291_mode_0"), val = tensor("EXACT")]; + tensor input_291_cast_fp16 = gelu(mode = input_291_mode_0, x = input_289_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor var_7916_pad_type_0 = const()[name = tensor("op_7916_pad_type_0"), val = tensor("valid")]; + tensor var_7916_strides_0 = const()[name = tensor("op_7916_strides_0"), val = tensor([1, 1])]; + tensor var_7916_pad_0 = const()[name = tensor("op_7916_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7916_dilations_0 = const()[name = tensor("op_7916_dilations_0"), val = tensor([1, 1])]; + tensor var_7916_groups_0 = const()[name = tensor("op_7916_groups_0"), val = tensor(1)]; + tensor blocks_28_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142425536)))]; + tensor blocks_28_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155532800)))]; + tensor var_7916_cast_fp16 = conv(bias = blocks_28_mlp_2_bias_to_fp16, dilations = var_7916_dilations_0, groups = var_7916_groups_0, pad = var_7916_pad_0, pad_type = var_7916_pad_type_0, strides = var_7916_strides_0, weight = blocks_28_mlp_2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("op_7916_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = var_7916_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_7925 = const()[name = tensor("op_7925"), val = tensor(1)]; + tensor input_293_axes_0 = const()[name = tensor("input_293_axes_0"), val = tensor([1])]; + tensor input_293_gamma_0_to_fp16 = const()[name = tensor("input_293_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155535424)))]; + tensor input_293_beta_0_to_fp16 = const()[name = tensor("input_293_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155538048)))]; + tensor var_7941_to_fp16 = const()[name = tensor("op_7941_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = input_293_beta_0_to_fp16, epsilon = var_7941_to_fp16, gamma = input_293_gamma_0_to_fp16, x = inputs_117_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor q_59_pad_type_0 = const()[name = tensor("q_59_pad_type_0"), val = tensor("valid")]; + tensor q_59_strides_0 = const()[name = tensor("q_59_strides_0"), val = tensor([1, 1])]; + tensor q_59_pad_0 = const()[name = tensor("q_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_59_dilations_0 = const()[name = tensor("q_59_dilations_0"), val = tensor([1, 1])]; + tensor q_59_groups_0 = const()[name = tensor("q_59_groups_0"), val = tensor(1)]; + tensor var_7976_weight_0_to_fp16 = const()[name = tensor("op_7976_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155540672)))]; + tensor var_7976_bias_0_to_fp16 = const()[name = tensor("op_7976_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158817536)))]; + tensor var_7976_cast_fp16 = conv(bias = var_7976_bias_0_to_fp16, dilations = q_59_dilations_0, groups = q_59_groups_0, pad = q_59_pad_0, pad_type = q_59_pad_type_0, strides = q_59_strides_0, weight = var_7976_weight_0_to_fp16, x = input_293_cast_fp16)[name = tensor("op_7976_cast_fp16")]; + tensor k_59_pad_type_0 = const()[name = tensor("k_59_pad_type_0"), val = tensor("valid")]; + tensor k_59_strides_0 = const()[name = tensor("k_59_strides_0"), val = tensor([1, 1])]; + tensor k_59_pad_0 = const()[name = tensor("k_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_59_dilations_0 = const()[name = tensor("k_59_dilations_0"), val = tensor([1, 1])]; + tensor k_59_groups_0 = const()[name = tensor("k_59_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_key_weight_to_fp16 = const()[name = tensor("blocks_29_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158820160)))]; + tensor k_59_cast_fp16 = conv(dilations = k_59_dilations_0, groups = k_59_groups_0, pad = k_59_pad_0, pad_type = k_59_pad_type_0, strides = k_59_strides_0, weight = blocks_29_attn_key_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor var_7974_pad_type_0 = const()[name = tensor("op_7974_pad_type_0"), val = tensor("valid")]; + tensor var_7974_strides_0 = const()[name = tensor("op_7974_strides_0"), val = tensor([1, 1])]; + tensor var_7974_pad_0 = const()[name = tensor("op_7974_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7974_dilations_0 = const()[name = tensor("op_7974_dilations_0"), val = tensor([1, 1])]; + tensor var_7974_groups_0 = const()[name = tensor("op_7974_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_value_weight_to_fp16 = const()[name = tensor("blocks_29_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162097024)))]; + tensor blocks_29_attn_value_bias_to_fp16 = const()[name = tensor("blocks_29_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165373888)))]; + tensor var_7974_cast_fp16 = conv(bias = blocks_29_attn_value_bias_to_fp16, dilations = var_7974_dilations_0, groups = var_7974_groups_0, pad = var_7974_pad_0, pad_type = var_7974_pad_type_0, strides = var_7974_strides_0, weight = blocks_29_attn_value_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("op_7974_cast_fp16")]; + tensor tile_87 = const()[name = tensor("tile_87"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7977_axis_0 = const()[name = tensor("op_7977_axis_0"), val = tensor(1)]; + tensor var_7977_cast_fp16_0, tensor var_7977_cast_fp16_1, tensor var_7977_cast_fp16_2, tensor var_7977_cast_fp16_3, tensor var_7977_cast_fp16_4, tensor var_7977_cast_fp16_5, tensor var_7977_cast_fp16_6, tensor var_7977_cast_fp16_7, tensor var_7977_cast_fp16_8, tensor var_7977_cast_fp16_9, tensor var_7977_cast_fp16_10, tensor var_7977_cast_fp16_11, tensor var_7977_cast_fp16_12, tensor var_7977_cast_fp16_13, tensor var_7977_cast_fp16_14, tensor var_7977_cast_fp16_15, tensor var_7977_cast_fp16_16, tensor var_7977_cast_fp16_17, tensor var_7977_cast_fp16_18, tensor var_7977_cast_fp16_19 = split(axis = var_7977_axis_0, split_sizes = tile_87, x = var_7976_cast_fp16)[name = tensor("op_7977_cast_fp16")]; + tensor var_7998_perm_0 = const()[name = tensor("op_7998_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_88 = const()[name = tensor("tile_88"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7999_axis_0 = const()[name = tensor("op_7999_axis_0"), val = tensor(3)]; + tensor var_7998_cast_fp16 = transpose(perm = var_7998_perm_0, x = k_59_cast_fp16)[name = tensor("transpose_3")]; + tensor var_7999_cast_fp16_0, tensor var_7999_cast_fp16_1, tensor var_7999_cast_fp16_2, tensor var_7999_cast_fp16_3, tensor var_7999_cast_fp16_4, tensor var_7999_cast_fp16_5, tensor var_7999_cast_fp16_6, tensor var_7999_cast_fp16_7, tensor var_7999_cast_fp16_8, tensor var_7999_cast_fp16_9, tensor var_7999_cast_fp16_10, tensor var_7999_cast_fp16_11, tensor var_7999_cast_fp16_12, tensor var_7999_cast_fp16_13, tensor var_7999_cast_fp16_14, tensor var_7999_cast_fp16_15, tensor var_7999_cast_fp16_16, tensor var_7999_cast_fp16_17, tensor var_7999_cast_fp16_18, tensor var_7999_cast_fp16_19 = split(axis = var_7999_axis_0, split_sizes = tile_88, x = var_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor tile_89 = const()[name = tensor("tile_89"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8020_axis_0 = const()[name = tensor("op_8020_axis_0"), val = tensor(1)]; + tensor var_8020_cast_fp16_0, tensor var_8020_cast_fp16_1, tensor var_8020_cast_fp16_2, tensor var_8020_cast_fp16_3, tensor var_8020_cast_fp16_4, tensor var_8020_cast_fp16_5, tensor var_8020_cast_fp16_6, tensor var_8020_cast_fp16_7, tensor var_8020_cast_fp16_8, tensor var_8020_cast_fp16_9, tensor var_8020_cast_fp16_10, tensor var_8020_cast_fp16_11, tensor var_8020_cast_fp16_12, tensor var_8020_cast_fp16_13, tensor var_8020_cast_fp16_14, tensor var_8020_cast_fp16_15, tensor var_8020_cast_fp16_16, tensor var_8020_cast_fp16_17, tensor var_8020_cast_fp16_18, tensor var_8020_cast_fp16_19 = split(axis = var_8020_axis_0, split_sizes = tile_89, x = var_7974_cast_fp16)[name = tensor("op_8020_cast_fp16")]; + tensor aw_1161_equation_0 = const()[name = tensor("aw_1161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1161_cast_fp16 = einsum(equation = aw_1161_equation_0, values = (var_7999_cast_fp16_0, var_7977_cast_fp16_0))[name = tensor("aw_1161_cast_fp16")]; + tensor aw_1163_equation_0 = const()[name = tensor("aw_1163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1163_cast_fp16 = einsum(equation = aw_1163_equation_0, values = (var_7999_cast_fp16_1, var_7977_cast_fp16_1))[name = tensor("aw_1163_cast_fp16")]; + tensor aw_1165_equation_0 = const()[name = tensor("aw_1165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1165_cast_fp16 = einsum(equation = aw_1165_equation_0, values = (var_7999_cast_fp16_2, var_7977_cast_fp16_2))[name = tensor("aw_1165_cast_fp16")]; + tensor aw_1167_equation_0 = const()[name = tensor("aw_1167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1167_cast_fp16 = einsum(equation = aw_1167_equation_0, values = (var_7999_cast_fp16_3, var_7977_cast_fp16_3))[name = tensor("aw_1167_cast_fp16")]; + tensor aw_1169_equation_0 = const()[name = tensor("aw_1169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1169_cast_fp16 = einsum(equation = aw_1169_equation_0, values = (var_7999_cast_fp16_4, var_7977_cast_fp16_4))[name = tensor("aw_1169_cast_fp16")]; + tensor aw_1171_equation_0 = const()[name = tensor("aw_1171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1171_cast_fp16 = einsum(equation = aw_1171_equation_0, values = (var_7999_cast_fp16_5, var_7977_cast_fp16_5))[name = tensor("aw_1171_cast_fp16")]; + tensor aw_1173_equation_0 = const()[name = tensor("aw_1173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1173_cast_fp16 = einsum(equation = aw_1173_equation_0, values = (var_7999_cast_fp16_6, var_7977_cast_fp16_6))[name = tensor("aw_1173_cast_fp16")]; + tensor aw_1175_equation_0 = const()[name = tensor("aw_1175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1175_cast_fp16 = einsum(equation = aw_1175_equation_0, values = (var_7999_cast_fp16_7, var_7977_cast_fp16_7))[name = tensor("aw_1175_cast_fp16")]; + tensor aw_1177_equation_0 = const()[name = tensor("aw_1177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1177_cast_fp16 = einsum(equation = aw_1177_equation_0, values = (var_7999_cast_fp16_8, var_7977_cast_fp16_8))[name = tensor("aw_1177_cast_fp16")]; + tensor aw_1179_equation_0 = const()[name = tensor("aw_1179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1179_cast_fp16 = einsum(equation = aw_1179_equation_0, values = (var_7999_cast_fp16_9, var_7977_cast_fp16_9))[name = tensor("aw_1179_cast_fp16")]; + tensor aw_1181_equation_0 = const()[name = tensor("aw_1181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1181_cast_fp16 = einsum(equation = aw_1181_equation_0, values = (var_7999_cast_fp16_10, var_7977_cast_fp16_10))[name = tensor("aw_1181_cast_fp16")]; + tensor aw_1183_equation_0 = const()[name = tensor("aw_1183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1183_cast_fp16 = einsum(equation = aw_1183_equation_0, values = (var_7999_cast_fp16_11, var_7977_cast_fp16_11))[name = tensor("aw_1183_cast_fp16")]; + tensor aw_1185_equation_0 = const()[name = tensor("aw_1185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1185_cast_fp16 = einsum(equation = aw_1185_equation_0, values = (var_7999_cast_fp16_12, var_7977_cast_fp16_12))[name = tensor("aw_1185_cast_fp16")]; + tensor aw_1187_equation_0 = const()[name = tensor("aw_1187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1187_cast_fp16 = einsum(equation = aw_1187_equation_0, values = (var_7999_cast_fp16_13, var_7977_cast_fp16_13))[name = tensor("aw_1187_cast_fp16")]; + tensor aw_1189_equation_0 = const()[name = tensor("aw_1189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1189_cast_fp16 = einsum(equation = aw_1189_equation_0, values = (var_7999_cast_fp16_14, var_7977_cast_fp16_14))[name = tensor("aw_1189_cast_fp16")]; + tensor aw_1191_equation_0 = const()[name = tensor("aw_1191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1191_cast_fp16 = einsum(equation = aw_1191_equation_0, values = (var_7999_cast_fp16_15, var_7977_cast_fp16_15))[name = tensor("aw_1191_cast_fp16")]; + tensor aw_1193_equation_0 = const()[name = tensor("aw_1193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1193_cast_fp16 = einsum(equation = aw_1193_equation_0, values = (var_7999_cast_fp16_16, var_7977_cast_fp16_16))[name = tensor("aw_1193_cast_fp16")]; + tensor aw_1195_equation_0 = const()[name = tensor("aw_1195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1195_cast_fp16 = einsum(equation = aw_1195_equation_0, values = (var_7999_cast_fp16_17, var_7977_cast_fp16_17))[name = tensor("aw_1195_cast_fp16")]; + tensor aw_1197_equation_0 = const()[name = tensor("aw_1197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1197_cast_fp16 = einsum(equation = aw_1197_equation_0, values = (var_7999_cast_fp16_18, var_7977_cast_fp16_18))[name = tensor("aw_1197_cast_fp16")]; + tensor aw_1199_equation_0 = const()[name = tensor("aw_1199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1199_cast_fp16 = einsum(equation = aw_1199_equation_0, values = (var_7999_cast_fp16_19, var_7977_cast_fp16_19))[name = tensor("aw_1199_cast_fp16")]; + tensor var_8081_cast_fp16 = softmax(axis = var_7925, x = aw_1161_cast_fp16)[name = tensor("op_8081_cast_fp16")]; + tensor var_8082_cast_fp16 = softmax(axis = var_7925, x = aw_1163_cast_fp16)[name = tensor("op_8082_cast_fp16")]; + tensor var_8083_cast_fp16 = softmax(axis = var_7925, x = aw_1165_cast_fp16)[name = tensor("op_8083_cast_fp16")]; + tensor var_8084_cast_fp16 = softmax(axis = var_7925, x = aw_1167_cast_fp16)[name = tensor("op_8084_cast_fp16")]; + tensor var_8085_cast_fp16 = softmax(axis = var_7925, x = aw_1169_cast_fp16)[name = tensor("op_8085_cast_fp16")]; + tensor var_8086_cast_fp16 = softmax(axis = var_7925, x = aw_1171_cast_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8087_cast_fp16 = softmax(axis = var_7925, x = aw_1173_cast_fp16)[name = tensor("op_8087_cast_fp16")]; + tensor var_8088_cast_fp16 = softmax(axis = var_7925, x = aw_1175_cast_fp16)[name = tensor("op_8088_cast_fp16")]; + tensor var_8089_cast_fp16 = softmax(axis = var_7925, x = aw_1177_cast_fp16)[name = tensor("op_8089_cast_fp16")]; + tensor var_8090_cast_fp16 = softmax(axis = var_7925, x = aw_1179_cast_fp16)[name = tensor("op_8090_cast_fp16")]; + tensor var_8091_cast_fp16 = softmax(axis = var_7925, x = aw_1181_cast_fp16)[name = tensor("op_8091_cast_fp16")]; + tensor var_8092_cast_fp16 = softmax(axis = var_7925, x = aw_1183_cast_fp16)[name = tensor("op_8092_cast_fp16")]; + tensor var_8093_cast_fp16 = softmax(axis = var_7925, x = aw_1185_cast_fp16)[name = tensor("op_8093_cast_fp16")]; + tensor var_8094_cast_fp16 = softmax(axis = var_7925, x = aw_1187_cast_fp16)[name = tensor("op_8094_cast_fp16")]; + tensor var_8095_cast_fp16 = softmax(axis = var_7925, x = aw_1189_cast_fp16)[name = tensor("op_8095_cast_fp16")]; + tensor var_8096_cast_fp16 = softmax(axis = var_7925, x = aw_1191_cast_fp16)[name = tensor("op_8096_cast_fp16")]; + tensor var_8097_cast_fp16 = softmax(axis = var_7925, x = aw_1193_cast_fp16)[name = tensor("op_8097_cast_fp16")]; + tensor var_8098_cast_fp16 = softmax(axis = var_7925, x = aw_1195_cast_fp16)[name = tensor("op_8098_cast_fp16")]; + tensor var_8099_cast_fp16 = softmax(axis = var_7925, x = aw_1197_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8100_cast_fp16 = softmax(axis = var_7925, x = aw_1199_cast_fp16)[name = tensor("op_8100_cast_fp16")]; + tensor var_8102_equation_0 = const()[name = tensor("op_8102_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8102_cast_fp16 = einsum(equation = var_8102_equation_0, values = (var_8020_cast_fp16_0, var_8081_cast_fp16))[name = tensor("op_8102_cast_fp16")]; + tensor var_8104_equation_0 = const()[name = tensor("op_8104_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8104_cast_fp16 = einsum(equation = var_8104_equation_0, values = (var_8020_cast_fp16_1, var_8082_cast_fp16))[name = tensor("op_8104_cast_fp16")]; + tensor var_8106_equation_0 = const()[name = tensor("op_8106_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8106_cast_fp16 = einsum(equation = var_8106_equation_0, values = (var_8020_cast_fp16_2, var_8083_cast_fp16))[name = tensor("op_8106_cast_fp16")]; + tensor var_8108_equation_0 = const()[name = tensor("op_8108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8108_cast_fp16 = einsum(equation = var_8108_equation_0, values = (var_8020_cast_fp16_3, var_8084_cast_fp16))[name = tensor("op_8108_cast_fp16")]; + tensor var_8110_equation_0 = const()[name = tensor("op_8110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8110_cast_fp16 = einsum(equation = var_8110_equation_0, values = (var_8020_cast_fp16_4, var_8085_cast_fp16))[name = tensor("op_8110_cast_fp16")]; + tensor var_8112_equation_0 = const()[name = tensor("op_8112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8112_cast_fp16 = einsum(equation = var_8112_equation_0, values = (var_8020_cast_fp16_5, var_8086_cast_fp16))[name = tensor("op_8112_cast_fp16")]; + tensor var_8114_equation_0 = const()[name = tensor("op_8114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8114_cast_fp16 = einsum(equation = var_8114_equation_0, values = (var_8020_cast_fp16_6, var_8087_cast_fp16))[name = tensor("op_8114_cast_fp16")]; + tensor var_8116_equation_0 = const()[name = tensor("op_8116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8116_cast_fp16 = einsum(equation = var_8116_equation_0, values = (var_8020_cast_fp16_7, var_8088_cast_fp16))[name = tensor("op_8116_cast_fp16")]; + tensor var_8118_equation_0 = const()[name = tensor("op_8118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8118_cast_fp16 = einsum(equation = var_8118_equation_0, values = (var_8020_cast_fp16_8, var_8089_cast_fp16))[name = tensor("op_8118_cast_fp16")]; + tensor var_8120_equation_0 = const()[name = tensor("op_8120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8120_cast_fp16 = einsum(equation = var_8120_equation_0, values = (var_8020_cast_fp16_9, var_8090_cast_fp16))[name = tensor("op_8120_cast_fp16")]; + tensor var_8122_equation_0 = const()[name = tensor("op_8122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8122_cast_fp16 = einsum(equation = var_8122_equation_0, values = (var_8020_cast_fp16_10, var_8091_cast_fp16))[name = tensor("op_8122_cast_fp16")]; + tensor var_8124_equation_0 = const()[name = tensor("op_8124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8124_cast_fp16 = einsum(equation = var_8124_equation_0, values = (var_8020_cast_fp16_11, var_8092_cast_fp16))[name = tensor("op_8124_cast_fp16")]; + tensor var_8126_equation_0 = const()[name = tensor("op_8126_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8126_cast_fp16 = einsum(equation = var_8126_equation_0, values = (var_8020_cast_fp16_12, var_8093_cast_fp16))[name = tensor("op_8126_cast_fp16")]; + tensor var_8128_equation_0 = const()[name = tensor("op_8128_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8128_cast_fp16 = einsum(equation = var_8128_equation_0, values = (var_8020_cast_fp16_13, var_8094_cast_fp16))[name = tensor("op_8128_cast_fp16")]; + tensor var_8130_equation_0 = const()[name = tensor("op_8130_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8130_cast_fp16 = einsum(equation = var_8130_equation_0, values = (var_8020_cast_fp16_14, var_8095_cast_fp16))[name = tensor("op_8130_cast_fp16")]; + tensor var_8132_equation_0 = const()[name = tensor("op_8132_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8132_cast_fp16 = einsum(equation = var_8132_equation_0, values = (var_8020_cast_fp16_15, var_8096_cast_fp16))[name = tensor("op_8132_cast_fp16")]; + tensor var_8134_equation_0 = const()[name = tensor("op_8134_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8134_cast_fp16 = einsum(equation = var_8134_equation_0, values = (var_8020_cast_fp16_16, var_8097_cast_fp16))[name = tensor("op_8134_cast_fp16")]; + tensor var_8136_equation_0 = const()[name = tensor("op_8136_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8136_cast_fp16 = einsum(equation = var_8136_equation_0, values = (var_8020_cast_fp16_17, var_8098_cast_fp16))[name = tensor("op_8136_cast_fp16")]; + tensor var_8138_equation_0 = const()[name = tensor("op_8138_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8138_cast_fp16 = einsum(equation = var_8138_equation_0, values = (var_8020_cast_fp16_18, var_8099_cast_fp16))[name = tensor("op_8138_cast_fp16")]; + tensor var_8140_equation_0 = const()[name = tensor("op_8140_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8140_cast_fp16 = einsum(equation = var_8140_equation_0, values = (var_8020_cast_fp16_19, var_8100_cast_fp16))[name = tensor("op_8140_cast_fp16")]; + tensor input_295_interleave_0 = const()[name = tensor("input_295_interleave_0"), val = tensor(false)]; + tensor input_295_cast_fp16 = concat(axis = var_7925, interleave = input_295_interleave_0, values = (var_8102_cast_fp16, var_8104_cast_fp16, var_8106_cast_fp16, var_8108_cast_fp16, var_8110_cast_fp16, var_8112_cast_fp16, var_8114_cast_fp16, var_8116_cast_fp16, var_8118_cast_fp16, var_8120_cast_fp16, var_8122_cast_fp16, var_8124_cast_fp16, var_8126_cast_fp16, var_8128_cast_fp16, var_8130_cast_fp16, var_8132_cast_fp16, var_8134_cast_fp16, var_8136_cast_fp16, var_8138_cast_fp16, var_8140_cast_fp16))[name = tensor("input_295_cast_fp16")]; + tensor var_8149_pad_type_0 = const()[name = tensor("op_8149_pad_type_0"), val = tensor("valid")]; + tensor var_8149_strides_0 = const()[name = tensor("op_8149_strides_0"), val = tensor([1, 1])]; + tensor var_8149_pad_0 = const()[name = tensor("op_8149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8149_dilations_0 = const()[name = tensor("op_8149_dilations_0"), val = tensor([1, 1])]; + tensor var_8149_groups_0 = const()[name = tensor("op_8149_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_out_weight_to_fp16 = const()[name = tensor("blocks_29_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165376512)))]; + tensor blocks_29_attn_out_bias_to_fp16 = const()[name = tensor("blocks_29_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168653376)))]; + tensor var_8149_cast_fp16 = conv(bias = blocks_29_attn_out_bias_to_fp16, dilations = var_8149_dilations_0, groups = var_8149_groups_0, pad = var_8149_pad_0, pad_type = var_8149_pad_type_0, strides = var_8149_strides_0, weight = blocks_29_attn_out_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("op_8149_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_8149_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor input_297_axes_0 = const()[name = tensor("input_297_axes_0"), val = tensor([1])]; + tensor input_297_gamma_0_to_fp16 = const()[name = tensor("input_297_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168656000)))]; + tensor input_297_beta_0_to_fp16 = const()[name = tensor("input_297_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168658624)))]; + tensor var_8159_to_fp16 = const()[name = tensor("op_8159_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_297_cast_fp16 = layer_norm(axes = input_297_axes_0, beta = input_297_beta_0_to_fp16, epsilon = var_8159_to_fp16, gamma = input_297_gamma_0_to_fp16, x = inputs_119_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_pad_type_0 = const()[name = tensor("input_299_pad_type_0"), val = tensor("valid")]; + tensor input_299_strides_0 = const()[name = tensor("input_299_strides_0"), val = tensor([1, 1])]; + tensor input_299_pad_0 = const()[name = tensor("input_299_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_299_dilations_0 = const()[name = tensor("input_299_dilations_0"), val = tensor([1, 1])]; + tensor input_299_groups_0 = const()[name = tensor("input_299_groups_0"), val = tensor(1)]; + tensor blocks_29_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168661248)))]; + tensor blocks_29_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181768512)))]; + tensor input_299_cast_fp16 = conv(bias = blocks_29_mlp_0_bias_to_fp16, dilations = input_299_dilations_0, groups = input_299_groups_0, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = input_299_strides_0, weight = blocks_29_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor input_301_mode_0 = const()[name = tensor("input_301_mode_0"), val = tensor("EXACT")]; + tensor input_301_cast_fp16 = gelu(mode = input_301_mode_0, x = input_299_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_8185_pad_type_0 = const()[name = tensor("op_8185_pad_type_0"), val = tensor("valid")]; + tensor var_8185_strides_0 = const()[name = tensor("op_8185_strides_0"), val = tensor([1, 1])]; + tensor var_8185_pad_0 = const()[name = tensor("op_8185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8185_dilations_0 = const()[name = tensor("op_8185_dilations_0"), val = tensor([1, 1])]; + tensor var_8185_groups_0 = const()[name = tensor("op_8185_groups_0"), val = tensor(1)]; + tensor blocks_29_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181778816)))]; + tensor blocks_29_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194886080)))]; + tensor var_8185_cast_fp16 = conv(bias = blocks_29_mlp_2_bias_to_fp16, dilations = var_8185_dilations_0, groups = var_8185_groups_0, pad = var_8185_pad_0, pad_type = var_8185_pad_type_0, strides = var_8185_strides_0, weight = blocks_29_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = var_8185_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_8194 = const()[name = tensor("op_8194"), val = tensor(1)]; + tensor input_303_axes_0 = const()[name = tensor("input_303_axes_0"), val = tensor([1])]; + tensor input_303_gamma_0_to_fp16 = const()[name = tensor("input_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194888704)))]; + tensor input_303_beta_0_to_fp16 = const()[name = tensor("input_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194891328)))]; + tensor var_8210_to_fp16 = const()[name = tensor("op_8210_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_303_cast_fp16 = layer_norm(axes = input_303_axes_0, beta = input_303_beta_0_to_fp16, epsilon = var_8210_to_fp16, gamma = input_303_gamma_0_to_fp16, x = inputs_121_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor q_61_pad_type_0 = const()[name = tensor("q_61_pad_type_0"), val = tensor("valid")]; + tensor q_61_strides_0 = const()[name = tensor("q_61_strides_0"), val = tensor([1, 1])]; + tensor q_61_pad_0 = const()[name = tensor("q_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_61_dilations_0 = const()[name = tensor("q_61_dilations_0"), val = tensor([1, 1])]; + tensor q_61_groups_0 = const()[name = tensor("q_61_groups_0"), val = tensor(1)]; + tensor var_8245_weight_0_to_fp16 = const()[name = tensor("op_8245_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194893952)))]; + tensor var_8245_bias_0_to_fp16 = const()[name = tensor("op_8245_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198170816)))]; + tensor var_8245_cast_fp16 = conv(bias = var_8245_bias_0_to_fp16, dilations = q_61_dilations_0, groups = q_61_groups_0, pad = q_61_pad_0, pad_type = q_61_pad_type_0, strides = q_61_strides_0, weight = var_8245_weight_0_to_fp16, x = input_303_cast_fp16)[name = tensor("op_8245_cast_fp16")]; + tensor k_61_pad_type_0 = const()[name = tensor("k_61_pad_type_0"), val = tensor("valid")]; + tensor k_61_strides_0 = const()[name = tensor("k_61_strides_0"), val = tensor([1, 1])]; + tensor k_61_pad_0 = const()[name = tensor("k_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_61_dilations_0 = const()[name = tensor("k_61_dilations_0"), val = tensor([1, 1])]; + tensor k_61_groups_0 = const()[name = tensor("k_61_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_key_weight_to_fp16 = const()[name = tensor("blocks_30_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198173440)))]; + tensor k_61_cast_fp16 = conv(dilations = k_61_dilations_0, groups = k_61_groups_0, pad = k_61_pad_0, pad_type = k_61_pad_type_0, strides = k_61_strides_0, weight = blocks_30_attn_key_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor var_8243_pad_type_0 = const()[name = tensor("op_8243_pad_type_0"), val = tensor("valid")]; + tensor var_8243_strides_0 = const()[name = tensor("op_8243_strides_0"), val = tensor([1, 1])]; + tensor var_8243_pad_0 = const()[name = tensor("op_8243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8243_dilations_0 = const()[name = tensor("op_8243_dilations_0"), val = tensor([1, 1])]; + tensor var_8243_groups_0 = const()[name = tensor("op_8243_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_value_weight_to_fp16 = const()[name = tensor("blocks_30_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201450304)))]; + tensor blocks_30_attn_value_bias_to_fp16 = const()[name = tensor("blocks_30_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204727168)))]; + tensor var_8243_cast_fp16 = conv(bias = blocks_30_attn_value_bias_to_fp16, dilations = var_8243_dilations_0, groups = var_8243_groups_0, pad = var_8243_pad_0, pad_type = var_8243_pad_type_0, strides = var_8243_strides_0, weight = blocks_30_attn_value_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("op_8243_cast_fp16")]; + tensor tile_90 = const()[name = tensor("tile_90"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8246_axis_0 = const()[name = tensor("op_8246_axis_0"), val = tensor(1)]; + tensor var_8246_cast_fp16_0, tensor var_8246_cast_fp16_1, tensor var_8246_cast_fp16_2, tensor var_8246_cast_fp16_3, tensor var_8246_cast_fp16_4, tensor var_8246_cast_fp16_5, tensor var_8246_cast_fp16_6, tensor var_8246_cast_fp16_7, tensor var_8246_cast_fp16_8, tensor var_8246_cast_fp16_9, tensor var_8246_cast_fp16_10, tensor var_8246_cast_fp16_11, tensor var_8246_cast_fp16_12, tensor var_8246_cast_fp16_13, tensor var_8246_cast_fp16_14, tensor var_8246_cast_fp16_15, tensor var_8246_cast_fp16_16, tensor var_8246_cast_fp16_17, tensor var_8246_cast_fp16_18, tensor var_8246_cast_fp16_19 = split(axis = var_8246_axis_0, split_sizes = tile_90, x = var_8245_cast_fp16)[name = tensor("op_8246_cast_fp16")]; + tensor var_8267_perm_0 = const()[name = tensor("op_8267_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_91 = const()[name = tensor("tile_91"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8268_axis_0 = const()[name = tensor("op_8268_axis_0"), val = tensor(3)]; + tensor var_8267_cast_fp16 = transpose(perm = var_8267_perm_0, x = k_61_cast_fp16)[name = tensor("transpose_2")]; + tensor var_8268_cast_fp16_0, tensor var_8268_cast_fp16_1, tensor var_8268_cast_fp16_2, tensor var_8268_cast_fp16_3, tensor var_8268_cast_fp16_4, tensor var_8268_cast_fp16_5, tensor var_8268_cast_fp16_6, tensor var_8268_cast_fp16_7, tensor var_8268_cast_fp16_8, tensor var_8268_cast_fp16_9, tensor var_8268_cast_fp16_10, tensor var_8268_cast_fp16_11, tensor var_8268_cast_fp16_12, tensor var_8268_cast_fp16_13, tensor var_8268_cast_fp16_14, tensor var_8268_cast_fp16_15, tensor var_8268_cast_fp16_16, tensor var_8268_cast_fp16_17, tensor var_8268_cast_fp16_18, tensor var_8268_cast_fp16_19 = split(axis = var_8268_axis_0, split_sizes = tile_91, x = var_8267_cast_fp16)[name = tensor("op_8268_cast_fp16")]; + tensor tile_92 = const()[name = tensor("tile_92"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8289_axis_0 = const()[name = tensor("op_8289_axis_0"), val = tensor(1)]; + tensor var_8289_cast_fp16_0, tensor var_8289_cast_fp16_1, tensor var_8289_cast_fp16_2, tensor var_8289_cast_fp16_3, tensor var_8289_cast_fp16_4, tensor var_8289_cast_fp16_5, tensor var_8289_cast_fp16_6, tensor var_8289_cast_fp16_7, tensor var_8289_cast_fp16_8, tensor var_8289_cast_fp16_9, tensor var_8289_cast_fp16_10, tensor var_8289_cast_fp16_11, tensor var_8289_cast_fp16_12, tensor var_8289_cast_fp16_13, tensor var_8289_cast_fp16_14, tensor var_8289_cast_fp16_15, tensor var_8289_cast_fp16_16, tensor var_8289_cast_fp16_17, tensor var_8289_cast_fp16_18, tensor var_8289_cast_fp16_19 = split(axis = var_8289_axis_0, split_sizes = tile_92, x = var_8243_cast_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor aw_1201_equation_0 = const()[name = tensor("aw_1201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1201_cast_fp16 = einsum(equation = aw_1201_equation_0, values = (var_8268_cast_fp16_0, var_8246_cast_fp16_0))[name = tensor("aw_1201_cast_fp16")]; + tensor aw_1203_equation_0 = const()[name = tensor("aw_1203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1203_cast_fp16 = einsum(equation = aw_1203_equation_0, values = (var_8268_cast_fp16_1, var_8246_cast_fp16_1))[name = tensor("aw_1203_cast_fp16")]; + tensor aw_1205_equation_0 = const()[name = tensor("aw_1205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1205_cast_fp16 = einsum(equation = aw_1205_equation_0, values = (var_8268_cast_fp16_2, var_8246_cast_fp16_2))[name = tensor("aw_1205_cast_fp16")]; + tensor aw_1207_equation_0 = const()[name = tensor("aw_1207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1207_cast_fp16 = einsum(equation = aw_1207_equation_0, values = (var_8268_cast_fp16_3, var_8246_cast_fp16_3))[name = tensor("aw_1207_cast_fp16")]; + tensor aw_1209_equation_0 = const()[name = tensor("aw_1209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1209_cast_fp16 = einsum(equation = aw_1209_equation_0, values = (var_8268_cast_fp16_4, var_8246_cast_fp16_4))[name = tensor("aw_1209_cast_fp16")]; + tensor aw_1211_equation_0 = const()[name = tensor("aw_1211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1211_cast_fp16 = einsum(equation = aw_1211_equation_0, values = (var_8268_cast_fp16_5, var_8246_cast_fp16_5))[name = tensor("aw_1211_cast_fp16")]; + tensor aw_1213_equation_0 = const()[name = tensor("aw_1213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1213_cast_fp16 = einsum(equation = aw_1213_equation_0, values = (var_8268_cast_fp16_6, var_8246_cast_fp16_6))[name = tensor("aw_1213_cast_fp16")]; + tensor aw_1215_equation_0 = const()[name = tensor("aw_1215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1215_cast_fp16 = einsum(equation = aw_1215_equation_0, values = (var_8268_cast_fp16_7, var_8246_cast_fp16_7))[name = tensor("aw_1215_cast_fp16")]; + tensor aw_1217_equation_0 = const()[name = tensor("aw_1217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1217_cast_fp16 = einsum(equation = aw_1217_equation_0, values = (var_8268_cast_fp16_8, var_8246_cast_fp16_8))[name = tensor("aw_1217_cast_fp16")]; + tensor aw_1219_equation_0 = const()[name = tensor("aw_1219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1219_cast_fp16 = einsum(equation = aw_1219_equation_0, values = (var_8268_cast_fp16_9, var_8246_cast_fp16_9))[name = tensor("aw_1219_cast_fp16")]; + tensor aw_1221_equation_0 = const()[name = tensor("aw_1221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1221_cast_fp16 = einsum(equation = aw_1221_equation_0, values = (var_8268_cast_fp16_10, var_8246_cast_fp16_10))[name = tensor("aw_1221_cast_fp16")]; + tensor aw_1223_equation_0 = const()[name = tensor("aw_1223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1223_cast_fp16 = einsum(equation = aw_1223_equation_0, values = (var_8268_cast_fp16_11, var_8246_cast_fp16_11))[name = tensor("aw_1223_cast_fp16")]; + tensor aw_1225_equation_0 = const()[name = tensor("aw_1225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1225_cast_fp16 = einsum(equation = aw_1225_equation_0, values = (var_8268_cast_fp16_12, var_8246_cast_fp16_12))[name = tensor("aw_1225_cast_fp16")]; + tensor aw_1227_equation_0 = const()[name = tensor("aw_1227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1227_cast_fp16 = einsum(equation = aw_1227_equation_0, values = (var_8268_cast_fp16_13, var_8246_cast_fp16_13))[name = tensor("aw_1227_cast_fp16")]; + tensor aw_1229_equation_0 = const()[name = tensor("aw_1229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1229_cast_fp16 = einsum(equation = aw_1229_equation_0, values = (var_8268_cast_fp16_14, var_8246_cast_fp16_14))[name = tensor("aw_1229_cast_fp16")]; + tensor aw_1231_equation_0 = const()[name = tensor("aw_1231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1231_cast_fp16 = einsum(equation = aw_1231_equation_0, values = (var_8268_cast_fp16_15, var_8246_cast_fp16_15))[name = tensor("aw_1231_cast_fp16")]; + tensor aw_1233_equation_0 = const()[name = tensor("aw_1233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1233_cast_fp16 = einsum(equation = aw_1233_equation_0, values = (var_8268_cast_fp16_16, var_8246_cast_fp16_16))[name = tensor("aw_1233_cast_fp16")]; + tensor aw_1235_equation_0 = const()[name = tensor("aw_1235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1235_cast_fp16 = einsum(equation = aw_1235_equation_0, values = (var_8268_cast_fp16_17, var_8246_cast_fp16_17))[name = tensor("aw_1235_cast_fp16")]; + tensor aw_1237_equation_0 = const()[name = tensor("aw_1237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1237_cast_fp16 = einsum(equation = aw_1237_equation_0, values = (var_8268_cast_fp16_18, var_8246_cast_fp16_18))[name = tensor("aw_1237_cast_fp16")]; + tensor aw_1239_equation_0 = const()[name = tensor("aw_1239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1239_cast_fp16 = einsum(equation = aw_1239_equation_0, values = (var_8268_cast_fp16_19, var_8246_cast_fp16_19))[name = tensor("aw_1239_cast_fp16")]; + tensor var_8350_cast_fp16 = softmax(axis = var_8194, x = aw_1201_cast_fp16)[name = tensor("op_8350_cast_fp16")]; + tensor var_8351_cast_fp16 = softmax(axis = var_8194, x = aw_1203_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = softmax(axis = var_8194, x = aw_1205_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor var_8353_cast_fp16 = softmax(axis = var_8194, x = aw_1207_cast_fp16)[name = tensor("op_8353_cast_fp16")]; + tensor var_8354_cast_fp16 = softmax(axis = var_8194, x = aw_1209_cast_fp16)[name = tensor("op_8354_cast_fp16")]; + tensor var_8355_cast_fp16 = softmax(axis = var_8194, x = aw_1211_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = softmax(axis = var_8194, x = aw_1213_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor var_8357_cast_fp16 = softmax(axis = var_8194, x = aw_1215_cast_fp16)[name = tensor("op_8357_cast_fp16")]; + tensor var_8358_cast_fp16 = softmax(axis = var_8194, x = aw_1217_cast_fp16)[name = tensor("op_8358_cast_fp16")]; + tensor var_8359_cast_fp16 = softmax(axis = var_8194, x = aw_1219_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = softmax(axis = var_8194, x = aw_1221_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor var_8361_cast_fp16 = softmax(axis = var_8194, x = aw_1223_cast_fp16)[name = tensor("op_8361_cast_fp16")]; + tensor var_8362_cast_fp16 = softmax(axis = var_8194, x = aw_1225_cast_fp16)[name = tensor("op_8362_cast_fp16")]; + tensor var_8363_cast_fp16 = softmax(axis = var_8194, x = aw_1227_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = softmax(axis = var_8194, x = aw_1229_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor var_8365_cast_fp16 = softmax(axis = var_8194, x = aw_1231_cast_fp16)[name = tensor("op_8365_cast_fp16")]; + tensor var_8366_cast_fp16 = softmax(axis = var_8194, x = aw_1233_cast_fp16)[name = tensor("op_8366_cast_fp16")]; + tensor var_8367_cast_fp16 = softmax(axis = var_8194, x = aw_1235_cast_fp16)[name = tensor("op_8367_cast_fp16")]; + tensor var_8368_cast_fp16 = softmax(axis = var_8194, x = aw_1237_cast_fp16)[name = tensor("op_8368_cast_fp16")]; + tensor var_8369_cast_fp16 = softmax(axis = var_8194, x = aw_1239_cast_fp16)[name = tensor("op_8369_cast_fp16")]; + tensor var_8371_equation_0 = const()[name = tensor("op_8371_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8371_cast_fp16 = einsum(equation = var_8371_equation_0, values = (var_8289_cast_fp16_0, var_8350_cast_fp16))[name = tensor("op_8371_cast_fp16")]; + tensor var_8373_equation_0 = const()[name = tensor("op_8373_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8373_cast_fp16 = einsum(equation = var_8373_equation_0, values = (var_8289_cast_fp16_1, var_8351_cast_fp16))[name = tensor("op_8373_cast_fp16")]; + tensor var_8375_equation_0 = const()[name = tensor("op_8375_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8375_cast_fp16 = einsum(equation = var_8375_equation_0, values = (var_8289_cast_fp16_2, var_8352_cast_fp16))[name = tensor("op_8375_cast_fp16")]; + tensor var_8377_equation_0 = const()[name = tensor("op_8377_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8377_cast_fp16 = einsum(equation = var_8377_equation_0, values = (var_8289_cast_fp16_3, var_8353_cast_fp16))[name = tensor("op_8377_cast_fp16")]; + tensor var_8379_equation_0 = const()[name = tensor("op_8379_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8379_cast_fp16 = einsum(equation = var_8379_equation_0, values = (var_8289_cast_fp16_4, var_8354_cast_fp16))[name = tensor("op_8379_cast_fp16")]; + tensor var_8381_equation_0 = const()[name = tensor("op_8381_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8381_cast_fp16 = einsum(equation = var_8381_equation_0, values = (var_8289_cast_fp16_5, var_8355_cast_fp16))[name = tensor("op_8381_cast_fp16")]; + tensor var_8383_equation_0 = const()[name = tensor("op_8383_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8383_cast_fp16 = einsum(equation = var_8383_equation_0, values = (var_8289_cast_fp16_6, var_8356_cast_fp16))[name = tensor("op_8383_cast_fp16")]; + tensor var_8385_equation_0 = const()[name = tensor("op_8385_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8385_cast_fp16 = einsum(equation = var_8385_equation_0, values = (var_8289_cast_fp16_7, var_8357_cast_fp16))[name = tensor("op_8385_cast_fp16")]; + tensor var_8387_equation_0 = const()[name = tensor("op_8387_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8387_cast_fp16 = einsum(equation = var_8387_equation_0, values = (var_8289_cast_fp16_8, var_8358_cast_fp16))[name = tensor("op_8387_cast_fp16")]; + tensor var_8389_equation_0 = const()[name = tensor("op_8389_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8389_cast_fp16 = einsum(equation = var_8389_equation_0, values = (var_8289_cast_fp16_9, var_8359_cast_fp16))[name = tensor("op_8389_cast_fp16")]; + tensor var_8391_equation_0 = const()[name = tensor("op_8391_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8391_cast_fp16 = einsum(equation = var_8391_equation_0, values = (var_8289_cast_fp16_10, var_8360_cast_fp16))[name = tensor("op_8391_cast_fp16")]; + tensor var_8393_equation_0 = const()[name = tensor("op_8393_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8393_cast_fp16 = einsum(equation = var_8393_equation_0, values = (var_8289_cast_fp16_11, var_8361_cast_fp16))[name = tensor("op_8393_cast_fp16")]; + tensor var_8395_equation_0 = const()[name = tensor("op_8395_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8395_cast_fp16 = einsum(equation = var_8395_equation_0, values = (var_8289_cast_fp16_12, var_8362_cast_fp16))[name = tensor("op_8395_cast_fp16")]; + tensor var_8397_equation_0 = const()[name = tensor("op_8397_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8397_cast_fp16 = einsum(equation = var_8397_equation_0, values = (var_8289_cast_fp16_13, var_8363_cast_fp16))[name = tensor("op_8397_cast_fp16")]; + tensor var_8399_equation_0 = const()[name = tensor("op_8399_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8399_cast_fp16 = einsum(equation = var_8399_equation_0, values = (var_8289_cast_fp16_14, var_8364_cast_fp16))[name = tensor("op_8399_cast_fp16")]; + tensor var_8401_equation_0 = const()[name = tensor("op_8401_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8401_cast_fp16 = einsum(equation = var_8401_equation_0, values = (var_8289_cast_fp16_15, var_8365_cast_fp16))[name = tensor("op_8401_cast_fp16")]; + tensor var_8403_equation_0 = const()[name = tensor("op_8403_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8403_cast_fp16 = einsum(equation = var_8403_equation_0, values = (var_8289_cast_fp16_16, var_8366_cast_fp16))[name = tensor("op_8403_cast_fp16")]; + tensor var_8405_equation_0 = const()[name = tensor("op_8405_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8405_cast_fp16 = einsum(equation = var_8405_equation_0, values = (var_8289_cast_fp16_17, var_8367_cast_fp16))[name = tensor("op_8405_cast_fp16")]; + tensor var_8407_equation_0 = const()[name = tensor("op_8407_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8407_cast_fp16 = einsum(equation = var_8407_equation_0, values = (var_8289_cast_fp16_18, var_8368_cast_fp16))[name = tensor("op_8407_cast_fp16")]; + tensor var_8409_equation_0 = const()[name = tensor("op_8409_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8409_cast_fp16 = einsum(equation = var_8409_equation_0, values = (var_8289_cast_fp16_19, var_8369_cast_fp16))[name = tensor("op_8409_cast_fp16")]; + tensor input_305_interleave_0 = const()[name = tensor("input_305_interleave_0"), val = tensor(false)]; + tensor input_305_cast_fp16 = concat(axis = var_8194, interleave = input_305_interleave_0, values = (var_8371_cast_fp16, var_8373_cast_fp16, var_8375_cast_fp16, var_8377_cast_fp16, var_8379_cast_fp16, var_8381_cast_fp16, var_8383_cast_fp16, var_8385_cast_fp16, var_8387_cast_fp16, var_8389_cast_fp16, var_8391_cast_fp16, var_8393_cast_fp16, var_8395_cast_fp16, var_8397_cast_fp16, var_8399_cast_fp16, var_8401_cast_fp16, var_8403_cast_fp16, var_8405_cast_fp16, var_8407_cast_fp16, var_8409_cast_fp16))[name = tensor("input_305_cast_fp16")]; + tensor var_8418_pad_type_0 = const()[name = tensor("op_8418_pad_type_0"), val = tensor("valid")]; + tensor var_8418_strides_0 = const()[name = tensor("op_8418_strides_0"), val = tensor([1, 1])]; + tensor var_8418_pad_0 = const()[name = tensor("op_8418_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8418_dilations_0 = const()[name = tensor("op_8418_dilations_0"), val = tensor([1, 1])]; + tensor var_8418_groups_0 = const()[name = tensor("op_8418_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_out_weight_to_fp16 = const()[name = tensor("blocks_30_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204729792)))]; + tensor blocks_30_attn_out_bias_to_fp16 = const()[name = tensor("blocks_30_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208006656)))]; + tensor var_8418_cast_fp16 = conv(bias = blocks_30_attn_out_bias_to_fp16, dilations = var_8418_dilations_0, groups = var_8418_groups_0, pad = var_8418_pad_0, pad_type = var_8418_pad_type_0, strides = var_8418_strides_0, weight = blocks_30_attn_out_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_8418_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor input_307_axes_0 = const()[name = tensor("input_307_axes_0"), val = tensor([1])]; + tensor input_307_gamma_0_to_fp16 = const()[name = tensor("input_307_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208009280)))]; + tensor input_307_beta_0_to_fp16 = const()[name = tensor("input_307_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208011904)))]; + tensor var_8428_to_fp16 = const()[name = tensor("op_8428_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = input_307_beta_0_to_fp16, epsilon = var_8428_to_fp16, gamma = input_307_gamma_0_to_fp16, x = inputs_123_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor input_309_pad_type_0 = const()[name = tensor("input_309_pad_type_0"), val = tensor("valid")]; + tensor input_309_strides_0 = const()[name = tensor("input_309_strides_0"), val = tensor([1, 1])]; + tensor input_309_pad_0 = const()[name = tensor("input_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_309_dilations_0 = const()[name = tensor("input_309_dilations_0"), val = tensor([1, 1])]; + tensor input_309_groups_0 = const()[name = tensor("input_309_groups_0"), val = tensor(1)]; + tensor blocks_30_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208014528)))]; + tensor blocks_30_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221121792)))]; + tensor input_309_cast_fp16 = conv(bias = blocks_30_mlp_0_bias_to_fp16, dilations = input_309_dilations_0, groups = input_309_groups_0, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = input_309_strides_0, weight = blocks_30_mlp_0_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor input_311_mode_0 = const()[name = tensor("input_311_mode_0"), val = tensor("EXACT")]; + tensor input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = input_309_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor var_8454_pad_type_0 = const()[name = tensor("op_8454_pad_type_0"), val = tensor("valid")]; + tensor var_8454_strides_0 = const()[name = tensor("op_8454_strides_0"), val = tensor([1, 1])]; + tensor var_8454_pad_0 = const()[name = tensor("op_8454_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8454_dilations_0 = const()[name = tensor("op_8454_dilations_0"), val = tensor([1, 1])]; + tensor var_8454_groups_0 = const()[name = tensor("op_8454_groups_0"), val = tensor(1)]; + tensor blocks_30_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221132096)))]; + tensor blocks_30_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234239360)))]; + tensor var_8454_cast_fp16 = conv(bias = blocks_30_mlp_2_bias_to_fp16, dilations = var_8454_dilations_0, groups = var_8454_groups_0, pad = var_8454_pad_0, pad_type = var_8454_pad_type_0, strides = var_8454_strides_0, weight = blocks_30_mlp_2_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = var_8454_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_8463 = const()[name = tensor("op_8463"), val = tensor(1)]; + tensor input_313_axes_0 = const()[name = tensor("input_313_axes_0"), val = tensor([1])]; + tensor input_313_gamma_0_to_fp16 = const()[name = tensor("input_313_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234241984)))]; + tensor input_313_beta_0_to_fp16 = const()[name = tensor("input_313_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234244608)))]; + tensor var_8479_to_fp16 = const()[name = tensor("op_8479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_313_cast_fp16 = layer_norm(axes = input_313_axes_0, beta = input_313_beta_0_to_fp16, epsilon = var_8479_to_fp16, gamma = input_313_gamma_0_to_fp16, x = inputs_125_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_8514_weight_0_to_fp16 = const()[name = tensor("op_8514_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234247232)))]; + tensor var_8514_bias_0_to_fp16 = const()[name = tensor("op_8514_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237524096)))]; + tensor var_8514_cast_fp16 = conv(bias = var_8514_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_8514_weight_0_to_fp16, x = input_313_cast_fp16)[name = tensor("op_8514_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_key_weight_to_fp16 = const()[name = tensor("blocks_31_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237526720)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_31_attn_key_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_8512_pad_type_0 = const()[name = tensor("op_8512_pad_type_0"), val = tensor("valid")]; + tensor var_8512_strides_0 = const()[name = tensor("op_8512_strides_0"), val = tensor([1, 1])]; + tensor var_8512_pad_0 = const()[name = tensor("op_8512_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8512_dilations_0 = const()[name = tensor("op_8512_dilations_0"), val = tensor([1, 1])]; + tensor var_8512_groups_0 = const()[name = tensor("op_8512_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_value_weight_to_fp16 = const()[name = tensor("blocks_31_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1240803584)))]; + tensor blocks_31_attn_value_bias_to_fp16 = const()[name = tensor("blocks_31_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244080448)))]; + tensor var_8512_cast_fp16 = conv(bias = blocks_31_attn_value_bias_to_fp16, dilations = var_8512_dilations_0, groups = var_8512_groups_0, pad = var_8512_pad_0, pad_type = var_8512_pad_type_0, strides = var_8512_strides_0, weight = blocks_31_attn_value_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("op_8512_cast_fp16")]; + tensor tile_93 = const()[name = tensor("tile_93"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8515_axis_0 = const()[name = tensor("op_8515_axis_0"), val = tensor(1)]; + tensor var_8515_cast_fp16_0, tensor var_8515_cast_fp16_1, tensor var_8515_cast_fp16_2, tensor var_8515_cast_fp16_3, tensor var_8515_cast_fp16_4, tensor var_8515_cast_fp16_5, tensor var_8515_cast_fp16_6, tensor var_8515_cast_fp16_7, tensor var_8515_cast_fp16_8, tensor var_8515_cast_fp16_9, tensor var_8515_cast_fp16_10, tensor var_8515_cast_fp16_11, tensor var_8515_cast_fp16_12, tensor var_8515_cast_fp16_13, tensor var_8515_cast_fp16_14, tensor var_8515_cast_fp16_15, tensor var_8515_cast_fp16_16, tensor var_8515_cast_fp16_17, tensor var_8515_cast_fp16_18, tensor var_8515_cast_fp16_19 = split(axis = var_8515_axis_0, split_sizes = tile_93, x = var_8514_cast_fp16)[name = tensor("op_8515_cast_fp16")]; + tensor var_8536_perm_0 = const()[name = tensor("op_8536_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_94 = const()[name = tensor("tile_94"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8537_axis_0 = const()[name = tensor("op_8537_axis_0"), val = tensor(3)]; + tensor var_8536_cast_fp16 = transpose(perm = var_8536_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_8537_cast_fp16_0, tensor var_8537_cast_fp16_1, tensor var_8537_cast_fp16_2, tensor var_8537_cast_fp16_3, tensor var_8537_cast_fp16_4, tensor var_8537_cast_fp16_5, tensor var_8537_cast_fp16_6, tensor var_8537_cast_fp16_7, tensor var_8537_cast_fp16_8, tensor var_8537_cast_fp16_9, tensor var_8537_cast_fp16_10, tensor var_8537_cast_fp16_11, tensor var_8537_cast_fp16_12, tensor var_8537_cast_fp16_13, tensor var_8537_cast_fp16_14, tensor var_8537_cast_fp16_15, tensor var_8537_cast_fp16_16, tensor var_8537_cast_fp16_17, tensor var_8537_cast_fp16_18, tensor var_8537_cast_fp16_19 = split(axis = var_8537_axis_0, split_sizes = tile_94, x = var_8536_cast_fp16)[name = tensor("op_8537_cast_fp16")]; + tensor tile_95 = const()[name = tensor("tile_95"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8558_axis_0 = const()[name = tensor("op_8558_axis_0"), val = tensor(1)]; + tensor var_8558_cast_fp16_0, tensor var_8558_cast_fp16_1, tensor var_8558_cast_fp16_2, tensor var_8558_cast_fp16_3, tensor var_8558_cast_fp16_4, tensor var_8558_cast_fp16_5, tensor var_8558_cast_fp16_6, tensor var_8558_cast_fp16_7, tensor var_8558_cast_fp16_8, tensor var_8558_cast_fp16_9, tensor var_8558_cast_fp16_10, tensor var_8558_cast_fp16_11, tensor var_8558_cast_fp16_12, tensor var_8558_cast_fp16_13, tensor var_8558_cast_fp16_14, tensor var_8558_cast_fp16_15, tensor var_8558_cast_fp16_16, tensor var_8558_cast_fp16_17, tensor var_8558_cast_fp16_18, tensor var_8558_cast_fp16_19 = split(axis = var_8558_axis_0, split_sizes = tile_95, x = var_8512_cast_fp16)[name = tensor("op_8558_cast_fp16")]; + tensor aw_1241_equation_0 = const()[name = tensor("aw_1241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1241_cast_fp16 = einsum(equation = aw_1241_equation_0, values = (var_8537_cast_fp16_0, var_8515_cast_fp16_0))[name = tensor("aw_1241_cast_fp16")]; + tensor aw_1243_equation_0 = const()[name = tensor("aw_1243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1243_cast_fp16 = einsum(equation = aw_1243_equation_0, values = (var_8537_cast_fp16_1, var_8515_cast_fp16_1))[name = tensor("aw_1243_cast_fp16")]; + tensor aw_1245_equation_0 = const()[name = tensor("aw_1245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1245_cast_fp16 = einsum(equation = aw_1245_equation_0, values = (var_8537_cast_fp16_2, var_8515_cast_fp16_2))[name = tensor("aw_1245_cast_fp16")]; + tensor aw_1247_equation_0 = const()[name = tensor("aw_1247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1247_cast_fp16 = einsum(equation = aw_1247_equation_0, values = (var_8537_cast_fp16_3, var_8515_cast_fp16_3))[name = tensor("aw_1247_cast_fp16")]; + tensor aw_1249_equation_0 = const()[name = tensor("aw_1249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1249_cast_fp16 = einsum(equation = aw_1249_equation_0, values = (var_8537_cast_fp16_4, var_8515_cast_fp16_4))[name = tensor("aw_1249_cast_fp16")]; + tensor aw_1251_equation_0 = const()[name = tensor("aw_1251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1251_cast_fp16 = einsum(equation = aw_1251_equation_0, values = (var_8537_cast_fp16_5, var_8515_cast_fp16_5))[name = tensor("aw_1251_cast_fp16")]; + tensor aw_1253_equation_0 = const()[name = tensor("aw_1253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1253_cast_fp16 = einsum(equation = aw_1253_equation_0, values = (var_8537_cast_fp16_6, var_8515_cast_fp16_6))[name = tensor("aw_1253_cast_fp16")]; + tensor aw_1255_equation_0 = const()[name = tensor("aw_1255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1255_cast_fp16 = einsum(equation = aw_1255_equation_0, values = (var_8537_cast_fp16_7, var_8515_cast_fp16_7))[name = tensor("aw_1255_cast_fp16")]; + tensor aw_1257_equation_0 = const()[name = tensor("aw_1257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1257_cast_fp16 = einsum(equation = aw_1257_equation_0, values = (var_8537_cast_fp16_8, var_8515_cast_fp16_8))[name = tensor("aw_1257_cast_fp16")]; + tensor aw_1259_equation_0 = const()[name = tensor("aw_1259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1259_cast_fp16 = einsum(equation = aw_1259_equation_0, values = (var_8537_cast_fp16_9, var_8515_cast_fp16_9))[name = tensor("aw_1259_cast_fp16")]; + tensor aw_1261_equation_0 = const()[name = tensor("aw_1261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1261_cast_fp16 = einsum(equation = aw_1261_equation_0, values = (var_8537_cast_fp16_10, var_8515_cast_fp16_10))[name = tensor("aw_1261_cast_fp16")]; + tensor aw_1263_equation_0 = const()[name = tensor("aw_1263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1263_cast_fp16 = einsum(equation = aw_1263_equation_0, values = (var_8537_cast_fp16_11, var_8515_cast_fp16_11))[name = tensor("aw_1263_cast_fp16")]; + tensor aw_1265_equation_0 = const()[name = tensor("aw_1265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1265_cast_fp16 = einsum(equation = aw_1265_equation_0, values = (var_8537_cast_fp16_12, var_8515_cast_fp16_12))[name = tensor("aw_1265_cast_fp16")]; + tensor aw_1267_equation_0 = const()[name = tensor("aw_1267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1267_cast_fp16 = einsum(equation = aw_1267_equation_0, values = (var_8537_cast_fp16_13, var_8515_cast_fp16_13))[name = tensor("aw_1267_cast_fp16")]; + tensor aw_1269_equation_0 = const()[name = tensor("aw_1269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1269_cast_fp16 = einsum(equation = aw_1269_equation_0, values = (var_8537_cast_fp16_14, var_8515_cast_fp16_14))[name = tensor("aw_1269_cast_fp16")]; + tensor aw_1271_equation_0 = const()[name = tensor("aw_1271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1271_cast_fp16 = einsum(equation = aw_1271_equation_0, values = (var_8537_cast_fp16_15, var_8515_cast_fp16_15))[name = tensor("aw_1271_cast_fp16")]; + tensor aw_1273_equation_0 = const()[name = tensor("aw_1273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1273_cast_fp16 = einsum(equation = aw_1273_equation_0, values = (var_8537_cast_fp16_16, var_8515_cast_fp16_16))[name = tensor("aw_1273_cast_fp16")]; + tensor aw_1275_equation_0 = const()[name = tensor("aw_1275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1275_cast_fp16 = einsum(equation = aw_1275_equation_0, values = (var_8537_cast_fp16_17, var_8515_cast_fp16_17))[name = tensor("aw_1275_cast_fp16")]; + tensor aw_1277_equation_0 = const()[name = tensor("aw_1277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1277_cast_fp16 = einsum(equation = aw_1277_equation_0, values = (var_8537_cast_fp16_18, var_8515_cast_fp16_18))[name = tensor("aw_1277_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_8537_cast_fp16_19, var_8515_cast_fp16_19))[name = tensor("aw_cast_fp16")]; + tensor var_8619_cast_fp16 = softmax(axis = var_8463, x = aw_1241_cast_fp16)[name = tensor("op_8619_cast_fp16")]; + tensor var_8620_cast_fp16 = softmax(axis = var_8463, x = aw_1243_cast_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor var_8621_cast_fp16 = softmax(axis = var_8463, x = aw_1245_cast_fp16)[name = tensor("op_8621_cast_fp16")]; + tensor var_8622_cast_fp16 = softmax(axis = var_8463, x = aw_1247_cast_fp16)[name = tensor("op_8622_cast_fp16")]; + tensor var_8623_cast_fp16 = softmax(axis = var_8463, x = aw_1249_cast_fp16)[name = tensor("op_8623_cast_fp16")]; + tensor var_8624_cast_fp16 = softmax(axis = var_8463, x = aw_1251_cast_fp16)[name = tensor("op_8624_cast_fp16")]; + tensor var_8625_cast_fp16 = softmax(axis = var_8463, x = aw_1253_cast_fp16)[name = tensor("op_8625_cast_fp16")]; + tensor var_8626_cast_fp16 = softmax(axis = var_8463, x = aw_1255_cast_fp16)[name = tensor("op_8626_cast_fp16")]; + tensor var_8627_cast_fp16 = softmax(axis = var_8463, x = aw_1257_cast_fp16)[name = tensor("op_8627_cast_fp16")]; + tensor var_8628_cast_fp16 = softmax(axis = var_8463, x = aw_1259_cast_fp16)[name = tensor("op_8628_cast_fp16")]; + tensor var_8629_cast_fp16 = softmax(axis = var_8463, x = aw_1261_cast_fp16)[name = tensor("op_8629_cast_fp16")]; + tensor var_8630_cast_fp16 = softmax(axis = var_8463, x = aw_1263_cast_fp16)[name = tensor("op_8630_cast_fp16")]; + tensor var_8631_cast_fp16 = softmax(axis = var_8463, x = aw_1265_cast_fp16)[name = tensor("op_8631_cast_fp16")]; + tensor var_8632_cast_fp16 = softmax(axis = var_8463, x = aw_1267_cast_fp16)[name = tensor("op_8632_cast_fp16")]; + tensor var_8633_cast_fp16 = softmax(axis = var_8463, x = aw_1269_cast_fp16)[name = tensor("op_8633_cast_fp16")]; + tensor var_8634_cast_fp16 = softmax(axis = var_8463, x = aw_1271_cast_fp16)[name = tensor("op_8634_cast_fp16")]; + tensor var_8635_cast_fp16 = softmax(axis = var_8463, x = aw_1273_cast_fp16)[name = tensor("op_8635_cast_fp16")]; + tensor var_8636_cast_fp16 = softmax(axis = var_8463, x = aw_1275_cast_fp16)[name = tensor("op_8636_cast_fp16")]; + tensor var_8637_cast_fp16 = softmax(axis = var_8463, x = aw_1277_cast_fp16)[name = tensor("op_8637_cast_fp16")]; + tensor var_8638_cast_fp16 = softmax(axis = var_8463, x = aw_cast_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor var_8640_equation_0 = const()[name = tensor("op_8640_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8640_cast_fp16 = einsum(equation = var_8640_equation_0, values = (var_8558_cast_fp16_0, var_8619_cast_fp16))[name = tensor("op_8640_cast_fp16")]; + tensor var_8642_equation_0 = const()[name = tensor("op_8642_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8642_cast_fp16 = einsum(equation = var_8642_equation_0, values = (var_8558_cast_fp16_1, var_8620_cast_fp16))[name = tensor("op_8642_cast_fp16")]; + tensor var_8644_equation_0 = const()[name = tensor("op_8644_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8644_cast_fp16 = einsum(equation = var_8644_equation_0, values = (var_8558_cast_fp16_2, var_8621_cast_fp16))[name = tensor("op_8644_cast_fp16")]; + tensor var_8646_equation_0 = const()[name = tensor("op_8646_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8646_cast_fp16 = einsum(equation = var_8646_equation_0, values = (var_8558_cast_fp16_3, var_8622_cast_fp16))[name = tensor("op_8646_cast_fp16")]; + tensor var_8648_equation_0 = const()[name = tensor("op_8648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8648_cast_fp16 = einsum(equation = var_8648_equation_0, values = (var_8558_cast_fp16_4, var_8623_cast_fp16))[name = tensor("op_8648_cast_fp16")]; + tensor var_8650_equation_0 = const()[name = tensor("op_8650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8650_cast_fp16 = einsum(equation = var_8650_equation_0, values = (var_8558_cast_fp16_5, var_8624_cast_fp16))[name = tensor("op_8650_cast_fp16")]; + tensor var_8652_equation_0 = const()[name = tensor("op_8652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8652_cast_fp16 = einsum(equation = var_8652_equation_0, values = (var_8558_cast_fp16_6, var_8625_cast_fp16))[name = tensor("op_8652_cast_fp16")]; + tensor var_8654_equation_0 = const()[name = tensor("op_8654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8654_cast_fp16 = einsum(equation = var_8654_equation_0, values = (var_8558_cast_fp16_7, var_8626_cast_fp16))[name = tensor("op_8654_cast_fp16")]; + tensor var_8656_equation_0 = const()[name = tensor("op_8656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8656_cast_fp16 = einsum(equation = var_8656_equation_0, values = (var_8558_cast_fp16_8, var_8627_cast_fp16))[name = tensor("op_8656_cast_fp16")]; + tensor var_8658_equation_0 = const()[name = tensor("op_8658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8658_cast_fp16 = einsum(equation = var_8658_equation_0, values = (var_8558_cast_fp16_9, var_8628_cast_fp16))[name = tensor("op_8658_cast_fp16")]; + tensor var_8660_equation_0 = const()[name = tensor("op_8660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8660_cast_fp16 = einsum(equation = var_8660_equation_0, values = (var_8558_cast_fp16_10, var_8629_cast_fp16))[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_equation_0 = const()[name = tensor("op_8662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8662_cast_fp16 = einsum(equation = var_8662_equation_0, values = (var_8558_cast_fp16_11, var_8630_cast_fp16))[name = tensor("op_8662_cast_fp16")]; + tensor var_8664_equation_0 = const()[name = tensor("op_8664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8664_cast_fp16 = einsum(equation = var_8664_equation_0, values = (var_8558_cast_fp16_12, var_8631_cast_fp16))[name = tensor("op_8664_cast_fp16")]; + tensor var_8666_equation_0 = const()[name = tensor("op_8666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8666_cast_fp16 = einsum(equation = var_8666_equation_0, values = (var_8558_cast_fp16_13, var_8632_cast_fp16))[name = tensor("op_8666_cast_fp16")]; + tensor var_8668_equation_0 = const()[name = tensor("op_8668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8668_cast_fp16 = einsum(equation = var_8668_equation_0, values = (var_8558_cast_fp16_14, var_8633_cast_fp16))[name = tensor("op_8668_cast_fp16")]; + tensor var_8670_equation_0 = const()[name = tensor("op_8670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8670_cast_fp16 = einsum(equation = var_8670_equation_0, values = (var_8558_cast_fp16_15, var_8634_cast_fp16))[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_equation_0 = const()[name = tensor("op_8672_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8672_cast_fp16 = einsum(equation = var_8672_equation_0, values = (var_8558_cast_fp16_16, var_8635_cast_fp16))[name = tensor("op_8672_cast_fp16")]; + tensor var_8674_equation_0 = const()[name = tensor("op_8674_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8674_cast_fp16 = einsum(equation = var_8674_equation_0, values = (var_8558_cast_fp16_17, var_8636_cast_fp16))[name = tensor("op_8674_cast_fp16")]; + tensor var_8676_equation_0 = const()[name = tensor("op_8676_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8676_cast_fp16 = einsum(equation = var_8676_equation_0, values = (var_8558_cast_fp16_18, var_8637_cast_fp16))[name = tensor("op_8676_cast_fp16")]; + tensor var_8678_equation_0 = const()[name = tensor("op_8678_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8678_cast_fp16 = einsum(equation = var_8678_equation_0, values = (var_8558_cast_fp16_19, var_8638_cast_fp16))[name = tensor("op_8678_cast_fp16")]; + tensor input_315_interleave_0 = const()[name = tensor("input_315_interleave_0"), val = tensor(false)]; + tensor input_315_cast_fp16 = concat(axis = var_8463, interleave = input_315_interleave_0, values = (var_8640_cast_fp16, var_8642_cast_fp16, var_8644_cast_fp16, var_8646_cast_fp16, var_8648_cast_fp16, var_8650_cast_fp16, var_8652_cast_fp16, var_8654_cast_fp16, var_8656_cast_fp16, var_8658_cast_fp16, var_8660_cast_fp16, var_8662_cast_fp16, var_8664_cast_fp16, var_8666_cast_fp16, var_8668_cast_fp16, var_8670_cast_fp16, var_8672_cast_fp16, var_8674_cast_fp16, var_8676_cast_fp16, var_8678_cast_fp16))[name = tensor("input_315_cast_fp16")]; + tensor var_8687_pad_type_0 = const()[name = tensor("op_8687_pad_type_0"), val = tensor("valid")]; + tensor var_8687_strides_0 = const()[name = tensor("op_8687_strides_0"), val = tensor([1, 1])]; + tensor var_8687_pad_0 = const()[name = tensor("op_8687_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8687_dilations_0 = const()[name = tensor("op_8687_dilations_0"), val = tensor([1, 1])]; + tensor var_8687_groups_0 = const()[name = tensor("op_8687_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_out_weight_to_fp16 = const()[name = tensor("blocks_31_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244083072)))]; + tensor blocks_31_attn_out_bias_to_fp16 = const()[name = tensor("blocks_31_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247359936)))]; + tensor var_8687_cast_fp16 = conv(bias = blocks_31_attn_out_bias_to_fp16, dilations = var_8687_dilations_0, groups = var_8687_groups_0, pad = var_8687_pad_0, pad_type = var_8687_pad_type_0, strides = var_8687_strides_0, weight = blocks_31_attn_out_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("op_8687_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = var_8687_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor input_317_axes_0 = const()[name = tensor("input_317_axes_0"), val = tensor([1])]; + tensor input_317_gamma_0_to_fp16 = const()[name = tensor("input_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247362560)))]; + tensor input_317_beta_0_to_fp16 = const()[name = tensor("input_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247365184)))]; + tensor var_8697_to_fp16 = const()[name = tensor("op_8697_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = input_317_beta_0_to_fp16, epsilon = var_8697_to_fp16, gamma = input_317_gamma_0_to_fp16, x = inputs_127_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_319_pad_type_0 = const()[name = tensor("input_319_pad_type_0"), val = tensor("valid")]; + tensor input_319_strides_0 = const()[name = tensor("input_319_strides_0"), val = tensor([1, 1])]; + tensor input_319_pad_0 = const()[name = tensor("input_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_319_dilations_0 = const()[name = tensor("input_319_dilations_0"), val = tensor([1, 1])]; + tensor input_319_groups_0 = const()[name = tensor("input_319_groups_0"), val = tensor(1)]; + tensor blocks_31_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247367808)))]; + tensor blocks_31_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260475072)))]; + tensor input_319_cast_fp16 = conv(bias = blocks_31_mlp_0_bias_to_fp16, dilations = input_319_dilations_0, groups = input_319_groups_0, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = input_319_strides_0, weight = blocks_31_mlp_0_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_319_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_8723_pad_type_0 = const()[name = tensor("op_8723_pad_type_0"), val = tensor("valid")]; + tensor var_8723_strides_0 = const()[name = tensor("op_8723_strides_0"), val = tensor([1, 1])]; + tensor var_8723_pad_0 = const()[name = tensor("op_8723_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8723_dilations_0 = const()[name = tensor("op_8723_dilations_0"), val = tensor([1, 1])]; + tensor var_8723_groups_0 = const()[name = tensor("op_8723_groups_0"), val = tensor(1)]; + tensor blocks_31_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260485376)))]; + tensor blocks_31_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273592640)))]; + tensor var_8723_cast_fp16 = conv(bias = blocks_31_mlp_2_bias_to_fp16, dilations = var_8723_dilations_0, groups = var_8723_groups_0, pad = var_8723_pad_0, pad_type = var_8723_pad_type_0, strides = var_8723_strides_0, weight = blocks_31_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_8723_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_8723_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273595264)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273597888)))]; + tensor var_8737_to_fp16 = const()[name = tensor("op_8737_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_8737_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_8748_axes_0 = const()[name = tensor("op_8748_axes_0"), val = tensor([2])]; + tensor var_8748_cast_fp16 = squeeze(axes = var_8748_axes_0, x = x_cast_fp16)[name = tensor("op_8748_cast_fp16")]; + tensor var_8751_perm_0 = const()[name = tensor("op_8751_perm_0"), val = tensor([0, 2, 1])]; + tensor var_8751_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8751_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_8751_cast_fp16 = transpose(perm = var_8751_perm_0, x = var_8748_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_8751_cast_fp16_to_fp32_dtype_0, x = var_8751_cast_fp16)[name = tensor("cast_131")]; + } -> (output); +} \ No newline at end of file diff --git a/large-v1/ggml-large-v1-encoder.mlmodelc/weights/weight.bin b/large-v1/ggml-large-v1-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2a379fed2102fb3298baa3bb24316bb02adb9800 --- /dev/null +++ b/large-v1/ggml-large-v1-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac5cd426142b4d97debcc46d2f536ae257ad80cc168790faf2ad3dcc862462a6 +size 1273600512 diff --git a/large-v1/ggml-large-v1.bin b/large-v1/ggml-large-v1.bin new file mode 100644 index 0000000000000000000000000000000000000000..047db73ddfd7789113dfc94c20b22bc1d044586f --- /dev/null +++ b/large-v1/ggml-large-v1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d99f41a10525d0206bddadd86760181fa920438b6b33237e3118ff6c83bb53d +size 3094623691 diff --git a/large-v2/.DS_Store b/large-v2/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..931d2f48b3a38feb68e3c2dd9805e9f49f666377 Binary files /dev/null and b/large-v2/.DS_Store differ diff --git a/large-v2/ggml-large-v2-encoder.mlmodelc/analytics/coremldata.bin b/large-v2/ggml-large-v2-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6dc984a46d0a814f3f51e1b9a397639a2cae2e9e --- /dev/null +++ b/large-v2/ggml-large-v2-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ab6e6c98293af2c12bf584cdda19a464709cef895b408c568445cd111d206b9 +size 243 diff --git a/large-v2/ggml-large-v2-encoder.mlmodelc/coremldata.bin b/large-v2/ggml-large-v2-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b1e529c8edbd7450c061f5421f9af2c3995ecfa3 --- /dev/null +++ b/large-v2/ggml-large-v2-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e10fcaa761a372b098ec16f7dabc66ca8334e6970a99cffba37006dd5f8decb +size 320 diff --git a/large-v2/ggml-large-v2-encoder.mlmodelc/metadata.json b/large-v2/ggml-large-v2-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..dc3b811679306c0d52d790854aa514c6a0efe9ca --- /dev/null +++ b/large-v2/ggml-large-v2-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 1280)", + "shortDescription" : "", + "shape" : "[1, 1500, 1280]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 32, + "Gelu" : 34, + "LayerNorm" : 65, + "Transpose" : 33, + "Softmax" : 640, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 65, + "Einsum" : 1280, + "ExpandDims" : 1, + "Split" : 96, + "Conv" : 194 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_large_v2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/large-v2/ggml-large-v2-encoder.mlmodelc/model.mil b/large-v2/ggml-large-v2-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8ab8cd9e6674b7e4122ebb511e9de5d9978c2bb6 --- /dev/null +++ b/large-v2/ggml-large-v2-encoder.mlmodelc/model.mil @@ -0,0 +1,5643 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_84_pad_type_0 = const()[name = tensor("op_84_pad_type_0"), val = tensor("custom")]; + tensor var_84_pad_0 = const()[name = tensor("op_84_pad_0"), val = tensor([1, 1])]; + tensor var_84_strides_0 = const()[name = tensor("op_84_strides_0"), val = tensor([1])]; + tensor var_84_dilations_0 = const()[name = tensor("op_84_dilations_0"), val = tensor([1])]; + tensor var_84_groups_0 = const()[name = tensor("op_84_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614528)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_132")]; + tensor var_84_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_84_dilations_0, groups = var_84_groups_0, pad = var_84_pad_0, pad_type = var_84_pad_type_0, strides = var_84_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_84_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_102_pad_type_0 = const()[name = tensor("op_102_pad_type_0"), val = tensor("custom")]; + tensor var_102_pad_0 = const()[name = tensor("op_102_pad_0"), val = tensor([1, 1])]; + tensor var_102_strides_0 = const()[name = tensor("op_102_strides_0"), val = tensor([2])]; + tensor var_102_dilations_0 = const()[name = tensor("op_102_dilations_0"), val = tensor([1])]; + tensor var_102_groups_0 = const()[name = tensor("op_102_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617152)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10447616)))]; + tensor var_102_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_102_dilations_0, groups = var_102_groups_0, pad = var_102_pad_0, pad_type = var_102_pad_type_0, strides = var_102_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_102_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_102_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_107_to_fp16 = const()[name = tensor("op_107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10450240)))]; + tensor var_109_cast_fp16 = add(x = x_3_cast_fp16, y = var_107_to_fp16)[name = tensor("op_109_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_109_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_124 = const()[name = tensor("op_124"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14290304)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14292928)))]; + tensor var_140_to_fp16 = const()[name = tensor("op_140_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_140_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_175_weight_0_to_fp16 = const()[name = tensor("op_175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14295552)))]; + tensor var_175_bias_0_to_fp16 = const()[name = tensor("op_175_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17572416)))]; + tensor var_175_cast_fp16 = conv(bias = var_175_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_175_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_175_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17575040)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_173_pad_type_0 = const()[name = tensor("op_173_pad_type_0"), val = tensor("valid")]; + tensor var_173_strides_0 = const()[name = tensor("op_173_strides_0"), val = tensor([1, 1])]; + tensor var_173_pad_0 = const()[name = tensor("op_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_173_dilations_0 = const()[name = tensor("op_173_dilations_0"), val = tensor([1, 1])]; + tensor var_173_groups_0 = const()[name = tensor("op_173_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20851904)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24128768)))]; + tensor var_173_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_173_dilations_0, groups = var_173_groups_0, pad = var_173_pad_0, pad_type = var_173_pad_type_0, strides = var_173_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_173_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_176_axis_0 = const()[name = tensor("op_176_axis_0"), val = tensor(1)]; + tensor var_176_cast_fp16_0, tensor var_176_cast_fp16_1, tensor var_176_cast_fp16_2, tensor var_176_cast_fp16_3, tensor var_176_cast_fp16_4, tensor var_176_cast_fp16_5, tensor var_176_cast_fp16_6, tensor var_176_cast_fp16_7, tensor var_176_cast_fp16_8, tensor var_176_cast_fp16_9, tensor var_176_cast_fp16_10, tensor var_176_cast_fp16_11, tensor var_176_cast_fp16_12, tensor var_176_cast_fp16_13, tensor var_176_cast_fp16_14, tensor var_176_cast_fp16_15, tensor var_176_cast_fp16_16, tensor var_176_cast_fp16_17, tensor var_176_cast_fp16_18, tensor var_176_cast_fp16_19 = split(axis = var_176_axis_0, split_sizes = tile_0, x = var_175_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor var_197_perm_0 = const()[name = tensor("op_197_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_198_axis_0 = const()[name = tensor("op_198_axis_0"), val = tensor(3)]; + tensor var_197_cast_fp16 = transpose(perm = var_197_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_32")]; + tensor var_198_cast_fp16_0, tensor var_198_cast_fp16_1, tensor var_198_cast_fp16_2, tensor var_198_cast_fp16_3, tensor var_198_cast_fp16_4, tensor var_198_cast_fp16_5, tensor var_198_cast_fp16_6, tensor var_198_cast_fp16_7, tensor var_198_cast_fp16_8, tensor var_198_cast_fp16_9, tensor var_198_cast_fp16_10, tensor var_198_cast_fp16_11, tensor var_198_cast_fp16_12, tensor var_198_cast_fp16_13, tensor var_198_cast_fp16_14, tensor var_198_cast_fp16_15, tensor var_198_cast_fp16_16, tensor var_198_cast_fp16_17, tensor var_198_cast_fp16_18, tensor var_198_cast_fp16_19 = split(axis = var_198_axis_0, split_sizes = tile_1, x = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_219_axis_0 = const()[name = tensor("op_219_axis_0"), val = tensor(1)]; + tensor var_219_cast_fp16_0, tensor var_219_cast_fp16_1, tensor var_219_cast_fp16_2, tensor var_219_cast_fp16_3, tensor var_219_cast_fp16_4, tensor var_219_cast_fp16_5, tensor var_219_cast_fp16_6, tensor var_219_cast_fp16_7, tensor var_219_cast_fp16_8, tensor var_219_cast_fp16_9, tensor var_219_cast_fp16_10, tensor var_219_cast_fp16_11, tensor var_219_cast_fp16_12, tensor var_219_cast_fp16_13, tensor var_219_cast_fp16_14, tensor var_219_cast_fp16_15, tensor var_219_cast_fp16_16, tensor var_219_cast_fp16_17, tensor var_219_cast_fp16_18, tensor var_219_cast_fp16_19 = split(axis = var_219_axis_0, split_sizes = tile_2, x = var_173_cast_fp16)[name = tensor("op_219_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_198_cast_fp16_0, var_176_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_198_cast_fp16_1, var_176_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_198_cast_fp16_2, var_176_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_198_cast_fp16_3, var_176_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_198_cast_fp16_4, var_176_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_198_cast_fp16_5, var_176_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_198_cast_fp16_6, var_176_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_198_cast_fp16_7, var_176_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_198_cast_fp16_8, var_176_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_198_cast_fp16_9, var_176_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_198_cast_fp16_10, var_176_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_198_cast_fp16_11, var_176_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_198_cast_fp16_12, var_176_cast_fp16_12))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_198_cast_fp16_13, var_176_cast_fp16_13))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_198_cast_fp16_14, var_176_cast_fp16_14))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_198_cast_fp16_15, var_176_cast_fp16_15))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_198_cast_fp16_16, var_176_cast_fp16_16))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_198_cast_fp16_17, var_176_cast_fp16_17))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_198_cast_fp16_18, var_176_cast_fp16_18))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_198_cast_fp16_19, var_176_cast_fp16_19))[name = tensor("aw_39_cast_fp16")]; + tensor var_280_cast_fp16 = softmax(axis = var_124, x = aw_1_cast_fp16)[name = tensor("op_280_cast_fp16")]; + tensor var_281_cast_fp16 = softmax(axis = var_124, x = aw_3_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor var_282_cast_fp16 = softmax(axis = var_124, x = aw_5_cast_fp16)[name = tensor("op_282_cast_fp16")]; + tensor var_283_cast_fp16 = softmax(axis = var_124, x = aw_7_cast_fp16)[name = tensor("op_283_cast_fp16")]; + tensor var_284_cast_fp16 = softmax(axis = var_124, x = aw_9_cast_fp16)[name = tensor("op_284_cast_fp16")]; + tensor var_285_cast_fp16 = softmax(axis = var_124, x = aw_11_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_286_cast_fp16 = softmax(axis = var_124, x = aw_13_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = softmax(axis = var_124, x = aw_15_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor var_288_cast_fp16 = softmax(axis = var_124, x = aw_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_289_cast_fp16 = softmax(axis = var_124, x = aw_19_cast_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_290_cast_fp16 = softmax(axis = var_124, x = aw_21_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = softmax(axis = var_124, x = aw_23_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor var_292_cast_fp16 = softmax(axis = var_124, x = aw_25_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_293_cast_fp16 = softmax(axis = var_124, x = aw_27_cast_fp16)[name = tensor("op_293_cast_fp16")]; + tensor var_294_cast_fp16 = softmax(axis = var_124, x = aw_29_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_295_cast_fp16 = softmax(axis = var_124, x = aw_31_cast_fp16)[name = tensor("op_295_cast_fp16")]; + tensor var_296_cast_fp16 = softmax(axis = var_124, x = aw_33_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_297_cast_fp16 = softmax(axis = var_124, x = aw_35_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_298_cast_fp16 = softmax(axis = var_124, x = aw_37_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_299_cast_fp16 = softmax(axis = var_124, x = aw_39_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_equation_0 = const()[name = tensor("op_301_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_301_cast_fp16 = einsum(equation = var_301_equation_0, values = (var_219_cast_fp16_0, var_280_cast_fp16))[name = tensor("op_301_cast_fp16")]; + tensor var_303_equation_0 = const()[name = tensor("op_303_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_303_cast_fp16 = einsum(equation = var_303_equation_0, values = (var_219_cast_fp16_1, var_281_cast_fp16))[name = tensor("op_303_cast_fp16")]; + tensor var_305_equation_0 = const()[name = tensor("op_305_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_305_cast_fp16 = einsum(equation = var_305_equation_0, values = (var_219_cast_fp16_2, var_282_cast_fp16))[name = tensor("op_305_cast_fp16")]; + tensor var_307_equation_0 = const()[name = tensor("op_307_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_307_cast_fp16 = einsum(equation = var_307_equation_0, values = (var_219_cast_fp16_3, var_283_cast_fp16))[name = tensor("op_307_cast_fp16")]; + tensor var_309_equation_0 = const()[name = tensor("op_309_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_309_cast_fp16 = einsum(equation = var_309_equation_0, values = (var_219_cast_fp16_4, var_284_cast_fp16))[name = tensor("op_309_cast_fp16")]; + tensor var_311_equation_0 = const()[name = tensor("op_311_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_311_cast_fp16 = einsum(equation = var_311_equation_0, values = (var_219_cast_fp16_5, var_285_cast_fp16))[name = tensor("op_311_cast_fp16")]; + tensor var_313_equation_0 = const()[name = tensor("op_313_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_313_cast_fp16 = einsum(equation = var_313_equation_0, values = (var_219_cast_fp16_6, var_286_cast_fp16))[name = tensor("op_313_cast_fp16")]; + tensor var_315_equation_0 = const()[name = tensor("op_315_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_315_cast_fp16 = einsum(equation = var_315_equation_0, values = (var_219_cast_fp16_7, var_287_cast_fp16))[name = tensor("op_315_cast_fp16")]; + tensor var_317_equation_0 = const()[name = tensor("op_317_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_317_cast_fp16 = einsum(equation = var_317_equation_0, values = (var_219_cast_fp16_8, var_288_cast_fp16))[name = tensor("op_317_cast_fp16")]; + tensor var_319_equation_0 = const()[name = tensor("op_319_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_319_cast_fp16 = einsum(equation = var_319_equation_0, values = (var_219_cast_fp16_9, var_289_cast_fp16))[name = tensor("op_319_cast_fp16")]; + tensor var_321_equation_0 = const()[name = tensor("op_321_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_321_cast_fp16 = einsum(equation = var_321_equation_0, values = (var_219_cast_fp16_10, var_290_cast_fp16))[name = tensor("op_321_cast_fp16")]; + tensor var_323_equation_0 = const()[name = tensor("op_323_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_323_cast_fp16 = einsum(equation = var_323_equation_0, values = (var_219_cast_fp16_11, var_291_cast_fp16))[name = tensor("op_323_cast_fp16")]; + tensor var_325_equation_0 = const()[name = tensor("op_325_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_325_cast_fp16 = einsum(equation = var_325_equation_0, values = (var_219_cast_fp16_12, var_292_cast_fp16))[name = tensor("op_325_cast_fp16")]; + tensor var_327_equation_0 = const()[name = tensor("op_327_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_327_cast_fp16 = einsum(equation = var_327_equation_0, values = (var_219_cast_fp16_13, var_293_cast_fp16))[name = tensor("op_327_cast_fp16")]; + tensor var_329_equation_0 = const()[name = tensor("op_329_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_329_cast_fp16 = einsum(equation = var_329_equation_0, values = (var_219_cast_fp16_14, var_294_cast_fp16))[name = tensor("op_329_cast_fp16")]; + tensor var_331_equation_0 = const()[name = tensor("op_331_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_331_cast_fp16 = einsum(equation = var_331_equation_0, values = (var_219_cast_fp16_15, var_295_cast_fp16))[name = tensor("op_331_cast_fp16")]; + tensor var_333_equation_0 = const()[name = tensor("op_333_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_333_cast_fp16 = einsum(equation = var_333_equation_0, values = (var_219_cast_fp16_16, var_296_cast_fp16))[name = tensor("op_333_cast_fp16")]; + tensor var_335_equation_0 = const()[name = tensor("op_335_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_335_cast_fp16 = einsum(equation = var_335_equation_0, values = (var_219_cast_fp16_17, var_297_cast_fp16))[name = tensor("op_335_cast_fp16")]; + tensor var_337_equation_0 = const()[name = tensor("op_337_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_337_cast_fp16 = einsum(equation = var_337_equation_0, values = (var_219_cast_fp16_18, var_298_cast_fp16))[name = tensor("op_337_cast_fp16")]; + tensor var_339_equation_0 = const()[name = tensor("op_339_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_339_cast_fp16 = einsum(equation = var_339_equation_0, values = (var_219_cast_fp16_19, var_299_cast_fp16))[name = tensor("op_339_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_124, interleave = input_5_interleave_0, values = (var_301_cast_fp16, var_303_cast_fp16, var_305_cast_fp16, var_307_cast_fp16, var_309_cast_fp16, var_311_cast_fp16, var_313_cast_fp16, var_315_cast_fp16, var_317_cast_fp16, var_319_cast_fp16, var_321_cast_fp16, var_323_cast_fp16, var_325_cast_fp16, var_327_cast_fp16, var_329_cast_fp16, var_331_cast_fp16, var_333_cast_fp16, var_335_cast_fp16, var_337_cast_fp16, var_339_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_348_pad_type_0 = const()[name = tensor("op_348_pad_type_0"), val = tensor("valid")]; + tensor var_348_strides_0 = const()[name = tensor("op_348_strides_0"), val = tensor([1, 1])]; + tensor var_348_pad_0 = const()[name = tensor("op_348_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_348_dilations_0 = const()[name = tensor("op_348_dilations_0"), val = tensor([1, 1])]; + tensor var_348_groups_0 = const()[name = tensor("op_348_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24131392)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27408256)))]; + tensor var_348_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_348_dilations_0, groups = var_348_groups_0, pad = var_348_pad_0, pad_type = var_348_pad_type_0, strides = var_348_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_348_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27410880)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27413504)))]; + tensor var_358_to_fp16 = const()[name = tensor("op_358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_358_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27416128)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40523392)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_384_pad_type_0 = const()[name = tensor("op_384_pad_type_0"), val = tensor("valid")]; + tensor var_384_strides_0 = const()[name = tensor("op_384_strides_0"), val = tensor([1, 1])]; + tensor var_384_pad_0 = const()[name = tensor("op_384_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_384_dilations_0 = const()[name = tensor("op_384_dilations_0"), val = tensor([1, 1])]; + tensor var_384_groups_0 = const()[name = tensor("op_384_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40533696)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53640960)))]; + tensor var_384_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_384_dilations_0, groups = var_384_groups_0, pad = var_384_pad_0, pad_type = var_384_pad_type_0, strides = var_384_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_384_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53643584)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53646208)))]; + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_409_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_444_weight_0_to_fp16 = const()[name = tensor("op_444_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53648832)))]; + tensor var_444_bias_0_to_fp16 = const()[name = tensor("op_444_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56925696)))]; + tensor var_444_cast_fp16 = conv(bias = var_444_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_444_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_444_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56928320)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_442_pad_type_0 = const()[name = tensor("op_442_pad_type_0"), val = tensor("valid")]; + tensor var_442_strides_0 = const()[name = tensor("op_442_strides_0"), val = tensor([1, 1])]; + tensor var_442_pad_0 = const()[name = tensor("op_442_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_442_dilations_0 = const()[name = tensor("op_442_dilations_0"), val = tensor([1, 1])]; + tensor var_442_groups_0 = const()[name = tensor("op_442_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60205184)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63482048)))]; + tensor var_442_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_442_dilations_0, groups = var_442_groups_0, pad = var_442_pad_0, pad_type = var_442_pad_type_0, strides = var_442_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_cast_fp16_0, tensor var_445_cast_fp16_1, tensor var_445_cast_fp16_2, tensor var_445_cast_fp16_3, tensor var_445_cast_fp16_4, tensor var_445_cast_fp16_5, tensor var_445_cast_fp16_6, tensor var_445_cast_fp16_7, tensor var_445_cast_fp16_8, tensor var_445_cast_fp16_9, tensor var_445_cast_fp16_10, tensor var_445_cast_fp16_11, tensor var_445_cast_fp16_12, tensor var_445_cast_fp16_13, tensor var_445_cast_fp16_14, tensor var_445_cast_fp16_15, tensor var_445_cast_fp16_16, tensor var_445_cast_fp16_17, tensor var_445_cast_fp16_18, tensor var_445_cast_fp16_19 = split(axis = var_445_axis_0, split_sizes = tile_3, x = var_444_cast_fp16)[name = tensor("op_445_cast_fp16")]; + tensor var_466_perm_0 = const()[name = tensor("op_466_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_467_axis_0 = const()[name = tensor("op_467_axis_0"), val = tensor(3)]; + tensor var_466_cast_fp16 = transpose(perm = var_466_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_31")]; + tensor var_467_cast_fp16_0, tensor var_467_cast_fp16_1, tensor var_467_cast_fp16_2, tensor var_467_cast_fp16_3, tensor var_467_cast_fp16_4, tensor var_467_cast_fp16_5, tensor var_467_cast_fp16_6, tensor var_467_cast_fp16_7, tensor var_467_cast_fp16_8, tensor var_467_cast_fp16_9, tensor var_467_cast_fp16_10, tensor var_467_cast_fp16_11, tensor var_467_cast_fp16_12, tensor var_467_cast_fp16_13, tensor var_467_cast_fp16_14, tensor var_467_cast_fp16_15, tensor var_467_cast_fp16_16, tensor var_467_cast_fp16_17, tensor var_467_cast_fp16_18, tensor var_467_cast_fp16_19 = split(axis = var_467_axis_0, split_sizes = tile_4, x = var_466_cast_fp16)[name = tensor("op_467_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_488_axis_0 = const()[name = tensor("op_488_axis_0"), val = tensor(1)]; + tensor var_488_cast_fp16_0, tensor var_488_cast_fp16_1, tensor var_488_cast_fp16_2, tensor var_488_cast_fp16_3, tensor var_488_cast_fp16_4, tensor var_488_cast_fp16_5, tensor var_488_cast_fp16_6, tensor var_488_cast_fp16_7, tensor var_488_cast_fp16_8, tensor var_488_cast_fp16_9, tensor var_488_cast_fp16_10, tensor var_488_cast_fp16_11, tensor var_488_cast_fp16_12, tensor var_488_cast_fp16_13, tensor var_488_cast_fp16_14, tensor var_488_cast_fp16_15, tensor var_488_cast_fp16_16, tensor var_488_cast_fp16_17, tensor var_488_cast_fp16_18, tensor var_488_cast_fp16_19 = split(axis = var_488_axis_0, split_sizes = tile_5, x = var_442_cast_fp16)[name = tensor("op_488_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_467_cast_fp16_0, var_445_cast_fp16_0))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_467_cast_fp16_1, var_445_cast_fp16_1))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_467_cast_fp16_2, var_445_cast_fp16_2))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_467_cast_fp16_3, var_445_cast_fp16_3))[name = tensor("aw_47_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_467_cast_fp16_4, var_445_cast_fp16_4))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_467_cast_fp16_5, var_445_cast_fp16_5))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_467_cast_fp16_6, var_445_cast_fp16_6))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_467_cast_fp16_7, var_445_cast_fp16_7))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_467_cast_fp16_8, var_445_cast_fp16_8))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_467_cast_fp16_9, var_445_cast_fp16_9))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_467_cast_fp16_10, var_445_cast_fp16_10))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_467_cast_fp16_11, var_445_cast_fp16_11))[name = tensor("aw_63_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_467_cast_fp16_12, var_445_cast_fp16_12))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_467_cast_fp16_13, var_445_cast_fp16_13))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_467_cast_fp16_14, var_445_cast_fp16_14))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_467_cast_fp16_15, var_445_cast_fp16_15))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_467_cast_fp16_16, var_445_cast_fp16_16))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_467_cast_fp16_17, var_445_cast_fp16_17))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_467_cast_fp16_18, var_445_cast_fp16_18))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_467_cast_fp16_19, var_445_cast_fp16_19))[name = tensor("aw_79_cast_fp16")]; + tensor var_549_cast_fp16 = softmax(axis = var_393, x = aw_41_cast_fp16)[name = tensor("op_549_cast_fp16")]; + tensor var_550_cast_fp16 = softmax(axis = var_393, x = aw_43_cast_fp16)[name = tensor("op_550_cast_fp16")]; + tensor var_551_cast_fp16 = softmax(axis = var_393, x = aw_45_cast_fp16)[name = tensor("op_551_cast_fp16")]; + tensor var_552_cast_fp16 = softmax(axis = var_393, x = aw_47_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = softmax(axis = var_393, x = aw_49_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor var_554_cast_fp16 = softmax(axis = var_393, x = aw_51_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_555_cast_fp16 = softmax(axis = var_393, x = aw_53_cast_fp16)[name = tensor("op_555_cast_fp16")]; + tensor var_556_cast_fp16 = softmax(axis = var_393, x = aw_55_cast_fp16)[name = tensor("op_556_cast_fp16")]; + tensor var_557_cast_fp16 = softmax(axis = var_393, x = aw_57_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_558_cast_fp16 = softmax(axis = var_393, x = aw_59_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor var_559_cast_fp16 = softmax(axis = var_393, x = aw_61_cast_fp16)[name = tensor("op_559_cast_fp16")]; + tensor var_560_cast_fp16 = softmax(axis = var_393, x = aw_63_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_561_cast_fp16 = softmax(axis = var_393, x = aw_65_cast_fp16)[name = tensor("op_561_cast_fp16")]; + tensor var_562_cast_fp16 = softmax(axis = var_393, x = aw_67_cast_fp16)[name = tensor("op_562_cast_fp16")]; + tensor var_563_cast_fp16 = softmax(axis = var_393, x = aw_69_cast_fp16)[name = tensor("op_563_cast_fp16")]; + tensor var_564_cast_fp16 = softmax(axis = var_393, x = aw_71_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor var_565_cast_fp16 = softmax(axis = var_393, x = aw_73_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_566_cast_fp16 = softmax(axis = var_393, x = aw_75_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = softmax(axis = var_393, x = aw_77_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor var_568_cast_fp16 = softmax(axis = var_393, x = aw_79_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor var_570_equation_0 = const()[name = tensor("op_570_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_570_cast_fp16 = einsum(equation = var_570_equation_0, values = (var_488_cast_fp16_0, var_549_cast_fp16))[name = tensor("op_570_cast_fp16")]; + tensor var_572_equation_0 = const()[name = tensor("op_572_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_572_cast_fp16 = einsum(equation = var_572_equation_0, values = (var_488_cast_fp16_1, var_550_cast_fp16))[name = tensor("op_572_cast_fp16")]; + tensor var_574_equation_0 = const()[name = tensor("op_574_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_574_cast_fp16 = einsum(equation = var_574_equation_0, values = (var_488_cast_fp16_2, var_551_cast_fp16))[name = tensor("op_574_cast_fp16")]; + tensor var_576_equation_0 = const()[name = tensor("op_576_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_576_cast_fp16 = einsum(equation = var_576_equation_0, values = (var_488_cast_fp16_3, var_552_cast_fp16))[name = tensor("op_576_cast_fp16")]; + tensor var_578_equation_0 = const()[name = tensor("op_578_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_578_cast_fp16 = einsum(equation = var_578_equation_0, values = (var_488_cast_fp16_4, var_553_cast_fp16))[name = tensor("op_578_cast_fp16")]; + tensor var_580_equation_0 = const()[name = tensor("op_580_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_580_cast_fp16 = einsum(equation = var_580_equation_0, values = (var_488_cast_fp16_5, var_554_cast_fp16))[name = tensor("op_580_cast_fp16")]; + tensor var_582_equation_0 = const()[name = tensor("op_582_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_582_cast_fp16 = einsum(equation = var_582_equation_0, values = (var_488_cast_fp16_6, var_555_cast_fp16))[name = tensor("op_582_cast_fp16")]; + tensor var_584_equation_0 = const()[name = tensor("op_584_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_584_cast_fp16 = einsum(equation = var_584_equation_0, values = (var_488_cast_fp16_7, var_556_cast_fp16))[name = tensor("op_584_cast_fp16")]; + tensor var_586_equation_0 = const()[name = tensor("op_586_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_586_cast_fp16 = einsum(equation = var_586_equation_0, values = (var_488_cast_fp16_8, var_557_cast_fp16))[name = tensor("op_586_cast_fp16")]; + tensor var_588_equation_0 = const()[name = tensor("op_588_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_588_cast_fp16 = einsum(equation = var_588_equation_0, values = (var_488_cast_fp16_9, var_558_cast_fp16))[name = tensor("op_588_cast_fp16")]; + tensor var_590_equation_0 = const()[name = tensor("op_590_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_590_cast_fp16 = einsum(equation = var_590_equation_0, values = (var_488_cast_fp16_10, var_559_cast_fp16))[name = tensor("op_590_cast_fp16")]; + tensor var_592_equation_0 = const()[name = tensor("op_592_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_592_cast_fp16 = einsum(equation = var_592_equation_0, values = (var_488_cast_fp16_11, var_560_cast_fp16))[name = tensor("op_592_cast_fp16")]; + tensor var_594_equation_0 = const()[name = tensor("op_594_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_594_cast_fp16 = einsum(equation = var_594_equation_0, values = (var_488_cast_fp16_12, var_561_cast_fp16))[name = tensor("op_594_cast_fp16")]; + tensor var_596_equation_0 = const()[name = tensor("op_596_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_596_cast_fp16 = einsum(equation = var_596_equation_0, values = (var_488_cast_fp16_13, var_562_cast_fp16))[name = tensor("op_596_cast_fp16")]; + tensor var_598_equation_0 = const()[name = tensor("op_598_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_598_cast_fp16 = einsum(equation = var_598_equation_0, values = (var_488_cast_fp16_14, var_563_cast_fp16))[name = tensor("op_598_cast_fp16")]; + tensor var_600_equation_0 = const()[name = tensor("op_600_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_600_cast_fp16 = einsum(equation = var_600_equation_0, values = (var_488_cast_fp16_15, var_564_cast_fp16))[name = tensor("op_600_cast_fp16")]; + tensor var_602_equation_0 = const()[name = tensor("op_602_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_602_cast_fp16 = einsum(equation = var_602_equation_0, values = (var_488_cast_fp16_16, var_565_cast_fp16))[name = tensor("op_602_cast_fp16")]; + tensor var_604_equation_0 = const()[name = tensor("op_604_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_604_cast_fp16 = einsum(equation = var_604_equation_0, values = (var_488_cast_fp16_17, var_566_cast_fp16))[name = tensor("op_604_cast_fp16")]; + tensor var_606_equation_0 = const()[name = tensor("op_606_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_606_cast_fp16 = einsum(equation = var_606_equation_0, values = (var_488_cast_fp16_18, var_567_cast_fp16))[name = tensor("op_606_cast_fp16")]; + tensor var_608_equation_0 = const()[name = tensor("op_608_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_608_cast_fp16 = einsum(equation = var_608_equation_0, values = (var_488_cast_fp16_19, var_568_cast_fp16))[name = tensor("op_608_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_393, interleave = input_15_interleave_0, values = (var_570_cast_fp16, var_572_cast_fp16, var_574_cast_fp16, var_576_cast_fp16, var_578_cast_fp16, var_580_cast_fp16, var_582_cast_fp16, var_584_cast_fp16, var_586_cast_fp16, var_588_cast_fp16, var_590_cast_fp16, var_592_cast_fp16, var_594_cast_fp16, var_596_cast_fp16, var_598_cast_fp16, var_600_cast_fp16, var_602_cast_fp16, var_604_cast_fp16, var_606_cast_fp16, var_608_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_617_pad_type_0 = const()[name = tensor("op_617_pad_type_0"), val = tensor("valid")]; + tensor var_617_strides_0 = const()[name = tensor("op_617_strides_0"), val = tensor([1, 1])]; + tensor var_617_pad_0 = const()[name = tensor("op_617_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_617_dilations_0 = const()[name = tensor("op_617_dilations_0"), val = tensor([1, 1])]; + tensor var_617_groups_0 = const()[name = tensor("op_617_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63484672)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66761536)))]; + tensor var_617_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_617_dilations_0, groups = var_617_groups_0, pad = var_617_pad_0, pad_type = var_617_pad_type_0, strides = var_617_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_617_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66764160)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66766784)))]; + tensor var_627_to_fp16 = const()[name = tensor("op_627_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_627_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66769408)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79876672)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_653_pad_type_0 = const()[name = tensor("op_653_pad_type_0"), val = tensor("valid")]; + tensor var_653_strides_0 = const()[name = tensor("op_653_strides_0"), val = tensor([1, 1])]; + tensor var_653_pad_0 = const()[name = tensor("op_653_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_653_dilations_0 = const()[name = tensor("op_653_dilations_0"), val = tensor([1, 1])]; + tensor var_653_groups_0 = const()[name = tensor("op_653_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79886976)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92994240)))]; + tensor var_653_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_653_dilations_0, groups = var_653_groups_0, pad = var_653_pad_0, pad_type = var_653_pad_type_0, strides = var_653_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_653_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_662 = const()[name = tensor("op_662"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92996864)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92999488)))]; + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_678_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_713_weight_0_to_fp16 = const()[name = tensor("op_713_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93002112)))]; + tensor var_713_bias_0_to_fp16 = const()[name = tensor("op_713_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96278976)))]; + tensor var_713_cast_fp16 = conv(bias = var_713_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_713_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96281600)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_711_pad_type_0 = const()[name = tensor("op_711_pad_type_0"), val = tensor("valid")]; + tensor var_711_strides_0 = const()[name = tensor("op_711_strides_0"), val = tensor([1, 1])]; + tensor var_711_pad_0 = const()[name = tensor("op_711_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_711_dilations_0 = const()[name = tensor("op_711_dilations_0"), val = tensor([1, 1])]; + tensor var_711_groups_0 = const()[name = tensor("op_711_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99558464)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102835328)))]; + tensor var_711_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_711_dilations_0, groups = var_711_groups_0, pad = var_711_pad_0, pad_type = var_711_pad_type_0, strides = var_711_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_711_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_714_axis_0 = const()[name = tensor("op_714_axis_0"), val = tensor(1)]; + tensor var_714_cast_fp16_0, tensor var_714_cast_fp16_1, tensor var_714_cast_fp16_2, tensor var_714_cast_fp16_3, tensor var_714_cast_fp16_4, tensor var_714_cast_fp16_5, tensor var_714_cast_fp16_6, tensor var_714_cast_fp16_7, tensor var_714_cast_fp16_8, tensor var_714_cast_fp16_9, tensor var_714_cast_fp16_10, tensor var_714_cast_fp16_11, tensor var_714_cast_fp16_12, tensor var_714_cast_fp16_13, tensor var_714_cast_fp16_14, tensor var_714_cast_fp16_15, tensor var_714_cast_fp16_16, tensor var_714_cast_fp16_17, tensor var_714_cast_fp16_18, tensor var_714_cast_fp16_19 = split(axis = var_714_axis_0, split_sizes = tile_6, x = var_713_cast_fp16)[name = tensor("op_714_cast_fp16")]; + tensor var_735_perm_0 = const()[name = tensor("op_735_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_736_axis_0 = const()[name = tensor("op_736_axis_0"), val = tensor(3)]; + tensor var_735_cast_fp16 = transpose(perm = var_735_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_30")]; + tensor var_736_cast_fp16_0, tensor var_736_cast_fp16_1, tensor var_736_cast_fp16_2, tensor var_736_cast_fp16_3, tensor var_736_cast_fp16_4, tensor var_736_cast_fp16_5, tensor var_736_cast_fp16_6, tensor var_736_cast_fp16_7, tensor var_736_cast_fp16_8, tensor var_736_cast_fp16_9, tensor var_736_cast_fp16_10, tensor var_736_cast_fp16_11, tensor var_736_cast_fp16_12, tensor var_736_cast_fp16_13, tensor var_736_cast_fp16_14, tensor var_736_cast_fp16_15, tensor var_736_cast_fp16_16, tensor var_736_cast_fp16_17, tensor var_736_cast_fp16_18, tensor var_736_cast_fp16_19 = split(axis = var_736_axis_0, split_sizes = tile_7, x = var_735_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_cast_fp16_0, tensor var_757_cast_fp16_1, tensor var_757_cast_fp16_2, tensor var_757_cast_fp16_3, tensor var_757_cast_fp16_4, tensor var_757_cast_fp16_5, tensor var_757_cast_fp16_6, tensor var_757_cast_fp16_7, tensor var_757_cast_fp16_8, tensor var_757_cast_fp16_9, tensor var_757_cast_fp16_10, tensor var_757_cast_fp16_11, tensor var_757_cast_fp16_12, tensor var_757_cast_fp16_13, tensor var_757_cast_fp16_14, tensor var_757_cast_fp16_15, tensor var_757_cast_fp16_16, tensor var_757_cast_fp16_17, tensor var_757_cast_fp16_18, tensor var_757_cast_fp16_19 = split(axis = var_757_axis_0, split_sizes = tile_8, x = var_711_cast_fp16)[name = tensor("op_757_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_736_cast_fp16_0, var_714_cast_fp16_0))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_736_cast_fp16_1, var_714_cast_fp16_1))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_736_cast_fp16_2, var_714_cast_fp16_2))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_736_cast_fp16_3, var_714_cast_fp16_3))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_736_cast_fp16_4, var_714_cast_fp16_4))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_736_cast_fp16_5, var_714_cast_fp16_5))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_736_cast_fp16_6, var_714_cast_fp16_6))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_736_cast_fp16_7, var_714_cast_fp16_7))[name = tensor("aw_95_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_736_cast_fp16_8, var_714_cast_fp16_8))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_736_cast_fp16_9, var_714_cast_fp16_9))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_736_cast_fp16_10, var_714_cast_fp16_10))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_736_cast_fp16_11, var_714_cast_fp16_11))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_736_cast_fp16_12, var_714_cast_fp16_12))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_736_cast_fp16_13, var_714_cast_fp16_13))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_736_cast_fp16_14, var_714_cast_fp16_14))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_736_cast_fp16_15, var_714_cast_fp16_15))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_736_cast_fp16_16, var_714_cast_fp16_16))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_736_cast_fp16_17, var_714_cast_fp16_17))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_736_cast_fp16_18, var_714_cast_fp16_18))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_736_cast_fp16_19, var_714_cast_fp16_19))[name = tensor("aw_119_cast_fp16")]; + tensor var_818_cast_fp16 = softmax(axis = var_662, x = aw_81_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_819_cast_fp16 = softmax(axis = var_662, x = aw_83_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820_cast_fp16 = softmax(axis = var_662, x = aw_85_cast_fp16)[name = tensor("op_820_cast_fp16")]; + tensor var_821_cast_fp16 = softmax(axis = var_662, x = aw_87_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = softmax(axis = var_662, x = aw_89_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor var_823_cast_fp16 = softmax(axis = var_662, x = aw_91_cast_fp16)[name = tensor("op_823_cast_fp16")]; + tensor var_824_cast_fp16 = softmax(axis = var_662, x = aw_93_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor var_825_cast_fp16 = softmax(axis = var_662, x = aw_95_cast_fp16)[name = tensor("op_825_cast_fp16")]; + tensor var_826_cast_fp16 = softmax(axis = var_662, x = aw_97_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_827_cast_fp16 = softmax(axis = var_662, x = aw_99_cast_fp16)[name = tensor("op_827_cast_fp16")]; + tensor var_828_cast_fp16 = softmax(axis = var_662, x = aw_101_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_cast_fp16 = softmax(axis = var_662, x = aw_103_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_830_cast_fp16 = softmax(axis = var_662, x = aw_105_cast_fp16)[name = tensor("op_830_cast_fp16")]; + tensor var_831_cast_fp16 = softmax(axis = var_662, x = aw_107_cast_fp16)[name = tensor("op_831_cast_fp16")]; + tensor var_832_cast_fp16 = softmax(axis = var_662, x = aw_109_cast_fp16)[name = tensor("op_832_cast_fp16")]; + tensor var_833_cast_fp16 = softmax(axis = var_662, x = aw_111_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_834_cast_fp16 = softmax(axis = var_662, x = aw_113_cast_fp16)[name = tensor("op_834_cast_fp16")]; + tensor var_835_cast_fp16 = softmax(axis = var_662, x = aw_115_cast_fp16)[name = tensor("op_835_cast_fp16")]; + tensor var_836_cast_fp16 = softmax(axis = var_662, x = aw_117_cast_fp16)[name = tensor("op_836_cast_fp16")]; + tensor var_837_cast_fp16 = softmax(axis = var_662, x = aw_119_cast_fp16)[name = tensor("op_837_cast_fp16")]; + tensor var_839_equation_0 = const()[name = tensor("op_839_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_839_cast_fp16 = einsum(equation = var_839_equation_0, values = (var_757_cast_fp16_0, var_818_cast_fp16))[name = tensor("op_839_cast_fp16")]; + tensor var_841_equation_0 = const()[name = tensor("op_841_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_841_cast_fp16 = einsum(equation = var_841_equation_0, values = (var_757_cast_fp16_1, var_819_cast_fp16))[name = tensor("op_841_cast_fp16")]; + tensor var_843_equation_0 = const()[name = tensor("op_843_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_843_cast_fp16 = einsum(equation = var_843_equation_0, values = (var_757_cast_fp16_2, var_820_cast_fp16))[name = tensor("op_843_cast_fp16")]; + tensor var_845_equation_0 = const()[name = tensor("op_845_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_845_cast_fp16 = einsum(equation = var_845_equation_0, values = (var_757_cast_fp16_3, var_821_cast_fp16))[name = tensor("op_845_cast_fp16")]; + tensor var_847_equation_0 = const()[name = tensor("op_847_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_847_cast_fp16 = einsum(equation = var_847_equation_0, values = (var_757_cast_fp16_4, var_822_cast_fp16))[name = tensor("op_847_cast_fp16")]; + tensor var_849_equation_0 = const()[name = tensor("op_849_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_849_cast_fp16 = einsum(equation = var_849_equation_0, values = (var_757_cast_fp16_5, var_823_cast_fp16))[name = tensor("op_849_cast_fp16")]; + tensor var_851_equation_0 = const()[name = tensor("op_851_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_851_cast_fp16 = einsum(equation = var_851_equation_0, values = (var_757_cast_fp16_6, var_824_cast_fp16))[name = tensor("op_851_cast_fp16")]; + tensor var_853_equation_0 = const()[name = tensor("op_853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_853_cast_fp16 = einsum(equation = var_853_equation_0, values = (var_757_cast_fp16_7, var_825_cast_fp16))[name = tensor("op_853_cast_fp16")]; + tensor var_855_equation_0 = const()[name = tensor("op_855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_855_cast_fp16 = einsum(equation = var_855_equation_0, values = (var_757_cast_fp16_8, var_826_cast_fp16))[name = tensor("op_855_cast_fp16")]; + tensor var_857_equation_0 = const()[name = tensor("op_857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_857_cast_fp16 = einsum(equation = var_857_equation_0, values = (var_757_cast_fp16_9, var_827_cast_fp16))[name = tensor("op_857_cast_fp16")]; + tensor var_859_equation_0 = const()[name = tensor("op_859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_859_cast_fp16 = einsum(equation = var_859_equation_0, values = (var_757_cast_fp16_10, var_828_cast_fp16))[name = tensor("op_859_cast_fp16")]; + tensor var_861_equation_0 = const()[name = tensor("op_861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_861_cast_fp16 = einsum(equation = var_861_equation_0, values = (var_757_cast_fp16_11, var_829_cast_fp16))[name = tensor("op_861_cast_fp16")]; + tensor var_863_equation_0 = const()[name = tensor("op_863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_863_cast_fp16 = einsum(equation = var_863_equation_0, values = (var_757_cast_fp16_12, var_830_cast_fp16))[name = tensor("op_863_cast_fp16")]; + tensor var_865_equation_0 = const()[name = tensor("op_865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_865_cast_fp16 = einsum(equation = var_865_equation_0, values = (var_757_cast_fp16_13, var_831_cast_fp16))[name = tensor("op_865_cast_fp16")]; + tensor var_867_equation_0 = const()[name = tensor("op_867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_867_cast_fp16 = einsum(equation = var_867_equation_0, values = (var_757_cast_fp16_14, var_832_cast_fp16))[name = tensor("op_867_cast_fp16")]; + tensor var_869_equation_0 = const()[name = tensor("op_869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_869_cast_fp16 = einsum(equation = var_869_equation_0, values = (var_757_cast_fp16_15, var_833_cast_fp16))[name = tensor("op_869_cast_fp16")]; + tensor var_871_equation_0 = const()[name = tensor("op_871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_871_cast_fp16 = einsum(equation = var_871_equation_0, values = (var_757_cast_fp16_16, var_834_cast_fp16))[name = tensor("op_871_cast_fp16")]; + tensor var_873_equation_0 = const()[name = tensor("op_873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_873_cast_fp16 = einsum(equation = var_873_equation_0, values = (var_757_cast_fp16_17, var_835_cast_fp16))[name = tensor("op_873_cast_fp16")]; + tensor var_875_equation_0 = const()[name = tensor("op_875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_875_cast_fp16 = einsum(equation = var_875_equation_0, values = (var_757_cast_fp16_18, var_836_cast_fp16))[name = tensor("op_875_cast_fp16")]; + tensor var_877_equation_0 = const()[name = tensor("op_877_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_877_cast_fp16 = einsum(equation = var_877_equation_0, values = (var_757_cast_fp16_19, var_837_cast_fp16))[name = tensor("op_877_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_662, interleave = input_25_interleave_0, values = (var_839_cast_fp16, var_841_cast_fp16, var_843_cast_fp16, var_845_cast_fp16, var_847_cast_fp16, var_849_cast_fp16, var_851_cast_fp16, var_853_cast_fp16, var_855_cast_fp16, var_857_cast_fp16, var_859_cast_fp16, var_861_cast_fp16, var_863_cast_fp16, var_865_cast_fp16, var_867_cast_fp16, var_869_cast_fp16, var_871_cast_fp16, var_873_cast_fp16, var_875_cast_fp16, var_877_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_886_pad_type_0 = const()[name = tensor("op_886_pad_type_0"), val = tensor("valid")]; + tensor var_886_strides_0 = const()[name = tensor("op_886_strides_0"), val = tensor([1, 1])]; + tensor var_886_pad_0 = const()[name = tensor("op_886_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_886_dilations_0 = const()[name = tensor("op_886_dilations_0"), val = tensor([1, 1])]; + tensor var_886_groups_0 = const()[name = tensor("op_886_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102837952)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106114816)))]; + tensor var_886_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_886_dilations_0, groups = var_886_groups_0, pad = var_886_pad_0, pad_type = var_886_pad_type_0, strides = var_886_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_886_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106117440)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106120064)))]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_896_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106122688)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119229952)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_922_pad_type_0 = const()[name = tensor("op_922_pad_type_0"), val = tensor("valid")]; + tensor var_922_strides_0 = const()[name = tensor("op_922_strides_0"), val = tensor([1, 1])]; + tensor var_922_pad_0 = const()[name = tensor("op_922_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_922_dilations_0 = const()[name = tensor("op_922_dilations_0"), val = tensor([1, 1])]; + tensor var_922_groups_0 = const()[name = tensor("op_922_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119240256)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132347520)))]; + tensor var_922_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_922_dilations_0, groups = var_922_groups_0, pad = var_922_pad_0, pad_type = var_922_pad_type_0, strides = var_922_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_922_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132350144)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132352768)))]; + tensor var_947_to_fp16 = const()[name = tensor("op_947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_947_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_982_weight_0_to_fp16 = const()[name = tensor("op_982_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132355392)))]; + tensor var_982_bias_0_to_fp16 = const()[name = tensor("op_982_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135632256)))]; + tensor var_982_cast_fp16 = conv(bias = var_982_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_982_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_982_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135634880)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_980_pad_type_0 = const()[name = tensor("op_980_pad_type_0"), val = tensor("valid")]; + tensor var_980_strides_0 = const()[name = tensor("op_980_strides_0"), val = tensor([1, 1])]; + tensor var_980_pad_0 = const()[name = tensor("op_980_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_980_dilations_0 = const()[name = tensor("op_980_dilations_0"), val = tensor([1, 1])]; + tensor var_980_groups_0 = const()[name = tensor("op_980_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138911744)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142188608)))]; + tensor var_980_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_980_dilations_0, groups = var_980_groups_0, pad = var_980_pad_0, pad_type = var_980_pad_type_0, strides = var_980_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_980_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_983_axis_0 = const()[name = tensor("op_983_axis_0"), val = tensor(1)]; + tensor var_983_cast_fp16_0, tensor var_983_cast_fp16_1, tensor var_983_cast_fp16_2, tensor var_983_cast_fp16_3, tensor var_983_cast_fp16_4, tensor var_983_cast_fp16_5, tensor var_983_cast_fp16_6, tensor var_983_cast_fp16_7, tensor var_983_cast_fp16_8, tensor var_983_cast_fp16_9, tensor var_983_cast_fp16_10, tensor var_983_cast_fp16_11, tensor var_983_cast_fp16_12, tensor var_983_cast_fp16_13, tensor var_983_cast_fp16_14, tensor var_983_cast_fp16_15, tensor var_983_cast_fp16_16, tensor var_983_cast_fp16_17, tensor var_983_cast_fp16_18, tensor var_983_cast_fp16_19 = split(axis = var_983_axis_0, split_sizes = tile_9, x = var_982_cast_fp16)[name = tensor("op_983_cast_fp16")]; + tensor var_1004_perm_0 = const()[name = tensor("op_1004_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1005_axis_0 = const()[name = tensor("op_1005_axis_0"), val = tensor(3)]; + tensor var_1004_cast_fp16 = transpose(perm = var_1004_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_29")]; + tensor var_1005_cast_fp16_0, tensor var_1005_cast_fp16_1, tensor var_1005_cast_fp16_2, tensor var_1005_cast_fp16_3, tensor var_1005_cast_fp16_4, tensor var_1005_cast_fp16_5, tensor var_1005_cast_fp16_6, tensor var_1005_cast_fp16_7, tensor var_1005_cast_fp16_8, tensor var_1005_cast_fp16_9, tensor var_1005_cast_fp16_10, tensor var_1005_cast_fp16_11, tensor var_1005_cast_fp16_12, tensor var_1005_cast_fp16_13, tensor var_1005_cast_fp16_14, tensor var_1005_cast_fp16_15, tensor var_1005_cast_fp16_16, tensor var_1005_cast_fp16_17, tensor var_1005_cast_fp16_18, tensor var_1005_cast_fp16_19 = split(axis = var_1005_axis_0, split_sizes = tile_10, x = var_1004_cast_fp16)[name = tensor("op_1005_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1026_axis_0 = const()[name = tensor("op_1026_axis_0"), val = tensor(1)]; + tensor var_1026_cast_fp16_0, tensor var_1026_cast_fp16_1, tensor var_1026_cast_fp16_2, tensor var_1026_cast_fp16_3, tensor var_1026_cast_fp16_4, tensor var_1026_cast_fp16_5, tensor var_1026_cast_fp16_6, tensor var_1026_cast_fp16_7, tensor var_1026_cast_fp16_8, tensor var_1026_cast_fp16_9, tensor var_1026_cast_fp16_10, tensor var_1026_cast_fp16_11, tensor var_1026_cast_fp16_12, tensor var_1026_cast_fp16_13, tensor var_1026_cast_fp16_14, tensor var_1026_cast_fp16_15, tensor var_1026_cast_fp16_16, tensor var_1026_cast_fp16_17, tensor var_1026_cast_fp16_18, tensor var_1026_cast_fp16_19 = split(axis = var_1026_axis_0, split_sizes = tile_11, x = var_980_cast_fp16)[name = tensor("op_1026_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_1005_cast_fp16_0, var_983_cast_fp16_0))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_1005_cast_fp16_1, var_983_cast_fp16_1))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_1005_cast_fp16_2, var_983_cast_fp16_2))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_1005_cast_fp16_3, var_983_cast_fp16_3))[name = tensor("aw_127_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1005_cast_fp16_4, var_983_cast_fp16_4))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1005_cast_fp16_5, var_983_cast_fp16_5))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1005_cast_fp16_6, var_983_cast_fp16_6))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1005_cast_fp16_7, var_983_cast_fp16_7))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1005_cast_fp16_8, var_983_cast_fp16_8))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1005_cast_fp16_9, var_983_cast_fp16_9))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1005_cast_fp16_10, var_983_cast_fp16_10))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1005_cast_fp16_11, var_983_cast_fp16_11))[name = tensor("aw_143_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1005_cast_fp16_12, var_983_cast_fp16_12))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1005_cast_fp16_13, var_983_cast_fp16_13))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1005_cast_fp16_14, var_983_cast_fp16_14))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1005_cast_fp16_15, var_983_cast_fp16_15))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1005_cast_fp16_16, var_983_cast_fp16_16))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1005_cast_fp16_17, var_983_cast_fp16_17))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1005_cast_fp16_18, var_983_cast_fp16_18))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1005_cast_fp16_19, var_983_cast_fp16_19))[name = tensor("aw_159_cast_fp16")]; + tensor var_1087_cast_fp16 = softmax(axis = var_931, x = aw_121_cast_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor var_1088_cast_fp16 = softmax(axis = var_931, x = aw_123_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = softmax(axis = var_931, x = aw_125_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor var_1090_cast_fp16 = softmax(axis = var_931, x = aw_127_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1091_cast_fp16 = softmax(axis = var_931, x = aw_129_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor var_1092_cast_fp16 = softmax(axis = var_931, x = aw_131_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = softmax(axis = var_931, x = aw_133_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor var_1094_cast_fp16 = softmax(axis = var_931, x = aw_135_cast_fp16)[name = tensor("op_1094_cast_fp16")]; + tensor var_1095_cast_fp16 = softmax(axis = var_931, x = aw_137_cast_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = softmax(axis = var_931, x = aw_139_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = softmax(axis = var_931, x = aw_141_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor var_1098_cast_fp16 = softmax(axis = var_931, x = aw_143_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = softmax(axis = var_931, x = aw_145_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor var_1100_cast_fp16 = softmax(axis = var_931, x = aw_147_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = softmax(axis = var_931, x = aw_149_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = softmax(axis = var_931, x = aw_151_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor var_1103_cast_fp16 = softmax(axis = var_931, x = aw_153_cast_fp16)[name = tensor("op_1103_cast_fp16")]; + tensor var_1104_cast_fp16 = softmax(axis = var_931, x = aw_155_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = softmax(axis = var_931, x = aw_157_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = softmax(axis = var_931, x = aw_159_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor var_1108_equation_0 = const()[name = tensor("op_1108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1108_cast_fp16 = einsum(equation = var_1108_equation_0, values = (var_1026_cast_fp16_0, var_1087_cast_fp16))[name = tensor("op_1108_cast_fp16")]; + tensor var_1110_equation_0 = const()[name = tensor("op_1110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1110_cast_fp16 = einsum(equation = var_1110_equation_0, values = (var_1026_cast_fp16_1, var_1088_cast_fp16))[name = tensor("op_1110_cast_fp16")]; + tensor var_1112_equation_0 = const()[name = tensor("op_1112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1112_cast_fp16 = einsum(equation = var_1112_equation_0, values = (var_1026_cast_fp16_2, var_1089_cast_fp16))[name = tensor("op_1112_cast_fp16")]; + tensor var_1114_equation_0 = const()[name = tensor("op_1114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1114_cast_fp16 = einsum(equation = var_1114_equation_0, values = (var_1026_cast_fp16_3, var_1090_cast_fp16))[name = tensor("op_1114_cast_fp16")]; + tensor var_1116_equation_0 = const()[name = tensor("op_1116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1116_cast_fp16 = einsum(equation = var_1116_equation_0, values = (var_1026_cast_fp16_4, var_1091_cast_fp16))[name = tensor("op_1116_cast_fp16")]; + tensor var_1118_equation_0 = const()[name = tensor("op_1118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1118_cast_fp16 = einsum(equation = var_1118_equation_0, values = (var_1026_cast_fp16_5, var_1092_cast_fp16))[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_equation_0 = const()[name = tensor("op_1120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1120_cast_fp16 = einsum(equation = var_1120_equation_0, values = (var_1026_cast_fp16_6, var_1093_cast_fp16))[name = tensor("op_1120_cast_fp16")]; + tensor var_1122_equation_0 = const()[name = tensor("op_1122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1122_cast_fp16 = einsum(equation = var_1122_equation_0, values = (var_1026_cast_fp16_7, var_1094_cast_fp16))[name = tensor("op_1122_cast_fp16")]; + tensor var_1124_equation_0 = const()[name = tensor("op_1124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1124_cast_fp16 = einsum(equation = var_1124_equation_0, values = (var_1026_cast_fp16_8, var_1095_cast_fp16))[name = tensor("op_1124_cast_fp16")]; + tensor var_1126_equation_0 = const()[name = tensor("op_1126_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1126_cast_fp16 = einsum(equation = var_1126_equation_0, values = (var_1026_cast_fp16_9, var_1096_cast_fp16))[name = tensor("op_1126_cast_fp16")]; + tensor var_1128_equation_0 = const()[name = tensor("op_1128_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1128_cast_fp16 = einsum(equation = var_1128_equation_0, values = (var_1026_cast_fp16_10, var_1097_cast_fp16))[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_equation_0 = const()[name = tensor("op_1130_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1130_cast_fp16 = einsum(equation = var_1130_equation_0, values = (var_1026_cast_fp16_11, var_1098_cast_fp16))[name = tensor("op_1130_cast_fp16")]; + tensor var_1132_equation_0 = const()[name = tensor("op_1132_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1132_cast_fp16 = einsum(equation = var_1132_equation_0, values = (var_1026_cast_fp16_12, var_1099_cast_fp16))[name = tensor("op_1132_cast_fp16")]; + tensor var_1134_equation_0 = const()[name = tensor("op_1134_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1134_cast_fp16 = einsum(equation = var_1134_equation_0, values = (var_1026_cast_fp16_13, var_1100_cast_fp16))[name = tensor("op_1134_cast_fp16")]; + tensor var_1136_equation_0 = const()[name = tensor("op_1136_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1136_cast_fp16 = einsum(equation = var_1136_equation_0, values = (var_1026_cast_fp16_14, var_1101_cast_fp16))[name = tensor("op_1136_cast_fp16")]; + tensor var_1138_equation_0 = const()[name = tensor("op_1138_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1138_cast_fp16 = einsum(equation = var_1138_equation_0, values = (var_1026_cast_fp16_15, var_1102_cast_fp16))[name = tensor("op_1138_cast_fp16")]; + tensor var_1140_equation_0 = const()[name = tensor("op_1140_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1140_cast_fp16 = einsum(equation = var_1140_equation_0, values = (var_1026_cast_fp16_16, var_1103_cast_fp16))[name = tensor("op_1140_cast_fp16")]; + tensor var_1142_equation_0 = const()[name = tensor("op_1142_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1142_cast_fp16 = einsum(equation = var_1142_equation_0, values = (var_1026_cast_fp16_17, var_1104_cast_fp16))[name = tensor("op_1142_cast_fp16")]; + tensor var_1144_equation_0 = const()[name = tensor("op_1144_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1144_cast_fp16 = einsum(equation = var_1144_equation_0, values = (var_1026_cast_fp16_18, var_1105_cast_fp16))[name = tensor("op_1144_cast_fp16")]; + tensor var_1146_equation_0 = const()[name = tensor("op_1146_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1146_cast_fp16 = einsum(equation = var_1146_equation_0, values = (var_1026_cast_fp16_19, var_1106_cast_fp16))[name = tensor("op_1146_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_931, interleave = input_35_interleave_0, values = (var_1108_cast_fp16, var_1110_cast_fp16, var_1112_cast_fp16, var_1114_cast_fp16, var_1116_cast_fp16, var_1118_cast_fp16, var_1120_cast_fp16, var_1122_cast_fp16, var_1124_cast_fp16, var_1126_cast_fp16, var_1128_cast_fp16, var_1130_cast_fp16, var_1132_cast_fp16, var_1134_cast_fp16, var_1136_cast_fp16, var_1138_cast_fp16, var_1140_cast_fp16, var_1142_cast_fp16, var_1144_cast_fp16, var_1146_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_1155_pad_type_0 = const()[name = tensor("op_1155_pad_type_0"), val = tensor("valid")]; + tensor var_1155_strides_0 = const()[name = tensor("op_1155_strides_0"), val = tensor([1, 1])]; + tensor var_1155_pad_0 = const()[name = tensor("op_1155_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1155_dilations_0 = const()[name = tensor("op_1155_dilations_0"), val = tensor([1, 1])]; + tensor var_1155_groups_0 = const()[name = tensor("op_1155_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142191232)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145468096)))]; + tensor var_1155_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_1155_dilations_0, groups = var_1155_groups_0, pad = var_1155_pad_0, pad_type = var_1155_pad_type_0, strides = var_1155_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_1155_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145470720)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145473344)))]; + tensor var_1165_to_fp16 = const()[name = tensor("op_1165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_1165_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145475968)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158583232)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1191_pad_type_0 = const()[name = tensor("op_1191_pad_type_0"), val = tensor("valid")]; + tensor var_1191_strides_0 = const()[name = tensor("op_1191_strides_0"), val = tensor([1, 1])]; + tensor var_1191_pad_0 = const()[name = tensor("op_1191_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1191_dilations_0 = const()[name = tensor("op_1191_dilations_0"), val = tensor([1, 1])]; + tensor var_1191_groups_0 = const()[name = tensor("op_1191_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158593536)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171700800)))]; + tensor var_1191_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_1191_dilations_0, groups = var_1191_groups_0, pad = var_1191_pad_0, pad_type = var_1191_pad_type_0, strides = var_1191_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_1191_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171703424)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171706048)))]; + tensor var_1216_to_fp16 = const()[name = tensor("op_1216_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_1216_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_1251_weight_0_to_fp16 = const()[name = tensor("op_1251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171708672)))]; + tensor var_1251_bias_0_to_fp16 = const()[name = tensor("op_1251_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174985536)))]; + tensor var_1251_cast_fp16 = conv(bias = var_1251_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_1251_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1251_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174988160)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_1249_pad_type_0 = const()[name = tensor("op_1249_pad_type_0"), val = tensor("valid")]; + tensor var_1249_strides_0 = const()[name = tensor("op_1249_strides_0"), val = tensor([1, 1])]; + tensor var_1249_pad_0 = const()[name = tensor("op_1249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1249_dilations_0 = const()[name = tensor("op_1249_dilations_0"), val = tensor([1, 1])]; + tensor var_1249_groups_0 = const()[name = tensor("op_1249_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178265024)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181541888)))]; + tensor var_1249_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_1249_dilations_0, groups = var_1249_groups_0, pad = var_1249_pad_0, pad_type = var_1249_pad_type_0, strides = var_1249_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1249_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1252_axis_0 = const()[name = tensor("op_1252_axis_0"), val = tensor(1)]; + tensor var_1252_cast_fp16_0, tensor var_1252_cast_fp16_1, tensor var_1252_cast_fp16_2, tensor var_1252_cast_fp16_3, tensor var_1252_cast_fp16_4, tensor var_1252_cast_fp16_5, tensor var_1252_cast_fp16_6, tensor var_1252_cast_fp16_7, tensor var_1252_cast_fp16_8, tensor var_1252_cast_fp16_9, tensor var_1252_cast_fp16_10, tensor var_1252_cast_fp16_11, tensor var_1252_cast_fp16_12, tensor var_1252_cast_fp16_13, tensor var_1252_cast_fp16_14, tensor var_1252_cast_fp16_15, tensor var_1252_cast_fp16_16, tensor var_1252_cast_fp16_17, tensor var_1252_cast_fp16_18, tensor var_1252_cast_fp16_19 = split(axis = var_1252_axis_0, split_sizes = tile_12, x = var_1251_cast_fp16)[name = tensor("op_1252_cast_fp16")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(3)]; + tensor var_1273_cast_fp16 = transpose(perm = var_1273_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_28")]; + tensor var_1274_cast_fp16_0, tensor var_1274_cast_fp16_1, tensor var_1274_cast_fp16_2, tensor var_1274_cast_fp16_3, tensor var_1274_cast_fp16_4, tensor var_1274_cast_fp16_5, tensor var_1274_cast_fp16_6, tensor var_1274_cast_fp16_7, tensor var_1274_cast_fp16_8, tensor var_1274_cast_fp16_9, tensor var_1274_cast_fp16_10, tensor var_1274_cast_fp16_11, tensor var_1274_cast_fp16_12, tensor var_1274_cast_fp16_13, tensor var_1274_cast_fp16_14, tensor var_1274_cast_fp16_15, tensor var_1274_cast_fp16_16, tensor var_1274_cast_fp16_17, tensor var_1274_cast_fp16_18, tensor var_1274_cast_fp16_19 = split(axis = var_1274_axis_0, split_sizes = tile_13, x = var_1273_cast_fp16)[name = tensor("op_1274_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1295_axis_0 = const()[name = tensor("op_1295_axis_0"), val = tensor(1)]; + tensor var_1295_cast_fp16_0, tensor var_1295_cast_fp16_1, tensor var_1295_cast_fp16_2, tensor var_1295_cast_fp16_3, tensor var_1295_cast_fp16_4, tensor var_1295_cast_fp16_5, tensor var_1295_cast_fp16_6, tensor var_1295_cast_fp16_7, tensor var_1295_cast_fp16_8, tensor var_1295_cast_fp16_9, tensor var_1295_cast_fp16_10, tensor var_1295_cast_fp16_11, tensor var_1295_cast_fp16_12, tensor var_1295_cast_fp16_13, tensor var_1295_cast_fp16_14, tensor var_1295_cast_fp16_15, tensor var_1295_cast_fp16_16, tensor var_1295_cast_fp16_17, tensor var_1295_cast_fp16_18, tensor var_1295_cast_fp16_19 = split(axis = var_1295_axis_0, split_sizes = tile_14, x = var_1249_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1274_cast_fp16_0, var_1252_cast_fp16_0))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1274_cast_fp16_1, var_1252_cast_fp16_1))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1274_cast_fp16_2, var_1252_cast_fp16_2))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1274_cast_fp16_3, var_1252_cast_fp16_3))[name = tensor("aw_167_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1274_cast_fp16_4, var_1252_cast_fp16_4))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1274_cast_fp16_5, var_1252_cast_fp16_5))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1274_cast_fp16_6, var_1252_cast_fp16_6))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1274_cast_fp16_7, var_1252_cast_fp16_7))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1274_cast_fp16_8, var_1252_cast_fp16_8))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1274_cast_fp16_9, var_1252_cast_fp16_9))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1274_cast_fp16_10, var_1252_cast_fp16_10))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1274_cast_fp16_11, var_1252_cast_fp16_11))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1274_cast_fp16_12, var_1252_cast_fp16_12))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1274_cast_fp16_13, var_1252_cast_fp16_13))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1274_cast_fp16_14, var_1252_cast_fp16_14))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1274_cast_fp16_15, var_1252_cast_fp16_15))[name = tensor("aw_191_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1274_cast_fp16_16, var_1252_cast_fp16_16))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1274_cast_fp16_17, var_1252_cast_fp16_17))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1274_cast_fp16_18, var_1252_cast_fp16_18))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1274_cast_fp16_19, var_1252_cast_fp16_19))[name = tensor("aw_199_cast_fp16")]; + tensor var_1356_cast_fp16 = softmax(axis = var_1200, x = aw_161_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor var_1357_cast_fp16 = softmax(axis = var_1200, x = aw_163_cast_fp16)[name = tensor("op_1357_cast_fp16")]; + tensor var_1358_cast_fp16 = softmax(axis = var_1200, x = aw_165_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = softmax(axis = var_1200, x = aw_167_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor var_1360_cast_fp16 = softmax(axis = var_1200, x = aw_169_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor var_1361_cast_fp16 = softmax(axis = var_1200, x = aw_171_cast_fp16)[name = tensor("op_1361_cast_fp16")]; + tensor var_1362_cast_fp16 = softmax(axis = var_1200, x = aw_173_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = softmax(axis = var_1200, x = aw_175_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = softmax(axis = var_1200, x = aw_177_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor var_1365_cast_fp16 = softmax(axis = var_1200, x = aw_179_cast_fp16)[name = tensor("op_1365_cast_fp16")]; + tensor var_1366_cast_fp16 = softmax(axis = var_1200, x = aw_181_cast_fp16)[name = tensor("op_1366_cast_fp16")]; + tensor var_1367_cast_fp16 = softmax(axis = var_1200, x = aw_183_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = softmax(axis = var_1200, x = aw_185_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor var_1369_cast_fp16 = softmax(axis = var_1200, x = aw_187_cast_fp16)[name = tensor("op_1369_cast_fp16")]; + tensor var_1370_cast_fp16 = softmax(axis = var_1200, x = aw_189_cast_fp16)[name = tensor("op_1370_cast_fp16")]; + tensor var_1371_cast_fp16 = softmax(axis = var_1200, x = aw_191_cast_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = softmax(axis = var_1200, x = aw_193_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor var_1373_cast_fp16 = softmax(axis = var_1200, x = aw_195_cast_fp16)[name = tensor("op_1373_cast_fp16")]; + tensor var_1374_cast_fp16 = softmax(axis = var_1200, x = aw_197_cast_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor var_1375_cast_fp16 = softmax(axis = var_1200, x = aw_199_cast_fp16)[name = tensor("op_1375_cast_fp16")]; + tensor var_1377_equation_0 = const()[name = tensor("op_1377_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1377_cast_fp16 = einsum(equation = var_1377_equation_0, values = (var_1295_cast_fp16_0, var_1356_cast_fp16))[name = tensor("op_1377_cast_fp16")]; + tensor var_1379_equation_0 = const()[name = tensor("op_1379_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1379_cast_fp16 = einsum(equation = var_1379_equation_0, values = (var_1295_cast_fp16_1, var_1357_cast_fp16))[name = tensor("op_1379_cast_fp16")]; + tensor var_1381_equation_0 = const()[name = tensor("op_1381_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1381_cast_fp16 = einsum(equation = var_1381_equation_0, values = (var_1295_cast_fp16_2, var_1358_cast_fp16))[name = tensor("op_1381_cast_fp16")]; + tensor var_1383_equation_0 = const()[name = tensor("op_1383_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1383_cast_fp16 = einsum(equation = var_1383_equation_0, values = (var_1295_cast_fp16_3, var_1359_cast_fp16))[name = tensor("op_1383_cast_fp16")]; + tensor var_1385_equation_0 = const()[name = tensor("op_1385_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1385_cast_fp16 = einsum(equation = var_1385_equation_0, values = (var_1295_cast_fp16_4, var_1360_cast_fp16))[name = tensor("op_1385_cast_fp16")]; + tensor var_1387_equation_0 = const()[name = tensor("op_1387_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1387_cast_fp16 = einsum(equation = var_1387_equation_0, values = (var_1295_cast_fp16_5, var_1361_cast_fp16))[name = tensor("op_1387_cast_fp16")]; + tensor var_1389_equation_0 = const()[name = tensor("op_1389_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1389_cast_fp16 = einsum(equation = var_1389_equation_0, values = (var_1295_cast_fp16_6, var_1362_cast_fp16))[name = tensor("op_1389_cast_fp16")]; + tensor var_1391_equation_0 = const()[name = tensor("op_1391_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1391_cast_fp16 = einsum(equation = var_1391_equation_0, values = (var_1295_cast_fp16_7, var_1363_cast_fp16))[name = tensor("op_1391_cast_fp16")]; + tensor var_1393_equation_0 = const()[name = tensor("op_1393_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1393_cast_fp16 = einsum(equation = var_1393_equation_0, values = (var_1295_cast_fp16_8, var_1364_cast_fp16))[name = tensor("op_1393_cast_fp16")]; + tensor var_1395_equation_0 = const()[name = tensor("op_1395_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1395_cast_fp16 = einsum(equation = var_1395_equation_0, values = (var_1295_cast_fp16_9, var_1365_cast_fp16))[name = tensor("op_1395_cast_fp16")]; + tensor var_1397_equation_0 = const()[name = tensor("op_1397_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1397_cast_fp16 = einsum(equation = var_1397_equation_0, values = (var_1295_cast_fp16_10, var_1366_cast_fp16))[name = tensor("op_1397_cast_fp16")]; + tensor var_1399_equation_0 = const()[name = tensor("op_1399_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1399_cast_fp16 = einsum(equation = var_1399_equation_0, values = (var_1295_cast_fp16_11, var_1367_cast_fp16))[name = tensor("op_1399_cast_fp16")]; + tensor var_1401_equation_0 = const()[name = tensor("op_1401_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1401_cast_fp16 = einsum(equation = var_1401_equation_0, values = (var_1295_cast_fp16_12, var_1368_cast_fp16))[name = tensor("op_1401_cast_fp16")]; + tensor var_1403_equation_0 = const()[name = tensor("op_1403_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1403_cast_fp16 = einsum(equation = var_1403_equation_0, values = (var_1295_cast_fp16_13, var_1369_cast_fp16))[name = tensor("op_1403_cast_fp16")]; + tensor var_1405_equation_0 = const()[name = tensor("op_1405_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1405_cast_fp16 = einsum(equation = var_1405_equation_0, values = (var_1295_cast_fp16_14, var_1370_cast_fp16))[name = tensor("op_1405_cast_fp16")]; + tensor var_1407_equation_0 = const()[name = tensor("op_1407_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1407_cast_fp16 = einsum(equation = var_1407_equation_0, values = (var_1295_cast_fp16_15, var_1371_cast_fp16))[name = tensor("op_1407_cast_fp16")]; + tensor var_1409_equation_0 = const()[name = tensor("op_1409_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1409_cast_fp16 = einsum(equation = var_1409_equation_0, values = (var_1295_cast_fp16_16, var_1372_cast_fp16))[name = tensor("op_1409_cast_fp16")]; + tensor var_1411_equation_0 = const()[name = tensor("op_1411_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1411_cast_fp16 = einsum(equation = var_1411_equation_0, values = (var_1295_cast_fp16_17, var_1373_cast_fp16))[name = tensor("op_1411_cast_fp16")]; + tensor var_1413_equation_0 = const()[name = tensor("op_1413_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1413_cast_fp16 = einsum(equation = var_1413_equation_0, values = (var_1295_cast_fp16_18, var_1374_cast_fp16))[name = tensor("op_1413_cast_fp16")]; + tensor var_1415_equation_0 = const()[name = tensor("op_1415_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1415_cast_fp16 = einsum(equation = var_1415_equation_0, values = (var_1295_cast_fp16_19, var_1375_cast_fp16))[name = tensor("op_1415_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_1200, interleave = input_45_interleave_0, values = (var_1377_cast_fp16, var_1379_cast_fp16, var_1381_cast_fp16, var_1383_cast_fp16, var_1385_cast_fp16, var_1387_cast_fp16, var_1389_cast_fp16, var_1391_cast_fp16, var_1393_cast_fp16, var_1395_cast_fp16, var_1397_cast_fp16, var_1399_cast_fp16, var_1401_cast_fp16, var_1403_cast_fp16, var_1405_cast_fp16, var_1407_cast_fp16, var_1409_cast_fp16, var_1411_cast_fp16, var_1413_cast_fp16, var_1415_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1424_pad_type_0 = const()[name = tensor("op_1424_pad_type_0"), val = tensor("valid")]; + tensor var_1424_strides_0 = const()[name = tensor("op_1424_strides_0"), val = tensor([1, 1])]; + tensor var_1424_pad_0 = const()[name = tensor("op_1424_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1424_dilations_0 = const()[name = tensor("op_1424_dilations_0"), val = tensor([1, 1])]; + tensor var_1424_groups_0 = const()[name = tensor("op_1424_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181544512)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184821376)))]; + tensor var_1424_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1424_dilations_0, groups = var_1424_groups_0, pad = var_1424_pad_0, pad_type = var_1424_pad_type_0, strides = var_1424_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1424_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184824000)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184826624)))]; + tensor var_1434_to_fp16 = const()[name = tensor("op_1434_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1434_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184829248)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197936512)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1460_pad_type_0 = const()[name = tensor("op_1460_pad_type_0"), val = tensor("valid")]; + tensor var_1460_strides_0 = const()[name = tensor("op_1460_strides_0"), val = tensor([1, 1])]; + tensor var_1460_pad_0 = const()[name = tensor("op_1460_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1460_dilations_0 = const()[name = tensor("op_1460_dilations_0"), val = tensor([1, 1])]; + tensor var_1460_groups_0 = const()[name = tensor("op_1460_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197946816)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211054080)))]; + tensor var_1460_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1460_dilations_0, groups = var_1460_groups_0, pad = var_1460_pad_0, pad_type = var_1460_pad_type_0, strides = var_1460_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1460_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1469 = const()[name = tensor("op_1469"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211056704)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211059328)))]; + tensor var_1485_to_fp16 = const()[name = tensor("op_1485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1485_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1520_weight_0_to_fp16 = const()[name = tensor("op_1520_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211061952)))]; + tensor var_1520_bias_0_to_fp16 = const()[name = tensor("op_1520_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214338816)))]; + tensor var_1520_cast_fp16 = conv(bias = var_1520_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1520_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214341440)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1518_pad_type_0 = const()[name = tensor("op_1518_pad_type_0"), val = tensor("valid")]; + tensor var_1518_strides_0 = const()[name = tensor("op_1518_strides_0"), val = tensor([1, 1])]; + tensor var_1518_pad_0 = const()[name = tensor("op_1518_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1518_dilations_0 = const()[name = tensor("op_1518_dilations_0"), val = tensor([1, 1])]; + tensor var_1518_groups_0 = const()[name = tensor("op_1518_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217618304)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220895168)))]; + tensor var_1518_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1518_dilations_0, groups = var_1518_groups_0, pad = var_1518_pad_0, pad_type = var_1518_pad_type_0, strides = var_1518_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1518_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1521_axis_0 = const()[name = tensor("op_1521_axis_0"), val = tensor(1)]; + tensor var_1521_cast_fp16_0, tensor var_1521_cast_fp16_1, tensor var_1521_cast_fp16_2, tensor var_1521_cast_fp16_3, tensor var_1521_cast_fp16_4, tensor var_1521_cast_fp16_5, tensor var_1521_cast_fp16_6, tensor var_1521_cast_fp16_7, tensor var_1521_cast_fp16_8, tensor var_1521_cast_fp16_9, tensor var_1521_cast_fp16_10, tensor var_1521_cast_fp16_11, tensor var_1521_cast_fp16_12, tensor var_1521_cast_fp16_13, tensor var_1521_cast_fp16_14, tensor var_1521_cast_fp16_15, tensor var_1521_cast_fp16_16, tensor var_1521_cast_fp16_17, tensor var_1521_cast_fp16_18, tensor var_1521_cast_fp16_19 = split(axis = var_1521_axis_0, split_sizes = tile_15, x = var_1520_cast_fp16)[name = tensor("op_1521_cast_fp16")]; + tensor var_1542_perm_0 = const()[name = tensor("op_1542_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1543_axis_0 = const()[name = tensor("op_1543_axis_0"), val = tensor(3)]; + tensor var_1542_cast_fp16 = transpose(perm = var_1542_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1543_cast_fp16_0, tensor var_1543_cast_fp16_1, tensor var_1543_cast_fp16_2, tensor var_1543_cast_fp16_3, tensor var_1543_cast_fp16_4, tensor var_1543_cast_fp16_5, tensor var_1543_cast_fp16_6, tensor var_1543_cast_fp16_7, tensor var_1543_cast_fp16_8, tensor var_1543_cast_fp16_9, tensor var_1543_cast_fp16_10, tensor var_1543_cast_fp16_11, tensor var_1543_cast_fp16_12, tensor var_1543_cast_fp16_13, tensor var_1543_cast_fp16_14, tensor var_1543_cast_fp16_15, tensor var_1543_cast_fp16_16, tensor var_1543_cast_fp16_17, tensor var_1543_cast_fp16_18, tensor var_1543_cast_fp16_19 = split(axis = var_1543_axis_0, split_sizes = tile_16, x = var_1542_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1564_axis_0 = const()[name = tensor("op_1564_axis_0"), val = tensor(1)]; + tensor var_1564_cast_fp16_0, tensor var_1564_cast_fp16_1, tensor var_1564_cast_fp16_2, tensor var_1564_cast_fp16_3, tensor var_1564_cast_fp16_4, tensor var_1564_cast_fp16_5, tensor var_1564_cast_fp16_6, tensor var_1564_cast_fp16_7, tensor var_1564_cast_fp16_8, tensor var_1564_cast_fp16_9, tensor var_1564_cast_fp16_10, tensor var_1564_cast_fp16_11, tensor var_1564_cast_fp16_12, tensor var_1564_cast_fp16_13, tensor var_1564_cast_fp16_14, tensor var_1564_cast_fp16_15, tensor var_1564_cast_fp16_16, tensor var_1564_cast_fp16_17, tensor var_1564_cast_fp16_18, tensor var_1564_cast_fp16_19 = split(axis = var_1564_axis_0, split_sizes = tile_17, x = var_1518_cast_fp16)[name = tensor("op_1564_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1543_cast_fp16_0, var_1521_cast_fp16_0))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1543_cast_fp16_1, var_1521_cast_fp16_1))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1543_cast_fp16_2, var_1521_cast_fp16_2))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1543_cast_fp16_3, var_1521_cast_fp16_3))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1543_cast_fp16_4, var_1521_cast_fp16_4))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1543_cast_fp16_5, var_1521_cast_fp16_5))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1543_cast_fp16_6, var_1521_cast_fp16_6))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1543_cast_fp16_7, var_1521_cast_fp16_7))[name = tensor("aw_215_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1543_cast_fp16_8, var_1521_cast_fp16_8))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1543_cast_fp16_9, var_1521_cast_fp16_9))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1543_cast_fp16_10, var_1521_cast_fp16_10))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1543_cast_fp16_11, var_1521_cast_fp16_11))[name = tensor("aw_223_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1543_cast_fp16_12, var_1521_cast_fp16_12))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1543_cast_fp16_13, var_1521_cast_fp16_13))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1543_cast_fp16_14, var_1521_cast_fp16_14))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1543_cast_fp16_15, var_1521_cast_fp16_15))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1543_cast_fp16_16, var_1521_cast_fp16_16))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1543_cast_fp16_17, var_1521_cast_fp16_17))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1543_cast_fp16_18, var_1521_cast_fp16_18))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1543_cast_fp16_19, var_1521_cast_fp16_19))[name = tensor("aw_239_cast_fp16")]; + tensor var_1625_cast_fp16 = softmax(axis = var_1469, x = aw_201_cast_fp16)[name = tensor("op_1625_cast_fp16")]; + tensor var_1626_cast_fp16 = softmax(axis = var_1469, x = aw_203_cast_fp16)[name = tensor("op_1626_cast_fp16")]; + tensor var_1627_cast_fp16 = softmax(axis = var_1469, x = aw_205_cast_fp16)[name = tensor("op_1627_cast_fp16")]; + tensor var_1628_cast_fp16 = softmax(axis = var_1469, x = aw_207_cast_fp16)[name = tensor("op_1628_cast_fp16")]; + tensor var_1629_cast_fp16 = softmax(axis = var_1469, x = aw_209_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1630_cast_fp16 = softmax(axis = var_1469, x = aw_211_cast_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = softmax(axis = var_1469, x = aw_213_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_cast_fp16 = softmax(axis = var_1469, x = aw_215_cast_fp16)[name = tensor("op_1632_cast_fp16")]; + tensor var_1633_cast_fp16 = softmax(axis = var_1469, x = aw_217_cast_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor var_1634_cast_fp16 = softmax(axis = var_1469, x = aw_219_cast_fp16)[name = tensor("op_1634_cast_fp16")]; + tensor var_1635_cast_fp16 = softmax(axis = var_1469, x = aw_221_cast_fp16)[name = tensor("op_1635_cast_fp16")]; + tensor var_1636_cast_fp16 = softmax(axis = var_1469, x = aw_223_cast_fp16)[name = tensor("op_1636_cast_fp16")]; + tensor var_1637_cast_fp16 = softmax(axis = var_1469, x = aw_225_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor var_1638_cast_fp16 = softmax(axis = var_1469, x = aw_227_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_cast_fp16 = softmax(axis = var_1469, x = aw_229_cast_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1640_cast_fp16 = softmax(axis = var_1469, x = aw_231_cast_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor var_1641_cast_fp16 = softmax(axis = var_1469, x = aw_233_cast_fp16)[name = tensor("op_1641_cast_fp16")]; + tensor var_1642_cast_fp16 = softmax(axis = var_1469, x = aw_235_cast_fp16)[name = tensor("op_1642_cast_fp16")]; + tensor var_1643_cast_fp16 = softmax(axis = var_1469, x = aw_237_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor var_1644_cast_fp16 = softmax(axis = var_1469, x = aw_239_cast_fp16)[name = tensor("op_1644_cast_fp16")]; + tensor var_1646_equation_0 = const()[name = tensor("op_1646_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1646_cast_fp16 = einsum(equation = var_1646_equation_0, values = (var_1564_cast_fp16_0, var_1625_cast_fp16))[name = tensor("op_1646_cast_fp16")]; + tensor var_1648_equation_0 = const()[name = tensor("op_1648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1648_cast_fp16 = einsum(equation = var_1648_equation_0, values = (var_1564_cast_fp16_1, var_1626_cast_fp16))[name = tensor("op_1648_cast_fp16")]; + tensor var_1650_equation_0 = const()[name = tensor("op_1650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1650_cast_fp16 = einsum(equation = var_1650_equation_0, values = (var_1564_cast_fp16_2, var_1627_cast_fp16))[name = tensor("op_1650_cast_fp16")]; + tensor var_1652_equation_0 = const()[name = tensor("op_1652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1652_cast_fp16 = einsum(equation = var_1652_equation_0, values = (var_1564_cast_fp16_3, var_1628_cast_fp16))[name = tensor("op_1652_cast_fp16")]; + tensor var_1654_equation_0 = const()[name = tensor("op_1654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1654_cast_fp16 = einsum(equation = var_1654_equation_0, values = (var_1564_cast_fp16_4, var_1629_cast_fp16))[name = tensor("op_1654_cast_fp16")]; + tensor var_1656_equation_0 = const()[name = tensor("op_1656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1656_cast_fp16 = einsum(equation = var_1656_equation_0, values = (var_1564_cast_fp16_5, var_1630_cast_fp16))[name = tensor("op_1656_cast_fp16")]; + tensor var_1658_equation_0 = const()[name = tensor("op_1658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1658_cast_fp16 = einsum(equation = var_1658_equation_0, values = (var_1564_cast_fp16_6, var_1631_cast_fp16))[name = tensor("op_1658_cast_fp16")]; + tensor var_1660_equation_0 = const()[name = tensor("op_1660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1660_cast_fp16 = einsum(equation = var_1660_equation_0, values = (var_1564_cast_fp16_7, var_1632_cast_fp16))[name = tensor("op_1660_cast_fp16")]; + tensor var_1662_equation_0 = const()[name = tensor("op_1662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1662_cast_fp16 = einsum(equation = var_1662_equation_0, values = (var_1564_cast_fp16_8, var_1633_cast_fp16))[name = tensor("op_1662_cast_fp16")]; + tensor var_1664_equation_0 = const()[name = tensor("op_1664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1664_cast_fp16 = einsum(equation = var_1664_equation_0, values = (var_1564_cast_fp16_9, var_1634_cast_fp16))[name = tensor("op_1664_cast_fp16")]; + tensor var_1666_equation_0 = const()[name = tensor("op_1666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1666_cast_fp16 = einsum(equation = var_1666_equation_0, values = (var_1564_cast_fp16_10, var_1635_cast_fp16))[name = tensor("op_1666_cast_fp16")]; + tensor var_1668_equation_0 = const()[name = tensor("op_1668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1668_cast_fp16 = einsum(equation = var_1668_equation_0, values = (var_1564_cast_fp16_11, var_1636_cast_fp16))[name = tensor("op_1668_cast_fp16")]; + tensor var_1670_equation_0 = const()[name = tensor("op_1670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1670_cast_fp16 = einsum(equation = var_1670_equation_0, values = (var_1564_cast_fp16_12, var_1637_cast_fp16))[name = tensor("op_1670_cast_fp16")]; + tensor var_1672_equation_0 = const()[name = tensor("op_1672_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1672_cast_fp16 = einsum(equation = var_1672_equation_0, values = (var_1564_cast_fp16_13, var_1638_cast_fp16))[name = tensor("op_1672_cast_fp16")]; + tensor var_1674_equation_0 = const()[name = tensor("op_1674_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1674_cast_fp16 = einsum(equation = var_1674_equation_0, values = (var_1564_cast_fp16_14, var_1639_cast_fp16))[name = tensor("op_1674_cast_fp16")]; + tensor var_1676_equation_0 = const()[name = tensor("op_1676_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1676_cast_fp16 = einsum(equation = var_1676_equation_0, values = (var_1564_cast_fp16_15, var_1640_cast_fp16))[name = tensor("op_1676_cast_fp16")]; + tensor var_1678_equation_0 = const()[name = tensor("op_1678_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1678_cast_fp16 = einsum(equation = var_1678_equation_0, values = (var_1564_cast_fp16_16, var_1641_cast_fp16))[name = tensor("op_1678_cast_fp16")]; + tensor var_1680_equation_0 = const()[name = tensor("op_1680_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1680_cast_fp16 = einsum(equation = var_1680_equation_0, values = (var_1564_cast_fp16_17, var_1642_cast_fp16))[name = tensor("op_1680_cast_fp16")]; + tensor var_1682_equation_0 = const()[name = tensor("op_1682_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1682_cast_fp16 = einsum(equation = var_1682_equation_0, values = (var_1564_cast_fp16_18, var_1643_cast_fp16))[name = tensor("op_1682_cast_fp16")]; + tensor var_1684_equation_0 = const()[name = tensor("op_1684_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1684_cast_fp16 = einsum(equation = var_1684_equation_0, values = (var_1564_cast_fp16_19, var_1644_cast_fp16))[name = tensor("op_1684_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1469, interleave = input_55_interleave_0, values = (var_1646_cast_fp16, var_1648_cast_fp16, var_1650_cast_fp16, var_1652_cast_fp16, var_1654_cast_fp16, var_1656_cast_fp16, var_1658_cast_fp16, var_1660_cast_fp16, var_1662_cast_fp16, var_1664_cast_fp16, var_1666_cast_fp16, var_1668_cast_fp16, var_1670_cast_fp16, var_1672_cast_fp16, var_1674_cast_fp16, var_1676_cast_fp16, var_1678_cast_fp16, var_1680_cast_fp16, var_1682_cast_fp16, var_1684_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1693_pad_type_0 = const()[name = tensor("op_1693_pad_type_0"), val = tensor("valid")]; + tensor var_1693_strides_0 = const()[name = tensor("op_1693_strides_0"), val = tensor([1, 1])]; + tensor var_1693_pad_0 = const()[name = tensor("op_1693_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1693_dilations_0 = const()[name = tensor("op_1693_dilations_0"), val = tensor([1, 1])]; + tensor var_1693_groups_0 = const()[name = tensor("op_1693_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220897792)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224174656)))]; + tensor var_1693_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1693_dilations_0, groups = var_1693_groups_0, pad = var_1693_pad_0, pad_type = var_1693_pad_type_0, strides = var_1693_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1693_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1693_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224177280)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224179904)))]; + tensor var_1703_to_fp16 = const()[name = tensor("op_1703_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1703_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224182528)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237289792)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1729_pad_type_0 = const()[name = tensor("op_1729_pad_type_0"), val = tensor("valid")]; + tensor var_1729_strides_0 = const()[name = tensor("op_1729_strides_0"), val = tensor([1, 1])]; + tensor var_1729_pad_0 = const()[name = tensor("op_1729_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1729_dilations_0 = const()[name = tensor("op_1729_dilations_0"), val = tensor([1, 1])]; + tensor var_1729_groups_0 = const()[name = tensor("op_1729_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237300096)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250407360)))]; + tensor var_1729_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1729_dilations_0, groups = var_1729_groups_0, pad = var_1729_pad_0, pad_type = var_1729_pad_type_0, strides = var_1729_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1729_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1729_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1738 = const()[name = tensor("op_1738"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250409984)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250412608)))]; + tensor var_1754_to_fp16 = const()[name = tensor("op_1754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1754_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1789_weight_0_to_fp16 = const()[name = tensor("op_1789_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250415232)))]; + tensor var_1789_bias_0_to_fp16 = const()[name = tensor("op_1789_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253692096)))]; + tensor var_1789_cast_fp16 = conv(bias = var_1789_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1789_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1789_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253694720)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1787_pad_type_0 = const()[name = tensor("op_1787_pad_type_0"), val = tensor("valid")]; + tensor var_1787_strides_0 = const()[name = tensor("op_1787_strides_0"), val = tensor([1, 1])]; + tensor var_1787_pad_0 = const()[name = tensor("op_1787_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1787_dilations_0 = const()[name = tensor("op_1787_dilations_0"), val = tensor([1, 1])]; + tensor var_1787_groups_0 = const()[name = tensor("op_1787_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256971584)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260248448)))]; + tensor var_1787_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1787_dilations_0, groups = var_1787_groups_0, pad = var_1787_pad_0, pad_type = var_1787_pad_type_0, strides = var_1787_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1790_axis_0 = const()[name = tensor("op_1790_axis_0"), val = tensor(1)]; + tensor var_1790_cast_fp16_0, tensor var_1790_cast_fp16_1, tensor var_1790_cast_fp16_2, tensor var_1790_cast_fp16_3, tensor var_1790_cast_fp16_4, tensor var_1790_cast_fp16_5, tensor var_1790_cast_fp16_6, tensor var_1790_cast_fp16_7, tensor var_1790_cast_fp16_8, tensor var_1790_cast_fp16_9, tensor var_1790_cast_fp16_10, tensor var_1790_cast_fp16_11, tensor var_1790_cast_fp16_12, tensor var_1790_cast_fp16_13, tensor var_1790_cast_fp16_14, tensor var_1790_cast_fp16_15, tensor var_1790_cast_fp16_16, tensor var_1790_cast_fp16_17, tensor var_1790_cast_fp16_18, tensor var_1790_cast_fp16_19 = split(axis = var_1790_axis_0, split_sizes = tile_18, x = var_1789_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1811_perm_0 = const()[name = tensor("op_1811_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1812_axis_0 = const()[name = tensor("op_1812_axis_0"), val = tensor(3)]; + tensor var_1811_cast_fp16 = transpose(perm = var_1811_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_26")]; + tensor var_1812_cast_fp16_0, tensor var_1812_cast_fp16_1, tensor var_1812_cast_fp16_2, tensor var_1812_cast_fp16_3, tensor var_1812_cast_fp16_4, tensor var_1812_cast_fp16_5, tensor var_1812_cast_fp16_6, tensor var_1812_cast_fp16_7, tensor var_1812_cast_fp16_8, tensor var_1812_cast_fp16_9, tensor var_1812_cast_fp16_10, tensor var_1812_cast_fp16_11, tensor var_1812_cast_fp16_12, tensor var_1812_cast_fp16_13, tensor var_1812_cast_fp16_14, tensor var_1812_cast_fp16_15, tensor var_1812_cast_fp16_16, tensor var_1812_cast_fp16_17, tensor var_1812_cast_fp16_18, tensor var_1812_cast_fp16_19 = split(axis = var_1812_axis_0, split_sizes = tile_19, x = var_1811_cast_fp16)[name = tensor("op_1812_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1833_axis_0 = const()[name = tensor("op_1833_axis_0"), val = tensor(1)]; + tensor var_1833_cast_fp16_0, tensor var_1833_cast_fp16_1, tensor var_1833_cast_fp16_2, tensor var_1833_cast_fp16_3, tensor var_1833_cast_fp16_4, tensor var_1833_cast_fp16_5, tensor var_1833_cast_fp16_6, tensor var_1833_cast_fp16_7, tensor var_1833_cast_fp16_8, tensor var_1833_cast_fp16_9, tensor var_1833_cast_fp16_10, tensor var_1833_cast_fp16_11, tensor var_1833_cast_fp16_12, tensor var_1833_cast_fp16_13, tensor var_1833_cast_fp16_14, tensor var_1833_cast_fp16_15, tensor var_1833_cast_fp16_16, tensor var_1833_cast_fp16_17, tensor var_1833_cast_fp16_18, tensor var_1833_cast_fp16_19 = split(axis = var_1833_axis_0, split_sizes = tile_20, x = var_1787_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_1812_cast_fp16_0, var_1790_cast_fp16_0))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_1812_cast_fp16_1, var_1790_cast_fp16_1))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_1812_cast_fp16_2, var_1790_cast_fp16_2))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_1812_cast_fp16_3, var_1790_cast_fp16_3))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_1812_cast_fp16_4, var_1790_cast_fp16_4))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_1812_cast_fp16_5, var_1790_cast_fp16_5))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_1812_cast_fp16_6, var_1790_cast_fp16_6))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_1812_cast_fp16_7, var_1790_cast_fp16_7))[name = tensor("aw_255_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_1812_cast_fp16_8, var_1790_cast_fp16_8))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_1812_cast_fp16_9, var_1790_cast_fp16_9))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_1812_cast_fp16_10, var_1790_cast_fp16_10))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_1812_cast_fp16_11, var_1790_cast_fp16_11))[name = tensor("aw_263_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_1812_cast_fp16_12, var_1790_cast_fp16_12))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_1812_cast_fp16_13, var_1790_cast_fp16_13))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_1812_cast_fp16_14, var_1790_cast_fp16_14))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_1812_cast_fp16_15, var_1790_cast_fp16_15))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_1812_cast_fp16_16, var_1790_cast_fp16_16))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_1812_cast_fp16_17, var_1790_cast_fp16_17))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_1812_cast_fp16_18, var_1790_cast_fp16_18))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_1812_cast_fp16_19, var_1790_cast_fp16_19))[name = tensor("aw_279_cast_fp16")]; + tensor var_1894_cast_fp16 = softmax(axis = var_1738, x = aw_241_cast_fp16)[name = tensor("op_1894_cast_fp16")]; + tensor var_1895_cast_fp16 = softmax(axis = var_1738, x = aw_243_cast_fp16)[name = tensor("op_1895_cast_fp16")]; + tensor var_1896_cast_fp16 = softmax(axis = var_1738, x = aw_245_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1897_cast_fp16 = softmax(axis = var_1738, x = aw_247_cast_fp16)[name = tensor("op_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = softmax(axis = var_1738, x = aw_249_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_cast_fp16 = softmax(axis = var_1738, x = aw_251_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor var_1900_cast_fp16 = softmax(axis = var_1738, x = aw_253_cast_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor var_1901_cast_fp16 = softmax(axis = var_1738, x = aw_255_cast_fp16)[name = tensor("op_1901_cast_fp16")]; + tensor var_1902_cast_fp16 = softmax(axis = var_1738, x = aw_257_cast_fp16)[name = tensor("op_1902_cast_fp16")]; + tensor var_1903_cast_fp16 = softmax(axis = var_1738, x = aw_259_cast_fp16)[name = tensor("op_1903_cast_fp16")]; + tensor var_1904_cast_fp16 = softmax(axis = var_1738, x = aw_261_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1905_cast_fp16 = softmax(axis = var_1738, x = aw_263_cast_fp16)[name = tensor("op_1905_cast_fp16")]; + tensor var_1906_cast_fp16 = softmax(axis = var_1738, x = aw_265_cast_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1907_cast_fp16 = softmax(axis = var_1738, x = aw_267_cast_fp16)[name = tensor("op_1907_cast_fp16")]; + tensor var_1908_cast_fp16 = softmax(axis = var_1738, x = aw_269_cast_fp16)[name = tensor("op_1908_cast_fp16")]; + tensor var_1909_cast_fp16 = softmax(axis = var_1738, x = aw_271_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor var_1910_cast_fp16 = softmax(axis = var_1738, x = aw_273_cast_fp16)[name = tensor("op_1910_cast_fp16")]; + tensor var_1911_cast_fp16 = softmax(axis = var_1738, x = aw_275_cast_fp16)[name = tensor("op_1911_cast_fp16")]; + tensor var_1912_cast_fp16 = softmax(axis = var_1738, x = aw_277_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1913_cast_fp16 = softmax(axis = var_1738, x = aw_279_cast_fp16)[name = tensor("op_1913_cast_fp16")]; + tensor var_1915_equation_0 = const()[name = tensor("op_1915_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1915_cast_fp16 = einsum(equation = var_1915_equation_0, values = (var_1833_cast_fp16_0, var_1894_cast_fp16))[name = tensor("op_1915_cast_fp16")]; + tensor var_1917_equation_0 = const()[name = tensor("op_1917_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1917_cast_fp16 = einsum(equation = var_1917_equation_0, values = (var_1833_cast_fp16_1, var_1895_cast_fp16))[name = tensor("op_1917_cast_fp16")]; + tensor var_1919_equation_0 = const()[name = tensor("op_1919_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1919_cast_fp16 = einsum(equation = var_1919_equation_0, values = (var_1833_cast_fp16_2, var_1896_cast_fp16))[name = tensor("op_1919_cast_fp16")]; + tensor var_1921_equation_0 = const()[name = tensor("op_1921_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1921_cast_fp16 = einsum(equation = var_1921_equation_0, values = (var_1833_cast_fp16_3, var_1897_cast_fp16))[name = tensor("op_1921_cast_fp16")]; + tensor var_1923_equation_0 = const()[name = tensor("op_1923_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1923_cast_fp16 = einsum(equation = var_1923_equation_0, values = (var_1833_cast_fp16_4, var_1898_cast_fp16))[name = tensor("op_1923_cast_fp16")]; + tensor var_1925_equation_0 = const()[name = tensor("op_1925_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1925_cast_fp16 = einsum(equation = var_1925_equation_0, values = (var_1833_cast_fp16_5, var_1899_cast_fp16))[name = tensor("op_1925_cast_fp16")]; + tensor var_1927_equation_0 = const()[name = tensor("op_1927_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1927_cast_fp16 = einsum(equation = var_1927_equation_0, values = (var_1833_cast_fp16_6, var_1900_cast_fp16))[name = tensor("op_1927_cast_fp16")]; + tensor var_1929_equation_0 = const()[name = tensor("op_1929_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1929_cast_fp16 = einsum(equation = var_1929_equation_0, values = (var_1833_cast_fp16_7, var_1901_cast_fp16))[name = tensor("op_1929_cast_fp16")]; + tensor var_1931_equation_0 = const()[name = tensor("op_1931_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1931_cast_fp16 = einsum(equation = var_1931_equation_0, values = (var_1833_cast_fp16_8, var_1902_cast_fp16))[name = tensor("op_1931_cast_fp16")]; + tensor var_1933_equation_0 = const()[name = tensor("op_1933_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1933_cast_fp16 = einsum(equation = var_1933_equation_0, values = (var_1833_cast_fp16_9, var_1903_cast_fp16))[name = tensor("op_1933_cast_fp16")]; + tensor var_1935_equation_0 = const()[name = tensor("op_1935_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1935_cast_fp16 = einsum(equation = var_1935_equation_0, values = (var_1833_cast_fp16_10, var_1904_cast_fp16))[name = tensor("op_1935_cast_fp16")]; + tensor var_1937_equation_0 = const()[name = tensor("op_1937_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1937_cast_fp16 = einsum(equation = var_1937_equation_0, values = (var_1833_cast_fp16_11, var_1905_cast_fp16))[name = tensor("op_1937_cast_fp16")]; + tensor var_1939_equation_0 = const()[name = tensor("op_1939_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1939_cast_fp16 = einsum(equation = var_1939_equation_0, values = (var_1833_cast_fp16_12, var_1906_cast_fp16))[name = tensor("op_1939_cast_fp16")]; + tensor var_1941_equation_0 = const()[name = tensor("op_1941_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1941_cast_fp16 = einsum(equation = var_1941_equation_0, values = (var_1833_cast_fp16_13, var_1907_cast_fp16))[name = tensor("op_1941_cast_fp16")]; + tensor var_1943_equation_0 = const()[name = tensor("op_1943_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1943_cast_fp16 = einsum(equation = var_1943_equation_0, values = (var_1833_cast_fp16_14, var_1908_cast_fp16))[name = tensor("op_1943_cast_fp16")]; + tensor var_1945_equation_0 = const()[name = tensor("op_1945_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1945_cast_fp16 = einsum(equation = var_1945_equation_0, values = (var_1833_cast_fp16_15, var_1909_cast_fp16))[name = tensor("op_1945_cast_fp16")]; + tensor var_1947_equation_0 = const()[name = tensor("op_1947_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1947_cast_fp16 = einsum(equation = var_1947_equation_0, values = (var_1833_cast_fp16_16, var_1910_cast_fp16))[name = tensor("op_1947_cast_fp16")]; + tensor var_1949_equation_0 = const()[name = tensor("op_1949_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1949_cast_fp16 = einsum(equation = var_1949_equation_0, values = (var_1833_cast_fp16_17, var_1911_cast_fp16))[name = tensor("op_1949_cast_fp16")]; + tensor var_1951_equation_0 = const()[name = tensor("op_1951_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1951_cast_fp16 = einsum(equation = var_1951_equation_0, values = (var_1833_cast_fp16_18, var_1912_cast_fp16))[name = tensor("op_1951_cast_fp16")]; + tensor var_1953_equation_0 = const()[name = tensor("op_1953_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1953_cast_fp16 = einsum(equation = var_1953_equation_0, values = (var_1833_cast_fp16_19, var_1913_cast_fp16))[name = tensor("op_1953_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1738, interleave = input_65_interleave_0, values = (var_1915_cast_fp16, var_1917_cast_fp16, var_1919_cast_fp16, var_1921_cast_fp16, var_1923_cast_fp16, var_1925_cast_fp16, var_1927_cast_fp16, var_1929_cast_fp16, var_1931_cast_fp16, var_1933_cast_fp16, var_1935_cast_fp16, var_1937_cast_fp16, var_1939_cast_fp16, var_1941_cast_fp16, var_1943_cast_fp16, var_1945_cast_fp16, var_1947_cast_fp16, var_1949_cast_fp16, var_1951_cast_fp16, var_1953_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1962_pad_type_0 = const()[name = tensor("op_1962_pad_type_0"), val = tensor("valid")]; + tensor var_1962_strides_0 = const()[name = tensor("op_1962_strides_0"), val = tensor([1, 1])]; + tensor var_1962_pad_0 = const()[name = tensor("op_1962_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1962_dilations_0 = const()[name = tensor("op_1962_dilations_0"), val = tensor([1, 1])]; + tensor var_1962_groups_0 = const()[name = tensor("op_1962_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260251072)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263527936)))]; + tensor var_1962_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1962_dilations_0, groups = var_1962_groups_0, pad = var_1962_pad_0, pad_type = var_1962_pad_type_0, strides = var_1962_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1962_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263530560)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263533184)))]; + tensor var_1972_to_fp16 = const()[name = tensor("op_1972_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1972_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263535808)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276643072)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1998_pad_type_0 = const()[name = tensor("op_1998_pad_type_0"), val = tensor("valid")]; + tensor var_1998_strides_0 = const()[name = tensor("op_1998_strides_0"), val = tensor([1, 1])]; + tensor var_1998_pad_0 = const()[name = tensor("op_1998_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1998_dilations_0 = const()[name = tensor("op_1998_dilations_0"), val = tensor([1, 1])]; + tensor var_1998_groups_0 = const()[name = tensor("op_1998_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276653376)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289760640)))]; + tensor var_1998_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1998_dilations_0, groups = var_1998_groups_0, pad = var_1998_pad_0, pad_type = var_1998_pad_type_0, strides = var_1998_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1998_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_2007 = const()[name = tensor("op_2007"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289763264)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289765888)))]; + tensor var_2023_to_fp16 = const()[name = tensor("op_2023_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_2023_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_2058_weight_0_to_fp16 = const()[name = tensor("op_2058_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289768512)))]; + tensor var_2058_bias_0_to_fp16 = const()[name = tensor("op_2058_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293045376)))]; + tensor var_2058_cast_fp16 = conv(bias = var_2058_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_2058_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_2058_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293048000)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_2056_pad_type_0 = const()[name = tensor("op_2056_pad_type_0"), val = tensor("valid")]; + tensor var_2056_strides_0 = const()[name = tensor("op_2056_strides_0"), val = tensor([1, 1])]; + tensor var_2056_pad_0 = const()[name = tensor("op_2056_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2056_dilations_0 = const()[name = tensor("op_2056_dilations_0"), val = tensor([1, 1])]; + tensor var_2056_groups_0 = const()[name = tensor("op_2056_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296324864)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299601728)))]; + tensor var_2056_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_2056_dilations_0, groups = var_2056_groups_0, pad = var_2056_pad_0, pad_type = var_2056_pad_type_0, strides = var_2056_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2059_axis_0 = const()[name = tensor("op_2059_axis_0"), val = tensor(1)]; + tensor var_2059_cast_fp16_0, tensor var_2059_cast_fp16_1, tensor var_2059_cast_fp16_2, tensor var_2059_cast_fp16_3, tensor var_2059_cast_fp16_4, tensor var_2059_cast_fp16_5, tensor var_2059_cast_fp16_6, tensor var_2059_cast_fp16_7, tensor var_2059_cast_fp16_8, tensor var_2059_cast_fp16_9, tensor var_2059_cast_fp16_10, tensor var_2059_cast_fp16_11, tensor var_2059_cast_fp16_12, tensor var_2059_cast_fp16_13, tensor var_2059_cast_fp16_14, tensor var_2059_cast_fp16_15, tensor var_2059_cast_fp16_16, tensor var_2059_cast_fp16_17, tensor var_2059_cast_fp16_18, tensor var_2059_cast_fp16_19 = split(axis = var_2059_axis_0, split_sizes = tile_21, x = var_2058_cast_fp16)[name = tensor("op_2059_cast_fp16")]; + tensor var_2080_perm_0 = const()[name = tensor("op_2080_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2081_axis_0 = const()[name = tensor("op_2081_axis_0"), val = tensor(3)]; + tensor var_2080_cast_fp16 = transpose(perm = var_2080_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_25")]; + tensor var_2081_cast_fp16_0, tensor var_2081_cast_fp16_1, tensor var_2081_cast_fp16_2, tensor var_2081_cast_fp16_3, tensor var_2081_cast_fp16_4, tensor var_2081_cast_fp16_5, tensor var_2081_cast_fp16_6, tensor var_2081_cast_fp16_7, tensor var_2081_cast_fp16_8, tensor var_2081_cast_fp16_9, tensor var_2081_cast_fp16_10, tensor var_2081_cast_fp16_11, tensor var_2081_cast_fp16_12, tensor var_2081_cast_fp16_13, tensor var_2081_cast_fp16_14, tensor var_2081_cast_fp16_15, tensor var_2081_cast_fp16_16, tensor var_2081_cast_fp16_17, tensor var_2081_cast_fp16_18, tensor var_2081_cast_fp16_19 = split(axis = var_2081_axis_0, split_sizes = tile_22, x = var_2080_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2102_axis_0 = const()[name = tensor("op_2102_axis_0"), val = tensor(1)]; + tensor var_2102_cast_fp16_0, tensor var_2102_cast_fp16_1, tensor var_2102_cast_fp16_2, tensor var_2102_cast_fp16_3, tensor var_2102_cast_fp16_4, tensor var_2102_cast_fp16_5, tensor var_2102_cast_fp16_6, tensor var_2102_cast_fp16_7, tensor var_2102_cast_fp16_8, tensor var_2102_cast_fp16_9, tensor var_2102_cast_fp16_10, tensor var_2102_cast_fp16_11, tensor var_2102_cast_fp16_12, tensor var_2102_cast_fp16_13, tensor var_2102_cast_fp16_14, tensor var_2102_cast_fp16_15, tensor var_2102_cast_fp16_16, tensor var_2102_cast_fp16_17, tensor var_2102_cast_fp16_18, tensor var_2102_cast_fp16_19 = split(axis = var_2102_axis_0, split_sizes = tile_23, x = var_2056_cast_fp16)[name = tensor("op_2102_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2081_cast_fp16_0, var_2059_cast_fp16_0))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2081_cast_fp16_1, var_2059_cast_fp16_1))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2081_cast_fp16_2, var_2059_cast_fp16_2))[name = tensor("aw_285_cast_fp16")]; + tensor aw_287_equation_0 = const()[name = tensor("aw_287_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_287_cast_fp16 = einsum(equation = aw_287_equation_0, values = (var_2081_cast_fp16_3, var_2059_cast_fp16_3))[name = tensor("aw_287_cast_fp16")]; + tensor aw_289_equation_0 = const()[name = tensor("aw_289_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_289_cast_fp16 = einsum(equation = aw_289_equation_0, values = (var_2081_cast_fp16_4, var_2059_cast_fp16_4))[name = tensor("aw_289_cast_fp16")]; + tensor aw_291_equation_0 = const()[name = tensor("aw_291_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_291_cast_fp16 = einsum(equation = aw_291_equation_0, values = (var_2081_cast_fp16_5, var_2059_cast_fp16_5))[name = tensor("aw_291_cast_fp16")]; + tensor aw_293_equation_0 = const()[name = tensor("aw_293_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_293_cast_fp16 = einsum(equation = aw_293_equation_0, values = (var_2081_cast_fp16_6, var_2059_cast_fp16_6))[name = tensor("aw_293_cast_fp16")]; + tensor aw_295_equation_0 = const()[name = tensor("aw_295_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_295_cast_fp16 = einsum(equation = aw_295_equation_0, values = (var_2081_cast_fp16_7, var_2059_cast_fp16_7))[name = tensor("aw_295_cast_fp16")]; + tensor aw_297_equation_0 = const()[name = tensor("aw_297_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_297_cast_fp16 = einsum(equation = aw_297_equation_0, values = (var_2081_cast_fp16_8, var_2059_cast_fp16_8))[name = tensor("aw_297_cast_fp16")]; + tensor aw_299_equation_0 = const()[name = tensor("aw_299_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_299_cast_fp16 = einsum(equation = aw_299_equation_0, values = (var_2081_cast_fp16_9, var_2059_cast_fp16_9))[name = tensor("aw_299_cast_fp16")]; + tensor aw_301_equation_0 = const()[name = tensor("aw_301_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_301_cast_fp16 = einsum(equation = aw_301_equation_0, values = (var_2081_cast_fp16_10, var_2059_cast_fp16_10))[name = tensor("aw_301_cast_fp16")]; + tensor aw_303_equation_0 = const()[name = tensor("aw_303_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_303_cast_fp16 = einsum(equation = aw_303_equation_0, values = (var_2081_cast_fp16_11, var_2059_cast_fp16_11))[name = tensor("aw_303_cast_fp16")]; + tensor aw_305_equation_0 = const()[name = tensor("aw_305_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_305_cast_fp16 = einsum(equation = aw_305_equation_0, values = (var_2081_cast_fp16_12, var_2059_cast_fp16_12))[name = tensor("aw_305_cast_fp16")]; + tensor aw_307_equation_0 = const()[name = tensor("aw_307_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_307_cast_fp16 = einsum(equation = aw_307_equation_0, values = (var_2081_cast_fp16_13, var_2059_cast_fp16_13))[name = tensor("aw_307_cast_fp16")]; + tensor aw_309_equation_0 = const()[name = tensor("aw_309_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_309_cast_fp16 = einsum(equation = aw_309_equation_0, values = (var_2081_cast_fp16_14, var_2059_cast_fp16_14))[name = tensor("aw_309_cast_fp16")]; + tensor aw_311_equation_0 = const()[name = tensor("aw_311_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_311_cast_fp16 = einsum(equation = aw_311_equation_0, values = (var_2081_cast_fp16_15, var_2059_cast_fp16_15))[name = tensor("aw_311_cast_fp16")]; + tensor aw_313_equation_0 = const()[name = tensor("aw_313_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_313_cast_fp16 = einsum(equation = aw_313_equation_0, values = (var_2081_cast_fp16_16, var_2059_cast_fp16_16))[name = tensor("aw_313_cast_fp16")]; + tensor aw_315_equation_0 = const()[name = tensor("aw_315_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_315_cast_fp16 = einsum(equation = aw_315_equation_0, values = (var_2081_cast_fp16_17, var_2059_cast_fp16_17))[name = tensor("aw_315_cast_fp16")]; + tensor aw_317_equation_0 = const()[name = tensor("aw_317_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_317_cast_fp16 = einsum(equation = aw_317_equation_0, values = (var_2081_cast_fp16_18, var_2059_cast_fp16_18))[name = tensor("aw_317_cast_fp16")]; + tensor aw_319_equation_0 = const()[name = tensor("aw_319_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_319_cast_fp16 = einsum(equation = aw_319_equation_0, values = (var_2081_cast_fp16_19, var_2059_cast_fp16_19))[name = tensor("aw_319_cast_fp16")]; + tensor var_2163_cast_fp16 = softmax(axis = var_2007, x = aw_281_cast_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = softmax(axis = var_2007, x = aw_283_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor var_2165_cast_fp16 = softmax(axis = var_2007, x = aw_285_cast_fp16)[name = tensor("op_2165_cast_fp16")]; + tensor var_2166_cast_fp16 = softmax(axis = var_2007, x = aw_287_cast_fp16)[name = tensor("op_2166_cast_fp16")]; + tensor var_2167_cast_fp16 = softmax(axis = var_2007, x = aw_289_cast_fp16)[name = tensor("op_2167_cast_fp16")]; + tensor var_2168_cast_fp16 = softmax(axis = var_2007, x = aw_291_cast_fp16)[name = tensor("op_2168_cast_fp16")]; + tensor var_2169_cast_fp16 = softmax(axis = var_2007, x = aw_293_cast_fp16)[name = tensor("op_2169_cast_fp16")]; + tensor var_2170_cast_fp16 = softmax(axis = var_2007, x = aw_295_cast_fp16)[name = tensor("op_2170_cast_fp16")]; + tensor var_2171_cast_fp16 = softmax(axis = var_2007, x = aw_297_cast_fp16)[name = tensor("op_2171_cast_fp16")]; + tensor var_2172_cast_fp16 = softmax(axis = var_2007, x = aw_299_cast_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor var_2173_cast_fp16 = softmax(axis = var_2007, x = aw_301_cast_fp16)[name = tensor("op_2173_cast_fp16")]; + tensor var_2174_cast_fp16 = softmax(axis = var_2007, x = aw_303_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor var_2175_cast_fp16 = softmax(axis = var_2007, x = aw_305_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor var_2176_cast_fp16 = softmax(axis = var_2007, x = aw_307_cast_fp16)[name = tensor("op_2176_cast_fp16")]; + tensor var_2177_cast_fp16 = softmax(axis = var_2007, x = aw_309_cast_fp16)[name = tensor("op_2177_cast_fp16")]; + tensor var_2178_cast_fp16 = softmax(axis = var_2007, x = aw_311_cast_fp16)[name = tensor("op_2178_cast_fp16")]; + tensor var_2179_cast_fp16 = softmax(axis = var_2007, x = aw_313_cast_fp16)[name = tensor("op_2179_cast_fp16")]; + tensor var_2180_cast_fp16 = softmax(axis = var_2007, x = aw_315_cast_fp16)[name = tensor("op_2180_cast_fp16")]; + tensor var_2181_cast_fp16 = softmax(axis = var_2007, x = aw_317_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2182_cast_fp16 = softmax(axis = var_2007, x = aw_319_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor var_2184_equation_0 = const()[name = tensor("op_2184_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2184_cast_fp16 = einsum(equation = var_2184_equation_0, values = (var_2102_cast_fp16_0, var_2163_cast_fp16))[name = tensor("op_2184_cast_fp16")]; + tensor var_2186_equation_0 = const()[name = tensor("op_2186_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2186_cast_fp16 = einsum(equation = var_2186_equation_0, values = (var_2102_cast_fp16_1, var_2164_cast_fp16))[name = tensor("op_2186_cast_fp16")]; + tensor var_2188_equation_0 = const()[name = tensor("op_2188_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2188_cast_fp16 = einsum(equation = var_2188_equation_0, values = (var_2102_cast_fp16_2, var_2165_cast_fp16))[name = tensor("op_2188_cast_fp16")]; + tensor var_2190_equation_0 = const()[name = tensor("op_2190_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2190_cast_fp16 = einsum(equation = var_2190_equation_0, values = (var_2102_cast_fp16_3, var_2166_cast_fp16))[name = tensor("op_2190_cast_fp16")]; + tensor var_2192_equation_0 = const()[name = tensor("op_2192_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2192_cast_fp16 = einsum(equation = var_2192_equation_0, values = (var_2102_cast_fp16_4, var_2167_cast_fp16))[name = tensor("op_2192_cast_fp16")]; + tensor var_2194_equation_0 = const()[name = tensor("op_2194_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2194_cast_fp16 = einsum(equation = var_2194_equation_0, values = (var_2102_cast_fp16_5, var_2168_cast_fp16))[name = tensor("op_2194_cast_fp16")]; + tensor var_2196_equation_0 = const()[name = tensor("op_2196_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2196_cast_fp16 = einsum(equation = var_2196_equation_0, values = (var_2102_cast_fp16_6, var_2169_cast_fp16))[name = tensor("op_2196_cast_fp16")]; + tensor var_2198_equation_0 = const()[name = tensor("op_2198_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2198_cast_fp16 = einsum(equation = var_2198_equation_0, values = (var_2102_cast_fp16_7, var_2170_cast_fp16))[name = tensor("op_2198_cast_fp16")]; + tensor var_2200_equation_0 = const()[name = tensor("op_2200_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2200_cast_fp16 = einsum(equation = var_2200_equation_0, values = (var_2102_cast_fp16_8, var_2171_cast_fp16))[name = tensor("op_2200_cast_fp16")]; + tensor var_2202_equation_0 = const()[name = tensor("op_2202_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2202_cast_fp16 = einsum(equation = var_2202_equation_0, values = (var_2102_cast_fp16_9, var_2172_cast_fp16))[name = tensor("op_2202_cast_fp16")]; + tensor var_2204_equation_0 = const()[name = tensor("op_2204_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2204_cast_fp16 = einsum(equation = var_2204_equation_0, values = (var_2102_cast_fp16_10, var_2173_cast_fp16))[name = tensor("op_2204_cast_fp16")]; + tensor var_2206_equation_0 = const()[name = tensor("op_2206_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2206_cast_fp16 = einsum(equation = var_2206_equation_0, values = (var_2102_cast_fp16_11, var_2174_cast_fp16))[name = tensor("op_2206_cast_fp16")]; + tensor var_2208_equation_0 = const()[name = tensor("op_2208_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2208_cast_fp16 = einsum(equation = var_2208_equation_0, values = (var_2102_cast_fp16_12, var_2175_cast_fp16))[name = tensor("op_2208_cast_fp16")]; + tensor var_2210_equation_0 = const()[name = tensor("op_2210_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2210_cast_fp16 = einsum(equation = var_2210_equation_0, values = (var_2102_cast_fp16_13, var_2176_cast_fp16))[name = tensor("op_2210_cast_fp16")]; + tensor var_2212_equation_0 = const()[name = tensor("op_2212_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2212_cast_fp16 = einsum(equation = var_2212_equation_0, values = (var_2102_cast_fp16_14, var_2177_cast_fp16))[name = tensor("op_2212_cast_fp16")]; + tensor var_2214_equation_0 = const()[name = tensor("op_2214_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2214_cast_fp16 = einsum(equation = var_2214_equation_0, values = (var_2102_cast_fp16_15, var_2178_cast_fp16))[name = tensor("op_2214_cast_fp16")]; + tensor var_2216_equation_0 = const()[name = tensor("op_2216_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2216_cast_fp16 = einsum(equation = var_2216_equation_0, values = (var_2102_cast_fp16_16, var_2179_cast_fp16))[name = tensor("op_2216_cast_fp16")]; + tensor var_2218_equation_0 = const()[name = tensor("op_2218_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2218_cast_fp16 = einsum(equation = var_2218_equation_0, values = (var_2102_cast_fp16_17, var_2180_cast_fp16))[name = tensor("op_2218_cast_fp16")]; + tensor var_2220_equation_0 = const()[name = tensor("op_2220_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2220_cast_fp16 = einsum(equation = var_2220_equation_0, values = (var_2102_cast_fp16_18, var_2181_cast_fp16))[name = tensor("op_2220_cast_fp16")]; + tensor var_2222_equation_0 = const()[name = tensor("op_2222_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2222_cast_fp16 = einsum(equation = var_2222_equation_0, values = (var_2102_cast_fp16_19, var_2182_cast_fp16))[name = tensor("op_2222_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_2007, interleave = input_75_interleave_0, values = (var_2184_cast_fp16, var_2186_cast_fp16, var_2188_cast_fp16, var_2190_cast_fp16, var_2192_cast_fp16, var_2194_cast_fp16, var_2196_cast_fp16, var_2198_cast_fp16, var_2200_cast_fp16, var_2202_cast_fp16, var_2204_cast_fp16, var_2206_cast_fp16, var_2208_cast_fp16, var_2210_cast_fp16, var_2212_cast_fp16, var_2214_cast_fp16, var_2216_cast_fp16, var_2218_cast_fp16, var_2220_cast_fp16, var_2222_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_2231_pad_type_0 = const()[name = tensor("op_2231_pad_type_0"), val = tensor("valid")]; + tensor var_2231_strides_0 = const()[name = tensor("op_2231_strides_0"), val = tensor([1, 1])]; + tensor var_2231_pad_0 = const()[name = tensor("op_2231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2231_dilations_0 = const()[name = tensor("op_2231_dilations_0"), val = tensor([1, 1])]; + tensor var_2231_groups_0 = const()[name = tensor("op_2231_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299604352)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302881216)))]; + tensor var_2231_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_2231_dilations_0, groups = var_2231_groups_0, pad = var_2231_pad_0, pad_type = var_2231_pad_type_0, strides = var_2231_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_2231_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302883840)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302886464)))]; + tensor var_2241_to_fp16 = const()[name = tensor("op_2241_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_2241_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302889088)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315996352)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_2267_pad_type_0 = const()[name = tensor("op_2267_pad_type_0"), val = tensor("valid")]; + tensor var_2267_strides_0 = const()[name = tensor("op_2267_strides_0"), val = tensor([1, 1])]; + tensor var_2267_pad_0 = const()[name = tensor("op_2267_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2267_dilations_0 = const()[name = tensor("op_2267_dilations_0"), val = tensor([1, 1])]; + tensor var_2267_groups_0 = const()[name = tensor("op_2267_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316006656)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329113920)))]; + tensor var_2267_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_2267_dilations_0, groups = var_2267_groups_0, pad = var_2267_pad_0, pad_type = var_2267_pad_type_0, strides = var_2267_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_2267_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_2276 = const()[name = tensor("op_2276"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329116544)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329119168)))]; + tensor var_2292_to_fp16 = const()[name = tensor("op_2292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_2292_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_2327_weight_0_to_fp16 = const()[name = tensor("op_2327_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329121792)))]; + tensor var_2327_bias_0_to_fp16 = const()[name = tensor("op_2327_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332398656)))]; + tensor var_2327_cast_fp16 = conv(bias = var_2327_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_2327_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2327_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332401280)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_2325_pad_type_0 = const()[name = tensor("op_2325_pad_type_0"), val = tensor("valid")]; + tensor var_2325_strides_0 = const()[name = tensor("op_2325_strides_0"), val = tensor([1, 1])]; + tensor var_2325_pad_0 = const()[name = tensor("op_2325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2325_dilations_0 = const()[name = tensor("op_2325_dilations_0"), val = tensor([1, 1])]; + tensor var_2325_groups_0 = const()[name = tensor("op_2325_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335678144)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338955008)))]; + tensor var_2325_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_2325_dilations_0, groups = var_2325_groups_0, pad = var_2325_pad_0, pad_type = var_2325_pad_type_0, strides = var_2325_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2328_axis_0 = const()[name = tensor("op_2328_axis_0"), val = tensor(1)]; + tensor var_2328_cast_fp16_0, tensor var_2328_cast_fp16_1, tensor var_2328_cast_fp16_2, tensor var_2328_cast_fp16_3, tensor var_2328_cast_fp16_4, tensor var_2328_cast_fp16_5, tensor var_2328_cast_fp16_6, tensor var_2328_cast_fp16_7, tensor var_2328_cast_fp16_8, tensor var_2328_cast_fp16_9, tensor var_2328_cast_fp16_10, tensor var_2328_cast_fp16_11, tensor var_2328_cast_fp16_12, tensor var_2328_cast_fp16_13, tensor var_2328_cast_fp16_14, tensor var_2328_cast_fp16_15, tensor var_2328_cast_fp16_16, tensor var_2328_cast_fp16_17, tensor var_2328_cast_fp16_18, tensor var_2328_cast_fp16_19 = split(axis = var_2328_axis_0, split_sizes = tile_24, x = var_2327_cast_fp16)[name = tensor("op_2328_cast_fp16")]; + tensor var_2349_perm_0 = const()[name = tensor("op_2349_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2350_axis_0 = const()[name = tensor("op_2350_axis_0"), val = tensor(3)]; + tensor var_2349_cast_fp16 = transpose(perm = var_2349_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_24")]; + tensor var_2350_cast_fp16_0, tensor var_2350_cast_fp16_1, tensor var_2350_cast_fp16_2, tensor var_2350_cast_fp16_3, tensor var_2350_cast_fp16_4, tensor var_2350_cast_fp16_5, tensor var_2350_cast_fp16_6, tensor var_2350_cast_fp16_7, tensor var_2350_cast_fp16_8, tensor var_2350_cast_fp16_9, tensor var_2350_cast_fp16_10, tensor var_2350_cast_fp16_11, tensor var_2350_cast_fp16_12, tensor var_2350_cast_fp16_13, tensor var_2350_cast_fp16_14, tensor var_2350_cast_fp16_15, tensor var_2350_cast_fp16_16, tensor var_2350_cast_fp16_17, tensor var_2350_cast_fp16_18, tensor var_2350_cast_fp16_19 = split(axis = var_2350_axis_0, split_sizes = tile_25, x = var_2349_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2371_axis_0 = const()[name = tensor("op_2371_axis_0"), val = tensor(1)]; + tensor var_2371_cast_fp16_0, tensor var_2371_cast_fp16_1, tensor var_2371_cast_fp16_2, tensor var_2371_cast_fp16_3, tensor var_2371_cast_fp16_4, tensor var_2371_cast_fp16_5, tensor var_2371_cast_fp16_6, tensor var_2371_cast_fp16_7, tensor var_2371_cast_fp16_8, tensor var_2371_cast_fp16_9, tensor var_2371_cast_fp16_10, tensor var_2371_cast_fp16_11, tensor var_2371_cast_fp16_12, tensor var_2371_cast_fp16_13, tensor var_2371_cast_fp16_14, tensor var_2371_cast_fp16_15, tensor var_2371_cast_fp16_16, tensor var_2371_cast_fp16_17, tensor var_2371_cast_fp16_18, tensor var_2371_cast_fp16_19 = split(axis = var_2371_axis_0, split_sizes = tile_26, x = var_2325_cast_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor aw_321_equation_0 = const()[name = tensor("aw_321_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_321_cast_fp16 = einsum(equation = aw_321_equation_0, values = (var_2350_cast_fp16_0, var_2328_cast_fp16_0))[name = tensor("aw_321_cast_fp16")]; + tensor aw_323_equation_0 = const()[name = tensor("aw_323_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_323_cast_fp16 = einsum(equation = aw_323_equation_0, values = (var_2350_cast_fp16_1, var_2328_cast_fp16_1))[name = tensor("aw_323_cast_fp16")]; + tensor aw_325_equation_0 = const()[name = tensor("aw_325_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_325_cast_fp16 = einsum(equation = aw_325_equation_0, values = (var_2350_cast_fp16_2, var_2328_cast_fp16_2))[name = tensor("aw_325_cast_fp16")]; + tensor aw_327_equation_0 = const()[name = tensor("aw_327_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_327_cast_fp16 = einsum(equation = aw_327_equation_0, values = (var_2350_cast_fp16_3, var_2328_cast_fp16_3))[name = tensor("aw_327_cast_fp16")]; + tensor aw_329_equation_0 = const()[name = tensor("aw_329_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_329_cast_fp16 = einsum(equation = aw_329_equation_0, values = (var_2350_cast_fp16_4, var_2328_cast_fp16_4))[name = tensor("aw_329_cast_fp16")]; + tensor aw_331_equation_0 = const()[name = tensor("aw_331_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_331_cast_fp16 = einsum(equation = aw_331_equation_0, values = (var_2350_cast_fp16_5, var_2328_cast_fp16_5))[name = tensor("aw_331_cast_fp16")]; + tensor aw_333_equation_0 = const()[name = tensor("aw_333_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_333_cast_fp16 = einsum(equation = aw_333_equation_0, values = (var_2350_cast_fp16_6, var_2328_cast_fp16_6))[name = tensor("aw_333_cast_fp16")]; + tensor aw_335_equation_0 = const()[name = tensor("aw_335_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_335_cast_fp16 = einsum(equation = aw_335_equation_0, values = (var_2350_cast_fp16_7, var_2328_cast_fp16_7))[name = tensor("aw_335_cast_fp16")]; + tensor aw_337_equation_0 = const()[name = tensor("aw_337_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_337_cast_fp16 = einsum(equation = aw_337_equation_0, values = (var_2350_cast_fp16_8, var_2328_cast_fp16_8))[name = tensor("aw_337_cast_fp16")]; + tensor aw_339_equation_0 = const()[name = tensor("aw_339_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_339_cast_fp16 = einsum(equation = aw_339_equation_0, values = (var_2350_cast_fp16_9, var_2328_cast_fp16_9))[name = tensor("aw_339_cast_fp16")]; + tensor aw_341_equation_0 = const()[name = tensor("aw_341_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_341_cast_fp16 = einsum(equation = aw_341_equation_0, values = (var_2350_cast_fp16_10, var_2328_cast_fp16_10))[name = tensor("aw_341_cast_fp16")]; + tensor aw_343_equation_0 = const()[name = tensor("aw_343_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_343_cast_fp16 = einsum(equation = aw_343_equation_0, values = (var_2350_cast_fp16_11, var_2328_cast_fp16_11))[name = tensor("aw_343_cast_fp16")]; + tensor aw_345_equation_0 = const()[name = tensor("aw_345_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_345_cast_fp16 = einsum(equation = aw_345_equation_0, values = (var_2350_cast_fp16_12, var_2328_cast_fp16_12))[name = tensor("aw_345_cast_fp16")]; + tensor aw_347_equation_0 = const()[name = tensor("aw_347_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_347_cast_fp16 = einsum(equation = aw_347_equation_0, values = (var_2350_cast_fp16_13, var_2328_cast_fp16_13))[name = tensor("aw_347_cast_fp16")]; + tensor aw_349_equation_0 = const()[name = tensor("aw_349_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_349_cast_fp16 = einsum(equation = aw_349_equation_0, values = (var_2350_cast_fp16_14, var_2328_cast_fp16_14))[name = tensor("aw_349_cast_fp16")]; + tensor aw_351_equation_0 = const()[name = tensor("aw_351_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_351_cast_fp16 = einsum(equation = aw_351_equation_0, values = (var_2350_cast_fp16_15, var_2328_cast_fp16_15))[name = tensor("aw_351_cast_fp16")]; + tensor aw_353_equation_0 = const()[name = tensor("aw_353_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_353_cast_fp16 = einsum(equation = aw_353_equation_0, values = (var_2350_cast_fp16_16, var_2328_cast_fp16_16))[name = tensor("aw_353_cast_fp16")]; + tensor aw_355_equation_0 = const()[name = tensor("aw_355_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_355_cast_fp16 = einsum(equation = aw_355_equation_0, values = (var_2350_cast_fp16_17, var_2328_cast_fp16_17))[name = tensor("aw_355_cast_fp16")]; + tensor aw_357_equation_0 = const()[name = tensor("aw_357_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_357_cast_fp16 = einsum(equation = aw_357_equation_0, values = (var_2350_cast_fp16_18, var_2328_cast_fp16_18))[name = tensor("aw_357_cast_fp16")]; + tensor aw_359_equation_0 = const()[name = tensor("aw_359_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_359_cast_fp16 = einsum(equation = aw_359_equation_0, values = (var_2350_cast_fp16_19, var_2328_cast_fp16_19))[name = tensor("aw_359_cast_fp16")]; + tensor var_2432_cast_fp16 = softmax(axis = var_2276, x = aw_321_cast_fp16)[name = tensor("op_2432_cast_fp16")]; + tensor var_2433_cast_fp16 = softmax(axis = var_2276, x = aw_323_cast_fp16)[name = tensor("op_2433_cast_fp16")]; + tensor var_2434_cast_fp16 = softmax(axis = var_2276, x = aw_325_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = softmax(axis = var_2276, x = aw_327_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = softmax(axis = var_2276, x = aw_329_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2437_cast_fp16 = softmax(axis = var_2276, x = aw_331_cast_fp16)[name = tensor("op_2437_cast_fp16")]; + tensor var_2438_cast_fp16 = softmax(axis = var_2276, x = aw_333_cast_fp16)[name = tensor("op_2438_cast_fp16")]; + tensor var_2439_cast_fp16 = softmax(axis = var_2276, x = aw_335_cast_fp16)[name = tensor("op_2439_cast_fp16")]; + tensor var_2440_cast_fp16 = softmax(axis = var_2276, x = aw_337_cast_fp16)[name = tensor("op_2440_cast_fp16")]; + tensor var_2441_cast_fp16 = softmax(axis = var_2276, x = aw_339_cast_fp16)[name = tensor("op_2441_cast_fp16")]; + tensor var_2442_cast_fp16 = softmax(axis = var_2276, x = aw_341_cast_fp16)[name = tensor("op_2442_cast_fp16")]; + tensor var_2443_cast_fp16 = softmax(axis = var_2276, x = aw_343_cast_fp16)[name = tensor("op_2443_cast_fp16")]; + tensor var_2444_cast_fp16 = softmax(axis = var_2276, x = aw_345_cast_fp16)[name = tensor("op_2444_cast_fp16")]; + tensor var_2445_cast_fp16 = softmax(axis = var_2276, x = aw_347_cast_fp16)[name = tensor("op_2445_cast_fp16")]; + tensor var_2446_cast_fp16 = softmax(axis = var_2276, x = aw_349_cast_fp16)[name = tensor("op_2446_cast_fp16")]; + tensor var_2447_cast_fp16 = softmax(axis = var_2276, x = aw_351_cast_fp16)[name = tensor("op_2447_cast_fp16")]; + tensor var_2448_cast_fp16 = softmax(axis = var_2276, x = aw_353_cast_fp16)[name = tensor("op_2448_cast_fp16")]; + tensor var_2449_cast_fp16 = softmax(axis = var_2276, x = aw_355_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor var_2450_cast_fp16 = softmax(axis = var_2276, x = aw_357_cast_fp16)[name = tensor("op_2450_cast_fp16")]; + tensor var_2451_cast_fp16 = softmax(axis = var_2276, x = aw_359_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2453_equation_0 = const()[name = tensor("op_2453_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2453_cast_fp16 = einsum(equation = var_2453_equation_0, values = (var_2371_cast_fp16_0, var_2432_cast_fp16))[name = tensor("op_2453_cast_fp16")]; + tensor var_2455_equation_0 = const()[name = tensor("op_2455_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2455_cast_fp16 = einsum(equation = var_2455_equation_0, values = (var_2371_cast_fp16_1, var_2433_cast_fp16))[name = tensor("op_2455_cast_fp16")]; + tensor var_2457_equation_0 = const()[name = tensor("op_2457_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2457_cast_fp16 = einsum(equation = var_2457_equation_0, values = (var_2371_cast_fp16_2, var_2434_cast_fp16))[name = tensor("op_2457_cast_fp16")]; + tensor var_2459_equation_0 = const()[name = tensor("op_2459_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2459_cast_fp16 = einsum(equation = var_2459_equation_0, values = (var_2371_cast_fp16_3, var_2435_cast_fp16))[name = tensor("op_2459_cast_fp16")]; + tensor var_2461_equation_0 = const()[name = tensor("op_2461_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2461_cast_fp16 = einsum(equation = var_2461_equation_0, values = (var_2371_cast_fp16_4, var_2436_cast_fp16))[name = tensor("op_2461_cast_fp16")]; + tensor var_2463_equation_0 = const()[name = tensor("op_2463_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2463_cast_fp16 = einsum(equation = var_2463_equation_0, values = (var_2371_cast_fp16_5, var_2437_cast_fp16))[name = tensor("op_2463_cast_fp16")]; + tensor var_2465_equation_0 = const()[name = tensor("op_2465_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2465_cast_fp16 = einsum(equation = var_2465_equation_0, values = (var_2371_cast_fp16_6, var_2438_cast_fp16))[name = tensor("op_2465_cast_fp16")]; + tensor var_2467_equation_0 = const()[name = tensor("op_2467_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2467_cast_fp16 = einsum(equation = var_2467_equation_0, values = (var_2371_cast_fp16_7, var_2439_cast_fp16))[name = tensor("op_2467_cast_fp16")]; + tensor var_2469_equation_0 = const()[name = tensor("op_2469_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2469_cast_fp16 = einsum(equation = var_2469_equation_0, values = (var_2371_cast_fp16_8, var_2440_cast_fp16))[name = tensor("op_2469_cast_fp16")]; + tensor var_2471_equation_0 = const()[name = tensor("op_2471_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2471_cast_fp16 = einsum(equation = var_2471_equation_0, values = (var_2371_cast_fp16_9, var_2441_cast_fp16))[name = tensor("op_2471_cast_fp16")]; + tensor var_2473_equation_0 = const()[name = tensor("op_2473_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2473_cast_fp16 = einsum(equation = var_2473_equation_0, values = (var_2371_cast_fp16_10, var_2442_cast_fp16))[name = tensor("op_2473_cast_fp16")]; + tensor var_2475_equation_0 = const()[name = tensor("op_2475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2475_cast_fp16 = einsum(equation = var_2475_equation_0, values = (var_2371_cast_fp16_11, var_2443_cast_fp16))[name = tensor("op_2475_cast_fp16")]; + tensor var_2477_equation_0 = const()[name = tensor("op_2477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2477_cast_fp16 = einsum(equation = var_2477_equation_0, values = (var_2371_cast_fp16_12, var_2444_cast_fp16))[name = tensor("op_2477_cast_fp16")]; + tensor var_2479_equation_0 = const()[name = tensor("op_2479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2479_cast_fp16 = einsum(equation = var_2479_equation_0, values = (var_2371_cast_fp16_13, var_2445_cast_fp16))[name = tensor("op_2479_cast_fp16")]; + tensor var_2481_equation_0 = const()[name = tensor("op_2481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2481_cast_fp16 = einsum(equation = var_2481_equation_0, values = (var_2371_cast_fp16_14, var_2446_cast_fp16))[name = tensor("op_2481_cast_fp16")]; + tensor var_2483_equation_0 = const()[name = tensor("op_2483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2483_cast_fp16 = einsum(equation = var_2483_equation_0, values = (var_2371_cast_fp16_15, var_2447_cast_fp16))[name = tensor("op_2483_cast_fp16")]; + tensor var_2485_equation_0 = const()[name = tensor("op_2485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2485_cast_fp16 = einsum(equation = var_2485_equation_0, values = (var_2371_cast_fp16_16, var_2448_cast_fp16))[name = tensor("op_2485_cast_fp16")]; + tensor var_2487_equation_0 = const()[name = tensor("op_2487_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2487_cast_fp16 = einsum(equation = var_2487_equation_0, values = (var_2371_cast_fp16_17, var_2449_cast_fp16))[name = tensor("op_2487_cast_fp16")]; + tensor var_2489_equation_0 = const()[name = tensor("op_2489_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2489_cast_fp16 = einsum(equation = var_2489_equation_0, values = (var_2371_cast_fp16_18, var_2450_cast_fp16))[name = tensor("op_2489_cast_fp16")]; + tensor var_2491_equation_0 = const()[name = tensor("op_2491_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2491_cast_fp16 = einsum(equation = var_2491_equation_0, values = (var_2371_cast_fp16_19, var_2451_cast_fp16))[name = tensor("op_2491_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_2276, interleave = input_85_interleave_0, values = (var_2453_cast_fp16, var_2455_cast_fp16, var_2457_cast_fp16, var_2459_cast_fp16, var_2461_cast_fp16, var_2463_cast_fp16, var_2465_cast_fp16, var_2467_cast_fp16, var_2469_cast_fp16, var_2471_cast_fp16, var_2473_cast_fp16, var_2475_cast_fp16, var_2477_cast_fp16, var_2479_cast_fp16, var_2481_cast_fp16, var_2483_cast_fp16, var_2485_cast_fp16, var_2487_cast_fp16, var_2489_cast_fp16, var_2491_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_2500_pad_type_0 = const()[name = tensor("op_2500_pad_type_0"), val = tensor("valid")]; + tensor var_2500_strides_0 = const()[name = tensor("op_2500_strides_0"), val = tensor([1, 1])]; + tensor var_2500_pad_0 = const()[name = tensor("op_2500_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2500_dilations_0 = const()[name = tensor("op_2500_dilations_0"), val = tensor([1, 1])]; + tensor var_2500_groups_0 = const()[name = tensor("op_2500_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338957632)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342234496)))]; + tensor var_2500_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_2500_dilations_0, groups = var_2500_groups_0, pad = var_2500_pad_0, pad_type = var_2500_pad_type_0, strides = var_2500_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_2500_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342237120)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342239744)))]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_2510_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342242368)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355349632)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2536_pad_type_0 = const()[name = tensor("op_2536_pad_type_0"), val = tensor("valid")]; + tensor var_2536_strides_0 = const()[name = tensor("op_2536_strides_0"), val = tensor([1, 1])]; + tensor var_2536_pad_0 = const()[name = tensor("op_2536_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2536_dilations_0 = const()[name = tensor("op_2536_dilations_0"), val = tensor([1, 1])]; + tensor var_2536_groups_0 = const()[name = tensor("op_2536_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355359936)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368467200)))]; + tensor var_2536_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_2536_dilations_0, groups = var_2536_groups_0, pad = var_2536_pad_0, pad_type = var_2536_pad_type_0, strides = var_2536_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_2536_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2545 = const()[name = tensor("op_2545"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368469824)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368472448)))]; + tensor var_2561_to_fp16 = const()[name = tensor("op_2561_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_2561_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_2596_weight_0_to_fp16 = const()[name = tensor("op_2596_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368475072)))]; + tensor var_2596_bias_0_to_fp16 = const()[name = tensor("op_2596_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371751936)))]; + tensor var_2596_cast_fp16 = conv(bias = var_2596_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_2596_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371754560)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_2594_pad_type_0 = const()[name = tensor("op_2594_pad_type_0"), val = tensor("valid")]; + tensor var_2594_strides_0 = const()[name = tensor("op_2594_strides_0"), val = tensor([1, 1])]; + tensor var_2594_pad_0 = const()[name = tensor("op_2594_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2594_dilations_0 = const()[name = tensor("op_2594_dilations_0"), val = tensor([1, 1])]; + tensor var_2594_groups_0 = const()[name = tensor("op_2594_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375031424)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378308288)))]; + tensor var_2594_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_2594_dilations_0, groups = var_2594_groups_0, pad = var_2594_pad_0, pad_type = var_2594_pad_type_0, strides = var_2594_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2594_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2597_axis_0 = const()[name = tensor("op_2597_axis_0"), val = tensor(1)]; + tensor var_2597_cast_fp16_0, tensor var_2597_cast_fp16_1, tensor var_2597_cast_fp16_2, tensor var_2597_cast_fp16_3, tensor var_2597_cast_fp16_4, tensor var_2597_cast_fp16_5, tensor var_2597_cast_fp16_6, tensor var_2597_cast_fp16_7, tensor var_2597_cast_fp16_8, tensor var_2597_cast_fp16_9, tensor var_2597_cast_fp16_10, tensor var_2597_cast_fp16_11, tensor var_2597_cast_fp16_12, tensor var_2597_cast_fp16_13, tensor var_2597_cast_fp16_14, tensor var_2597_cast_fp16_15, tensor var_2597_cast_fp16_16, tensor var_2597_cast_fp16_17, tensor var_2597_cast_fp16_18, tensor var_2597_cast_fp16_19 = split(axis = var_2597_axis_0, split_sizes = tile_27, x = var_2596_cast_fp16)[name = tensor("op_2597_cast_fp16")]; + tensor var_2618_perm_0 = const()[name = tensor("op_2618_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2619_axis_0 = const()[name = tensor("op_2619_axis_0"), val = tensor(3)]; + tensor var_2618_cast_fp16 = transpose(perm = var_2618_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_23")]; + tensor var_2619_cast_fp16_0, tensor var_2619_cast_fp16_1, tensor var_2619_cast_fp16_2, tensor var_2619_cast_fp16_3, tensor var_2619_cast_fp16_4, tensor var_2619_cast_fp16_5, tensor var_2619_cast_fp16_6, tensor var_2619_cast_fp16_7, tensor var_2619_cast_fp16_8, tensor var_2619_cast_fp16_9, tensor var_2619_cast_fp16_10, tensor var_2619_cast_fp16_11, tensor var_2619_cast_fp16_12, tensor var_2619_cast_fp16_13, tensor var_2619_cast_fp16_14, tensor var_2619_cast_fp16_15, tensor var_2619_cast_fp16_16, tensor var_2619_cast_fp16_17, tensor var_2619_cast_fp16_18, tensor var_2619_cast_fp16_19 = split(axis = var_2619_axis_0, split_sizes = tile_28, x = var_2618_cast_fp16)[name = tensor("op_2619_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2640_axis_0 = const()[name = tensor("op_2640_axis_0"), val = tensor(1)]; + tensor var_2640_cast_fp16_0, tensor var_2640_cast_fp16_1, tensor var_2640_cast_fp16_2, tensor var_2640_cast_fp16_3, tensor var_2640_cast_fp16_4, tensor var_2640_cast_fp16_5, tensor var_2640_cast_fp16_6, tensor var_2640_cast_fp16_7, tensor var_2640_cast_fp16_8, tensor var_2640_cast_fp16_9, tensor var_2640_cast_fp16_10, tensor var_2640_cast_fp16_11, tensor var_2640_cast_fp16_12, tensor var_2640_cast_fp16_13, tensor var_2640_cast_fp16_14, tensor var_2640_cast_fp16_15, tensor var_2640_cast_fp16_16, tensor var_2640_cast_fp16_17, tensor var_2640_cast_fp16_18, tensor var_2640_cast_fp16_19 = split(axis = var_2640_axis_0, split_sizes = tile_29, x = var_2594_cast_fp16)[name = tensor("op_2640_cast_fp16")]; + tensor aw_361_equation_0 = const()[name = tensor("aw_361_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_361_cast_fp16 = einsum(equation = aw_361_equation_0, values = (var_2619_cast_fp16_0, var_2597_cast_fp16_0))[name = tensor("aw_361_cast_fp16")]; + tensor aw_363_equation_0 = const()[name = tensor("aw_363_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_363_cast_fp16 = einsum(equation = aw_363_equation_0, values = (var_2619_cast_fp16_1, var_2597_cast_fp16_1))[name = tensor("aw_363_cast_fp16")]; + tensor aw_365_equation_0 = const()[name = tensor("aw_365_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_365_cast_fp16 = einsum(equation = aw_365_equation_0, values = (var_2619_cast_fp16_2, var_2597_cast_fp16_2))[name = tensor("aw_365_cast_fp16")]; + tensor aw_367_equation_0 = const()[name = tensor("aw_367_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_367_cast_fp16 = einsum(equation = aw_367_equation_0, values = (var_2619_cast_fp16_3, var_2597_cast_fp16_3))[name = tensor("aw_367_cast_fp16")]; + tensor aw_369_equation_0 = const()[name = tensor("aw_369_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_369_cast_fp16 = einsum(equation = aw_369_equation_0, values = (var_2619_cast_fp16_4, var_2597_cast_fp16_4))[name = tensor("aw_369_cast_fp16")]; + tensor aw_371_equation_0 = const()[name = tensor("aw_371_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_371_cast_fp16 = einsum(equation = aw_371_equation_0, values = (var_2619_cast_fp16_5, var_2597_cast_fp16_5))[name = tensor("aw_371_cast_fp16")]; + tensor aw_373_equation_0 = const()[name = tensor("aw_373_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_373_cast_fp16 = einsum(equation = aw_373_equation_0, values = (var_2619_cast_fp16_6, var_2597_cast_fp16_6))[name = tensor("aw_373_cast_fp16")]; + tensor aw_375_equation_0 = const()[name = tensor("aw_375_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_375_cast_fp16 = einsum(equation = aw_375_equation_0, values = (var_2619_cast_fp16_7, var_2597_cast_fp16_7))[name = tensor("aw_375_cast_fp16")]; + tensor aw_377_equation_0 = const()[name = tensor("aw_377_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_377_cast_fp16 = einsum(equation = aw_377_equation_0, values = (var_2619_cast_fp16_8, var_2597_cast_fp16_8))[name = tensor("aw_377_cast_fp16")]; + tensor aw_379_equation_0 = const()[name = tensor("aw_379_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_379_cast_fp16 = einsum(equation = aw_379_equation_0, values = (var_2619_cast_fp16_9, var_2597_cast_fp16_9))[name = tensor("aw_379_cast_fp16")]; + tensor aw_381_equation_0 = const()[name = tensor("aw_381_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_381_cast_fp16 = einsum(equation = aw_381_equation_0, values = (var_2619_cast_fp16_10, var_2597_cast_fp16_10))[name = tensor("aw_381_cast_fp16")]; + tensor aw_383_equation_0 = const()[name = tensor("aw_383_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_383_cast_fp16 = einsum(equation = aw_383_equation_0, values = (var_2619_cast_fp16_11, var_2597_cast_fp16_11))[name = tensor("aw_383_cast_fp16")]; + tensor aw_385_equation_0 = const()[name = tensor("aw_385_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_385_cast_fp16 = einsum(equation = aw_385_equation_0, values = (var_2619_cast_fp16_12, var_2597_cast_fp16_12))[name = tensor("aw_385_cast_fp16")]; + tensor aw_387_equation_0 = const()[name = tensor("aw_387_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_387_cast_fp16 = einsum(equation = aw_387_equation_0, values = (var_2619_cast_fp16_13, var_2597_cast_fp16_13))[name = tensor("aw_387_cast_fp16")]; + tensor aw_389_equation_0 = const()[name = tensor("aw_389_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_389_cast_fp16 = einsum(equation = aw_389_equation_0, values = (var_2619_cast_fp16_14, var_2597_cast_fp16_14))[name = tensor("aw_389_cast_fp16")]; + tensor aw_391_equation_0 = const()[name = tensor("aw_391_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_391_cast_fp16 = einsum(equation = aw_391_equation_0, values = (var_2619_cast_fp16_15, var_2597_cast_fp16_15))[name = tensor("aw_391_cast_fp16")]; + tensor aw_393_equation_0 = const()[name = tensor("aw_393_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_393_cast_fp16 = einsum(equation = aw_393_equation_0, values = (var_2619_cast_fp16_16, var_2597_cast_fp16_16))[name = tensor("aw_393_cast_fp16")]; + tensor aw_395_equation_0 = const()[name = tensor("aw_395_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_395_cast_fp16 = einsum(equation = aw_395_equation_0, values = (var_2619_cast_fp16_17, var_2597_cast_fp16_17))[name = tensor("aw_395_cast_fp16")]; + tensor aw_397_equation_0 = const()[name = tensor("aw_397_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_397_cast_fp16 = einsum(equation = aw_397_equation_0, values = (var_2619_cast_fp16_18, var_2597_cast_fp16_18))[name = tensor("aw_397_cast_fp16")]; + tensor aw_399_equation_0 = const()[name = tensor("aw_399_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_399_cast_fp16 = einsum(equation = aw_399_equation_0, values = (var_2619_cast_fp16_19, var_2597_cast_fp16_19))[name = tensor("aw_399_cast_fp16")]; + tensor var_2701_cast_fp16 = softmax(axis = var_2545, x = aw_361_cast_fp16)[name = tensor("op_2701_cast_fp16")]; + tensor var_2702_cast_fp16 = softmax(axis = var_2545, x = aw_363_cast_fp16)[name = tensor("op_2702_cast_fp16")]; + tensor var_2703_cast_fp16 = softmax(axis = var_2545, x = aw_365_cast_fp16)[name = tensor("op_2703_cast_fp16")]; + tensor var_2704_cast_fp16 = softmax(axis = var_2545, x = aw_367_cast_fp16)[name = tensor("op_2704_cast_fp16")]; + tensor var_2705_cast_fp16 = softmax(axis = var_2545, x = aw_369_cast_fp16)[name = tensor("op_2705_cast_fp16")]; + tensor var_2706_cast_fp16 = softmax(axis = var_2545, x = aw_371_cast_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = softmax(axis = var_2545, x = aw_373_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor var_2708_cast_fp16 = softmax(axis = var_2545, x = aw_375_cast_fp16)[name = tensor("op_2708_cast_fp16")]; + tensor var_2709_cast_fp16 = softmax(axis = var_2545, x = aw_377_cast_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor var_2710_cast_fp16 = softmax(axis = var_2545, x = aw_379_cast_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor var_2711_cast_fp16 = softmax(axis = var_2545, x = aw_381_cast_fp16)[name = tensor("op_2711_cast_fp16")]; + tensor var_2712_cast_fp16 = softmax(axis = var_2545, x = aw_383_cast_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = softmax(axis = var_2545, x = aw_385_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor var_2714_cast_fp16 = softmax(axis = var_2545, x = aw_387_cast_fp16)[name = tensor("op_2714_cast_fp16")]; + tensor var_2715_cast_fp16 = softmax(axis = var_2545, x = aw_389_cast_fp16)[name = tensor("op_2715_cast_fp16")]; + tensor var_2716_cast_fp16 = softmax(axis = var_2545, x = aw_391_cast_fp16)[name = tensor("op_2716_cast_fp16")]; + tensor var_2717_cast_fp16 = softmax(axis = var_2545, x = aw_393_cast_fp16)[name = tensor("op_2717_cast_fp16")]; + tensor var_2718_cast_fp16 = softmax(axis = var_2545, x = aw_395_cast_fp16)[name = tensor("op_2718_cast_fp16")]; + tensor var_2719_cast_fp16 = softmax(axis = var_2545, x = aw_397_cast_fp16)[name = tensor("op_2719_cast_fp16")]; + tensor var_2720_cast_fp16 = softmax(axis = var_2545, x = aw_399_cast_fp16)[name = tensor("op_2720_cast_fp16")]; + tensor var_2722_equation_0 = const()[name = tensor("op_2722_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2722_cast_fp16 = einsum(equation = var_2722_equation_0, values = (var_2640_cast_fp16_0, var_2701_cast_fp16))[name = tensor("op_2722_cast_fp16")]; + tensor var_2724_equation_0 = const()[name = tensor("op_2724_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2724_cast_fp16 = einsum(equation = var_2724_equation_0, values = (var_2640_cast_fp16_1, var_2702_cast_fp16))[name = tensor("op_2724_cast_fp16")]; + tensor var_2726_equation_0 = const()[name = tensor("op_2726_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2726_cast_fp16 = einsum(equation = var_2726_equation_0, values = (var_2640_cast_fp16_2, var_2703_cast_fp16))[name = tensor("op_2726_cast_fp16")]; + tensor var_2728_equation_0 = const()[name = tensor("op_2728_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2728_cast_fp16 = einsum(equation = var_2728_equation_0, values = (var_2640_cast_fp16_3, var_2704_cast_fp16))[name = tensor("op_2728_cast_fp16")]; + tensor var_2730_equation_0 = const()[name = tensor("op_2730_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2730_cast_fp16 = einsum(equation = var_2730_equation_0, values = (var_2640_cast_fp16_4, var_2705_cast_fp16))[name = tensor("op_2730_cast_fp16")]; + tensor var_2732_equation_0 = const()[name = tensor("op_2732_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2732_cast_fp16 = einsum(equation = var_2732_equation_0, values = (var_2640_cast_fp16_5, var_2706_cast_fp16))[name = tensor("op_2732_cast_fp16")]; + tensor var_2734_equation_0 = const()[name = tensor("op_2734_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2734_cast_fp16 = einsum(equation = var_2734_equation_0, values = (var_2640_cast_fp16_6, var_2707_cast_fp16))[name = tensor("op_2734_cast_fp16")]; + tensor var_2736_equation_0 = const()[name = tensor("op_2736_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2736_cast_fp16 = einsum(equation = var_2736_equation_0, values = (var_2640_cast_fp16_7, var_2708_cast_fp16))[name = tensor("op_2736_cast_fp16")]; + tensor var_2738_equation_0 = const()[name = tensor("op_2738_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2738_cast_fp16 = einsum(equation = var_2738_equation_0, values = (var_2640_cast_fp16_8, var_2709_cast_fp16))[name = tensor("op_2738_cast_fp16")]; + tensor var_2740_equation_0 = const()[name = tensor("op_2740_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2740_cast_fp16 = einsum(equation = var_2740_equation_0, values = (var_2640_cast_fp16_9, var_2710_cast_fp16))[name = tensor("op_2740_cast_fp16")]; + tensor var_2742_equation_0 = const()[name = tensor("op_2742_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2742_cast_fp16 = einsum(equation = var_2742_equation_0, values = (var_2640_cast_fp16_10, var_2711_cast_fp16))[name = tensor("op_2742_cast_fp16")]; + tensor var_2744_equation_0 = const()[name = tensor("op_2744_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2744_cast_fp16 = einsum(equation = var_2744_equation_0, values = (var_2640_cast_fp16_11, var_2712_cast_fp16))[name = tensor("op_2744_cast_fp16")]; + tensor var_2746_equation_0 = const()[name = tensor("op_2746_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2746_cast_fp16 = einsum(equation = var_2746_equation_0, values = (var_2640_cast_fp16_12, var_2713_cast_fp16))[name = tensor("op_2746_cast_fp16")]; + tensor var_2748_equation_0 = const()[name = tensor("op_2748_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2748_cast_fp16 = einsum(equation = var_2748_equation_0, values = (var_2640_cast_fp16_13, var_2714_cast_fp16))[name = tensor("op_2748_cast_fp16")]; + tensor var_2750_equation_0 = const()[name = tensor("op_2750_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2750_cast_fp16 = einsum(equation = var_2750_equation_0, values = (var_2640_cast_fp16_14, var_2715_cast_fp16))[name = tensor("op_2750_cast_fp16")]; + tensor var_2752_equation_0 = const()[name = tensor("op_2752_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2752_cast_fp16 = einsum(equation = var_2752_equation_0, values = (var_2640_cast_fp16_15, var_2716_cast_fp16))[name = tensor("op_2752_cast_fp16")]; + tensor var_2754_equation_0 = const()[name = tensor("op_2754_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2754_cast_fp16 = einsum(equation = var_2754_equation_0, values = (var_2640_cast_fp16_16, var_2717_cast_fp16))[name = tensor("op_2754_cast_fp16")]; + tensor var_2756_equation_0 = const()[name = tensor("op_2756_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2756_cast_fp16 = einsum(equation = var_2756_equation_0, values = (var_2640_cast_fp16_17, var_2718_cast_fp16))[name = tensor("op_2756_cast_fp16")]; + tensor var_2758_equation_0 = const()[name = tensor("op_2758_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2758_cast_fp16 = einsum(equation = var_2758_equation_0, values = (var_2640_cast_fp16_18, var_2719_cast_fp16))[name = tensor("op_2758_cast_fp16")]; + tensor var_2760_equation_0 = const()[name = tensor("op_2760_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2760_cast_fp16 = einsum(equation = var_2760_equation_0, values = (var_2640_cast_fp16_19, var_2720_cast_fp16))[name = tensor("op_2760_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_2545, interleave = input_95_interleave_0, values = (var_2722_cast_fp16, var_2724_cast_fp16, var_2726_cast_fp16, var_2728_cast_fp16, var_2730_cast_fp16, var_2732_cast_fp16, var_2734_cast_fp16, var_2736_cast_fp16, var_2738_cast_fp16, var_2740_cast_fp16, var_2742_cast_fp16, var_2744_cast_fp16, var_2746_cast_fp16, var_2748_cast_fp16, var_2750_cast_fp16, var_2752_cast_fp16, var_2754_cast_fp16, var_2756_cast_fp16, var_2758_cast_fp16, var_2760_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2769_pad_type_0 = const()[name = tensor("op_2769_pad_type_0"), val = tensor("valid")]; + tensor var_2769_strides_0 = const()[name = tensor("op_2769_strides_0"), val = tensor([1, 1])]; + tensor var_2769_pad_0 = const()[name = tensor("op_2769_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2769_dilations_0 = const()[name = tensor("op_2769_dilations_0"), val = tensor([1, 1])]; + tensor var_2769_groups_0 = const()[name = tensor("op_2769_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378310912)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381587776)))]; + tensor var_2769_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2769_dilations_0, groups = var_2769_groups_0, pad = var_2769_pad_0, pad_type = var_2769_pad_type_0, strides = var_2769_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2769_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2769_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381590400)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381593024)))]; + tensor var_2779_to_fp16 = const()[name = tensor("op_2779_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2779_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381595648)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394702912)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2805_pad_type_0 = const()[name = tensor("op_2805_pad_type_0"), val = tensor("valid")]; + tensor var_2805_strides_0 = const()[name = tensor("op_2805_strides_0"), val = tensor([1, 1])]; + tensor var_2805_pad_0 = const()[name = tensor("op_2805_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2805_dilations_0 = const()[name = tensor("op_2805_dilations_0"), val = tensor([1, 1])]; + tensor var_2805_groups_0 = const()[name = tensor("op_2805_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394713216)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407820480)))]; + tensor var_2805_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2805_dilations_0, groups = var_2805_groups_0, pad = var_2805_pad_0, pad_type = var_2805_pad_type_0, strides = var_2805_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2805_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2805_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407823104)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407825728)))]; + tensor var_2830_to_fp16 = const()[name = tensor("op_2830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2830_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2865_weight_0_to_fp16 = const()[name = tensor("op_2865_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407828352)))]; + tensor var_2865_bias_0_to_fp16 = const()[name = tensor("op_2865_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411105216)))]; + tensor var_2865_cast_fp16 = conv(bias = var_2865_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2865_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411107840)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2863_pad_type_0 = const()[name = tensor("op_2863_pad_type_0"), val = tensor("valid")]; + tensor var_2863_strides_0 = const()[name = tensor("op_2863_strides_0"), val = tensor([1, 1])]; + tensor var_2863_pad_0 = const()[name = tensor("op_2863_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2863_dilations_0 = const()[name = tensor("op_2863_dilations_0"), val = tensor([1, 1])]; + tensor var_2863_groups_0 = const()[name = tensor("op_2863_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414384704)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417661568)))]; + tensor var_2863_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2863_dilations_0, groups = var_2863_groups_0, pad = var_2863_pad_0, pad_type = var_2863_pad_type_0, strides = var_2863_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2863_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2866_axis_0 = const()[name = tensor("op_2866_axis_0"), val = tensor(1)]; + tensor var_2866_cast_fp16_0, tensor var_2866_cast_fp16_1, tensor var_2866_cast_fp16_2, tensor var_2866_cast_fp16_3, tensor var_2866_cast_fp16_4, tensor var_2866_cast_fp16_5, tensor var_2866_cast_fp16_6, tensor var_2866_cast_fp16_7, tensor var_2866_cast_fp16_8, tensor var_2866_cast_fp16_9, tensor var_2866_cast_fp16_10, tensor var_2866_cast_fp16_11, tensor var_2866_cast_fp16_12, tensor var_2866_cast_fp16_13, tensor var_2866_cast_fp16_14, tensor var_2866_cast_fp16_15, tensor var_2866_cast_fp16_16, tensor var_2866_cast_fp16_17, tensor var_2866_cast_fp16_18, tensor var_2866_cast_fp16_19 = split(axis = var_2866_axis_0, split_sizes = tile_30, x = var_2865_cast_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor var_2887_perm_0 = const()[name = tensor("op_2887_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2888_axis_0 = const()[name = tensor("op_2888_axis_0"), val = tensor(3)]; + tensor var_2887_cast_fp16 = transpose(perm = var_2887_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_22")]; + tensor var_2888_cast_fp16_0, tensor var_2888_cast_fp16_1, tensor var_2888_cast_fp16_2, tensor var_2888_cast_fp16_3, tensor var_2888_cast_fp16_4, tensor var_2888_cast_fp16_5, tensor var_2888_cast_fp16_6, tensor var_2888_cast_fp16_7, tensor var_2888_cast_fp16_8, tensor var_2888_cast_fp16_9, tensor var_2888_cast_fp16_10, tensor var_2888_cast_fp16_11, tensor var_2888_cast_fp16_12, tensor var_2888_cast_fp16_13, tensor var_2888_cast_fp16_14, tensor var_2888_cast_fp16_15, tensor var_2888_cast_fp16_16, tensor var_2888_cast_fp16_17, tensor var_2888_cast_fp16_18, tensor var_2888_cast_fp16_19 = split(axis = var_2888_axis_0, split_sizes = tile_31, x = var_2887_cast_fp16)[name = tensor("op_2888_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2909_axis_0 = const()[name = tensor("op_2909_axis_0"), val = tensor(1)]; + tensor var_2909_cast_fp16_0, tensor var_2909_cast_fp16_1, tensor var_2909_cast_fp16_2, tensor var_2909_cast_fp16_3, tensor var_2909_cast_fp16_4, tensor var_2909_cast_fp16_5, tensor var_2909_cast_fp16_6, tensor var_2909_cast_fp16_7, tensor var_2909_cast_fp16_8, tensor var_2909_cast_fp16_9, tensor var_2909_cast_fp16_10, tensor var_2909_cast_fp16_11, tensor var_2909_cast_fp16_12, tensor var_2909_cast_fp16_13, tensor var_2909_cast_fp16_14, tensor var_2909_cast_fp16_15, tensor var_2909_cast_fp16_16, tensor var_2909_cast_fp16_17, tensor var_2909_cast_fp16_18, tensor var_2909_cast_fp16_19 = split(axis = var_2909_axis_0, split_sizes = tile_32, x = var_2863_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor aw_401_equation_0 = const()[name = tensor("aw_401_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_401_cast_fp16 = einsum(equation = aw_401_equation_0, values = (var_2888_cast_fp16_0, var_2866_cast_fp16_0))[name = tensor("aw_401_cast_fp16")]; + tensor aw_403_equation_0 = const()[name = tensor("aw_403_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_403_cast_fp16 = einsum(equation = aw_403_equation_0, values = (var_2888_cast_fp16_1, var_2866_cast_fp16_1))[name = tensor("aw_403_cast_fp16")]; + tensor aw_405_equation_0 = const()[name = tensor("aw_405_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_405_cast_fp16 = einsum(equation = aw_405_equation_0, values = (var_2888_cast_fp16_2, var_2866_cast_fp16_2))[name = tensor("aw_405_cast_fp16")]; + tensor aw_407_equation_0 = const()[name = tensor("aw_407_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_407_cast_fp16 = einsum(equation = aw_407_equation_0, values = (var_2888_cast_fp16_3, var_2866_cast_fp16_3))[name = tensor("aw_407_cast_fp16")]; + tensor aw_409_equation_0 = const()[name = tensor("aw_409_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_409_cast_fp16 = einsum(equation = aw_409_equation_0, values = (var_2888_cast_fp16_4, var_2866_cast_fp16_4))[name = tensor("aw_409_cast_fp16")]; + tensor aw_411_equation_0 = const()[name = tensor("aw_411_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_411_cast_fp16 = einsum(equation = aw_411_equation_0, values = (var_2888_cast_fp16_5, var_2866_cast_fp16_5))[name = tensor("aw_411_cast_fp16")]; + tensor aw_413_equation_0 = const()[name = tensor("aw_413_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_413_cast_fp16 = einsum(equation = aw_413_equation_0, values = (var_2888_cast_fp16_6, var_2866_cast_fp16_6))[name = tensor("aw_413_cast_fp16")]; + tensor aw_415_equation_0 = const()[name = tensor("aw_415_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_415_cast_fp16 = einsum(equation = aw_415_equation_0, values = (var_2888_cast_fp16_7, var_2866_cast_fp16_7))[name = tensor("aw_415_cast_fp16")]; + tensor aw_417_equation_0 = const()[name = tensor("aw_417_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_417_cast_fp16 = einsum(equation = aw_417_equation_0, values = (var_2888_cast_fp16_8, var_2866_cast_fp16_8))[name = tensor("aw_417_cast_fp16")]; + tensor aw_419_equation_0 = const()[name = tensor("aw_419_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_419_cast_fp16 = einsum(equation = aw_419_equation_0, values = (var_2888_cast_fp16_9, var_2866_cast_fp16_9))[name = tensor("aw_419_cast_fp16")]; + tensor aw_421_equation_0 = const()[name = tensor("aw_421_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_421_cast_fp16 = einsum(equation = aw_421_equation_0, values = (var_2888_cast_fp16_10, var_2866_cast_fp16_10))[name = tensor("aw_421_cast_fp16")]; + tensor aw_423_equation_0 = const()[name = tensor("aw_423_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_423_cast_fp16 = einsum(equation = aw_423_equation_0, values = (var_2888_cast_fp16_11, var_2866_cast_fp16_11))[name = tensor("aw_423_cast_fp16")]; + tensor aw_425_equation_0 = const()[name = tensor("aw_425_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_425_cast_fp16 = einsum(equation = aw_425_equation_0, values = (var_2888_cast_fp16_12, var_2866_cast_fp16_12))[name = tensor("aw_425_cast_fp16")]; + tensor aw_427_equation_0 = const()[name = tensor("aw_427_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_427_cast_fp16 = einsum(equation = aw_427_equation_0, values = (var_2888_cast_fp16_13, var_2866_cast_fp16_13))[name = tensor("aw_427_cast_fp16")]; + tensor aw_429_equation_0 = const()[name = tensor("aw_429_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_429_cast_fp16 = einsum(equation = aw_429_equation_0, values = (var_2888_cast_fp16_14, var_2866_cast_fp16_14))[name = tensor("aw_429_cast_fp16")]; + tensor aw_431_equation_0 = const()[name = tensor("aw_431_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_431_cast_fp16 = einsum(equation = aw_431_equation_0, values = (var_2888_cast_fp16_15, var_2866_cast_fp16_15))[name = tensor("aw_431_cast_fp16")]; + tensor aw_433_equation_0 = const()[name = tensor("aw_433_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_433_cast_fp16 = einsum(equation = aw_433_equation_0, values = (var_2888_cast_fp16_16, var_2866_cast_fp16_16))[name = tensor("aw_433_cast_fp16")]; + tensor aw_435_equation_0 = const()[name = tensor("aw_435_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_435_cast_fp16 = einsum(equation = aw_435_equation_0, values = (var_2888_cast_fp16_17, var_2866_cast_fp16_17))[name = tensor("aw_435_cast_fp16")]; + tensor aw_437_equation_0 = const()[name = tensor("aw_437_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_437_cast_fp16 = einsum(equation = aw_437_equation_0, values = (var_2888_cast_fp16_18, var_2866_cast_fp16_18))[name = tensor("aw_437_cast_fp16")]; + tensor aw_439_equation_0 = const()[name = tensor("aw_439_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_439_cast_fp16 = einsum(equation = aw_439_equation_0, values = (var_2888_cast_fp16_19, var_2866_cast_fp16_19))[name = tensor("aw_439_cast_fp16")]; + tensor var_2970_cast_fp16 = softmax(axis = var_2814, x = aw_401_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = softmax(axis = var_2814, x = aw_403_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_cast_fp16 = softmax(axis = var_2814, x = aw_405_cast_fp16)[name = tensor("op_2972_cast_fp16")]; + tensor var_2973_cast_fp16 = softmax(axis = var_2814, x = aw_407_cast_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor var_2974_cast_fp16 = softmax(axis = var_2814, x = aw_409_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = softmax(axis = var_2814, x = aw_411_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor var_2976_cast_fp16 = softmax(axis = var_2814, x = aw_413_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor var_2977_cast_fp16 = softmax(axis = var_2814, x = aw_415_cast_fp16)[name = tensor("op_2977_cast_fp16")]; + tensor var_2978_cast_fp16 = softmax(axis = var_2814, x = aw_417_cast_fp16)[name = tensor("op_2978_cast_fp16")]; + tensor var_2979_cast_fp16 = softmax(axis = var_2814, x = aw_419_cast_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor var_2980_cast_fp16 = softmax(axis = var_2814, x = aw_421_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor var_2981_cast_fp16 = softmax(axis = var_2814, x = aw_423_cast_fp16)[name = tensor("op_2981_cast_fp16")]; + tensor var_2982_cast_fp16 = softmax(axis = var_2814, x = aw_425_cast_fp16)[name = tensor("op_2982_cast_fp16")]; + tensor var_2983_cast_fp16 = softmax(axis = var_2814, x = aw_427_cast_fp16)[name = tensor("op_2983_cast_fp16")]; + tensor var_2984_cast_fp16 = softmax(axis = var_2814, x = aw_429_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor var_2985_cast_fp16 = softmax(axis = var_2814, x = aw_431_cast_fp16)[name = tensor("op_2985_cast_fp16")]; + tensor var_2986_cast_fp16 = softmax(axis = var_2814, x = aw_433_cast_fp16)[name = tensor("op_2986_cast_fp16")]; + tensor var_2987_cast_fp16 = softmax(axis = var_2814, x = aw_435_cast_fp16)[name = tensor("op_2987_cast_fp16")]; + tensor var_2988_cast_fp16 = softmax(axis = var_2814, x = aw_437_cast_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2989_cast_fp16 = softmax(axis = var_2814, x = aw_439_cast_fp16)[name = tensor("op_2989_cast_fp16")]; + tensor var_2991_equation_0 = const()[name = tensor("op_2991_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2991_cast_fp16 = einsum(equation = var_2991_equation_0, values = (var_2909_cast_fp16_0, var_2970_cast_fp16))[name = tensor("op_2991_cast_fp16")]; + tensor var_2993_equation_0 = const()[name = tensor("op_2993_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2993_cast_fp16 = einsum(equation = var_2993_equation_0, values = (var_2909_cast_fp16_1, var_2971_cast_fp16))[name = tensor("op_2993_cast_fp16")]; + tensor var_2995_equation_0 = const()[name = tensor("op_2995_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2995_cast_fp16 = einsum(equation = var_2995_equation_0, values = (var_2909_cast_fp16_2, var_2972_cast_fp16))[name = tensor("op_2995_cast_fp16")]; + tensor var_2997_equation_0 = const()[name = tensor("op_2997_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2997_cast_fp16 = einsum(equation = var_2997_equation_0, values = (var_2909_cast_fp16_3, var_2973_cast_fp16))[name = tensor("op_2997_cast_fp16")]; + tensor var_2999_equation_0 = const()[name = tensor("op_2999_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2999_cast_fp16 = einsum(equation = var_2999_equation_0, values = (var_2909_cast_fp16_4, var_2974_cast_fp16))[name = tensor("op_2999_cast_fp16")]; + tensor var_3001_equation_0 = const()[name = tensor("op_3001_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3001_cast_fp16 = einsum(equation = var_3001_equation_0, values = (var_2909_cast_fp16_5, var_2975_cast_fp16))[name = tensor("op_3001_cast_fp16")]; + tensor var_3003_equation_0 = const()[name = tensor("op_3003_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3003_cast_fp16 = einsum(equation = var_3003_equation_0, values = (var_2909_cast_fp16_6, var_2976_cast_fp16))[name = tensor("op_3003_cast_fp16")]; + tensor var_3005_equation_0 = const()[name = tensor("op_3005_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3005_cast_fp16 = einsum(equation = var_3005_equation_0, values = (var_2909_cast_fp16_7, var_2977_cast_fp16))[name = tensor("op_3005_cast_fp16")]; + tensor var_3007_equation_0 = const()[name = tensor("op_3007_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3007_cast_fp16 = einsum(equation = var_3007_equation_0, values = (var_2909_cast_fp16_8, var_2978_cast_fp16))[name = tensor("op_3007_cast_fp16")]; + tensor var_3009_equation_0 = const()[name = tensor("op_3009_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3009_cast_fp16 = einsum(equation = var_3009_equation_0, values = (var_2909_cast_fp16_9, var_2979_cast_fp16))[name = tensor("op_3009_cast_fp16")]; + tensor var_3011_equation_0 = const()[name = tensor("op_3011_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3011_cast_fp16 = einsum(equation = var_3011_equation_0, values = (var_2909_cast_fp16_10, var_2980_cast_fp16))[name = tensor("op_3011_cast_fp16")]; + tensor var_3013_equation_0 = const()[name = tensor("op_3013_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3013_cast_fp16 = einsum(equation = var_3013_equation_0, values = (var_2909_cast_fp16_11, var_2981_cast_fp16))[name = tensor("op_3013_cast_fp16")]; + tensor var_3015_equation_0 = const()[name = tensor("op_3015_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3015_cast_fp16 = einsum(equation = var_3015_equation_0, values = (var_2909_cast_fp16_12, var_2982_cast_fp16))[name = tensor("op_3015_cast_fp16")]; + tensor var_3017_equation_0 = const()[name = tensor("op_3017_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3017_cast_fp16 = einsum(equation = var_3017_equation_0, values = (var_2909_cast_fp16_13, var_2983_cast_fp16))[name = tensor("op_3017_cast_fp16")]; + tensor var_3019_equation_0 = const()[name = tensor("op_3019_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3019_cast_fp16 = einsum(equation = var_3019_equation_0, values = (var_2909_cast_fp16_14, var_2984_cast_fp16))[name = tensor("op_3019_cast_fp16")]; + tensor var_3021_equation_0 = const()[name = tensor("op_3021_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3021_cast_fp16 = einsum(equation = var_3021_equation_0, values = (var_2909_cast_fp16_15, var_2985_cast_fp16))[name = tensor("op_3021_cast_fp16")]; + tensor var_3023_equation_0 = const()[name = tensor("op_3023_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3023_cast_fp16 = einsum(equation = var_3023_equation_0, values = (var_2909_cast_fp16_16, var_2986_cast_fp16))[name = tensor("op_3023_cast_fp16")]; + tensor var_3025_equation_0 = const()[name = tensor("op_3025_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3025_cast_fp16 = einsum(equation = var_3025_equation_0, values = (var_2909_cast_fp16_17, var_2987_cast_fp16))[name = tensor("op_3025_cast_fp16")]; + tensor var_3027_equation_0 = const()[name = tensor("op_3027_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3027_cast_fp16 = einsum(equation = var_3027_equation_0, values = (var_2909_cast_fp16_18, var_2988_cast_fp16))[name = tensor("op_3027_cast_fp16")]; + tensor var_3029_equation_0 = const()[name = tensor("op_3029_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3029_cast_fp16 = einsum(equation = var_3029_equation_0, values = (var_2909_cast_fp16_19, var_2989_cast_fp16))[name = tensor("op_3029_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2814, interleave = input_105_interleave_0, values = (var_2991_cast_fp16, var_2993_cast_fp16, var_2995_cast_fp16, var_2997_cast_fp16, var_2999_cast_fp16, var_3001_cast_fp16, var_3003_cast_fp16, var_3005_cast_fp16, var_3007_cast_fp16, var_3009_cast_fp16, var_3011_cast_fp16, var_3013_cast_fp16, var_3015_cast_fp16, var_3017_cast_fp16, var_3019_cast_fp16, var_3021_cast_fp16, var_3023_cast_fp16, var_3025_cast_fp16, var_3027_cast_fp16, var_3029_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_3038_pad_type_0 = const()[name = tensor("op_3038_pad_type_0"), val = tensor("valid")]; + tensor var_3038_strides_0 = const()[name = tensor("op_3038_strides_0"), val = tensor([1, 1])]; + tensor var_3038_pad_0 = const()[name = tensor("op_3038_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3038_dilations_0 = const()[name = tensor("op_3038_dilations_0"), val = tensor([1, 1])]; + tensor var_3038_groups_0 = const()[name = tensor("op_3038_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417664192)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420941056)))]; + tensor var_3038_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_3038_dilations_0, groups = var_3038_groups_0, pad = var_3038_pad_0, pad_type = var_3038_pad_type_0, strides = var_3038_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_3038_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420943680)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420946304)))]; + tensor var_3048_to_fp16 = const()[name = tensor("op_3048_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_3048_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420948928)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434056192)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_3074_pad_type_0 = const()[name = tensor("op_3074_pad_type_0"), val = tensor("valid")]; + tensor var_3074_strides_0 = const()[name = tensor("op_3074_strides_0"), val = tensor([1, 1])]; + tensor var_3074_pad_0 = const()[name = tensor("op_3074_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3074_dilations_0 = const()[name = tensor("op_3074_dilations_0"), val = tensor([1, 1])]; + tensor var_3074_groups_0 = const()[name = tensor("op_3074_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434066496)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447173760)))]; + tensor var_3074_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_3074_dilations_0, groups = var_3074_groups_0, pad = var_3074_pad_0, pad_type = var_3074_pad_type_0, strides = var_3074_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_3074_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_3074_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447176384)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447179008)))]; + tensor var_3099_to_fp16 = const()[name = tensor("op_3099_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_3099_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_23_pad_type_0 = const()[name = tensor("q_23_pad_type_0"), val = tensor("valid")]; + tensor q_23_strides_0 = const()[name = tensor("q_23_strides_0"), val = tensor([1, 1])]; + tensor q_23_pad_0 = const()[name = tensor("q_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_23_dilations_0 = const()[name = tensor("q_23_dilations_0"), val = tensor([1, 1])]; + tensor q_23_groups_0 = const()[name = tensor("q_23_groups_0"), val = tensor(1)]; + tensor var_3134_weight_0_to_fp16 = const()[name = tensor("op_3134_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447181632)))]; + tensor var_3134_bias_0_to_fp16 = const()[name = tensor("op_3134_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450458496)))]; + tensor var_3134_cast_fp16 = conv(bias = var_3134_bias_0_to_fp16, dilations = q_23_dilations_0, groups = q_23_groups_0, pad = q_23_pad_0, pad_type = q_23_pad_type_0, strides = q_23_strides_0, weight = var_3134_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor k_23_pad_type_0 = const()[name = tensor("k_23_pad_type_0"), val = tensor("valid")]; + tensor k_23_strides_0 = const()[name = tensor("k_23_strides_0"), val = tensor([1, 1])]; + tensor k_23_pad_0 = const()[name = tensor("k_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_23_dilations_0 = const()[name = tensor("k_23_dilations_0"), val = tensor([1, 1])]; + tensor k_23_groups_0 = const()[name = tensor("k_23_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450461120)))]; + tensor k_23_cast_fp16 = conv(dilations = k_23_dilations_0, groups = k_23_groups_0, pad = k_23_pad_0, pad_type = k_23_pad_type_0, strides = k_23_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_3132_pad_type_0 = const()[name = tensor("op_3132_pad_type_0"), val = tensor("valid")]; + tensor var_3132_strides_0 = const()[name = tensor("op_3132_strides_0"), val = tensor([1, 1])]; + tensor var_3132_pad_0 = const()[name = tensor("op_3132_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3132_dilations_0 = const()[name = tensor("op_3132_dilations_0"), val = tensor([1, 1])]; + tensor var_3132_groups_0 = const()[name = tensor("op_3132_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453737984)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457014848)))]; + tensor var_3132_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_3132_dilations_0, groups = var_3132_groups_0, pad = var_3132_pad_0, pad_type = var_3132_pad_type_0, strides = var_3132_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_3132_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3135_axis_0 = const()[name = tensor("op_3135_axis_0"), val = tensor(1)]; + tensor var_3135_cast_fp16_0, tensor var_3135_cast_fp16_1, tensor var_3135_cast_fp16_2, tensor var_3135_cast_fp16_3, tensor var_3135_cast_fp16_4, tensor var_3135_cast_fp16_5, tensor var_3135_cast_fp16_6, tensor var_3135_cast_fp16_7, tensor var_3135_cast_fp16_8, tensor var_3135_cast_fp16_9, tensor var_3135_cast_fp16_10, tensor var_3135_cast_fp16_11, tensor var_3135_cast_fp16_12, tensor var_3135_cast_fp16_13, tensor var_3135_cast_fp16_14, tensor var_3135_cast_fp16_15, tensor var_3135_cast_fp16_16, tensor var_3135_cast_fp16_17, tensor var_3135_cast_fp16_18, tensor var_3135_cast_fp16_19 = split(axis = var_3135_axis_0, split_sizes = tile_33, x = var_3134_cast_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3156_perm_0 = const()[name = tensor("op_3156_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3157_axis_0 = const()[name = tensor("op_3157_axis_0"), val = tensor(3)]; + tensor var_3156_cast_fp16 = transpose(perm = var_3156_perm_0, x = k_23_cast_fp16)[name = tensor("transpose_21")]; + tensor var_3157_cast_fp16_0, tensor var_3157_cast_fp16_1, tensor var_3157_cast_fp16_2, tensor var_3157_cast_fp16_3, tensor var_3157_cast_fp16_4, tensor var_3157_cast_fp16_5, tensor var_3157_cast_fp16_6, tensor var_3157_cast_fp16_7, tensor var_3157_cast_fp16_8, tensor var_3157_cast_fp16_9, tensor var_3157_cast_fp16_10, tensor var_3157_cast_fp16_11, tensor var_3157_cast_fp16_12, tensor var_3157_cast_fp16_13, tensor var_3157_cast_fp16_14, tensor var_3157_cast_fp16_15, tensor var_3157_cast_fp16_16, tensor var_3157_cast_fp16_17, tensor var_3157_cast_fp16_18, tensor var_3157_cast_fp16_19 = split(axis = var_3157_axis_0, split_sizes = tile_34, x = var_3156_cast_fp16)[name = tensor("op_3157_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3178_axis_0 = const()[name = tensor("op_3178_axis_0"), val = tensor(1)]; + tensor var_3178_cast_fp16_0, tensor var_3178_cast_fp16_1, tensor var_3178_cast_fp16_2, tensor var_3178_cast_fp16_3, tensor var_3178_cast_fp16_4, tensor var_3178_cast_fp16_5, tensor var_3178_cast_fp16_6, tensor var_3178_cast_fp16_7, tensor var_3178_cast_fp16_8, tensor var_3178_cast_fp16_9, tensor var_3178_cast_fp16_10, tensor var_3178_cast_fp16_11, tensor var_3178_cast_fp16_12, tensor var_3178_cast_fp16_13, tensor var_3178_cast_fp16_14, tensor var_3178_cast_fp16_15, tensor var_3178_cast_fp16_16, tensor var_3178_cast_fp16_17, tensor var_3178_cast_fp16_18, tensor var_3178_cast_fp16_19 = split(axis = var_3178_axis_0, split_sizes = tile_35, x = var_3132_cast_fp16)[name = tensor("op_3178_cast_fp16")]; + tensor aw_441_equation_0 = const()[name = tensor("aw_441_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_441_cast_fp16 = einsum(equation = aw_441_equation_0, values = (var_3157_cast_fp16_0, var_3135_cast_fp16_0))[name = tensor("aw_441_cast_fp16")]; + tensor aw_443_equation_0 = const()[name = tensor("aw_443_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_443_cast_fp16 = einsum(equation = aw_443_equation_0, values = (var_3157_cast_fp16_1, var_3135_cast_fp16_1))[name = tensor("aw_443_cast_fp16")]; + tensor aw_445_equation_0 = const()[name = tensor("aw_445_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_445_cast_fp16 = einsum(equation = aw_445_equation_0, values = (var_3157_cast_fp16_2, var_3135_cast_fp16_2))[name = tensor("aw_445_cast_fp16")]; + tensor aw_447_equation_0 = const()[name = tensor("aw_447_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_447_cast_fp16 = einsum(equation = aw_447_equation_0, values = (var_3157_cast_fp16_3, var_3135_cast_fp16_3))[name = tensor("aw_447_cast_fp16")]; + tensor aw_449_equation_0 = const()[name = tensor("aw_449_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_449_cast_fp16 = einsum(equation = aw_449_equation_0, values = (var_3157_cast_fp16_4, var_3135_cast_fp16_4))[name = tensor("aw_449_cast_fp16")]; + tensor aw_451_equation_0 = const()[name = tensor("aw_451_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_451_cast_fp16 = einsum(equation = aw_451_equation_0, values = (var_3157_cast_fp16_5, var_3135_cast_fp16_5))[name = tensor("aw_451_cast_fp16")]; + tensor aw_453_equation_0 = const()[name = tensor("aw_453_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_453_cast_fp16 = einsum(equation = aw_453_equation_0, values = (var_3157_cast_fp16_6, var_3135_cast_fp16_6))[name = tensor("aw_453_cast_fp16")]; + tensor aw_455_equation_0 = const()[name = tensor("aw_455_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_455_cast_fp16 = einsum(equation = aw_455_equation_0, values = (var_3157_cast_fp16_7, var_3135_cast_fp16_7))[name = tensor("aw_455_cast_fp16")]; + tensor aw_457_equation_0 = const()[name = tensor("aw_457_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_457_cast_fp16 = einsum(equation = aw_457_equation_0, values = (var_3157_cast_fp16_8, var_3135_cast_fp16_8))[name = tensor("aw_457_cast_fp16")]; + tensor aw_459_equation_0 = const()[name = tensor("aw_459_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_459_cast_fp16 = einsum(equation = aw_459_equation_0, values = (var_3157_cast_fp16_9, var_3135_cast_fp16_9))[name = tensor("aw_459_cast_fp16")]; + tensor aw_461_equation_0 = const()[name = tensor("aw_461_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_461_cast_fp16 = einsum(equation = aw_461_equation_0, values = (var_3157_cast_fp16_10, var_3135_cast_fp16_10))[name = tensor("aw_461_cast_fp16")]; + tensor aw_463_equation_0 = const()[name = tensor("aw_463_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_463_cast_fp16 = einsum(equation = aw_463_equation_0, values = (var_3157_cast_fp16_11, var_3135_cast_fp16_11))[name = tensor("aw_463_cast_fp16")]; + tensor aw_465_equation_0 = const()[name = tensor("aw_465_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_465_cast_fp16 = einsum(equation = aw_465_equation_0, values = (var_3157_cast_fp16_12, var_3135_cast_fp16_12))[name = tensor("aw_465_cast_fp16")]; + tensor aw_467_equation_0 = const()[name = tensor("aw_467_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_467_cast_fp16 = einsum(equation = aw_467_equation_0, values = (var_3157_cast_fp16_13, var_3135_cast_fp16_13))[name = tensor("aw_467_cast_fp16")]; + tensor aw_469_equation_0 = const()[name = tensor("aw_469_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_469_cast_fp16 = einsum(equation = aw_469_equation_0, values = (var_3157_cast_fp16_14, var_3135_cast_fp16_14))[name = tensor("aw_469_cast_fp16")]; + tensor aw_471_equation_0 = const()[name = tensor("aw_471_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_471_cast_fp16 = einsum(equation = aw_471_equation_0, values = (var_3157_cast_fp16_15, var_3135_cast_fp16_15))[name = tensor("aw_471_cast_fp16")]; + tensor aw_473_equation_0 = const()[name = tensor("aw_473_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_473_cast_fp16 = einsum(equation = aw_473_equation_0, values = (var_3157_cast_fp16_16, var_3135_cast_fp16_16))[name = tensor("aw_473_cast_fp16")]; + tensor aw_475_equation_0 = const()[name = tensor("aw_475_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_475_cast_fp16 = einsum(equation = aw_475_equation_0, values = (var_3157_cast_fp16_17, var_3135_cast_fp16_17))[name = tensor("aw_475_cast_fp16")]; + tensor aw_477_equation_0 = const()[name = tensor("aw_477_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_477_cast_fp16 = einsum(equation = aw_477_equation_0, values = (var_3157_cast_fp16_18, var_3135_cast_fp16_18))[name = tensor("aw_477_cast_fp16")]; + tensor aw_479_equation_0 = const()[name = tensor("aw_479_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_479_cast_fp16 = einsum(equation = aw_479_equation_0, values = (var_3157_cast_fp16_19, var_3135_cast_fp16_19))[name = tensor("aw_479_cast_fp16")]; + tensor var_3239_cast_fp16 = softmax(axis = var_3083, x = aw_441_cast_fp16)[name = tensor("op_3239_cast_fp16")]; + tensor var_3240_cast_fp16 = softmax(axis = var_3083, x = aw_443_cast_fp16)[name = tensor("op_3240_cast_fp16")]; + tensor var_3241_cast_fp16 = softmax(axis = var_3083, x = aw_445_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor var_3242_cast_fp16 = softmax(axis = var_3083, x = aw_447_cast_fp16)[name = tensor("op_3242_cast_fp16")]; + tensor var_3243_cast_fp16 = softmax(axis = var_3083, x = aw_449_cast_fp16)[name = tensor("op_3243_cast_fp16")]; + tensor var_3244_cast_fp16 = softmax(axis = var_3083, x = aw_451_cast_fp16)[name = tensor("op_3244_cast_fp16")]; + tensor var_3245_cast_fp16 = softmax(axis = var_3083, x = aw_453_cast_fp16)[name = tensor("op_3245_cast_fp16")]; + tensor var_3246_cast_fp16 = softmax(axis = var_3083, x = aw_455_cast_fp16)[name = tensor("op_3246_cast_fp16")]; + tensor var_3247_cast_fp16 = softmax(axis = var_3083, x = aw_457_cast_fp16)[name = tensor("op_3247_cast_fp16")]; + tensor var_3248_cast_fp16 = softmax(axis = var_3083, x = aw_459_cast_fp16)[name = tensor("op_3248_cast_fp16")]; + tensor var_3249_cast_fp16 = softmax(axis = var_3083, x = aw_461_cast_fp16)[name = tensor("op_3249_cast_fp16")]; + tensor var_3250_cast_fp16 = softmax(axis = var_3083, x = aw_463_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = softmax(axis = var_3083, x = aw_465_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor var_3252_cast_fp16 = softmax(axis = var_3083, x = aw_467_cast_fp16)[name = tensor("op_3252_cast_fp16")]; + tensor var_3253_cast_fp16 = softmax(axis = var_3083, x = aw_469_cast_fp16)[name = tensor("op_3253_cast_fp16")]; + tensor var_3254_cast_fp16 = softmax(axis = var_3083, x = aw_471_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = softmax(axis = var_3083, x = aw_473_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor var_3256_cast_fp16 = softmax(axis = var_3083, x = aw_475_cast_fp16)[name = tensor("op_3256_cast_fp16")]; + tensor var_3257_cast_fp16 = softmax(axis = var_3083, x = aw_477_cast_fp16)[name = tensor("op_3257_cast_fp16")]; + tensor var_3258_cast_fp16 = softmax(axis = var_3083, x = aw_479_cast_fp16)[name = tensor("op_3258_cast_fp16")]; + tensor var_3260_equation_0 = const()[name = tensor("op_3260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3260_cast_fp16 = einsum(equation = var_3260_equation_0, values = (var_3178_cast_fp16_0, var_3239_cast_fp16))[name = tensor("op_3260_cast_fp16")]; + tensor var_3262_equation_0 = const()[name = tensor("op_3262_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3262_cast_fp16 = einsum(equation = var_3262_equation_0, values = (var_3178_cast_fp16_1, var_3240_cast_fp16))[name = tensor("op_3262_cast_fp16")]; + tensor var_3264_equation_0 = const()[name = tensor("op_3264_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3264_cast_fp16 = einsum(equation = var_3264_equation_0, values = (var_3178_cast_fp16_2, var_3241_cast_fp16))[name = tensor("op_3264_cast_fp16")]; + tensor var_3266_equation_0 = const()[name = tensor("op_3266_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3266_cast_fp16 = einsum(equation = var_3266_equation_0, values = (var_3178_cast_fp16_3, var_3242_cast_fp16))[name = tensor("op_3266_cast_fp16")]; + tensor var_3268_equation_0 = const()[name = tensor("op_3268_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3268_cast_fp16 = einsum(equation = var_3268_equation_0, values = (var_3178_cast_fp16_4, var_3243_cast_fp16))[name = tensor("op_3268_cast_fp16")]; + tensor var_3270_equation_0 = const()[name = tensor("op_3270_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3270_cast_fp16 = einsum(equation = var_3270_equation_0, values = (var_3178_cast_fp16_5, var_3244_cast_fp16))[name = tensor("op_3270_cast_fp16")]; + tensor var_3272_equation_0 = const()[name = tensor("op_3272_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3272_cast_fp16 = einsum(equation = var_3272_equation_0, values = (var_3178_cast_fp16_6, var_3245_cast_fp16))[name = tensor("op_3272_cast_fp16")]; + tensor var_3274_equation_0 = const()[name = tensor("op_3274_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3274_cast_fp16 = einsum(equation = var_3274_equation_0, values = (var_3178_cast_fp16_7, var_3246_cast_fp16))[name = tensor("op_3274_cast_fp16")]; + tensor var_3276_equation_0 = const()[name = tensor("op_3276_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3276_cast_fp16 = einsum(equation = var_3276_equation_0, values = (var_3178_cast_fp16_8, var_3247_cast_fp16))[name = tensor("op_3276_cast_fp16")]; + tensor var_3278_equation_0 = const()[name = tensor("op_3278_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3278_cast_fp16 = einsum(equation = var_3278_equation_0, values = (var_3178_cast_fp16_9, var_3248_cast_fp16))[name = tensor("op_3278_cast_fp16")]; + tensor var_3280_equation_0 = const()[name = tensor("op_3280_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3280_cast_fp16 = einsum(equation = var_3280_equation_0, values = (var_3178_cast_fp16_10, var_3249_cast_fp16))[name = tensor("op_3280_cast_fp16")]; + tensor var_3282_equation_0 = const()[name = tensor("op_3282_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3282_cast_fp16 = einsum(equation = var_3282_equation_0, values = (var_3178_cast_fp16_11, var_3250_cast_fp16))[name = tensor("op_3282_cast_fp16")]; + tensor var_3284_equation_0 = const()[name = tensor("op_3284_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3284_cast_fp16 = einsum(equation = var_3284_equation_0, values = (var_3178_cast_fp16_12, var_3251_cast_fp16))[name = tensor("op_3284_cast_fp16")]; + tensor var_3286_equation_0 = const()[name = tensor("op_3286_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3286_cast_fp16 = einsum(equation = var_3286_equation_0, values = (var_3178_cast_fp16_13, var_3252_cast_fp16))[name = tensor("op_3286_cast_fp16")]; + tensor var_3288_equation_0 = const()[name = tensor("op_3288_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3288_cast_fp16 = einsum(equation = var_3288_equation_0, values = (var_3178_cast_fp16_14, var_3253_cast_fp16))[name = tensor("op_3288_cast_fp16")]; + tensor var_3290_equation_0 = const()[name = tensor("op_3290_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3290_cast_fp16 = einsum(equation = var_3290_equation_0, values = (var_3178_cast_fp16_15, var_3254_cast_fp16))[name = tensor("op_3290_cast_fp16")]; + tensor var_3292_equation_0 = const()[name = tensor("op_3292_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3292_cast_fp16 = einsum(equation = var_3292_equation_0, values = (var_3178_cast_fp16_16, var_3255_cast_fp16))[name = tensor("op_3292_cast_fp16")]; + tensor var_3294_equation_0 = const()[name = tensor("op_3294_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3294_cast_fp16 = einsum(equation = var_3294_equation_0, values = (var_3178_cast_fp16_17, var_3256_cast_fp16))[name = tensor("op_3294_cast_fp16")]; + tensor var_3296_equation_0 = const()[name = tensor("op_3296_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3296_cast_fp16 = einsum(equation = var_3296_equation_0, values = (var_3178_cast_fp16_18, var_3257_cast_fp16))[name = tensor("op_3296_cast_fp16")]; + tensor var_3298_equation_0 = const()[name = tensor("op_3298_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3298_cast_fp16 = einsum(equation = var_3298_equation_0, values = (var_3178_cast_fp16_19, var_3258_cast_fp16))[name = tensor("op_3298_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_3083, interleave = input_115_interleave_0, values = (var_3260_cast_fp16, var_3262_cast_fp16, var_3264_cast_fp16, var_3266_cast_fp16, var_3268_cast_fp16, var_3270_cast_fp16, var_3272_cast_fp16, var_3274_cast_fp16, var_3276_cast_fp16, var_3278_cast_fp16, var_3280_cast_fp16, var_3282_cast_fp16, var_3284_cast_fp16, var_3286_cast_fp16, var_3288_cast_fp16, var_3290_cast_fp16, var_3292_cast_fp16, var_3294_cast_fp16, var_3296_cast_fp16, var_3298_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_3307_pad_type_0 = const()[name = tensor("op_3307_pad_type_0"), val = tensor("valid")]; + tensor var_3307_strides_0 = const()[name = tensor("op_3307_strides_0"), val = tensor([1, 1])]; + tensor var_3307_pad_0 = const()[name = tensor("op_3307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3307_dilations_0 = const()[name = tensor("op_3307_dilations_0"), val = tensor([1, 1])]; + tensor var_3307_groups_0 = const()[name = tensor("op_3307_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457017472)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460294336)))]; + tensor var_3307_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_3307_dilations_0, groups = var_3307_groups_0, pad = var_3307_pad_0, pad_type = var_3307_pad_type_0, strides = var_3307_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_3307_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460296960)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460299584)))]; + tensor var_3317_to_fp16 = const()[name = tensor("op_3317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_3317_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460302208)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473409472)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_3343_pad_type_0 = const()[name = tensor("op_3343_pad_type_0"), val = tensor("valid")]; + tensor var_3343_strides_0 = const()[name = tensor("op_3343_strides_0"), val = tensor([1, 1])]; + tensor var_3343_pad_0 = const()[name = tensor("op_3343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3343_dilations_0 = const()[name = tensor("op_3343_dilations_0"), val = tensor([1, 1])]; + tensor var_3343_groups_0 = const()[name = tensor("op_3343_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473419776)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486527040)))]; + tensor var_3343_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_3343_dilations_0, groups = var_3343_groups_0, pad = var_3343_pad_0, pad_type = var_3343_pad_type_0, strides = var_3343_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("op_3343_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_3343_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor(1)]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([1])]; + tensor input_123_gamma_0_to_fp16 = const()[name = tensor("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486529664)))]; + tensor input_123_beta_0_to_fp16 = const()[name = tensor("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486532288)))]; + tensor var_3368_to_fp16 = const()[name = tensor("op_3368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = input_123_beta_0_to_fp16, epsilon = var_3368_to_fp16, gamma = input_123_gamma_0_to_fp16, x = inputs_49_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor q_25_pad_type_0 = const()[name = tensor("q_25_pad_type_0"), val = tensor("valid")]; + tensor q_25_strides_0 = const()[name = tensor("q_25_strides_0"), val = tensor([1, 1])]; + tensor q_25_pad_0 = const()[name = tensor("q_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_25_dilations_0 = const()[name = tensor("q_25_dilations_0"), val = tensor([1, 1])]; + tensor q_25_groups_0 = const()[name = tensor("q_25_groups_0"), val = tensor(1)]; + tensor var_3403_weight_0_to_fp16 = const()[name = tensor("op_3403_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486534912)))]; + tensor var_3403_bias_0_to_fp16 = const()[name = tensor("op_3403_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489811776)))]; + tensor var_3403_cast_fp16 = conv(bias = var_3403_bias_0_to_fp16, dilations = q_25_dilations_0, groups = q_25_groups_0, pad = q_25_pad_0, pad_type = q_25_pad_type_0, strides = q_25_strides_0, weight = var_3403_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3403_cast_fp16")]; + tensor k_25_pad_type_0 = const()[name = tensor("k_25_pad_type_0"), val = tensor("valid")]; + tensor k_25_strides_0 = const()[name = tensor("k_25_strides_0"), val = tensor([1, 1])]; + tensor k_25_pad_0 = const()[name = tensor("k_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_25_dilations_0 = const()[name = tensor("k_25_dilations_0"), val = tensor([1, 1])]; + tensor k_25_groups_0 = const()[name = tensor("k_25_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_key_weight_to_fp16 = const()[name = tensor("blocks_12_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489814400)))]; + tensor k_25_cast_fp16 = conv(dilations = k_25_dilations_0, groups = k_25_groups_0, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = k_25_strides_0, weight = blocks_12_attn_key_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor var_3401_pad_type_0 = const()[name = tensor("op_3401_pad_type_0"), val = tensor("valid")]; + tensor var_3401_strides_0 = const()[name = tensor("op_3401_strides_0"), val = tensor([1, 1])]; + tensor var_3401_pad_0 = const()[name = tensor("op_3401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3401_dilations_0 = const()[name = tensor("op_3401_dilations_0"), val = tensor([1, 1])]; + tensor var_3401_groups_0 = const()[name = tensor("op_3401_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_value_weight_to_fp16 = const()[name = tensor("blocks_12_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493091264)))]; + tensor blocks_12_attn_value_bias_to_fp16 = const()[name = tensor("blocks_12_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496368128)))]; + tensor var_3401_cast_fp16 = conv(bias = blocks_12_attn_value_bias_to_fp16, dilations = var_3401_dilations_0, groups = var_3401_groups_0, pad = var_3401_pad_0, pad_type = var_3401_pad_type_0, strides = var_3401_strides_0, weight = blocks_12_attn_value_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3401_cast_fp16")]; + tensor tile_36 = const()[name = tensor("tile_36"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3404_axis_0 = const()[name = tensor("op_3404_axis_0"), val = tensor(1)]; + tensor var_3404_cast_fp16_0, tensor var_3404_cast_fp16_1, tensor var_3404_cast_fp16_2, tensor var_3404_cast_fp16_3, tensor var_3404_cast_fp16_4, tensor var_3404_cast_fp16_5, tensor var_3404_cast_fp16_6, tensor var_3404_cast_fp16_7, tensor var_3404_cast_fp16_8, tensor var_3404_cast_fp16_9, tensor var_3404_cast_fp16_10, tensor var_3404_cast_fp16_11, tensor var_3404_cast_fp16_12, tensor var_3404_cast_fp16_13, tensor var_3404_cast_fp16_14, tensor var_3404_cast_fp16_15, tensor var_3404_cast_fp16_16, tensor var_3404_cast_fp16_17, tensor var_3404_cast_fp16_18, tensor var_3404_cast_fp16_19 = split(axis = var_3404_axis_0, split_sizes = tile_36, x = var_3403_cast_fp16)[name = tensor("op_3404_cast_fp16")]; + tensor var_3425_perm_0 = const()[name = tensor("op_3425_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_37 = const()[name = tensor("tile_37"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3426_axis_0 = const()[name = tensor("op_3426_axis_0"), val = tensor(3)]; + tensor var_3425_cast_fp16 = transpose(perm = var_3425_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_20")]; + tensor var_3426_cast_fp16_0, tensor var_3426_cast_fp16_1, tensor var_3426_cast_fp16_2, tensor var_3426_cast_fp16_3, tensor var_3426_cast_fp16_4, tensor var_3426_cast_fp16_5, tensor var_3426_cast_fp16_6, tensor var_3426_cast_fp16_7, tensor var_3426_cast_fp16_8, tensor var_3426_cast_fp16_9, tensor var_3426_cast_fp16_10, tensor var_3426_cast_fp16_11, tensor var_3426_cast_fp16_12, tensor var_3426_cast_fp16_13, tensor var_3426_cast_fp16_14, tensor var_3426_cast_fp16_15, tensor var_3426_cast_fp16_16, tensor var_3426_cast_fp16_17, tensor var_3426_cast_fp16_18, tensor var_3426_cast_fp16_19 = split(axis = var_3426_axis_0, split_sizes = tile_37, x = var_3425_cast_fp16)[name = tensor("op_3426_cast_fp16")]; + tensor tile_38 = const()[name = tensor("tile_38"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3447_axis_0 = const()[name = tensor("op_3447_axis_0"), val = tensor(1)]; + tensor var_3447_cast_fp16_0, tensor var_3447_cast_fp16_1, tensor var_3447_cast_fp16_2, tensor var_3447_cast_fp16_3, tensor var_3447_cast_fp16_4, tensor var_3447_cast_fp16_5, tensor var_3447_cast_fp16_6, tensor var_3447_cast_fp16_7, tensor var_3447_cast_fp16_8, tensor var_3447_cast_fp16_9, tensor var_3447_cast_fp16_10, tensor var_3447_cast_fp16_11, tensor var_3447_cast_fp16_12, tensor var_3447_cast_fp16_13, tensor var_3447_cast_fp16_14, tensor var_3447_cast_fp16_15, tensor var_3447_cast_fp16_16, tensor var_3447_cast_fp16_17, tensor var_3447_cast_fp16_18, tensor var_3447_cast_fp16_19 = split(axis = var_3447_axis_0, split_sizes = tile_38, x = var_3401_cast_fp16)[name = tensor("op_3447_cast_fp16")]; + tensor aw_481_equation_0 = const()[name = tensor("aw_481_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_481_cast_fp16 = einsum(equation = aw_481_equation_0, values = (var_3426_cast_fp16_0, var_3404_cast_fp16_0))[name = tensor("aw_481_cast_fp16")]; + tensor aw_483_equation_0 = const()[name = tensor("aw_483_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_483_cast_fp16 = einsum(equation = aw_483_equation_0, values = (var_3426_cast_fp16_1, var_3404_cast_fp16_1))[name = tensor("aw_483_cast_fp16")]; + tensor aw_485_equation_0 = const()[name = tensor("aw_485_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_485_cast_fp16 = einsum(equation = aw_485_equation_0, values = (var_3426_cast_fp16_2, var_3404_cast_fp16_2))[name = tensor("aw_485_cast_fp16")]; + tensor aw_487_equation_0 = const()[name = tensor("aw_487_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_487_cast_fp16 = einsum(equation = aw_487_equation_0, values = (var_3426_cast_fp16_3, var_3404_cast_fp16_3))[name = tensor("aw_487_cast_fp16")]; + tensor aw_489_equation_0 = const()[name = tensor("aw_489_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_489_cast_fp16 = einsum(equation = aw_489_equation_0, values = (var_3426_cast_fp16_4, var_3404_cast_fp16_4))[name = tensor("aw_489_cast_fp16")]; + tensor aw_491_equation_0 = const()[name = tensor("aw_491_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_491_cast_fp16 = einsum(equation = aw_491_equation_0, values = (var_3426_cast_fp16_5, var_3404_cast_fp16_5))[name = tensor("aw_491_cast_fp16")]; + tensor aw_493_equation_0 = const()[name = tensor("aw_493_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_493_cast_fp16 = einsum(equation = aw_493_equation_0, values = (var_3426_cast_fp16_6, var_3404_cast_fp16_6))[name = tensor("aw_493_cast_fp16")]; + tensor aw_495_equation_0 = const()[name = tensor("aw_495_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_495_cast_fp16 = einsum(equation = aw_495_equation_0, values = (var_3426_cast_fp16_7, var_3404_cast_fp16_7))[name = tensor("aw_495_cast_fp16")]; + tensor aw_497_equation_0 = const()[name = tensor("aw_497_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_497_cast_fp16 = einsum(equation = aw_497_equation_0, values = (var_3426_cast_fp16_8, var_3404_cast_fp16_8))[name = tensor("aw_497_cast_fp16")]; + tensor aw_499_equation_0 = const()[name = tensor("aw_499_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_499_cast_fp16 = einsum(equation = aw_499_equation_0, values = (var_3426_cast_fp16_9, var_3404_cast_fp16_9))[name = tensor("aw_499_cast_fp16")]; + tensor aw_501_equation_0 = const()[name = tensor("aw_501_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_501_cast_fp16 = einsum(equation = aw_501_equation_0, values = (var_3426_cast_fp16_10, var_3404_cast_fp16_10))[name = tensor("aw_501_cast_fp16")]; + tensor aw_503_equation_0 = const()[name = tensor("aw_503_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_503_cast_fp16 = einsum(equation = aw_503_equation_0, values = (var_3426_cast_fp16_11, var_3404_cast_fp16_11))[name = tensor("aw_503_cast_fp16")]; + tensor aw_505_equation_0 = const()[name = tensor("aw_505_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_505_cast_fp16 = einsum(equation = aw_505_equation_0, values = (var_3426_cast_fp16_12, var_3404_cast_fp16_12))[name = tensor("aw_505_cast_fp16")]; + tensor aw_507_equation_0 = const()[name = tensor("aw_507_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_507_cast_fp16 = einsum(equation = aw_507_equation_0, values = (var_3426_cast_fp16_13, var_3404_cast_fp16_13))[name = tensor("aw_507_cast_fp16")]; + tensor aw_509_equation_0 = const()[name = tensor("aw_509_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_509_cast_fp16 = einsum(equation = aw_509_equation_0, values = (var_3426_cast_fp16_14, var_3404_cast_fp16_14))[name = tensor("aw_509_cast_fp16")]; + tensor aw_511_equation_0 = const()[name = tensor("aw_511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_511_cast_fp16 = einsum(equation = aw_511_equation_0, values = (var_3426_cast_fp16_15, var_3404_cast_fp16_15))[name = tensor("aw_511_cast_fp16")]; + tensor aw_513_equation_0 = const()[name = tensor("aw_513_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_513_cast_fp16 = einsum(equation = aw_513_equation_0, values = (var_3426_cast_fp16_16, var_3404_cast_fp16_16))[name = tensor("aw_513_cast_fp16")]; + tensor aw_515_equation_0 = const()[name = tensor("aw_515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_515_cast_fp16 = einsum(equation = aw_515_equation_0, values = (var_3426_cast_fp16_17, var_3404_cast_fp16_17))[name = tensor("aw_515_cast_fp16")]; + tensor aw_517_equation_0 = const()[name = tensor("aw_517_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_517_cast_fp16 = einsum(equation = aw_517_equation_0, values = (var_3426_cast_fp16_18, var_3404_cast_fp16_18))[name = tensor("aw_517_cast_fp16")]; + tensor aw_519_equation_0 = const()[name = tensor("aw_519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_519_cast_fp16 = einsum(equation = aw_519_equation_0, values = (var_3426_cast_fp16_19, var_3404_cast_fp16_19))[name = tensor("aw_519_cast_fp16")]; + tensor var_3508_cast_fp16 = softmax(axis = var_3352, x = aw_481_cast_fp16)[name = tensor("op_3508_cast_fp16")]; + tensor var_3509_cast_fp16 = softmax(axis = var_3352, x = aw_483_cast_fp16)[name = tensor("op_3509_cast_fp16")]; + tensor var_3510_cast_fp16 = softmax(axis = var_3352, x = aw_485_cast_fp16)[name = tensor("op_3510_cast_fp16")]; + tensor var_3511_cast_fp16 = softmax(axis = var_3352, x = aw_487_cast_fp16)[name = tensor("op_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = softmax(axis = var_3352, x = aw_489_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_cast_fp16 = softmax(axis = var_3352, x = aw_491_cast_fp16)[name = tensor("op_3513_cast_fp16")]; + tensor var_3514_cast_fp16 = softmax(axis = var_3352, x = aw_493_cast_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor var_3515_cast_fp16 = softmax(axis = var_3352, x = aw_495_cast_fp16)[name = tensor("op_3515_cast_fp16")]; + tensor var_3516_cast_fp16 = softmax(axis = var_3352, x = aw_497_cast_fp16)[name = tensor("op_3516_cast_fp16")]; + tensor var_3517_cast_fp16 = softmax(axis = var_3352, x = aw_499_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3518_cast_fp16 = softmax(axis = var_3352, x = aw_501_cast_fp16)[name = tensor("op_3518_cast_fp16")]; + tensor var_3519_cast_fp16 = softmax(axis = var_3352, x = aw_503_cast_fp16)[name = tensor("op_3519_cast_fp16")]; + tensor var_3520_cast_fp16 = softmax(axis = var_3352, x = aw_505_cast_fp16)[name = tensor("op_3520_cast_fp16")]; + tensor var_3521_cast_fp16 = softmax(axis = var_3352, x = aw_507_cast_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3522_cast_fp16 = softmax(axis = var_3352, x = aw_509_cast_fp16)[name = tensor("op_3522_cast_fp16")]; + tensor var_3523_cast_fp16 = softmax(axis = var_3352, x = aw_511_cast_fp16)[name = tensor("op_3523_cast_fp16")]; + tensor var_3524_cast_fp16 = softmax(axis = var_3352, x = aw_513_cast_fp16)[name = tensor("op_3524_cast_fp16")]; + tensor var_3525_cast_fp16 = softmax(axis = var_3352, x = aw_515_cast_fp16)[name = tensor("op_3525_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_3352, x = aw_517_cast_fp16)[name = tensor("op_3526_cast_fp16")]; + tensor var_3527_cast_fp16 = softmax(axis = var_3352, x = aw_519_cast_fp16)[name = tensor("op_3527_cast_fp16")]; + tensor var_3529_equation_0 = const()[name = tensor("op_3529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3529_cast_fp16 = einsum(equation = var_3529_equation_0, values = (var_3447_cast_fp16_0, var_3508_cast_fp16))[name = tensor("op_3529_cast_fp16")]; + tensor var_3531_equation_0 = const()[name = tensor("op_3531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3531_cast_fp16 = einsum(equation = var_3531_equation_0, values = (var_3447_cast_fp16_1, var_3509_cast_fp16))[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_equation_0 = const()[name = tensor("op_3533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3533_cast_fp16 = einsum(equation = var_3533_equation_0, values = (var_3447_cast_fp16_2, var_3510_cast_fp16))[name = tensor("op_3533_cast_fp16")]; + tensor var_3535_equation_0 = const()[name = tensor("op_3535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3535_cast_fp16 = einsum(equation = var_3535_equation_0, values = (var_3447_cast_fp16_3, var_3511_cast_fp16))[name = tensor("op_3535_cast_fp16")]; + tensor var_3537_equation_0 = const()[name = tensor("op_3537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3537_cast_fp16 = einsum(equation = var_3537_equation_0, values = (var_3447_cast_fp16_4, var_3512_cast_fp16))[name = tensor("op_3537_cast_fp16")]; + tensor var_3539_equation_0 = const()[name = tensor("op_3539_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3539_cast_fp16 = einsum(equation = var_3539_equation_0, values = (var_3447_cast_fp16_5, var_3513_cast_fp16))[name = tensor("op_3539_cast_fp16")]; + tensor var_3541_equation_0 = const()[name = tensor("op_3541_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3541_cast_fp16 = einsum(equation = var_3541_equation_0, values = (var_3447_cast_fp16_6, var_3514_cast_fp16))[name = tensor("op_3541_cast_fp16")]; + tensor var_3543_equation_0 = const()[name = tensor("op_3543_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3543_cast_fp16 = einsum(equation = var_3543_equation_0, values = (var_3447_cast_fp16_7, var_3515_cast_fp16))[name = tensor("op_3543_cast_fp16")]; + tensor var_3545_equation_0 = const()[name = tensor("op_3545_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3545_cast_fp16 = einsum(equation = var_3545_equation_0, values = (var_3447_cast_fp16_8, var_3516_cast_fp16))[name = tensor("op_3545_cast_fp16")]; + tensor var_3547_equation_0 = const()[name = tensor("op_3547_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3547_cast_fp16 = einsum(equation = var_3547_equation_0, values = (var_3447_cast_fp16_9, var_3517_cast_fp16))[name = tensor("op_3547_cast_fp16")]; + tensor var_3549_equation_0 = const()[name = tensor("op_3549_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3549_cast_fp16 = einsum(equation = var_3549_equation_0, values = (var_3447_cast_fp16_10, var_3518_cast_fp16))[name = tensor("op_3549_cast_fp16")]; + tensor var_3551_equation_0 = const()[name = tensor("op_3551_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3551_cast_fp16 = einsum(equation = var_3551_equation_0, values = (var_3447_cast_fp16_11, var_3519_cast_fp16))[name = tensor("op_3551_cast_fp16")]; + tensor var_3553_equation_0 = const()[name = tensor("op_3553_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3553_cast_fp16 = einsum(equation = var_3553_equation_0, values = (var_3447_cast_fp16_12, var_3520_cast_fp16))[name = tensor("op_3553_cast_fp16")]; + tensor var_3555_equation_0 = const()[name = tensor("op_3555_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3555_cast_fp16 = einsum(equation = var_3555_equation_0, values = (var_3447_cast_fp16_13, var_3521_cast_fp16))[name = tensor("op_3555_cast_fp16")]; + tensor var_3557_equation_0 = const()[name = tensor("op_3557_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3557_cast_fp16 = einsum(equation = var_3557_equation_0, values = (var_3447_cast_fp16_14, var_3522_cast_fp16))[name = tensor("op_3557_cast_fp16")]; + tensor var_3559_equation_0 = const()[name = tensor("op_3559_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3559_cast_fp16 = einsum(equation = var_3559_equation_0, values = (var_3447_cast_fp16_15, var_3523_cast_fp16))[name = tensor("op_3559_cast_fp16")]; + tensor var_3561_equation_0 = const()[name = tensor("op_3561_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3561_cast_fp16 = einsum(equation = var_3561_equation_0, values = (var_3447_cast_fp16_16, var_3524_cast_fp16))[name = tensor("op_3561_cast_fp16")]; + tensor var_3563_equation_0 = const()[name = tensor("op_3563_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3563_cast_fp16 = einsum(equation = var_3563_equation_0, values = (var_3447_cast_fp16_17, var_3525_cast_fp16))[name = tensor("op_3563_cast_fp16")]; + tensor var_3565_equation_0 = const()[name = tensor("op_3565_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3565_cast_fp16 = einsum(equation = var_3565_equation_0, values = (var_3447_cast_fp16_18, var_3526_cast_fp16))[name = tensor("op_3565_cast_fp16")]; + tensor var_3567_equation_0 = const()[name = tensor("op_3567_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3567_cast_fp16 = einsum(equation = var_3567_equation_0, values = (var_3447_cast_fp16_19, var_3527_cast_fp16))[name = tensor("op_3567_cast_fp16")]; + tensor input_125_interleave_0 = const()[name = tensor("input_125_interleave_0"), val = tensor(false)]; + tensor input_125_cast_fp16 = concat(axis = var_3352, interleave = input_125_interleave_0, values = (var_3529_cast_fp16, var_3531_cast_fp16, var_3533_cast_fp16, var_3535_cast_fp16, var_3537_cast_fp16, var_3539_cast_fp16, var_3541_cast_fp16, var_3543_cast_fp16, var_3545_cast_fp16, var_3547_cast_fp16, var_3549_cast_fp16, var_3551_cast_fp16, var_3553_cast_fp16, var_3555_cast_fp16, var_3557_cast_fp16, var_3559_cast_fp16, var_3561_cast_fp16, var_3563_cast_fp16, var_3565_cast_fp16, var_3567_cast_fp16))[name = tensor("input_125_cast_fp16")]; + tensor var_3576_pad_type_0 = const()[name = tensor("op_3576_pad_type_0"), val = tensor("valid")]; + tensor var_3576_strides_0 = const()[name = tensor("op_3576_strides_0"), val = tensor([1, 1])]; + tensor var_3576_pad_0 = const()[name = tensor("op_3576_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3576_dilations_0 = const()[name = tensor("op_3576_dilations_0"), val = tensor([1, 1])]; + tensor var_3576_groups_0 = const()[name = tensor("op_3576_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_out_weight_to_fp16 = const()[name = tensor("blocks_12_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496370752)))]; + tensor blocks_12_attn_out_bias_to_fp16 = const()[name = tensor("blocks_12_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499647616)))]; + tensor var_3576_cast_fp16 = conv(bias = blocks_12_attn_out_bias_to_fp16, dilations = var_3576_dilations_0, groups = var_3576_groups_0, pad = var_3576_pad_0, pad_type = var_3576_pad_type_0, strides = var_3576_strides_0, weight = blocks_12_attn_out_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = var_3576_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([1])]; + tensor input_127_gamma_0_to_fp16 = const()[name = tensor("input_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499650240)))]; + tensor input_127_beta_0_to_fp16 = const()[name = tensor("input_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499652864)))]; + tensor var_3586_to_fp16 = const()[name = tensor("op_3586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = input_127_beta_0_to_fp16, epsilon = var_3586_to_fp16, gamma = input_127_gamma_0_to_fp16, x = inputs_51_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("valid")]; + tensor input_129_strides_0 = const()[name = tensor("input_129_strides_0"), val = tensor([1, 1])]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_129_dilations_0 = const()[name = tensor("input_129_dilations_0"), val = tensor([1, 1])]; + tensor input_129_groups_0 = const()[name = tensor("input_129_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499655488)))]; + tensor blocks_12_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512762752)))]; + tensor input_129_cast_fp16 = conv(bias = blocks_12_mlp_0_bias_to_fp16, dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = blocks_12_mlp_0_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3612_pad_type_0 = const()[name = tensor("op_3612_pad_type_0"), val = tensor("valid")]; + tensor var_3612_strides_0 = const()[name = tensor("op_3612_strides_0"), val = tensor([1, 1])]; + tensor var_3612_pad_0 = const()[name = tensor("op_3612_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3612_dilations_0 = const()[name = tensor("op_3612_dilations_0"), val = tensor([1, 1])]; + tensor var_3612_groups_0 = const()[name = tensor("op_3612_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512773056)))]; + tensor blocks_12_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525880320)))]; + tensor var_3612_cast_fp16 = conv(bias = blocks_12_mlp_2_bias_to_fp16, dilations = var_3612_dilations_0, groups = var_3612_groups_0, pad = var_3612_pad_0, pad_type = var_3612_pad_type_0, strides = var_3612_strides_0, weight = blocks_12_mlp_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_3612_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3621 = const()[name = tensor("op_3621"), val = tensor(1)]; + tensor input_133_axes_0 = const()[name = tensor("input_133_axes_0"), val = tensor([1])]; + tensor input_133_gamma_0_to_fp16 = const()[name = tensor("input_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525882944)))]; + tensor input_133_beta_0_to_fp16 = const()[name = tensor("input_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525885568)))]; + tensor var_3637_to_fp16 = const()[name = tensor("op_3637_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_133_cast_fp16 = layer_norm(axes = input_133_axes_0, beta = input_133_beta_0_to_fp16, epsilon = var_3637_to_fp16, gamma = input_133_gamma_0_to_fp16, x = inputs_53_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor q_27_pad_type_0 = const()[name = tensor("q_27_pad_type_0"), val = tensor("valid")]; + tensor q_27_strides_0 = const()[name = tensor("q_27_strides_0"), val = tensor([1, 1])]; + tensor q_27_pad_0 = const()[name = tensor("q_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_27_dilations_0 = const()[name = tensor("q_27_dilations_0"), val = tensor([1, 1])]; + tensor q_27_groups_0 = const()[name = tensor("q_27_groups_0"), val = tensor(1)]; + tensor var_3672_weight_0_to_fp16 = const()[name = tensor("op_3672_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525888192)))]; + tensor var_3672_bias_0_to_fp16 = const()[name = tensor("op_3672_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529165056)))]; + tensor var_3672_cast_fp16 = conv(bias = var_3672_bias_0_to_fp16, dilations = q_27_dilations_0, groups = q_27_groups_0, pad = q_27_pad_0, pad_type = q_27_pad_type_0, strides = q_27_strides_0, weight = var_3672_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3672_cast_fp16")]; + tensor k_27_pad_type_0 = const()[name = tensor("k_27_pad_type_0"), val = tensor("valid")]; + tensor k_27_strides_0 = const()[name = tensor("k_27_strides_0"), val = tensor([1, 1])]; + tensor k_27_pad_0 = const()[name = tensor("k_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_27_dilations_0 = const()[name = tensor("k_27_dilations_0"), val = tensor([1, 1])]; + tensor k_27_groups_0 = const()[name = tensor("k_27_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_key_weight_to_fp16 = const()[name = tensor("blocks_13_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529167680)))]; + tensor k_27_cast_fp16 = conv(dilations = k_27_dilations_0, groups = k_27_groups_0, pad = k_27_pad_0, pad_type = k_27_pad_type_0, strides = k_27_strides_0, weight = blocks_13_attn_key_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor var_3670_pad_type_0 = const()[name = tensor("op_3670_pad_type_0"), val = tensor("valid")]; + tensor var_3670_strides_0 = const()[name = tensor("op_3670_strides_0"), val = tensor([1, 1])]; + tensor var_3670_pad_0 = const()[name = tensor("op_3670_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3670_dilations_0 = const()[name = tensor("op_3670_dilations_0"), val = tensor([1, 1])]; + tensor var_3670_groups_0 = const()[name = tensor("op_3670_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_value_weight_to_fp16 = const()[name = tensor("blocks_13_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532444544)))]; + tensor blocks_13_attn_value_bias_to_fp16 = const()[name = tensor("blocks_13_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535721408)))]; + tensor var_3670_cast_fp16 = conv(bias = blocks_13_attn_value_bias_to_fp16, dilations = var_3670_dilations_0, groups = var_3670_groups_0, pad = var_3670_pad_0, pad_type = var_3670_pad_type_0, strides = var_3670_strides_0, weight = blocks_13_attn_value_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3670_cast_fp16")]; + tensor tile_39 = const()[name = tensor("tile_39"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3673_axis_0 = const()[name = tensor("op_3673_axis_0"), val = tensor(1)]; + tensor var_3673_cast_fp16_0, tensor var_3673_cast_fp16_1, tensor var_3673_cast_fp16_2, tensor var_3673_cast_fp16_3, tensor var_3673_cast_fp16_4, tensor var_3673_cast_fp16_5, tensor var_3673_cast_fp16_6, tensor var_3673_cast_fp16_7, tensor var_3673_cast_fp16_8, tensor var_3673_cast_fp16_9, tensor var_3673_cast_fp16_10, tensor var_3673_cast_fp16_11, tensor var_3673_cast_fp16_12, tensor var_3673_cast_fp16_13, tensor var_3673_cast_fp16_14, tensor var_3673_cast_fp16_15, tensor var_3673_cast_fp16_16, tensor var_3673_cast_fp16_17, tensor var_3673_cast_fp16_18, tensor var_3673_cast_fp16_19 = split(axis = var_3673_axis_0, split_sizes = tile_39, x = var_3672_cast_fp16)[name = tensor("op_3673_cast_fp16")]; + tensor var_3694_perm_0 = const()[name = tensor("op_3694_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_40 = const()[name = tensor("tile_40"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3695_axis_0 = const()[name = tensor("op_3695_axis_0"), val = tensor(3)]; + tensor var_3694_cast_fp16 = transpose(perm = var_3694_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_19")]; + tensor var_3695_cast_fp16_0, tensor var_3695_cast_fp16_1, tensor var_3695_cast_fp16_2, tensor var_3695_cast_fp16_3, tensor var_3695_cast_fp16_4, tensor var_3695_cast_fp16_5, tensor var_3695_cast_fp16_6, tensor var_3695_cast_fp16_7, tensor var_3695_cast_fp16_8, tensor var_3695_cast_fp16_9, tensor var_3695_cast_fp16_10, tensor var_3695_cast_fp16_11, tensor var_3695_cast_fp16_12, tensor var_3695_cast_fp16_13, tensor var_3695_cast_fp16_14, tensor var_3695_cast_fp16_15, tensor var_3695_cast_fp16_16, tensor var_3695_cast_fp16_17, tensor var_3695_cast_fp16_18, tensor var_3695_cast_fp16_19 = split(axis = var_3695_axis_0, split_sizes = tile_40, x = var_3694_cast_fp16)[name = tensor("op_3695_cast_fp16")]; + tensor tile_41 = const()[name = tensor("tile_41"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3716_axis_0 = const()[name = tensor("op_3716_axis_0"), val = tensor(1)]; + tensor var_3716_cast_fp16_0, tensor var_3716_cast_fp16_1, tensor var_3716_cast_fp16_2, tensor var_3716_cast_fp16_3, tensor var_3716_cast_fp16_4, tensor var_3716_cast_fp16_5, tensor var_3716_cast_fp16_6, tensor var_3716_cast_fp16_7, tensor var_3716_cast_fp16_8, tensor var_3716_cast_fp16_9, tensor var_3716_cast_fp16_10, tensor var_3716_cast_fp16_11, tensor var_3716_cast_fp16_12, tensor var_3716_cast_fp16_13, tensor var_3716_cast_fp16_14, tensor var_3716_cast_fp16_15, tensor var_3716_cast_fp16_16, tensor var_3716_cast_fp16_17, tensor var_3716_cast_fp16_18, tensor var_3716_cast_fp16_19 = split(axis = var_3716_axis_0, split_sizes = tile_41, x = var_3670_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor aw_521_equation_0 = const()[name = tensor("aw_521_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_521_cast_fp16 = einsum(equation = aw_521_equation_0, values = (var_3695_cast_fp16_0, var_3673_cast_fp16_0))[name = tensor("aw_521_cast_fp16")]; + tensor aw_523_equation_0 = const()[name = tensor("aw_523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_523_cast_fp16 = einsum(equation = aw_523_equation_0, values = (var_3695_cast_fp16_1, var_3673_cast_fp16_1))[name = tensor("aw_523_cast_fp16")]; + tensor aw_525_equation_0 = const()[name = tensor("aw_525_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_525_cast_fp16 = einsum(equation = aw_525_equation_0, values = (var_3695_cast_fp16_2, var_3673_cast_fp16_2))[name = tensor("aw_525_cast_fp16")]; + tensor aw_527_equation_0 = const()[name = tensor("aw_527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_527_cast_fp16 = einsum(equation = aw_527_equation_0, values = (var_3695_cast_fp16_3, var_3673_cast_fp16_3))[name = tensor("aw_527_cast_fp16")]; + tensor aw_529_equation_0 = const()[name = tensor("aw_529_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_529_cast_fp16 = einsum(equation = aw_529_equation_0, values = (var_3695_cast_fp16_4, var_3673_cast_fp16_4))[name = tensor("aw_529_cast_fp16")]; + tensor aw_531_equation_0 = const()[name = tensor("aw_531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_531_cast_fp16 = einsum(equation = aw_531_equation_0, values = (var_3695_cast_fp16_5, var_3673_cast_fp16_5))[name = tensor("aw_531_cast_fp16")]; + tensor aw_533_equation_0 = const()[name = tensor("aw_533_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_533_cast_fp16 = einsum(equation = aw_533_equation_0, values = (var_3695_cast_fp16_6, var_3673_cast_fp16_6))[name = tensor("aw_533_cast_fp16")]; + tensor aw_535_equation_0 = const()[name = tensor("aw_535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_535_cast_fp16 = einsum(equation = aw_535_equation_0, values = (var_3695_cast_fp16_7, var_3673_cast_fp16_7))[name = tensor("aw_535_cast_fp16")]; + tensor aw_537_equation_0 = const()[name = tensor("aw_537_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_537_cast_fp16 = einsum(equation = aw_537_equation_0, values = (var_3695_cast_fp16_8, var_3673_cast_fp16_8))[name = tensor("aw_537_cast_fp16")]; + tensor aw_539_equation_0 = const()[name = tensor("aw_539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_539_cast_fp16 = einsum(equation = aw_539_equation_0, values = (var_3695_cast_fp16_9, var_3673_cast_fp16_9))[name = tensor("aw_539_cast_fp16")]; + tensor aw_541_equation_0 = const()[name = tensor("aw_541_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_541_cast_fp16 = einsum(equation = aw_541_equation_0, values = (var_3695_cast_fp16_10, var_3673_cast_fp16_10))[name = tensor("aw_541_cast_fp16")]; + tensor aw_543_equation_0 = const()[name = tensor("aw_543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_543_cast_fp16 = einsum(equation = aw_543_equation_0, values = (var_3695_cast_fp16_11, var_3673_cast_fp16_11))[name = tensor("aw_543_cast_fp16")]; + tensor aw_545_equation_0 = const()[name = tensor("aw_545_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_545_cast_fp16 = einsum(equation = aw_545_equation_0, values = (var_3695_cast_fp16_12, var_3673_cast_fp16_12))[name = tensor("aw_545_cast_fp16")]; + tensor aw_547_equation_0 = const()[name = tensor("aw_547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_547_cast_fp16 = einsum(equation = aw_547_equation_0, values = (var_3695_cast_fp16_13, var_3673_cast_fp16_13))[name = tensor("aw_547_cast_fp16")]; + tensor aw_549_equation_0 = const()[name = tensor("aw_549_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_549_cast_fp16 = einsum(equation = aw_549_equation_0, values = (var_3695_cast_fp16_14, var_3673_cast_fp16_14))[name = tensor("aw_549_cast_fp16")]; + tensor aw_551_equation_0 = const()[name = tensor("aw_551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_551_cast_fp16 = einsum(equation = aw_551_equation_0, values = (var_3695_cast_fp16_15, var_3673_cast_fp16_15))[name = tensor("aw_551_cast_fp16")]; + tensor aw_553_equation_0 = const()[name = tensor("aw_553_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_553_cast_fp16 = einsum(equation = aw_553_equation_0, values = (var_3695_cast_fp16_16, var_3673_cast_fp16_16))[name = tensor("aw_553_cast_fp16")]; + tensor aw_555_equation_0 = const()[name = tensor("aw_555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_555_cast_fp16 = einsum(equation = aw_555_equation_0, values = (var_3695_cast_fp16_17, var_3673_cast_fp16_17))[name = tensor("aw_555_cast_fp16")]; + tensor aw_557_equation_0 = const()[name = tensor("aw_557_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_557_cast_fp16 = einsum(equation = aw_557_equation_0, values = (var_3695_cast_fp16_18, var_3673_cast_fp16_18))[name = tensor("aw_557_cast_fp16")]; + tensor aw_559_equation_0 = const()[name = tensor("aw_559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_559_cast_fp16 = einsum(equation = aw_559_equation_0, values = (var_3695_cast_fp16_19, var_3673_cast_fp16_19))[name = tensor("aw_559_cast_fp16")]; + tensor var_3777_cast_fp16 = softmax(axis = var_3621, x = aw_521_cast_fp16)[name = tensor("op_3777_cast_fp16")]; + tensor var_3778_cast_fp16 = softmax(axis = var_3621, x = aw_523_cast_fp16)[name = tensor("op_3778_cast_fp16")]; + tensor var_3779_cast_fp16 = softmax(axis = var_3621, x = aw_525_cast_fp16)[name = tensor("op_3779_cast_fp16")]; + tensor var_3780_cast_fp16 = softmax(axis = var_3621, x = aw_527_cast_fp16)[name = tensor("op_3780_cast_fp16")]; + tensor var_3781_cast_fp16 = softmax(axis = var_3621, x = aw_529_cast_fp16)[name = tensor("op_3781_cast_fp16")]; + tensor var_3782_cast_fp16 = softmax(axis = var_3621, x = aw_531_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = softmax(axis = var_3621, x = aw_533_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor var_3784_cast_fp16 = softmax(axis = var_3621, x = aw_535_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3785_cast_fp16 = softmax(axis = var_3621, x = aw_537_cast_fp16)[name = tensor("op_3785_cast_fp16")]; + tensor var_3786_cast_fp16 = softmax(axis = var_3621, x = aw_539_cast_fp16)[name = tensor("op_3786_cast_fp16")]; + tensor var_3787_cast_fp16 = softmax(axis = var_3621, x = aw_541_cast_fp16)[name = tensor("op_3787_cast_fp16")]; + tensor var_3788_cast_fp16 = softmax(axis = var_3621, x = aw_543_cast_fp16)[name = tensor("op_3788_cast_fp16")]; + tensor var_3789_cast_fp16 = softmax(axis = var_3621, x = aw_545_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3790_cast_fp16 = softmax(axis = var_3621, x = aw_547_cast_fp16)[name = tensor("op_3790_cast_fp16")]; + tensor var_3791_cast_fp16 = softmax(axis = var_3621, x = aw_549_cast_fp16)[name = tensor("op_3791_cast_fp16")]; + tensor var_3792_cast_fp16 = softmax(axis = var_3621, x = aw_551_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3793_cast_fp16 = softmax(axis = var_3621, x = aw_553_cast_fp16)[name = tensor("op_3793_cast_fp16")]; + tensor var_3794_cast_fp16 = softmax(axis = var_3621, x = aw_555_cast_fp16)[name = tensor("op_3794_cast_fp16")]; + tensor var_3795_cast_fp16 = softmax(axis = var_3621, x = aw_557_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = softmax(axis = var_3621, x = aw_559_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor var_3798_equation_0 = const()[name = tensor("op_3798_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3798_cast_fp16 = einsum(equation = var_3798_equation_0, values = (var_3716_cast_fp16_0, var_3777_cast_fp16))[name = tensor("op_3798_cast_fp16")]; + tensor var_3800_equation_0 = const()[name = tensor("op_3800_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3800_cast_fp16 = einsum(equation = var_3800_equation_0, values = (var_3716_cast_fp16_1, var_3778_cast_fp16))[name = tensor("op_3800_cast_fp16")]; + tensor var_3802_equation_0 = const()[name = tensor("op_3802_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3802_cast_fp16 = einsum(equation = var_3802_equation_0, values = (var_3716_cast_fp16_2, var_3779_cast_fp16))[name = tensor("op_3802_cast_fp16")]; + tensor var_3804_equation_0 = const()[name = tensor("op_3804_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3804_cast_fp16 = einsum(equation = var_3804_equation_0, values = (var_3716_cast_fp16_3, var_3780_cast_fp16))[name = tensor("op_3804_cast_fp16")]; + tensor var_3806_equation_0 = const()[name = tensor("op_3806_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3806_cast_fp16 = einsum(equation = var_3806_equation_0, values = (var_3716_cast_fp16_4, var_3781_cast_fp16))[name = tensor("op_3806_cast_fp16")]; + tensor var_3808_equation_0 = const()[name = tensor("op_3808_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3808_cast_fp16 = einsum(equation = var_3808_equation_0, values = (var_3716_cast_fp16_5, var_3782_cast_fp16))[name = tensor("op_3808_cast_fp16")]; + tensor var_3810_equation_0 = const()[name = tensor("op_3810_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3810_cast_fp16 = einsum(equation = var_3810_equation_0, values = (var_3716_cast_fp16_6, var_3783_cast_fp16))[name = tensor("op_3810_cast_fp16")]; + tensor var_3812_equation_0 = const()[name = tensor("op_3812_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3812_cast_fp16 = einsum(equation = var_3812_equation_0, values = (var_3716_cast_fp16_7, var_3784_cast_fp16))[name = tensor("op_3812_cast_fp16")]; + tensor var_3814_equation_0 = const()[name = tensor("op_3814_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3814_cast_fp16 = einsum(equation = var_3814_equation_0, values = (var_3716_cast_fp16_8, var_3785_cast_fp16))[name = tensor("op_3814_cast_fp16")]; + tensor var_3816_equation_0 = const()[name = tensor("op_3816_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3816_cast_fp16 = einsum(equation = var_3816_equation_0, values = (var_3716_cast_fp16_9, var_3786_cast_fp16))[name = tensor("op_3816_cast_fp16")]; + tensor var_3818_equation_0 = const()[name = tensor("op_3818_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3818_cast_fp16 = einsum(equation = var_3818_equation_0, values = (var_3716_cast_fp16_10, var_3787_cast_fp16))[name = tensor("op_3818_cast_fp16")]; + tensor var_3820_equation_0 = const()[name = tensor("op_3820_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3820_cast_fp16 = einsum(equation = var_3820_equation_0, values = (var_3716_cast_fp16_11, var_3788_cast_fp16))[name = tensor("op_3820_cast_fp16")]; + tensor var_3822_equation_0 = const()[name = tensor("op_3822_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3822_cast_fp16 = einsum(equation = var_3822_equation_0, values = (var_3716_cast_fp16_12, var_3789_cast_fp16))[name = tensor("op_3822_cast_fp16")]; + tensor var_3824_equation_0 = const()[name = tensor("op_3824_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3824_cast_fp16 = einsum(equation = var_3824_equation_0, values = (var_3716_cast_fp16_13, var_3790_cast_fp16))[name = tensor("op_3824_cast_fp16")]; + tensor var_3826_equation_0 = const()[name = tensor("op_3826_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3826_cast_fp16 = einsum(equation = var_3826_equation_0, values = (var_3716_cast_fp16_14, var_3791_cast_fp16))[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_equation_0 = const()[name = tensor("op_3828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3828_cast_fp16 = einsum(equation = var_3828_equation_0, values = (var_3716_cast_fp16_15, var_3792_cast_fp16))[name = tensor("op_3828_cast_fp16")]; + tensor var_3830_equation_0 = const()[name = tensor("op_3830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3830_cast_fp16 = einsum(equation = var_3830_equation_0, values = (var_3716_cast_fp16_16, var_3793_cast_fp16))[name = tensor("op_3830_cast_fp16")]; + tensor var_3832_equation_0 = const()[name = tensor("op_3832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3832_cast_fp16 = einsum(equation = var_3832_equation_0, values = (var_3716_cast_fp16_17, var_3794_cast_fp16))[name = tensor("op_3832_cast_fp16")]; + tensor var_3834_equation_0 = const()[name = tensor("op_3834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3834_cast_fp16 = einsum(equation = var_3834_equation_0, values = (var_3716_cast_fp16_18, var_3795_cast_fp16))[name = tensor("op_3834_cast_fp16")]; + tensor var_3836_equation_0 = const()[name = tensor("op_3836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3836_cast_fp16 = einsum(equation = var_3836_equation_0, values = (var_3716_cast_fp16_19, var_3796_cast_fp16))[name = tensor("op_3836_cast_fp16")]; + tensor input_135_interleave_0 = const()[name = tensor("input_135_interleave_0"), val = tensor(false)]; + tensor input_135_cast_fp16 = concat(axis = var_3621, interleave = input_135_interleave_0, values = (var_3798_cast_fp16, var_3800_cast_fp16, var_3802_cast_fp16, var_3804_cast_fp16, var_3806_cast_fp16, var_3808_cast_fp16, var_3810_cast_fp16, var_3812_cast_fp16, var_3814_cast_fp16, var_3816_cast_fp16, var_3818_cast_fp16, var_3820_cast_fp16, var_3822_cast_fp16, var_3824_cast_fp16, var_3826_cast_fp16, var_3828_cast_fp16, var_3830_cast_fp16, var_3832_cast_fp16, var_3834_cast_fp16, var_3836_cast_fp16))[name = tensor("input_135_cast_fp16")]; + tensor var_3845_pad_type_0 = const()[name = tensor("op_3845_pad_type_0"), val = tensor("valid")]; + tensor var_3845_strides_0 = const()[name = tensor("op_3845_strides_0"), val = tensor([1, 1])]; + tensor var_3845_pad_0 = const()[name = tensor("op_3845_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3845_dilations_0 = const()[name = tensor("op_3845_dilations_0"), val = tensor([1, 1])]; + tensor var_3845_groups_0 = const()[name = tensor("op_3845_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_out_weight_to_fp16 = const()[name = tensor("blocks_13_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535724032)))]; + tensor blocks_13_attn_out_bias_to_fp16 = const()[name = tensor("blocks_13_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539000896)))]; + tensor var_3845_cast_fp16 = conv(bias = blocks_13_attn_out_bias_to_fp16, dilations = var_3845_dilations_0, groups = var_3845_groups_0, pad = var_3845_pad_0, pad_type = var_3845_pad_type_0, strides = var_3845_strides_0, weight = blocks_13_attn_out_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("op_3845_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = var_3845_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([1])]; + tensor input_137_gamma_0_to_fp16 = const()[name = tensor("input_137_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539003520)))]; + tensor input_137_beta_0_to_fp16 = const()[name = tensor("input_137_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539006144)))]; + tensor var_3855_to_fp16 = const()[name = tensor("op_3855_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = input_137_beta_0_to_fp16, epsilon = var_3855_to_fp16, gamma = input_137_gamma_0_to_fp16, x = inputs_55_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("valid")]; + tensor input_139_strides_0 = const()[name = tensor("input_139_strides_0"), val = tensor([1, 1])]; + tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_139_dilations_0 = const()[name = tensor("input_139_dilations_0"), val = tensor([1, 1])]; + tensor input_139_groups_0 = const()[name = tensor("input_139_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539008768)))]; + tensor blocks_13_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552116032)))]; + tensor input_139_cast_fp16 = conv(bias = blocks_13_mlp_0_bias_to_fp16, dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = blocks_13_mlp_0_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3881_pad_type_0 = const()[name = tensor("op_3881_pad_type_0"), val = tensor("valid")]; + tensor var_3881_strides_0 = const()[name = tensor("op_3881_strides_0"), val = tensor([1, 1])]; + tensor var_3881_pad_0 = const()[name = tensor("op_3881_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3881_dilations_0 = const()[name = tensor("op_3881_dilations_0"), val = tensor([1, 1])]; + tensor var_3881_groups_0 = const()[name = tensor("op_3881_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552126336)))]; + tensor blocks_13_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565233600)))]; + tensor var_3881_cast_fp16 = conv(bias = blocks_13_mlp_2_bias_to_fp16, dilations = var_3881_dilations_0, groups = var_3881_groups_0, pad = var_3881_pad_0, pad_type = var_3881_pad_type_0, strides = var_3881_strides_0, weight = blocks_13_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = var_3881_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3890 = const()[name = tensor("op_3890"), val = tensor(1)]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([1])]; + tensor input_143_gamma_0_to_fp16 = const()[name = tensor("input_143_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565236224)))]; + tensor input_143_beta_0_to_fp16 = const()[name = tensor("input_143_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565238848)))]; + tensor var_3906_to_fp16 = const()[name = tensor("op_3906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = input_143_beta_0_to_fp16, epsilon = var_3906_to_fp16, gamma = input_143_gamma_0_to_fp16, x = inputs_57_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor q_29_pad_type_0 = const()[name = tensor("q_29_pad_type_0"), val = tensor("valid")]; + tensor q_29_strides_0 = const()[name = tensor("q_29_strides_0"), val = tensor([1, 1])]; + tensor q_29_pad_0 = const()[name = tensor("q_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_29_dilations_0 = const()[name = tensor("q_29_dilations_0"), val = tensor([1, 1])]; + tensor q_29_groups_0 = const()[name = tensor("q_29_groups_0"), val = tensor(1)]; + tensor var_3941_weight_0_to_fp16 = const()[name = tensor("op_3941_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565241472)))]; + tensor var_3941_bias_0_to_fp16 = const()[name = tensor("op_3941_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568518336)))]; + tensor var_3941_cast_fp16 = conv(bias = var_3941_bias_0_to_fp16, dilations = q_29_dilations_0, groups = q_29_groups_0, pad = q_29_pad_0, pad_type = q_29_pad_type_0, strides = q_29_strides_0, weight = var_3941_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3941_cast_fp16")]; + tensor k_29_pad_type_0 = const()[name = tensor("k_29_pad_type_0"), val = tensor("valid")]; + tensor k_29_strides_0 = const()[name = tensor("k_29_strides_0"), val = tensor([1, 1])]; + tensor k_29_pad_0 = const()[name = tensor("k_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_29_dilations_0 = const()[name = tensor("k_29_dilations_0"), val = tensor([1, 1])]; + tensor k_29_groups_0 = const()[name = tensor("k_29_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_key_weight_to_fp16 = const()[name = tensor("blocks_14_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568520960)))]; + tensor k_29_cast_fp16 = conv(dilations = k_29_dilations_0, groups = k_29_groups_0, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = k_29_strides_0, weight = blocks_14_attn_key_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor var_3939_pad_type_0 = const()[name = tensor("op_3939_pad_type_0"), val = tensor("valid")]; + tensor var_3939_strides_0 = const()[name = tensor("op_3939_strides_0"), val = tensor([1, 1])]; + tensor var_3939_pad_0 = const()[name = tensor("op_3939_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3939_dilations_0 = const()[name = tensor("op_3939_dilations_0"), val = tensor([1, 1])]; + tensor var_3939_groups_0 = const()[name = tensor("op_3939_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_value_weight_to_fp16 = const()[name = tensor("blocks_14_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571797824)))]; + tensor blocks_14_attn_value_bias_to_fp16 = const()[name = tensor("blocks_14_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575074688)))]; + tensor var_3939_cast_fp16 = conv(bias = blocks_14_attn_value_bias_to_fp16, dilations = var_3939_dilations_0, groups = var_3939_groups_0, pad = var_3939_pad_0, pad_type = var_3939_pad_type_0, strides = var_3939_strides_0, weight = blocks_14_attn_value_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor tile_42 = const()[name = tensor("tile_42"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3942_axis_0 = const()[name = tensor("op_3942_axis_0"), val = tensor(1)]; + tensor var_3942_cast_fp16_0, tensor var_3942_cast_fp16_1, tensor var_3942_cast_fp16_2, tensor var_3942_cast_fp16_3, tensor var_3942_cast_fp16_4, tensor var_3942_cast_fp16_5, tensor var_3942_cast_fp16_6, tensor var_3942_cast_fp16_7, tensor var_3942_cast_fp16_8, tensor var_3942_cast_fp16_9, tensor var_3942_cast_fp16_10, tensor var_3942_cast_fp16_11, tensor var_3942_cast_fp16_12, tensor var_3942_cast_fp16_13, tensor var_3942_cast_fp16_14, tensor var_3942_cast_fp16_15, tensor var_3942_cast_fp16_16, tensor var_3942_cast_fp16_17, tensor var_3942_cast_fp16_18, tensor var_3942_cast_fp16_19 = split(axis = var_3942_axis_0, split_sizes = tile_42, x = var_3941_cast_fp16)[name = tensor("op_3942_cast_fp16")]; + tensor var_3963_perm_0 = const()[name = tensor("op_3963_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_43 = const()[name = tensor("tile_43"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3964_axis_0 = const()[name = tensor("op_3964_axis_0"), val = tensor(3)]; + tensor var_3963_cast_fp16 = transpose(perm = var_3963_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_18")]; + tensor var_3964_cast_fp16_0, tensor var_3964_cast_fp16_1, tensor var_3964_cast_fp16_2, tensor var_3964_cast_fp16_3, tensor var_3964_cast_fp16_4, tensor var_3964_cast_fp16_5, tensor var_3964_cast_fp16_6, tensor var_3964_cast_fp16_7, tensor var_3964_cast_fp16_8, tensor var_3964_cast_fp16_9, tensor var_3964_cast_fp16_10, tensor var_3964_cast_fp16_11, tensor var_3964_cast_fp16_12, tensor var_3964_cast_fp16_13, tensor var_3964_cast_fp16_14, tensor var_3964_cast_fp16_15, tensor var_3964_cast_fp16_16, tensor var_3964_cast_fp16_17, tensor var_3964_cast_fp16_18, tensor var_3964_cast_fp16_19 = split(axis = var_3964_axis_0, split_sizes = tile_43, x = var_3963_cast_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor tile_44 = const()[name = tensor("tile_44"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3985_axis_0 = const()[name = tensor("op_3985_axis_0"), val = tensor(1)]; + tensor var_3985_cast_fp16_0, tensor var_3985_cast_fp16_1, tensor var_3985_cast_fp16_2, tensor var_3985_cast_fp16_3, tensor var_3985_cast_fp16_4, tensor var_3985_cast_fp16_5, tensor var_3985_cast_fp16_6, tensor var_3985_cast_fp16_7, tensor var_3985_cast_fp16_8, tensor var_3985_cast_fp16_9, tensor var_3985_cast_fp16_10, tensor var_3985_cast_fp16_11, tensor var_3985_cast_fp16_12, tensor var_3985_cast_fp16_13, tensor var_3985_cast_fp16_14, tensor var_3985_cast_fp16_15, tensor var_3985_cast_fp16_16, tensor var_3985_cast_fp16_17, tensor var_3985_cast_fp16_18, tensor var_3985_cast_fp16_19 = split(axis = var_3985_axis_0, split_sizes = tile_44, x = var_3939_cast_fp16)[name = tensor("op_3985_cast_fp16")]; + tensor aw_561_equation_0 = const()[name = tensor("aw_561_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_561_cast_fp16 = einsum(equation = aw_561_equation_0, values = (var_3964_cast_fp16_0, var_3942_cast_fp16_0))[name = tensor("aw_561_cast_fp16")]; + tensor aw_563_equation_0 = const()[name = tensor("aw_563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_563_cast_fp16 = einsum(equation = aw_563_equation_0, values = (var_3964_cast_fp16_1, var_3942_cast_fp16_1))[name = tensor("aw_563_cast_fp16")]; + tensor aw_565_equation_0 = const()[name = tensor("aw_565_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_565_cast_fp16 = einsum(equation = aw_565_equation_0, values = (var_3964_cast_fp16_2, var_3942_cast_fp16_2))[name = tensor("aw_565_cast_fp16")]; + tensor aw_567_equation_0 = const()[name = tensor("aw_567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_567_cast_fp16 = einsum(equation = aw_567_equation_0, values = (var_3964_cast_fp16_3, var_3942_cast_fp16_3))[name = tensor("aw_567_cast_fp16")]; + tensor aw_569_equation_0 = const()[name = tensor("aw_569_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_569_cast_fp16 = einsum(equation = aw_569_equation_0, values = (var_3964_cast_fp16_4, var_3942_cast_fp16_4))[name = tensor("aw_569_cast_fp16")]; + tensor aw_571_equation_0 = const()[name = tensor("aw_571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_571_cast_fp16 = einsum(equation = aw_571_equation_0, values = (var_3964_cast_fp16_5, var_3942_cast_fp16_5))[name = tensor("aw_571_cast_fp16")]; + tensor aw_573_equation_0 = const()[name = tensor("aw_573_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_573_cast_fp16 = einsum(equation = aw_573_equation_0, values = (var_3964_cast_fp16_6, var_3942_cast_fp16_6))[name = tensor("aw_573_cast_fp16")]; + tensor aw_575_equation_0 = const()[name = tensor("aw_575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_575_cast_fp16 = einsum(equation = aw_575_equation_0, values = (var_3964_cast_fp16_7, var_3942_cast_fp16_7))[name = tensor("aw_575_cast_fp16")]; + tensor aw_577_equation_0 = const()[name = tensor("aw_577_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_577_cast_fp16 = einsum(equation = aw_577_equation_0, values = (var_3964_cast_fp16_8, var_3942_cast_fp16_8))[name = tensor("aw_577_cast_fp16")]; + tensor aw_579_equation_0 = const()[name = tensor("aw_579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_579_cast_fp16 = einsum(equation = aw_579_equation_0, values = (var_3964_cast_fp16_9, var_3942_cast_fp16_9))[name = tensor("aw_579_cast_fp16")]; + tensor aw_581_equation_0 = const()[name = tensor("aw_581_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_581_cast_fp16 = einsum(equation = aw_581_equation_0, values = (var_3964_cast_fp16_10, var_3942_cast_fp16_10))[name = tensor("aw_581_cast_fp16")]; + tensor aw_583_equation_0 = const()[name = tensor("aw_583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_583_cast_fp16 = einsum(equation = aw_583_equation_0, values = (var_3964_cast_fp16_11, var_3942_cast_fp16_11))[name = tensor("aw_583_cast_fp16")]; + tensor aw_585_equation_0 = const()[name = tensor("aw_585_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_585_cast_fp16 = einsum(equation = aw_585_equation_0, values = (var_3964_cast_fp16_12, var_3942_cast_fp16_12))[name = tensor("aw_585_cast_fp16")]; + tensor aw_587_equation_0 = const()[name = tensor("aw_587_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_587_cast_fp16 = einsum(equation = aw_587_equation_0, values = (var_3964_cast_fp16_13, var_3942_cast_fp16_13))[name = tensor("aw_587_cast_fp16")]; + tensor aw_589_equation_0 = const()[name = tensor("aw_589_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_589_cast_fp16 = einsum(equation = aw_589_equation_0, values = (var_3964_cast_fp16_14, var_3942_cast_fp16_14))[name = tensor("aw_589_cast_fp16")]; + tensor aw_591_equation_0 = const()[name = tensor("aw_591_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_591_cast_fp16 = einsum(equation = aw_591_equation_0, values = (var_3964_cast_fp16_15, var_3942_cast_fp16_15))[name = tensor("aw_591_cast_fp16")]; + tensor aw_593_equation_0 = const()[name = tensor("aw_593_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_593_cast_fp16 = einsum(equation = aw_593_equation_0, values = (var_3964_cast_fp16_16, var_3942_cast_fp16_16))[name = tensor("aw_593_cast_fp16")]; + tensor aw_595_equation_0 = const()[name = tensor("aw_595_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_595_cast_fp16 = einsum(equation = aw_595_equation_0, values = (var_3964_cast_fp16_17, var_3942_cast_fp16_17))[name = tensor("aw_595_cast_fp16")]; + tensor aw_597_equation_0 = const()[name = tensor("aw_597_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_597_cast_fp16 = einsum(equation = aw_597_equation_0, values = (var_3964_cast_fp16_18, var_3942_cast_fp16_18))[name = tensor("aw_597_cast_fp16")]; + tensor aw_599_equation_0 = const()[name = tensor("aw_599_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_599_cast_fp16 = einsum(equation = aw_599_equation_0, values = (var_3964_cast_fp16_19, var_3942_cast_fp16_19))[name = tensor("aw_599_cast_fp16")]; + tensor var_4046_cast_fp16 = softmax(axis = var_3890, x = aw_561_cast_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor var_4047_cast_fp16 = softmax(axis = var_3890, x = aw_563_cast_fp16)[name = tensor("op_4047_cast_fp16")]; + tensor var_4048_cast_fp16 = softmax(axis = var_3890, x = aw_565_cast_fp16)[name = tensor("op_4048_cast_fp16")]; + tensor var_4049_cast_fp16 = softmax(axis = var_3890, x = aw_567_cast_fp16)[name = tensor("op_4049_cast_fp16")]; + tensor var_4050_cast_fp16 = softmax(axis = var_3890, x = aw_569_cast_fp16)[name = tensor("op_4050_cast_fp16")]; + tensor var_4051_cast_fp16 = softmax(axis = var_3890, x = aw_571_cast_fp16)[name = tensor("op_4051_cast_fp16")]; + tensor var_4052_cast_fp16 = softmax(axis = var_3890, x = aw_573_cast_fp16)[name = tensor("op_4052_cast_fp16")]; + tensor var_4053_cast_fp16 = softmax(axis = var_3890, x = aw_575_cast_fp16)[name = tensor("op_4053_cast_fp16")]; + tensor var_4054_cast_fp16 = softmax(axis = var_3890, x = aw_577_cast_fp16)[name = tensor("op_4054_cast_fp16")]; + tensor var_4055_cast_fp16 = softmax(axis = var_3890, x = aw_579_cast_fp16)[name = tensor("op_4055_cast_fp16")]; + tensor var_4056_cast_fp16 = softmax(axis = var_3890, x = aw_581_cast_fp16)[name = tensor("op_4056_cast_fp16")]; + tensor var_4057_cast_fp16 = softmax(axis = var_3890, x = aw_583_cast_fp16)[name = tensor("op_4057_cast_fp16")]; + tensor var_4058_cast_fp16 = softmax(axis = var_3890, x = aw_585_cast_fp16)[name = tensor("op_4058_cast_fp16")]; + tensor var_4059_cast_fp16 = softmax(axis = var_3890, x = aw_587_cast_fp16)[name = tensor("op_4059_cast_fp16")]; + tensor var_4060_cast_fp16 = softmax(axis = var_3890, x = aw_589_cast_fp16)[name = tensor("op_4060_cast_fp16")]; + tensor var_4061_cast_fp16 = softmax(axis = var_3890, x = aw_591_cast_fp16)[name = tensor("op_4061_cast_fp16")]; + tensor var_4062_cast_fp16 = softmax(axis = var_3890, x = aw_593_cast_fp16)[name = tensor("op_4062_cast_fp16")]; + tensor var_4063_cast_fp16 = softmax(axis = var_3890, x = aw_595_cast_fp16)[name = tensor("op_4063_cast_fp16")]; + tensor var_4064_cast_fp16 = softmax(axis = var_3890, x = aw_597_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_cast_fp16 = softmax(axis = var_3890, x = aw_599_cast_fp16)[name = tensor("op_4065_cast_fp16")]; + tensor var_4067_equation_0 = const()[name = tensor("op_4067_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4067_cast_fp16 = einsum(equation = var_4067_equation_0, values = (var_3985_cast_fp16_0, var_4046_cast_fp16))[name = tensor("op_4067_cast_fp16")]; + tensor var_4069_equation_0 = const()[name = tensor("op_4069_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4069_cast_fp16 = einsum(equation = var_4069_equation_0, values = (var_3985_cast_fp16_1, var_4047_cast_fp16))[name = tensor("op_4069_cast_fp16")]; + tensor var_4071_equation_0 = const()[name = tensor("op_4071_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4071_cast_fp16 = einsum(equation = var_4071_equation_0, values = (var_3985_cast_fp16_2, var_4048_cast_fp16))[name = tensor("op_4071_cast_fp16")]; + tensor var_4073_equation_0 = const()[name = tensor("op_4073_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4073_cast_fp16 = einsum(equation = var_4073_equation_0, values = (var_3985_cast_fp16_3, var_4049_cast_fp16))[name = tensor("op_4073_cast_fp16")]; + tensor var_4075_equation_0 = const()[name = tensor("op_4075_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4075_cast_fp16 = einsum(equation = var_4075_equation_0, values = (var_3985_cast_fp16_4, var_4050_cast_fp16))[name = tensor("op_4075_cast_fp16")]; + tensor var_4077_equation_0 = const()[name = tensor("op_4077_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4077_cast_fp16 = einsum(equation = var_4077_equation_0, values = (var_3985_cast_fp16_5, var_4051_cast_fp16))[name = tensor("op_4077_cast_fp16")]; + tensor var_4079_equation_0 = const()[name = tensor("op_4079_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4079_cast_fp16 = einsum(equation = var_4079_equation_0, values = (var_3985_cast_fp16_6, var_4052_cast_fp16))[name = tensor("op_4079_cast_fp16")]; + tensor var_4081_equation_0 = const()[name = tensor("op_4081_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4081_cast_fp16 = einsum(equation = var_4081_equation_0, values = (var_3985_cast_fp16_7, var_4053_cast_fp16))[name = tensor("op_4081_cast_fp16")]; + tensor var_4083_equation_0 = const()[name = tensor("op_4083_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4083_cast_fp16 = einsum(equation = var_4083_equation_0, values = (var_3985_cast_fp16_8, var_4054_cast_fp16))[name = tensor("op_4083_cast_fp16")]; + tensor var_4085_equation_0 = const()[name = tensor("op_4085_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4085_cast_fp16 = einsum(equation = var_4085_equation_0, values = (var_3985_cast_fp16_9, var_4055_cast_fp16))[name = tensor("op_4085_cast_fp16")]; + tensor var_4087_equation_0 = const()[name = tensor("op_4087_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4087_cast_fp16 = einsum(equation = var_4087_equation_0, values = (var_3985_cast_fp16_10, var_4056_cast_fp16))[name = tensor("op_4087_cast_fp16")]; + tensor var_4089_equation_0 = const()[name = tensor("op_4089_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4089_cast_fp16 = einsum(equation = var_4089_equation_0, values = (var_3985_cast_fp16_11, var_4057_cast_fp16))[name = tensor("op_4089_cast_fp16")]; + tensor var_4091_equation_0 = const()[name = tensor("op_4091_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4091_cast_fp16 = einsum(equation = var_4091_equation_0, values = (var_3985_cast_fp16_12, var_4058_cast_fp16))[name = tensor("op_4091_cast_fp16")]; + tensor var_4093_equation_0 = const()[name = tensor("op_4093_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4093_cast_fp16 = einsum(equation = var_4093_equation_0, values = (var_3985_cast_fp16_13, var_4059_cast_fp16))[name = tensor("op_4093_cast_fp16")]; + tensor var_4095_equation_0 = const()[name = tensor("op_4095_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4095_cast_fp16 = einsum(equation = var_4095_equation_0, values = (var_3985_cast_fp16_14, var_4060_cast_fp16))[name = tensor("op_4095_cast_fp16")]; + tensor var_4097_equation_0 = const()[name = tensor("op_4097_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4097_cast_fp16 = einsum(equation = var_4097_equation_0, values = (var_3985_cast_fp16_15, var_4061_cast_fp16))[name = tensor("op_4097_cast_fp16")]; + tensor var_4099_equation_0 = const()[name = tensor("op_4099_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4099_cast_fp16 = einsum(equation = var_4099_equation_0, values = (var_3985_cast_fp16_16, var_4062_cast_fp16))[name = tensor("op_4099_cast_fp16")]; + tensor var_4101_equation_0 = const()[name = tensor("op_4101_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4101_cast_fp16 = einsum(equation = var_4101_equation_0, values = (var_3985_cast_fp16_17, var_4063_cast_fp16))[name = tensor("op_4101_cast_fp16")]; + tensor var_4103_equation_0 = const()[name = tensor("op_4103_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4103_cast_fp16 = einsum(equation = var_4103_equation_0, values = (var_3985_cast_fp16_18, var_4064_cast_fp16))[name = tensor("op_4103_cast_fp16")]; + tensor var_4105_equation_0 = const()[name = tensor("op_4105_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4105_cast_fp16 = einsum(equation = var_4105_equation_0, values = (var_3985_cast_fp16_19, var_4065_cast_fp16))[name = tensor("op_4105_cast_fp16")]; + tensor input_145_interleave_0 = const()[name = tensor("input_145_interleave_0"), val = tensor(false)]; + tensor input_145_cast_fp16 = concat(axis = var_3890, interleave = input_145_interleave_0, values = (var_4067_cast_fp16, var_4069_cast_fp16, var_4071_cast_fp16, var_4073_cast_fp16, var_4075_cast_fp16, var_4077_cast_fp16, var_4079_cast_fp16, var_4081_cast_fp16, var_4083_cast_fp16, var_4085_cast_fp16, var_4087_cast_fp16, var_4089_cast_fp16, var_4091_cast_fp16, var_4093_cast_fp16, var_4095_cast_fp16, var_4097_cast_fp16, var_4099_cast_fp16, var_4101_cast_fp16, var_4103_cast_fp16, var_4105_cast_fp16))[name = tensor("input_145_cast_fp16")]; + tensor var_4114_pad_type_0 = const()[name = tensor("op_4114_pad_type_0"), val = tensor("valid")]; + tensor var_4114_strides_0 = const()[name = tensor("op_4114_strides_0"), val = tensor([1, 1])]; + tensor var_4114_pad_0 = const()[name = tensor("op_4114_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4114_dilations_0 = const()[name = tensor("op_4114_dilations_0"), val = tensor([1, 1])]; + tensor var_4114_groups_0 = const()[name = tensor("op_4114_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_out_weight_to_fp16 = const()[name = tensor("blocks_14_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575077312)))]; + tensor blocks_14_attn_out_bias_to_fp16 = const()[name = tensor("blocks_14_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578354176)))]; + tensor var_4114_cast_fp16 = conv(bias = blocks_14_attn_out_bias_to_fp16, dilations = var_4114_dilations_0, groups = var_4114_groups_0, pad = var_4114_pad_0, pad_type = var_4114_pad_type_0, strides = var_4114_strides_0, weight = blocks_14_attn_out_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("op_4114_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_4114_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([1])]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578356800)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578359424)))]; + tensor var_4124_to_fp16 = const()[name = tensor("op_4124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = input_147_beta_0_to_fp16, epsilon = var_4124_to_fp16, gamma = input_147_gamma_0_to_fp16, x = inputs_59_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578362048)))]; + tensor blocks_14_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591469312)))]; + tensor input_149_cast_fp16 = conv(bias = blocks_14_mlp_0_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = blocks_14_mlp_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_4150_pad_type_0 = const()[name = tensor("op_4150_pad_type_0"), val = tensor("valid")]; + tensor var_4150_strides_0 = const()[name = tensor("op_4150_strides_0"), val = tensor([1, 1])]; + tensor var_4150_pad_0 = const()[name = tensor("op_4150_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4150_dilations_0 = const()[name = tensor("op_4150_dilations_0"), val = tensor([1, 1])]; + tensor var_4150_groups_0 = const()[name = tensor("op_4150_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591479616)))]; + tensor blocks_14_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604586880)))]; + tensor var_4150_cast_fp16 = conv(bias = blocks_14_mlp_2_bias_to_fp16, dilations = var_4150_dilations_0, groups = var_4150_groups_0, pad = var_4150_pad_0, pad_type = var_4150_pad_type_0, strides = var_4150_strides_0, weight = blocks_14_mlp_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = var_4150_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_4159 = const()[name = tensor("op_4159"), val = tensor(1)]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([1])]; + tensor input_153_gamma_0_to_fp16 = const()[name = tensor("input_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604589504)))]; + tensor input_153_beta_0_to_fp16 = const()[name = tensor("input_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604592128)))]; + tensor var_4175_to_fp16 = const()[name = tensor("op_4175_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = input_153_beta_0_to_fp16, epsilon = var_4175_to_fp16, gamma = input_153_gamma_0_to_fp16, x = inputs_61_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor q_31_pad_type_0 = const()[name = tensor("q_31_pad_type_0"), val = tensor("valid")]; + tensor q_31_strides_0 = const()[name = tensor("q_31_strides_0"), val = tensor([1, 1])]; + tensor q_31_pad_0 = const()[name = tensor("q_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_31_dilations_0 = const()[name = tensor("q_31_dilations_0"), val = tensor([1, 1])]; + tensor q_31_groups_0 = const()[name = tensor("q_31_groups_0"), val = tensor(1)]; + tensor var_4210_weight_0_to_fp16 = const()[name = tensor("op_4210_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604594752)))]; + tensor var_4210_bias_0_to_fp16 = const()[name = tensor("op_4210_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607871616)))]; + tensor var_4210_cast_fp16 = conv(bias = var_4210_bias_0_to_fp16, dilations = q_31_dilations_0, groups = q_31_groups_0, pad = q_31_pad_0, pad_type = q_31_pad_type_0, strides = q_31_strides_0, weight = var_4210_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("op_4210_cast_fp16")]; + tensor k_31_pad_type_0 = const()[name = tensor("k_31_pad_type_0"), val = tensor("valid")]; + tensor k_31_strides_0 = const()[name = tensor("k_31_strides_0"), val = tensor([1, 1])]; + tensor k_31_pad_0 = const()[name = tensor("k_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_31_dilations_0 = const()[name = tensor("k_31_dilations_0"), val = tensor([1, 1])]; + tensor k_31_groups_0 = const()[name = tensor("k_31_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_key_weight_to_fp16 = const()[name = tensor("blocks_15_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607874240)))]; + tensor k_31_cast_fp16 = conv(dilations = k_31_dilations_0, groups = k_31_groups_0, pad = k_31_pad_0, pad_type = k_31_pad_type_0, strides = k_31_strides_0, weight = blocks_15_attn_key_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor var_4208_pad_type_0 = const()[name = tensor("op_4208_pad_type_0"), val = tensor("valid")]; + tensor var_4208_strides_0 = const()[name = tensor("op_4208_strides_0"), val = tensor([1, 1])]; + tensor var_4208_pad_0 = const()[name = tensor("op_4208_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4208_dilations_0 = const()[name = tensor("op_4208_dilations_0"), val = tensor([1, 1])]; + tensor var_4208_groups_0 = const()[name = tensor("op_4208_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_value_weight_to_fp16 = const()[name = tensor("blocks_15_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611151104)))]; + tensor blocks_15_attn_value_bias_to_fp16 = const()[name = tensor("blocks_15_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614427968)))]; + tensor var_4208_cast_fp16 = conv(bias = blocks_15_attn_value_bias_to_fp16, dilations = var_4208_dilations_0, groups = var_4208_groups_0, pad = var_4208_pad_0, pad_type = var_4208_pad_type_0, strides = var_4208_strides_0, weight = blocks_15_attn_value_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor tile_45 = const()[name = tensor("tile_45"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4211_axis_0 = const()[name = tensor("op_4211_axis_0"), val = tensor(1)]; + tensor var_4211_cast_fp16_0, tensor var_4211_cast_fp16_1, tensor var_4211_cast_fp16_2, tensor var_4211_cast_fp16_3, tensor var_4211_cast_fp16_4, tensor var_4211_cast_fp16_5, tensor var_4211_cast_fp16_6, tensor var_4211_cast_fp16_7, tensor var_4211_cast_fp16_8, tensor var_4211_cast_fp16_9, tensor var_4211_cast_fp16_10, tensor var_4211_cast_fp16_11, tensor var_4211_cast_fp16_12, tensor var_4211_cast_fp16_13, tensor var_4211_cast_fp16_14, tensor var_4211_cast_fp16_15, tensor var_4211_cast_fp16_16, tensor var_4211_cast_fp16_17, tensor var_4211_cast_fp16_18, tensor var_4211_cast_fp16_19 = split(axis = var_4211_axis_0, split_sizes = tile_45, x = var_4210_cast_fp16)[name = tensor("op_4211_cast_fp16")]; + tensor var_4232_perm_0 = const()[name = tensor("op_4232_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_46 = const()[name = tensor("tile_46"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4233_axis_0 = const()[name = tensor("op_4233_axis_0"), val = tensor(3)]; + tensor var_4232_cast_fp16 = transpose(perm = var_4232_perm_0, x = k_31_cast_fp16)[name = tensor("transpose_17")]; + tensor var_4233_cast_fp16_0, tensor var_4233_cast_fp16_1, tensor var_4233_cast_fp16_2, tensor var_4233_cast_fp16_3, tensor var_4233_cast_fp16_4, tensor var_4233_cast_fp16_5, tensor var_4233_cast_fp16_6, tensor var_4233_cast_fp16_7, tensor var_4233_cast_fp16_8, tensor var_4233_cast_fp16_9, tensor var_4233_cast_fp16_10, tensor var_4233_cast_fp16_11, tensor var_4233_cast_fp16_12, tensor var_4233_cast_fp16_13, tensor var_4233_cast_fp16_14, tensor var_4233_cast_fp16_15, tensor var_4233_cast_fp16_16, tensor var_4233_cast_fp16_17, tensor var_4233_cast_fp16_18, tensor var_4233_cast_fp16_19 = split(axis = var_4233_axis_0, split_sizes = tile_46, x = var_4232_cast_fp16)[name = tensor("op_4233_cast_fp16")]; + tensor tile_47 = const()[name = tensor("tile_47"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4254_axis_0 = const()[name = tensor("op_4254_axis_0"), val = tensor(1)]; + tensor var_4254_cast_fp16_0, tensor var_4254_cast_fp16_1, tensor var_4254_cast_fp16_2, tensor var_4254_cast_fp16_3, tensor var_4254_cast_fp16_4, tensor var_4254_cast_fp16_5, tensor var_4254_cast_fp16_6, tensor var_4254_cast_fp16_7, tensor var_4254_cast_fp16_8, tensor var_4254_cast_fp16_9, tensor var_4254_cast_fp16_10, tensor var_4254_cast_fp16_11, tensor var_4254_cast_fp16_12, tensor var_4254_cast_fp16_13, tensor var_4254_cast_fp16_14, tensor var_4254_cast_fp16_15, tensor var_4254_cast_fp16_16, tensor var_4254_cast_fp16_17, tensor var_4254_cast_fp16_18, tensor var_4254_cast_fp16_19 = split(axis = var_4254_axis_0, split_sizes = tile_47, x = var_4208_cast_fp16)[name = tensor("op_4254_cast_fp16")]; + tensor aw_601_equation_0 = const()[name = tensor("aw_601_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_601_cast_fp16 = einsum(equation = aw_601_equation_0, values = (var_4233_cast_fp16_0, var_4211_cast_fp16_0))[name = tensor("aw_601_cast_fp16")]; + tensor aw_603_equation_0 = const()[name = tensor("aw_603_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_603_cast_fp16 = einsum(equation = aw_603_equation_0, values = (var_4233_cast_fp16_1, var_4211_cast_fp16_1))[name = tensor("aw_603_cast_fp16")]; + tensor aw_605_equation_0 = const()[name = tensor("aw_605_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_605_cast_fp16 = einsum(equation = aw_605_equation_0, values = (var_4233_cast_fp16_2, var_4211_cast_fp16_2))[name = tensor("aw_605_cast_fp16")]; + tensor aw_607_equation_0 = const()[name = tensor("aw_607_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_607_cast_fp16 = einsum(equation = aw_607_equation_0, values = (var_4233_cast_fp16_3, var_4211_cast_fp16_3))[name = tensor("aw_607_cast_fp16")]; + tensor aw_609_equation_0 = const()[name = tensor("aw_609_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_609_cast_fp16 = einsum(equation = aw_609_equation_0, values = (var_4233_cast_fp16_4, var_4211_cast_fp16_4))[name = tensor("aw_609_cast_fp16")]; + tensor aw_611_equation_0 = const()[name = tensor("aw_611_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_611_cast_fp16 = einsum(equation = aw_611_equation_0, values = (var_4233_cast_fp16_5, var_4211_cast_fp16_5))[name = tensor("aw_611_cast_fp16")]; + tensor aw_613_equation_0 = const()[name = tensor("aw_613_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_613_cast_fp16 = einsum(equation = aw_613_equation_0, values = (var_4233_cast_fp16_6, var_4211_cast_fp16_6))[name = tensor("aw_613_cast_fp16")]; + tensor aw_615_equation_0 = const()[name = tensor("aw_615_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_615_cast_fp16 = einsum(equation = aw_615_equation_0, values = (var_4233_cast_fp16_7, var_4211_cast_fp16_7))[name = tensor("aw_615_cast_fp16")]; + tensor aw_617_equation_0 = const()[name = tensor("aw_617_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_617_cast_fp16 = einsum(equation = aw_617_equation_0, values = (var_4233_cast_fp16_8, var_4211_cast_fp16_8))[name = tensor("aw_617_cast_fp16")]; + tensor aw_619_equation_0 = const()[name = tensor("aw_619_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_619_cast_fp16 = einsum(equation = aw_619_equation_0, values = (var_4233_cast_fp16_9, var_4211_cast_fp16_9))[name = tensor("aw_619_cast_fp16")]; + tensor aw_621_equation_0 = const()[name = tensor("aw_621_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_621_cast_fp16 = einsum(equation = aw_621_equation_0, values = (var_4233_cast_fp16_10, var_4211_cast_fp16_10))[name = tensor("aw_621_cast_fp16")]; + tensor aw_623_equation_0 = const()[name = tensor("aw_623_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_623_cast_fp16 = einsum(equation = aw_623_equation_0, values = (var_4233_cast_fp16_11, var_4211_cast_fp16_11))[name = tensor("aw_623_cast_fp16")]; + tensor aw_625_equation_0 = const()[name = tensor("aw_625_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_625_cast_fp16 = einsum(equation = aw_625_equation_0, values = (var_4233_cast_fp16_12, var_4211_cast_fp16_12))[name = tensor("aw_625_cast_fp16")]; + tensor aw_627_equation_0 = const()[name = tensor("aw_627_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_627_cast_fp16 = einsum(equation = aw_627_equation_0, values = (var_4233_cast_fp16_13, var_4211_cast_fp16_13))[name = tensor("aw_627_cast_fp16")]; + tensor aw_629_equation_0 = const()[name = tensor("aw_629_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_629_cast_fp16 = einsum(equation = aw_629_equation_0, values = (var_4233_cast_fp16_14, var_4211_cast_fp16_14))[name = tensor("aw_629_cast_fp16")]; + tensor aw_631_equation_0 = const()[name = tensor("aw_631_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_631_cast_fp16 = einsum(equation = aw_631_equation_0, values = (var_4233_cast_fp16_15, var_4211_cast_fp16_15))[name = tensor("aw_631_cast_fp16")]; + tensor aw_633_equation_0 = const()[name = tensor("aw_633_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_633_cast_fp16 = einsum(equation = aw_633_equation_0, values = (var_4233_cast_fp16_16, var_4211_cast_fp16_16))[name = tensor("aw_633_cast_fp16")]; + tensor aw_635_equation_0 = const()[name = tensor("aw_635_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_635_cast_fp16 = einsum(equation = aw_635_equation_0, values = (var_4233_cast_fp16_17, var_4211_cast_fp16_17))[name = tensor("aw_635_cast_fp16")]; + tensor aw_637_equation_0 = const()[name = tensor("aw_637_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_637_cast_fp16 = einsum(equation = aw_637_equation_0, values = (var_4233_cast_fp16_18, var_4211_cast_fp16_18))[name = tensor("aw_637_cast_fp16")]; + tensor aw_639_equation_0 = const()[name = tensor("aw_639_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_639_cast_fp16 = einsum(equation = aw_639_equation_0, values = (var_4233_cast_fp16_19, var_4211_cast_fp16_19))[name = tensor("aw_639_cast_fp16")]; + tensor var_4315_cast_fp16 = softmax(axis = var_4159, x = aw_601_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4316_cast_fp16 = softmax(axis = var_4159, x = aw_603_cast_fp16)[name = tensor("op_4316_cast_fp16")]; + tensor var_4317_cast_fp16 = softmax(axis = var_4159, x = aw_605_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor var_4318_cast_fp16 = softmax(axis = var_4159, x = aw_607_cast_fp16)[name = tensor("op_4318_cast_fp16")]; + tensor var_4319_cast_fp16 = softmax(axis = var_4159, x = aw_609_cast_fp16)[name = tensor("op_4319_cast_fp16")]; + tensor var_4320_cast_fp16 = softmax(axis = var_4159, x = aw_611_cast_fp16)[name = tensor("op_4320_cast_fp16")]; + tensor var_4321_cast_fp16 = softmax(axis = var_4159, x = aw_613_cast_fp16)[name = tensor("op_4321_cast_fp16")]; + tensor var_4322_cast_fp16 = softmax(axis = var_4159, x = aw_615_cast_fp16)[name = tensor("op_4322_cast_fp16")]; + tensor var_4323_cast_fp16 = softmax(axis = var_4159, x = aw_617_cast_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4324_cast_fp16 = softmax(axis = var_4159, x = aw_619_cast_fp16)[name = tensor("op_4324_cast_fp16")]; + tensor var_4325_cast_fp16 = softmax(axis = var_4159, x = aw_621_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor var_4326_cast_fp16 = softmax(axis = var_4159, x = aw_623_cast_fp16)[name = tensor("op_4326_cast_fp16")]; + tensor var_4327_cast_fp16 = softmax(axis = var_4159, x = aw_625_cast_fp16)[name = tensor("op_4327_cast_fp16")]; + tensor var_4328_cast_fp16 = softmax(axis = var_4159, x = aw_627_cast_fp16)[name = tensor("op_4328_cast_fp16")]; + tensor var_4329_cast_fp16 = softmax(axis = var_4159, x = aw_629_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_cast_fp16 = softmax(axis = var_4159, x = aw_631_cast_fp16)[name = tensor("op_4330_cast_fp16")]; + tensor var_4331_cast_fp16 = softmax(axis = var_4159, x = aw_633_cast_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor var_4332_cast_fp16 = softmax(axis = var_4159, x = aw_635_cast_fp16)[name = tensor("op_4332_cast_fp16")]; + tensor var_4333_cast_fp16 = softmax(axis = var_4159, x = aw_637_cast_fp16)[name = tensor("op_4333_cast_fp16")]; + tensor var_4334_cast_fp16 = softmax(axis = var_4159, x = aw_639_cast_fp16)[name = tensor("op_4334_cast_fp16")]; + tensor var_4336_equation_0 = const()[name = tensor("op_4336_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4336_cast_fp16 = einsum(equation = var_4336_equation_0, values = (var_4254_cast_fp16_0, var_4315_cast_fp16))[name = tensor("op_4336_cast_fp16")]; + tensor var_4338_equation_0 = const()[name = tensor("op_4338_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4338_cast_fp16 = einsum(equation = var_4338_equation_0, values = (var_4254_cast_fp16_1, var_4316_cast_fp16))[name = tensor("op_4338_cast_fp16")]; + tensor var_4340_equation_0 = const()[name = tensor("op_4340_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4340_cast_fp16 = einsum(equation = var_4340_equation_0, values = (var_4254_cast_fp16_2, var_4317_cast_fp16))[name = tensor("op_4340_cast_fp16")]; + tensor var_4342_equation_0 = const()[name = tensor("op_4342_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4342_cast_fp16 = einsum(equation = var_4342_equation_0, values = (var_4254_cast_fp16_3, var_4318_cast_fp16))[name = tensor("op_4342_cast_fp16")]; + tensor var_4344_equation_0 = const()[name = tensor("op_4344_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4344_cast_fp16 = einsum(equation = var_4344_equation_0, values = (var_4254_cast_fp16_4, var_4319_cast_fp16))[name = tensor("op_4344_cast_fp16")]; + tensor var_4346_equation_0 = const()[name = tensor("op_4346_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4346_cast_fp16 = einsum(equation = var_4346_equation_0, values = (var_4254_cast_fp16_5, var_4320_cast_fp16))[name = tensor("op_4346_cast_fp16")]; + tensor var_4348_equation_0 = const()[name = tensor("op_4348_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4348_cast_fp16 = einsum(equation = var_4348_equation_0, values = (var_4254_cast_fp16_6, var_4321_cast_fp16))[name = tensor("op_4348_cast_fp16")]; + tensor var_4350_equation_0 = const()[name = tensor("op_4350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4350_cast_fp16 = einsum(equation = var_4350_equation_0, values = (var_4254_cast_fp16_7, var_4322_cast_fp16))[name = tensor("op_4350_cast_fp16")]; + tensor var_4352_equation_0 = const()[name = tensor("op_4352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4352_cast_fp16 = einsum(equation = var_4352_equation_0, values = (var_4254_cast_fp16_8, var_4323_cast_fp16))[name = tensor("op_4352_cast_fp16")]; + tensor var_4354_equation_0 = const()[name = tensor("op_4354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4354_cast_fp16 = einsum(equation = var_4354_equation_0, values = (var_4254_cast_fp16_9, var_4324_cast_fp16))[name = tensor("op_4354_cast_fp16")]; + tensor var_4356_equation_0 = const()[name = tensor("op_4356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4356_cast_fp16 = einsum(equation = var_4356_equation_0, values = (var_4254_cast_fp16_10, var_4325_cast_fp16))[name = tensor("op_4356_cast_fp16")]; + tensor var_4358_equation_0 = const()[name = tensor("op_4358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4358_cast_fp16 = einsum(equation = var_4358_equation_0, values = (var_4254_cast_fp16_11, var_4326_cast_fp16))[name = tensor("op_4358_cast_fp16")]; + tensor var_4360_equation_0 = const()[name = tensor("op_4360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4360_cast_fp16 = einsum(equation = var_4360_equation_0, values = (var_4254_cast_fp16_12, var_4327_cast_fp16))[name = tensor("op_4360_cast_fp16")]; + tensor var_4362_equation_0 = const()[name = tensor("op_4362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4362_cast_fp16 = einsum(equation = var_4362_equation_0, values = (var_4254_cast_fp16_13, var_4328_cast_fp16))[name = tensor("op_4362_cast_fp16")]; + tensor var_4364_equation_0 = const()[name = tensor("op_4364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4364_cast_fp16 = einsum(equation = var_4364_equation_0, values = (var_4254_cast_fp16_14, var_4329_cast_fp16))[name = tensor("op_4364_cast_fp16")]; + tensor var_4366_equation_0 = const()[name = tensor("op_4366_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4366_cast_fp16 = einsum(equation = var_4366_equation_0, values = (var_4254_cast_fp16_15, var_4330_cast_fp16))[name = tensor("op_4366_cast_fp16")]; + tensor var_4368_equation_0 = const()[name = tensor("op_4368_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4368_cast_fp16 = einsum(equation = var_4368_equation_0, values = (var_4254_cast_fp16_16, var_4331_cast_fp16))[name = tensor("op_4368_cast_fp16")]; + tensor var_4370_equation_0 = const()[name = tensor("op_4370_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4370_cast_fp16 = einsum(equation = var_4370_equation_0, values = (var_4254_cast_fp16_17, var_4332_cast_fp16))[name = tensor("op_4370_cast_fp16")]; + tensor var_4372_equation_0 = const()[name = tensor("op_4372_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4372_cast_fp16 = einsum(equation = var_4372_equation_0, values = (var_4254_cast_fp16_18, var_4333_cast_fp16))[name = tensor("op_4372_cast_fp16")]; + tensor var_4374_equation_0 = const()[name = tensor("op_4374_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4374_cast_fp16 = einsum(equation = var_4374_equation_0, values = (var_4254_cast_fp16_19, var_4334_cast_fp16))[name = tensor("op_4374_cast_fp16")]; + tensor input_155_interleave_0 = const()[name = tensor("input_155_interleave_0"), val = tensor(false)]; + tensor input_155_cast_fp16 = concat(axis = var_4159, interleave = input_155_interleave_0, values = (var_4336_cast_fp16, var_4338_cast_fp16, var_4340_cast_fp16, var_4342_cast_fp16, var_4344_cast_fp16, var_4346_cast_fp16, var_4348_cast_fp16, var_4350_cast_fp16, var_4352_cast_fp16, var_4354_cast_fp16, var_4356_cast_fp16, var_4358_cast_fp16, var_4360_cast_fp16, var_4362_cast_fp16, var_4364_cast_fp16, var_4366_cast_fp16, var_4368_cast_fp16, var_4370_cast_fp16, var_4372_cast_fp16, var_4374_cast_fp16))[name = tensor("input_155_cast_fp16")]; + tensor var_4383_pad_type_0 = const()[name = tensor("op_4383_pad_type_0"), val = tensor("valid")]; + tensor var_4383_strides_0 = const()[name = tensor("op_4383_strides_0"), val = tensor([1, 1])]; + tensor var_4383_pad_0 = const()[name = tensor("op_4383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4383_dilations_0 = const()[name = tensor("op_4383_dilations_0"), val = tensor([1, 1])]; + tensor var_4383_groups_0 = const()[name = tensor("op_4383_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_out_weight_to_fp16 = const()[name = tensor("blocks_15_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614430592)))]; + tensor blocks_15_attn_out_bias_to_fp16 = const()[name = tensor("blocks_15_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617707456)))]; + tensor var_4383_cast_fp16 = conv(bias = blocks_15_attn_out_bias_to_fp16, dilations = var_4383_dilations_0, groups = var_4383_groups_0, pad = var_4383_pad_0, pad_type = var_4383_pad_type_0, strides = var_4383_strides_0, weight = blocks_15_attn_out_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("op_4383_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_4383_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor input_157_axes_0 = const()[name = tensor("input_157_axes_0"), val = tensor([1])]; + tensor input_157_gamma_0_to_fp16 = const()[name = tensor("input_157_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617710080)))]; + tensor input_157_beta_0_to_fp16 = const()[name = tensor("input_157_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617712704)))]; + tensor var_4393_to_fp16 = const()[name = tensor("op_4393_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = input_157_beta_0_to_fp16, epsilon = var_4393_to_fp16, gamma = input_157_gamma_0_to_fp16, x = inputs_63_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("valid")]; + tensor input_159_strides_0 = const()[name = tensor("input_159_strides_0"), val = tensor([1, 1])]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_159_dilations_0 = const()[name = tensor("input_159_dilations_0"), val = tensor([1, 1])]; + tensor input_159_groups_0 = const()[name = tensor("input_159_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617715328)))]; + tensor blocks_15_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630822592)))]; + tensor input_159_cast_fp16 = conv(bias = blocks_15_mlp_0_bias_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = blocks_15_mlp_0_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_4419_pad_type_0 = const()[name = tensor("op_4419_pad_type_0"), val = tensor("valid")]; + tensor var_4419_strides_0 = const()[name = tensor("op_4419_strides_0"), val = tensor([1, 1])]; + tensor var_4419_pad_0 = const()[name = tensor("op_4419_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4419_dilations_0 = const()[name = tensor("op_4419_dilations_0"), val = tensor([1, 1])]; + tensor var_4419_groups_0 = const()[name = tensor("op_4419_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630832896)))]; + tensor blocks_15_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643940160)))]; + tensor var_4419_cast_fp16 = conv(bias = blocks_15_mlp_2_bias_to_fp16, dilations = var_4419_dilations_0, groups = var_4419_groups_0, pad = var_4419_pad_0, pad_type = var_4419_pad_type_0, strides = var_4419_strides_0, weight = blocks_15_mlp_2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("op_4419_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = var_4419_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_4428 = const()[name = tensor("op_4428"), val = tensor(1)]; + tensor input_163_axes_0 = const()[name = tensor("input_163_axes_0"), val = tensor([1])]; + tensor input_163_gamma_0_to_fp16 = const()[name = tensor("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643942784)))]; + tensor input_163_beta_0_to_fp16 = const()[name = tensor("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643945408)))]; + tensor var_4444_to_fp16 = const()[name = tensor("op_4444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = input_163_beta_0_to_fp16, epsilon = var_4444_to_fp16, gamma = input_163_gamma_0_to_fp16, x = inputs_65_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor q_33_pad_type_0 = const()[name = tensor("q_33_pad_type_0"), val = tensor("valid")]; + tensor q_33_strides_0 = const()[name = tensor("q_33_strides_0"), val = tensor([1, 1])]; + tensor q_33_pad_0 = const()[name = tensor("q_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_33_dilations_0 = const()[name = tensor("q_33_dilations_0"), val = tensor([1, 1])]; + tensor q_33_groups_0 = const()[name = tensor("q_33_groups_0"), val = tensor(1)]; + tensor var_4479_weight_0_to_fp16 = const()[name = tensor("op_4479_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643948032)))]; + tensor var_4479_bias_0_to_fp16 = const()[name = tensor("op_4479_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647224896)))]; + tensor var_4479_cast_fp16 = conv(bias = var_4479_bias_0_to_fp16, dilations = q_33_dilations_0, groups = q_33_groups_0, pad = q_33_pad_0, pad_type = q_33_pad_type_0, strides = q_33_strides_0, weight = var_4479_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("op_4479_cast_fp16")]; + tensor k_33_pad_type_0 = const()[name = tensor("k_33_pad_type_0"), val = tensor("valid")]; + tensor k_33_strides_0 = const()[name = tensor("k_33_strides_0"), val = tensor([1, 1])]; + tensor k_33_pad_0 = const()[name = tensor("k_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_33_dilations_0 = const()[name = tensor("k_33_dilations_0"), val = tensor([1, 1])]; + tensor k_33_groups_0 = const()[name = tensor("k_33_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_key_weight_to_fp16 = const()[name = tensor("blocks_16_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647227520)))]; + tensor k_33_cast_fp16 = conv(dilations = k_33_dilations_0, groups = k_33_groups_0, pad = k_33_pad_0, pad_type = k_33_pad_type_0, strides = k_33_strides_0, weight = blocks_16_attn_key_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor var_4477_pad_type_0 = const()[name = tensor("op_4477_pad_type_0"), val = tensor("valid")]; + tensor var_4477_strides_0 = const()[name = tensor("op_4477_strides_0"), val = tensor([1, 1])]; + tensor var_4477_pad_0 = const()[name = tensor("op_4477_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4477_dilations_0 = const()[name = tensor("op_4477_dilations_0"), val = tensor([1, 1])]; + tensor var_4477_groups_0 = const()[name = tensor("op_4477_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_value_weight_to_fp16 = const()[name = tensor("blocks_16_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650504384)))]; + tensor blocks_16_attn_value_bias_to_fp16 = const()[name = tensor("blocks_16_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653781248)))]; + tensor var_4477_cast_fp16 = conv(bias = blocks_16_attn_value_bias_to_fp16, dilations = var_4477_dilations_0, groups = var_4477_groups_0, pad = var_4477_pad_0, pad_type = var_4477_pad_type_0, strides = var_4477_strides_0, weight = blocks_16_attn_value_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("op_4477_cast_fp16")]; + tensor tile_48 = const()[name = tensor("tile_48"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4480_axis_0 = const()[name = tensor("op_4480_axis_0"), val = tensor(1)]; + tensor var_4480_cast_fp16_0, tensor var_4480_cast_fp16_1, tensor var_4480_cast_fp16_2, tensor var_4480_cast_fp16_3, tensor var_4480_cast_fp16_4, tensor var_4480_cast_fp16_5, tensor var_4480_cast_fp16_6, tensor var_4480_cast_fp16_7, tensor var_4480_cast_fp16_8, tensor var_4480_cast_fp16_9, tensor var_4480_cast_fp16_10, tensor var_4480_cast_fp16_11, tensor var_4480_cast_fp16_12, tensor var_4480_cast_fp16_13, tensor var_4480_cast_fp16_14, tensor var_4480_cast_fp16_15, tensor var_4480_cast_fp16_16, tensor var_4480_cast_fp16_17, tensor var_4480_cast_fp16_18, tensor var_4480_cast_fp16_19 = split(axis = var_4480_axis_0, split_sizes = tile_48, x = var_4479_cast_fp16)[name = tensor("op_4480_cast_fp16")]; + tensor var_4501_perm_0 = const()[name = tensor("op_4501_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_49 = const()[name = tensor("tile_49"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4502_axis_0 = const()[name = tensor("op_4502_axis_0"), val = tensor(3)]; + tensor var_4501_cast_fp16 = transpose(perm = var_4501_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_16")]; + tensor var_4502_cast_fp16_0, tensor var_4502_cast_fp16_1, tensor var_4502_cast_fp16_2, tensor var_4502_cast_fp16_3, tensor var_4502_cast_fp16_4, tensor var_4502_cast_fp16_5, tensor var_4502_cast_fp16_6, tensor var_4502_cast_fp16_7, tensor var_4502_cast_fp16_8, tensor var_4502_cast_fp16_9, tensor var_4502_cast_fp16_10, tensor var_4502_cast_fp16_11, tensor var_4502_cast_fp16_12, tensor var_4502_cast_fp16_13, tensor var_4502_cast_fp16_14, tensor var_4502_cast_fp16_15, tensor var_4502_cast_fp16_16, tensor var_4502_cast_fp16_17, tensor var_4502_cast_fp16_18, tensor var_4502_cast_fp16_19 = split(axis = var_4502_axis_0, split_sizes = tile_49, x = var_4501_cast_fp16)[name = tensor("op_4502_cast_fp16")]; + tensor tile_50 = const()[name = tensor("tile_50"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4523_axis_0 = const()[name = tensor("op_4523_axis_0"), val = tensor(1)]; + tensor var_4523_cast_fp16_0, tensor var_4523_cast_fp16_1, tensor var_4523_cast_fp16_2, tensor var_4523_cast_fp16_3, tensor var_4523_cast_fp16_4, tensor var_4523_cast_fp16_5, tensor var_4523_cast_fp16_6, tensor var_4523_cast_fp16_7, tensor var_4523_cast_fp16_8, tensor var_4523_cast_fp16_9, tensor var_4523_cast_fp16_10, tensor var_4523_cast_fp16_11, tensor var_4523_cast_fp16_12, tensor var_4523_cast_fp16_13, tensor var_4523_cast_fp16_14, tensor var_4523_cast_fp16_15, tensor var_4523_cast_fp16_16, tensor var_4523_cast_fp16_17, tensor var_4523_cast_fp16_18, tensor var_4523_cast_fp16_19 = split(axis = var_4523_axis_0, split_sizes = tile_50, x = var_4477_cast_fp16)[name = tensor("op_4523_cast_fp16")]; + tensor aw_641_equation_0 = const()[name = tensor("aw_641_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_641_cast_fp16 = einsum(equation = aw_641_equation_0, values = (var_4502_cast_fp16_0, var_4480_cast_fp16_0))[name = tensor("aw_641_cast_fp16")]; + tensor aw_643_equation_0 = const()[name = tensor("aw_643_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_643_cast_fp16 = einsum(equation = aw_643_equation_0, values = (var_4502_cast_fp16_1, var_4480_cast_fp16_1))[name = tensor("aw_643_cast_fp16")]; + tensor aw_645_equation_0 = const()[name = tensor("aw_645_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_645_cast_fp16 = einsum(equation = aw_645_equation_0, values = (var_4502_cast_fp16_2, var_4480_cast_fp16_2))[name = tensor("aw_645_cast_fp16")]; + tensor aw_647_equation_0 = const()[name = tensor("aw_647_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_647_cast_fp16 = einsum(equation = aw_647_equation_0, values = (var_4502_cast_fp16_3, var_4480_cast_fp16_3))[name = tensor("aw_647_cast_fp16")]; + tensor aw_649_equation_0 = const()[name = tensor("aw_649_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_649_cast_fp16 = einsum(equation = aw_649_equation_0, values = (var_4502_cast_fp16_4, var_4480_cast_fp16_4))[name = tensor("aw_649_cast_fp16")]; + tensor aw_651_equation_0 = const()[name = tensor("aw_651_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_651_cast_fp16 = einsum(equation = aw_651_equation_0, values = (var_4502_cast_fp16_5, var_4480_cast_fp16_5))[name = tensor("aw_651_cast_fp16")]; + tensor aw_653_equation_0 = const()[name = tensor("aw_653_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_653_cast_fp16 = einsum(equation = aw_653_equation_0, values = (var_4502_cast_fp16_6, var_4480_cast_fp16_6))[name = tensor("aw_653_cast_fp16")]; + tensor aw_655_equation_0 = const()[name = tensor("aw_655_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_655_cast_fp16 = einsum(equation = aw_655_equation_0, values = (var_4502_cast_fp16_7, var_4480_cast_fp16_7))[name = tensor("aw_655_cast_fp16")]; + tensor aw_657_equation_0 = const()[name = tensor("aw_657_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_657_cast_fp16 = einsum(equation = aw_657_equation_0, values = (var_4502_cast_fp16_8, var_4480_cast_fp16_8))[name = tensor("aw_657_cast_fp16")]; + tensor aw_659_equation_0 = const()[name = tensor("aw_659_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_659_cast_fp16 = einsum(equation = aw_659_equation_0, values = (var_4502_cast_fp16_9, var_4480_cast_fp16_9))[name = tensor("aw_659_cast_fp16")]; + tensor aw_661_equation_0 = const()[name = tensor("aw_661_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_661_cast_fp16 = einsum(equation = aw_661_equation_0, values = (var_4502_cast_fp16_10, var_4480_cast_fp16_10))[name = tensor("aw_661_cast_fp16")]; + tensor aw_663_equation_0 = const()[name = tensor("aw_663_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_663_cast_fp16 = einsum(equation = aw_663_equation_0, values = (var_4502_cast_fp16_11, var_4480_cast_fp16_11))[name = tensor("aw_663_cast_fp16")]; + tensor aw_665_equation_0 = const()[name = tensor("aw_665_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_665_cast_fp16 = einsum(equation = aw_665_equation_0, values = (var_4502_cast_fp16_12, var_4480_cast_fp16_12))[name = tensor("aw_665_cast_fp16")]; + tensor aw_667_equation_0 = const()[name = tensor("aw_667_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_667_cast_fp16 = einsum(equation = aw_667_equation_0, values = (var_4502_cast_fp16_13, var_4480_cast_fp16_13))[name = tensor("aw_667_cast_fp16")]; + tensor aw_669_equation_0 = const()[name = tensor("aw_669_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_669_cast_fp16 = einsum(equation = aw_669_equation_0, values = (var_4502_cast_fp16_14, var_4480_cast_fp16_14))[name = tensor("aw_669_cast_fp16")]; + tensor aw_671_equation_0 = const()[name = tensor("aw_671_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_671_cast_fp16 = einsum(equation = aw_671_equation_0, values = (var_4502_cast_fp16_15, var_4480_cast_fp16_15))[name = tensor("aw_671_cast_fp16")]; + tensor aw_673_equation_0 = const()[name = tensor("aw_673_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_673_cast_fp16 = einsum(equation = aw_673_equation_0, values = (var_4502_cast_fp16_16, var_4480_cast_fp16_16))[name = tensor("aw_673_cast_fp16")]; + tensor aw_675_equation_0 = const()[name = tensor("aw_675_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_675_cast_fp16 = einsum(equation = aw_675_equation_0, values = (var_4502_cast_fp16_17, var_4480_cast_fp16_17))[name = tensor("aw_675_cast_fp16")]; + tensor aw_677_equation_0 = const()[name = tensor("aw_677_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_677_cast_fp16 = einsum(equation = aw_677_equation_0, values = (var_4502_cast_fp16_18, var_4480_cast_fp16_18))[name = tensor("aw_677_cast_fp16")]; + tensor aw_679_equation_0 = const()[name = tensor("aw_679_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_679_cast_fp16 = einsum(equation = aw_679_equation_0, values = (var_4502_cast_fp16_19, var_4480_cast_fp16_19))[name = tensor("aw_679_cast_fp16")]; + tensor var_4584_cast_fp16 = softmax(axis = var_4428, x = aw_641_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = softmax(axis = var_4428, x = aw_643_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor var_4586_cast_fp16 = softmax(axis = var_4428, x = aw_645_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = softmax(axis = var_4428, x = aw_647_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor var_4588_cast_fp16 = softmax(axis = var_4428, x = aw_649_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = softmax(axis = var_4428, x = aw_651_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor var_4590_cast_fp16 = softmax(axis = var_4428, x = aw_653_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_4428, x = aw_655_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor var_4592_cast_fp16 = softmax(axis = var_4428, x = aw_657_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = softmax(axis = var_4428, x = aw_659_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor var_4594_cast_fp16 = softmax(axis = var_4428, x = aw_661_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = softmax(axis = var_4428, x = aw_663_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor var_4596_cast_fp16 = softmax(axis = var_4428, x = aw_665_cast_fp16)[name = tensor("op_4596_cast_fp16")]; + tensor var_4597_cast_fp16 = softmax(axis = var_4428, x = aw_667_cast_fp16)[name = tensor("op_4597_cast_fp16")]; + tensor var_4598_cast_fp16 = softmax(axis = var_4428, x = aw_669_cast_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4599_cast_fp16 = softmax(axis = var_4428, x = aw_671_cast_fp16)[name = tensor("op_4599_cast_fp16")]; + tensor var_4600_cast_fp16 = softmax(axis = var_4428, x = aw_673_cast_fp16)[name = tensor("op_4600_cast_fp16")]; + tensor var_4601_cast_fp16 = softmax(axis = var_4428, x = aw_675_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor var_4602_cast_fp16 = softmax(axis = var_4428, x = aw_677_cast_fp16)[name = tensor("op_4602_cast_fp16")]; + tensor var_4603_cast_fp16 = softmax(axis = var_4428, x = aw_679_cast_fp16)[name = tensor("op_4603_cast_fp16")]; + tensor var_4605_equation_0 = const()[name = tensor("op_4605_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4605_cast_fp16 = einsum(equation = var_4605_equation_0, values = (var_4523_cast_fp16_0, var_4584_cast_fp16))[name = tensor("op_4605_cast_fp16")]; + tensor var_4607_equation_0 = const()[name = tensor("op_4607_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4607_cast_fp16 = einsum(equation = var_4607_equation_0, values = (var_4523_cast_fp16_1, var_4585_cast_fp16))[name = tensor("op_4607_cast_fp16")]; + tensor var_4609_equation_0 = const()[name = tensor("op_4609_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4609_cast_fp16 = einsum(equation = var_4609_equation_0, values = (var_4523_cast_fp16_2, var_4586_cast_fp16))[name = tensor("op_4609_cast_fp16")]; + tensor var_4611_equation_0 = const()[name = tensor("op_4611_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4611_cast_fp16 = einsum(equation = var_4611_equation_0, values = (var_4523_cast_fp16_3, var_4587_cast_fp16))[name = tensor("op_4611_cast_fp16")]; + tensor var_4613_equation_0 = const()[name = tensor("op_4613_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4613_cast_fp16 = einsum(equation = var_4613_equation_0, values = (var_4523_cast_fp16_4, var_4588_cast_fp16))[name = tensor("op_4613_cast_fp16")]; + tensor var_4615_equation_0 = const()[name = tensor("op_4615_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4615_cast_fp16 = einsum(equation = var_4615_equation_0, values = (var_4523_cast_fp16_5, var_4589_cast_fp16))[name = tensor("op_4615_cast_fp16")]; + tensor var_4617_equation_0 = const()[name = tensor("op_4617_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4617_cast_fp16 = einsum(equation = var_4617_equation_0, values = (var_4523_cast_fp16_6, var_4590_cast_fp16))[name = tensor("op_4617_cast_fp16")]; + tensor var_4619_equation_0 = const()[name = tensor("op_4619_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4619_cast_fp16 = einsum(equation = var_4619_equation_0, values = (var_4523_cast_fp16_7, var_4591_cast_fp16))[name = tensor("op_4619_cast_fp16")]; + tensor var_4621_equation_0 = const()[name = tensor("op_4621_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4621_cast_fp16 = einsum(equation = var_4621_equation_0, values = (var_4523_cast_fp16_8, var_4592_cast_fp16))[name = tensor("op_4621_cast_fp16")]; + tensor var_4623_equation_0 = const()[name = tensor("op_4623_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4623_cast_fp16 = einsum(equation = var_4623_equation_0, values = (var_4523_cast_fp16_9, var_4593_cast_fp16))[name = tensor("op_4623_cast_fp16")]; + tensor var_4625_equation_0 = const()[name = tensor("op_4625_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4625_cast_fp16 = einsum(equation = var_4625_equation_0, values = (var_4523_cast_fp16_10, var_4594_cast_fp16))[name = tensor("op_4625_cast_fp16")]; + tensor var_4627_equation_0 = const()[name = tensor("op_4627_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4627_cast_fp16 = einsum(equation = var_4627_equation_0, values = (var_4523_cast_fp16_11, var_4595_cast_fp16))[name = tensor("op_4627_cast_fp16")]; + tensor var_4629_equation_0 = const()[name = tensor("op_4629_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4629_cast_fp16 = einsum(equation = var_4629_equation_0, values = (var_4523_cast_fp16_12, var_4596_cast_fp16))[name = tensor("op_4629_cast_fp16")]; + tensor var_4631_equation_0 = const()[name = tensor("op_4631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4631_cast_fp16 = einsum(equation = var_4631_equation_0, values = (var_4523_cast_fp16_13, var_4597_cast_fp16))[name = tensor("op_4631_cast_fp16")]; + tensor var_4633_equation_0 = const()[name = tensor("op_4633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4633_cast_fp16 = einsum(equation = var_4633_equation_0, values = (var_4523_cast_fp16_14, var_4598_cast_fp16))[name = tensor("op_4633_cast_fp16")]; + tensor var_4635_equation_0 = const()[name = tensor("op_4635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4635_cast_fp16 = einsum(equation = var_4635_equation_0, values = (var_4523_cast_fp16_15, var_4599_cast_fp16))[name = tensor("op_4635_cast_fp16")]; + tensor var_4637_equation_0 = const()[name = tensor("op_4637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4637_cast_fp16 = einsum(equation = var_4637_equation_0, values = (var_4523_cast_fp16_16, var_4600_cast_fp16))[name = tensor("op_4637_cast_fp16")]; + tensor var_4639_equation_0 = const()[name = tensor("op_4639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4639_cast_fp16 = einsum(equation = var_4639_equation_0, values = (var_4523_cast_fp16_17, var_4601_cast_fp16))[name = tensor("op_4639_cast_fp16")]; + tensor var_4641_equation_0 = const()[name = tensor("op_4641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4641_cast_fp16 = einsum(equation = var_4641_equation_0, values = (var_4523_cast_fp16_18, var_4602_cast_fp16))[name = tensor("op_4641_cast_fp16")]; + tensor var_4643_equation_0 = const()[name = tensor("op_4643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4643_cast_fp16 = einsum(equation = var_4643_equation_0, values = (var_4523_cast_fp16_19, var_4603_cast_fp16))[name = tensor("op_4643_cast_fp16")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165_cast_fp16 = concat(axis = var_4428, interleave = input_165_interleave_0, values = (var_4605_cast_fp16, var_4607_cast_fp16, var_4609_cast_fp16, var_4611_cast_fp16, var_4613_cast_fp16, var_4615_cast_fp16, var_4617_cast_fp16, var_4619_cast_fp16, var_4621_cast_fp16, var_4623_cast_fp16, var_4625_cast_fp16, var_4627_cast_fp16, var_4629_cast_fp16, var_4631_cast_fp16, var_4633_cast_fp16, var_4635_cast_fp16, var_4637_cast_fp16, var_4639_cast_fp16, var_4641_cast_fp16, var_4643_cast_fp16))[name = tensor("input_165_cast_fp16")]; + tensor var_4652_pad_type_0 = const()[name = tensor("op_4652_pad_type_0"), val = tensor("valid")]; + tensor var_4652_strides_0 = const()[name = tensor("op_4652_strides_0"), val = tensor([1, 1])]; + tensor var_4652_pad_0 = const()[name = tensor("op_4652_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4652_dilations_0 = const()[name = tensor("op_4652_dilations_0"), val = tensor([1, 1])]; + tensor var_4652_groups_0 = const()[name = tensor("op_4652_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_out_weight_to_fp16 = const()[name = tensor("blocks_16_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653783872)))]; + tensor blocks_16_attn_out_bias_to_fp16 = const()[name = tensor("blocks_16_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657060736)))]; + tensor var_4652_cast_fp16 = conv(bias = blocks_16_attn_out_bias_to_fp16, dilations = var_4652_dilations_0, groups = var_4652_groups_0, pad = var_4652_pad_0, pad_type = var_4652_pad_type_0, strides = var_4652_strides_0, weight = blocks_16_attn_out_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("op_4652_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = var_4652_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([1])]; + tensor input_167_gamma_0_to_fp16 = const()[name = tensor("input_167_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657063360)))]; + tensor input_167_beta_0_to_fp16 = const()[name = tensor("input_167_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657065984)))]; + tensor var_4662_to_fp16 = const()[name = tensor("op_4662_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = input_167_beta_0_to_fp16, epsilon = var_4662_to_fp16, gamma = input_167_gamma_0_to_fp16, x = inputs_67_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("valid")]; + tensor input_169_strides_0 = const()[name = tensor("input_169_strides_0"), val = tensor([1, 1])]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_169_dilations_0 = const()[name = tensor("input_169_dilations_0"), val = tensor([1, 1])]; + tensor input_169_groups_0 = const()[name = tensor("input_169_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657068608)))]; + tensor blocks_16_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670175872)))]; + tensor input_169_cast_fp16 = conv(bias = blocks_16_mlp_0_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = blocks_16_mlp_0_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_4688_pad_type_0 = const()[name = tensor("op_4688_pad_type_0"), val = tensor("valid")]; + tensor var_4688_strides_0 = const()[name = tensor("op_4688_strides_0"), val = tensor([1, 1])]; + tensor var_4688_pad_0 = const()[name = tensor("op_4688_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4688_dilations_0 = const()[name = tensor("op_4688_dilations_0"), val = tensor([1, 1])]; + tensor var_4688_groups_0 = const()[name = tensor("op_4688_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670186176)))]; + tensor blocks_16_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683293440)))]; + tensor var_4688_cast_fp16 = conv(bias = blocks_16_mlp_2_bias_to_fp16, dilations = var_4688_dilations_0, groups = var_4688_groups_0, pad = var_4688_pad_0, pad_type = var_4688_pad_type_0, strides = var_4688_strides_0, weight = blocks_16_mlp_2_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_4688_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4697 = const()[name = tensor("op_4697"), val = tensor(1)]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([1])]; + tensor input_173_gamma_0_to_fp16 = const()[name = tensor("input_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683296064)))]; + tensor input_173_beta_0_to_fp16 = const()[name = tensor("input_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683298688)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = input_173_beta_0_to_fp16, epsilon = var_4713_to_fp16, gamma = input_173_gamma_0_to_fp16, x = inputs_69_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor q_35_pad_type_0 = const()[name = tensor("q_35_pad_type_0"), val = tensor("valid")]; + tensor q_35_strides_0 = const()[name = tensor("q_35_strides_0"), val = tensor([1, 1])]; + tensor q_35_pad_0 = const()[name = tensor("q_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_35_dilations_0 = const()[name = tensor("q_35_dilations_0"), val = tensor([1, 1])]; + tensor q_35_groups_0 = const()[name = tensor("q_35_groups_0"), val = tensor(1)]; + tensor var_4748_weight_0_to_fp16 = const()[name = tensor("op_4748_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683301312)))]; + tensor var_4748_bias_0_to_fp16 = const()[name = tensor("op_4748_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686578176)))]; + tensor var_4748_cast_fp16 = conv(bias = var_4748_bias_0_to_fp16, dilations = q_35_dilations_0, groups = q_35_groups_0, pad = q_35_pad_0, pad_type = q_35_pad_type_0, strides = q_35_strides_0, weight = var_4748_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4748_cast_fp16")]; + tensor k_35_pad_type_0 = const()[name = tensor("k_35_pad_type_0"), val = tensor("valid")]; + tensor k_35_strides_0 = const()[name = tensor("k_35_strides_0"), val = tensor([1, 1])]; + tensor k_35_pad_0 = const()[name = tensor("k_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_35_dilations_0 = const()[name = tensor("k_35_dilations_0"), val = tensor([1, 1])]; + tensor k_35_groups_0 = const()[name = tensor("k_35_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_key_weight_to_fp16 = const()[name = tensor("blocks_17_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686580800)))]; + tensor k_35_cast_fp16 = conv(dilations = k_35_dilations_0, groups = k_35_groups_0, pad = k_35_pad_0, pad_type = k_35_pad_type_0, strides = k_35_strides_0, weight = blocks_17_attn_key_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_4746_pad_type_0 = const()[name = tensor("op_4746_pad_type_0"), val = tensor("valid")]; + tensor var_4746_strides_0 = const()[name = tensor("op_4746_strides_0"), val = tensor([1, 1])]; + tensor var_4746_pad_0 = const()[name = tensor("op_4746_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4746_dilations_0 = const()[name = tensor("op_4746_dilations_0"), val = tensor([1, 1])]; + tensor var_4746_groups_0 = const()[name = tensor("op_4746_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_value_weight_to_fp16 = const()[name = tensor("blocks_17_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689857664)))]; + tensor blocks_17_attn_value_bias_to_fp16 = const()[name = tensor("blocks_17_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693134528)))]; + tensor var_4746_cast_fp16 = conv(bias = blocks_17_attn_value_bias_to_fp16, dilations = var_4746_dilations_0, groups = var_4746_groups_0, pad = var_4746_pad_0, pad_type = var_4746_pad_type_0, strides = var_4746_strides_0, weight = blocks_17_attn_value_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4746_cast_fp16")]; + tensor tile_51 = const()[name = tensor("tile_51"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4749_axis_0 = const()[name = tensor("op_4749_axis_0"), val = tensor(1)]; + tensor var_4749_cast_fp16_0, tensor var_4749_cast_fp16_1, tensor var_4749_cast_fp16_2, tensor var_4749_cast_fp16_3, tensor var_4749_cast_fp16_4, tensor var_4749_cast_fp16_5, tensor var_4749_cast_fp16_6, tensor var_4749_cast_fp16_7, tensor var_4749_cast_fp16_8, tensor var_4749_cast_fp16_9, tensor var_4749_cast_fp16_10, tensor var_4749_cast_fp16_11, tensor var_4749_cast_fp16_12, tensor var_4749_cast_fp16_13, tensor var_4749_cast_fp16_14, tensor var_4749_cast_fp16_15, tensor var_4749_cast_fp16_16, tensor var_4749_cast_fp16_17, tensor var_4749_cast_fp16_18, tensor var_4749_cast_fp16_19 = split(axis = var_4749_axis_0, split_sizes = tile_51, x = var_4748_cast_fp16)[name = tensor("op_4749_cast_fp16")]; + tensor var_4770_perm_0 = const()[name = tensor("op_4770_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_52 = const()[name = tensor("tile_52"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4771_axis_0 = const()[name = tensor("op_4771_axis_0"), val = tensor(3)]; + tensor var_4770_cast_fp16 = transpose(perm = var_4770_perm_0, x = k_35_cast_fp16)[name = tensor("transpose_15")]; + tensor var_4771_cast_fp16_0, tensor var_4771_cast_fp16_1, tensor var_4771_cast_fp16_2, tensor var_4771_cast_fp16_3, tensor var_4771_cast_fp16_4, tensor var_4771_cast_fp16_5, tensor var_4771_cast_fp16_6, tensor var_4771_cast_fp16_7, tensor var_4771_cast_fp16_8, tensor var_4771_cast_fp16_9, tensor var_4771_cast_fp16_10, tensor var_4771_cast_fp16_11, tensor var_4771_cast_fp16_12, tensor var_4771_cast_fp16_13, tensor var_4771_cast_fp16_14, tensor var_4771_cast_fp16_15, tensor var_4771_cast_fp16_16, tensor var_4771_cast_fp16_17, tensor var_4771_cast_fp16_18, tensor var_4771_cast_fp16_19 = split(axis = var_4771_axis_0, split_sizes = tile_52, x = var_4770_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor tile_53 = const()[name = tensor("tile_53"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4792_axis_0 = const()[name = tensor("op_4792_axis_0"), val = tensor(1)]; + tensor var_4792_cast_fp16_0, tensor var_4792_cast_fp16_1, tensor var_4792_cast_fp16_2, tensor var_4792_cast_fp16_3, tensor var_4792_cast_fp16_4, tensor var_4792_cast_fp16_5, tensor var_4792_cast_fp16_6, tensor var_4792_cast_fp16_7, tensor var_4792_cast_fp16_8, tensor var_4792_cast_fp16_9, tensor var_4792_cast_fp16_10, tensor var_4792_cast_fp16_11, tensor var_4792_cast_fp16_12, tensor var_4792_cast_fp16_13, tensor var_4792_cast_fp16_14, tensor var_4792_cast_fp16_15, tensor var_4792_cast_fp16_16, tensor var_4792_cast_fp16_17, tensor var_4792_cast_fp16_18, tensor var_4792_cast_fp16_19 = split(axis = var_4792_axis_0, split_sizes = tile_53, x = var_4746_cast_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor aw_681_equation_0 = const()[name = tensor("aw_681_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_681_cast_fp16 = einsum(equation = aw_681_equation_0, values = (var_4771_cast_fp16_0, var_4749_cast_fp16_0))[name = tensor("aw_681_cast_fp16")]; + tensor aw_683_equation_0 = const()[name = tensor("aw_683_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_683_cast_fp16 = einsum(equation = aw_683_equation_0, values = (var_4771_cast_fp16_1, var_4749_cast_fp16_1))[name = tensor("aw_683_cast_fp16")]; + tensor aw_685_equation_0 = const()[name = tensor("aw_685_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_685_cast_fp16 = einsum(equation = aw_685_equation_0, values = (var_4771_cast_fp16_2, var_4749_cast_fp16_2))[name = tensor("aw_685_cast_fp16")]; + tensor aw_687_equation_0 = const()[name = tensor("aw_687_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_687_cast_fp16 = einsum(equation = aw_687_equation_0, values = (var_4771_cast_fp16_3, var_4749_cast_fp16_3))[name = tensor("aw_687_cast_fp16")]; + tensor aw_689_equation_0 = const()[name = tensor("aw_689_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_689_cast_fp16 = einsum(equation = aw_689_equation_0, values = (var_4771_cast_fp16_4, var_4749_cast_fp16_4))[name = tensor("aw_689_cast_fp16")]; + tensor aw_691_equation_0 = const()[name = tensor("aw_691_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_691_cast_fp16 = einsum(equation = aw_691_equation_0, values = (var_4771_cast_fp16_5, var_4749_cast_fp16_5))[name = tensor("aw_691_cast_fp16")]; + tensor aw_693_equation_0 = const()[name = tensor("aw_693_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_693_cast_fp16 = einsum(equation = aw_693_equation_0, values = (var_4771_cast_fp16_6, var_4749_cast_fp16_6))[name = tensor("aw_693_cast_fp16")]; + tensor aw_695_equation_0 = const()[name = tensor("aw_695_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_695_cast_fp16 = einsum(equation = aw_695_equation_0, values = (var_4771_cast_fp16_7, var_4749_cast_fp16_7))[name = tensor("aw_695_cast_fp16")]; + tensor aw_697_equation_0 = const()[name = tensor("aw_697_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_697_cast_fp16 = einsum(equation = aw_697_equation_0, values = (var_4771_cast_fp16_8, var_4749_cast_fp16_8))[name = tensor("aw_697_cast_fp16")]; + tensor aw_699_equation_0 = const()[name = tensor("aw_699_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_699_cast_fp16 = einsum(equation = aw_699_equation_0, values = (var_4771_cast_fp16_9, var_4749_cast_fp16_9))[name = tensor("aw_699_cast_fp16")]; + tensor aw_701_equation_0 = const()[name = tensor("aw_701_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_701_cast_fp16 = einsum(equation = aw_701_equation_0, values = (var_4771_cast_fp16_10, var_4749_cast_fp16_10))[name = tensor("aw_701_cast_fp16")]; + tensor aw_703_equation_0 = const()[name = tensor("aw_703_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_703_cast_fp16 = einsum(equation = aw_703_equation_0, values = (var_4771_cast_fp16_11, var_4749_cast_fp16_11))[name = tensor("aw_703_cast_fp16")]; + tensor aw_705_equation_0 = const()[name = tensor("aw_705_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_705_cast_fp16 = einsum(equation = aw_705_equation_0, values = (var_4771_cast_fp16_12, var_4749_cast_fp16_12))[name = tensor("aw_705_cast_fp16")]; + tensor aw_707_equation_0 = const()[name = tensor("aw_707_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_707_cast_fp16 = einsum(equation = aw_707_equation_0, values = (var_4771_cast_fp16_13, var_4749_cast_fp16_13))[name = tensor("aw_707_cast_fp16")]; + tensor aw_709_equation_0 = const()[name = tensor("aw_709_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_709_cast_fp16 = einsum(equation = aw_709_equation_0, values = (var_4771_cast_fp16_14, var_4749_cast_fp16_14))[name = tensor("aw_709_cast_fp16")]; + tensor aw_711_equation_0 = const()[name = tensor("aw_711_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_711_cast_fp16 = einsum(equation = aw_711_equation_0, values = (var_4771_cast_fp16_15, var_4749_cast_fp16_15))[name = tensor("aw_711_cast_fp16")]; + tensor aw_713_equation_0 = const()[name = tensor("aw_713_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_713_cast_fp16 = einsum(equation = aw_713_equation_0, values = (var_4771_cast_fp16_16, var_4749_cast_fp16_16))[name = tensor("aw_713_cast_fp16")]; + tensor aw_715_equation_0 = const()[name = tensor("aw_715_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_715_cast_fp16 = einsum(equation = aw_715_equation_0, values = (var_4771_cast_fp16_17, var_4749_cast_fp16_17))[name = tensor("aw_715_cast_fp16")]; + tensor aw_717_equation_0 = const()[name = tensor("aw_717_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_717_cast_fp16 = einsum(equation = aw_717_equation_0, values = (var_4771_cast_fp16_18, var_4749_cast_fp16_18))[name = tensor("aw_717_cast_fp16")]; + tensor aw_719_equation_0 = const()[name = tensor("aw_719_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_719_cast_fp16 = einsum(equation = aw_719_equation_0, values = (var_4771_cast_fp16_19, var_4749_cast_fp16_19))[name = tensor("aw_719_cast_fp16")]; + tensor var_4853_cast_fp16 = softmax(axis = var_4697, x = aw_681_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = softmax(axis = var_4697, x = aw_683_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor var_4855_cast_fp16 = softmax(axis = var_4697, x = aw_685_cast_fp16)[name = tensor("op_4855_cast_fp16")]; + tensor var_4856_cast_fp16 = softmax(axis = var_4697, x = aw_687_cast_fp16)[name = tensor("op_4856_cast_fp16")]; + tensor var_4857_cast_fp16 = softmax(axis = var_4697, x = aw_689_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = softmax(axis = var_4697, x = aw_691_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor var_4859_cast_fp16 = softmax(axis = var_4697, x = aw_693_cast_fp16)[name = tensor("op_4859_cast_fp16")]; + tensor var_4860_cast_fp16 = softmax(axis = var_4697, x = aw_695_cast_fp16)[name = tensor("op_4860_cast_fp16")]; + tensor var_4861_cast_fp16 = softmax(axis = var_4697, x = aw_697_cast_fp16)[name = tensor("op_4861_cast_fp16")]; + tensor var_4862_cast_fp16 = softmax(axis = var_4697, x = aw_699_cast_fp16)[name = tensor("op_4862_cast_fp16")]; + tensor var_4863_cast_fp16 = softmax(axis = var_4697, x = aw_701_cast_fp16)[name = tensor("op_4863_cast_fp16")]; + tensor var_4864_cast_fp16 = softmax(axis = var_4697, x = aw_703_cast_fp16)[name = tensor("op_4864_cast_fp16")]; + tensor var_4865_cast_fp16 = softmax(axis = var_4697, x = aw_705_cast_fp16)[name = tensor("op_4865_cast_fp16")]; + tensor var_4866_cast_fp16 = softmax(axis = var_4697, x = aw_707_cast_fp16)[name = tensor("op_4866_cast_fp16")]; + tensor var_4867_cast_fp16 = softmax(axis = var_4697, x = aw_709_cast_fp16)[name = tensor("op_4867_cast_fp16")]; + tensor var_4868_cast_fp16 = softmax(axis = var_4697, x = aw_711_cast_fp16)[name = tensor("op_4868_cast_fp16")]; + tensor var_4869_cast_fp16 = softmax(axis = var_4697, x = aw_713_cast_fp16)[name = tensor("op_4869_cast_fp16")]; + tensor var_4870_cast_fp16 = softmax(axis = var_4697, x = aw_715_cast_fp16)[name = tensor("op_4870_cast_fp16")]; + tensor var_4871_cast_fp16 = softmax(axis = var_4697, x = aw_717_cast_fp16)[name = tensor("op_4871_cast_fp16")]; + tensor var_4872_cast_fp16 = softmax(axis = var_4697, x = aw_719_cast_fp16)[name = tensor("op_4872_cast_fp16")]; + tensor var_4874_equation_0 = const()[name = tensor("op_4874_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4874_cast_fp16 = einsum(equation = var_4874_equation_0, values = (var_4792_cast_fp16_0, var_4853_cast_fp16))[name = tensor("op_4874_cast_fp16")]; + tensor var_4876_equation_0 = const()[name = tensor("op_4876_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4876_cast_fp16 = einsum(equation = var_4876_equation_0, values = (var_4792_cast_fp16_1, var_4854_cast_fp16))[name = tensor("op_4876_cast_fp16")]; + tensor var_4878_equation_0 = const()[name = tensor("op_4878_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4878_cast_fp16 = einsum(equation = var_4878_equation_0, values = (var_4792_cast_fp16_2, var_4855_cast_fp16))[name = tensor("op_4878_cast_fp16")]; + tensor var_4880_equation_0 = const()[name = tensor("op_4880_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4880_cast_fp16 = einsum(equation = var_4880_equation_0, values = (var_4792_cast_fp16_3, var_4856_cast_fp16))[name = tensor("op_4880_cast_fp16")]; + tensor var_4882_equation_0 = const()[name = tensor("op_4882_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4882_cast_fp16 = einsum(equation = var_4882_equation_0, values = (var_4792_cast_fp16_4, var_4857_cast_fp16))[name = tensor("op_4882_cast_fp16")]; + tensor var_4884_equation_0 = const()[name = tensor("op_4884_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4884_cast_fp16 = einsum(equation = var_4884_equation_0, values = (var_4792_cast_fp16_5, var_4858_cast_fp16))[name = tensor("op_4884_cast_fp16")]; + tensor var_4886_equation_0 = const()[name = tensor("op_4886_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4886_cast_fp16 = einsum(equation = var_4886_equation_0, values = (var_4792_cast_fp16_6, var_4859_cast_fp16))[name = tensor("op_4886_cast_fp16")]; + tensor var_4888_equation_0 = const()[name = tensor("op_4888_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4888_cast_fp16 = einsum(equation = var_4888_equation_0, values = (var_4792_cast_fp16_7, var_4860_cast_fp16))[name = tensor("op_4888_cast_fp16")]; + tensor var_4890_equation_0 = const()[name = tensor("op_4890_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4890_cast_fp16 = einsum(equation = var_4890_equation_0, values = (var_4792_cast_fp16_8, var_4861_cast_fp16))[name = tensor("op_4890_cast_fp16")]; + tensor var_4892_equation_0 = const()[name = tensor("op_4892_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4892_cast_fp16 = einsum(equation = var_4892_equation_0, values = (var_4792_cast_fp16_9, var_4862_cast_fp16))[name = tensor("op_4892_cast_fp16")]; + tensor var_4894_equation_0 = const()[name = tensor("op_4894_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4894_cast_fp16 = einsum(equation = var_4894_equation_0, values = (var_4792_cast_fp16_10, var_4863_cast_fp16))[name = tensor("op_4894_cast_fp16")]; + tensor var_4896_equation_0 = const()[name = tensor("op_4896_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4896_cast_fp16 = einsum(equation = var_4896_equation_0, values = (var_4792_cast_fp16_11, var_4864_cast_fp16))[name = tensor("op_4896_cast_fp16")]; + tensor var_4898_equation_0 = const()[name = tensor("op_4898_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4898_cast_fp16 = einsum(equation = var_4898_equation_0, values = (var_4792_cast_fp16_12, var_4865_cast_fp16))[name = tensor("op_4898_cast_fp16")]; + tensor var_4900_equation_0 = const()[name = tensor("op_4900_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4900_cast_fp16 = einsum(equation = var_4900_equation_0, values = (var_4792_cast_fp16_13, var_4866_cast_fp16))[name = tensor("op_4900_cast_fp16")]; + tensor var_4902_equation_0 = const()[name = tensor("op_4902_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4902_cast_fp16 = einsum(equation = var_4902_equation_0, values = (var_4792_cast_fp16_14, var_4867_cast_fp16))[name = tensor("op_4902_cast_fp16")]; + tensor var_4904_equation_0 = const()[name = tensor("op_4904_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4904_cast_fp16 = einsum(equation = var_4904_equation_0, values = (var_4792_cast_fp16_15, var_4868_cast_fp16))[name = tensor("op_4904_cast_fp16")]; + tensor var_4906_equation_0 = const()[name = tensor("op_4906_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4906_cast_fp16 = einsum(equation = var_4906_equation_0, values = (var_4792_cast_fp16_16, var_4869_cast_fp16))[name = tensor("op_4906_cast_fp16")]; + tensor var_4908_equation_0 = const()[name = tensor("op_4908_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4908_cast_fp16 = einsum(equation = var_4908_equation_0, values = (var_4792_cast_fp16_17, var_4870_cast_fp16))[name = tensor("op_4908_cast_fp16")]; + tensor var_4910_equation_0 = const()[name = tensor("op_4910_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4910_cast_fp16 = einsum(equation = var_4910_equation_0, values = (var_4792_cast_fp16_18, var_4871_cast_fp16))[name = tensor("op_4910_cast_fp16")]; + tensor var_4912_equation_0 = const()[name = tensor("op_4912_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4912_cast_fp16 = einsum(equation = var_4912_equation_0, values = (var_4792_cast_fp16_19, var_4872_cast_fp16))[name = tensor("op_4912_cast_fp16")]; + tensor input_175_interleave_0 = const()[name = tensor("input_175_interleave_0"), val = tensor(false)]; + tensor input_175_cast_fp16 = concat(axis = var_4697, interleave = input_175_interleave_0, values = (var_4874_cast_fp16, var_4876_cast_fp16, var_4878_cast_fp16, var_4880_cast_fp16, var_4882_cast_fp16, var_4884_cast_fp16, var_4886_cast_fp16, var_4888_cast_fp16, var_4890_cast_fp16, var_4892_cast_fp16, var_4894_cast_fp16, var_4896_cast_fp16, var_4898_cast_fp16, var_4900_cast_fp16, var_4902_cast_fp16, var_4904_cast_fp16, var_4906_cast_fp16, var_4908_cast_fp16, var_4910_cast_fp16, var_4912_cast_fp16))[name = tensor("input_175_cast_fp16")]; + tensor var_4921_pad_type_0 = const()[name = tensor("op_4921_pad_type_0"), val = tensor("valid")]; + tensor var_4921_strides_0 = const()[name = tensor("op_4921_strides_0"), val = tensor([1, 1])]; + tensor var_4921_pad_0 = const()[name = tensor("op_4921_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4921_dilations_0 = const()[name = tensor("op_4921_dilations_0"), val = tensor([1, 1])]; + tensor var_4921_groups_0 = const()[name = tensor("op_4921_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_out_weight_to_fp16 = const()[name = tensor("blocks_17_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693137152)))]; + tensor blocks_17_attn_out_bias_to_fp16 = const()[name = tensor("blocks_17_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696414016)))]; + tensor var_4921_cast_fp16 = conv(bias = blocks_17_attn_out_bias_to_fp16, dilations = var_4921_dilations_0, groups = var_4921_groups_0, pad = var_4921_pad_0, pad_type = var_4921_pad_type_0, strides = var_4921_strides_0, weight = blocks_17_attn_out_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = var_4921_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([1])]; + tensor input_177_gamma_0_to_fp16 = const()[name = tensor("input_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696416640)))]; + tensor input_177_beta_0_to_fp16 = const()[name = tensor("input_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696419264)))]; + tensor var_4931_to_fp16 = const()[name = tensor("op_4931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = input_177_beta_0_to_fp16, epsilon = var_4931_to_fp16, gamma = input_177_gamma_0_to_fp16, x = inputs_71_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("valid")]; + tensor input_179_strides_0 = const()[name = tensor("input_179_strides_0"), val = tensor([1, 1])]; + tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_179_dilations_0 = const()[name = tensor("input_179_dilations_0"), val = tensor([1, 1])]; + tensor input_179_groups_0 = const()[name = tensor("input_179_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696421888)))]; + tensor blocks_17_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709529152)))]; + tensor input_179_cast_fp16 = conv(bias = blocks_17_mlp_0_bias_to_fp16, dilations = input_179_dilations_0, groups = input_179_groups_0, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = input_179_strides_0, weight = blocks_17_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4957_pad_type_0 = const()[name = tensor("op_4957_pad_type_0"), val = tensor("valid")]; + tensor var_4957_strides_0 = const()[name = tensor("op_4957_strides_0"), val = tensor([1, 1])]; + tensor var_4957_pad_0 = const()[name = tensor("op_4957_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4957_dilations_0 = const()[name = tensor("op_4957_dilations_0"), val = tensor([1, 1])]; + tensor var_4957_groups_0 = const()[name = tensor("op_4957_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709539456)))]; + tensor blocks_17_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722646720)))]; + tensor var_4957_cast_fp16 = conv(bias = blocks_17_mlp_2_bias_to_fp16, dilations = var_4957_dilations_0, groups = var_4957_groups_0, pad = var_4957_pad_0, pad_type = var_4957_pad_type_0, strides = var_4957_strides_0, weight = blocks_17_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("op_4957_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_4957_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4966 = const()[name = tensor("op_4966"), val = tensor(1)]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([1])]; + tensor input_183_gamma_0_to_fp16 = const()[name = tensor("input_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722649344)))]; + tensor input_183_beta_0_to_fp16 = const()[name = tensor("input_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722651968)))]; + tensor var_4982_to_fp16 = const()[name = tensor("op_4982_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = input_183_beta_0_to_fp16, epsilon = var_4982_to_fp16, gamma = input_183_gamma_0_to_fp16, x = inputs_73_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor q_37_pad_type_0 = const()[name = tensor("q_37_pad_type_0"), val = tensor("valid")]; + tensor q_37_strides_0 = const()[name = tensor("q_37_strides_0"), val = tensor([1, 1])]; + tensor q_37_pad_0 = const()[name = tensor("q_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_37_dilations_0 = const()[name = tensor("q_37_dilations_0"), val = tensor([1, 1])]; + tensor q_37_groups_0 = const()[name = tensor("q_37_groups_0"), val = tensor(1)]; + tensor var_5017_weight_0_to_fp16 = const()[name = tensor("op_5017_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722654592)))]; + tensor var_5017_bias_0_to_fp16 = const()[name = tensor("op_5017_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725931456)))]; + tensor var_5017_cast_fp16 = conv(bias = var_5017_bias_0_to_fp16, dilations = q_37_dilations_0, groups = q_37_groups_0, pad = q_37_pad_0, pad_type = q_37_pad_type_0, strides = q_37_strides_0, weight = var_5017_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("op_5017_cast_fp16")]; + tensor k_37_pad_type_0 = const()[name = tensor("k_37_pad_type_0"), val = tensor("valid")]; + tensor k_37_strides_0 = const()[name = tensor("k_37_strides_0"), val = tensor([1, 1])]; + tensor k_37_pad_0 = const()[name = tensor("k_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_37_dilations_0 = const()[name = tensor("k_37_dilations_0"), val = tensor([1, 1])]; + tensor k_37_groups_0 = const()[name = tensor("k_37_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_key_weight_to_fp16 = const()[name = tensor("blocks_18_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725934080)))]; + tensor k_37_cast_fp16 = conv(dilations = k_37_dilations_0, groups = k_37_groups_0, pad = k_37_pad_0, pad_type = k_37_pad_type_0, strides = k_37_strides_0, weight = blocks_18_attn_key_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor var_5015_pad_type_0 = const()[name = tensor("op_5015_pad_type_0"), val = tensor("valid")]; + tensor var_5015_strides_0 = const()[name = tensor("op_5015_strides_0"), val = tensor([1, 1])]; + tensor var_5015_pad_0 = const()[name = tensor("op_5015_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5015_dilations_0 = const()[name = tensor("op_5015_dilations_0"), val = tensor([1, 1])]; + tensor var_5015_groups_0 = const()[name = tensor("op_5015_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_value_weight_to_fp16 = const()[name = tensor("blocks_18_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729210944)))]; + tensor blocks_18_attn_value_bias_to_fp16 = const()[name = tensor("blocks_18_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732487808)))]; + tensor var_5015_cast_fp16 = conv(bias = blocks_18_attn_value_bias_to_fp16, dilations = var_5015_dilations_0, groups = var_5015_groups_0, pad = var_5015_pad_0, pad_type = var_5015_pad_type_0, strides = var_5015_strides_0, weight = blocks_18_attn_value_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor tile_54 = const()[name = tensor("tile_54"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5018_axis_0 = const()[name = tensor("op_5018_axis_0"), val = tensor(1)]; + tensor var_5018_cast_fp16_0, tensor var_5018_cast_fp16_1, tensor var_5018_cast_fp16_2, tensor var_5018_cast_fp16_3, tensor var_5018_cast_fp16_4, tensor var_5018_cast_fp16_5, tensor var_5018_cast_fp16_6, tensor var_5018_cast_fp16_7, tensor var_5018_cast_fp16_8, tensor var_5018_cast_fp16_9, tensor var_5018_cast_fp16_10, tensor var_5018_cast_fp16_11, tensor var_5018_cast_fp16_12, tensor var_5018_cast_fp16_13, tensor var_5018_cast_fp16_14, tensor var_5018_cast_fp16_15, tensor var_5018_cast_fp16_16, tensor var_5018_cast_fp16_17, tensor var_5018_cast_fp16_18, tensor var_5018_cast_fp16_19 = split(axis = var_5018_axis_0, split_sizes = tile_54, x = var_5017_cast_fp16)[name = tensor("op_5018_cast_fp16")]; + tensor var_5039_perm_0 = const()[name = tensor("op_5039_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_55 = const()[name = tensor("tile_55"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5040_axis_0 = const()[name = tensor("op_5040_axis_0"), val = tensor(3)]; + tensor var_5039_cast_fp16 = transpose(perm = var_5039_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_14")]; + tensor var_5040_cast_fp16_0, tensor var_5040_cast_fp16_1, tensor var_5040_cast_fp16_2, tensor var_5040_cast_fp16_3, tensor var_5040_cast_fp16_4, tensor var_5040_cast_fp16_5, tensor var_5040_cast_fp16_6, tensor var_5040_cast_fp16_7, tensor var_5040_cast_fp16_8, tensor var_5040_cast_fp16_9, tensor var_5040_cast_fp16_10, tensor var_5040_cast_fp16_11, tensor var_5040_cast_fp16_12, tensor var_5040_cast_fp16_13, tensor var_5040_cast_fp16_14, tensor var_5040_cast_fp16_15, tensor var_5040_cast_fp16_16, tensor var_5040_cast_fp16_17, tensor var_5040_cast_fp16_18, tensor var_5040_cast_fp16_19 = split(axis = var_5040_axis_0, split_sizes = tile_55, x = var_5039_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor tile_56 = const()[name = tensor("tile_56"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5061_axis_0 = const()[name = tensor("op_5061_axis_0"), val = tensor(1)]; + tensor var_5061_cast_fp16_0, tensor var_5061_cast_fp16_1, tensor var_5061_cast_fp16_2, tensor var_5061_cast_fp16_3, tensor var_5061_cast_fp16_4, tensor var_5061_cast_fp16_5, tensor var_5061_cast_fp16_6, tensor var_5061_cast_fp16_7, tensor var_5061_cast_fp16_8, tensor var_5061_cast_fp16_9, tensor var_5061_cast_fp16_10, tensor var_5061_cast_fp16_11, tensor var_5061_cast_fp16_12, tensor var_5061_cast_fp16_13, tensor var_5061_cast_fp16_14, tensor var_5061_cast_fp16_15, tensor var_5061_cast_fp16_16, tensor var_5061_cast_fp16_17, tensor var_5061_cast_fp16_18, tensor var_5061_cast_fp16_19 = split(axis = var_5061_axis_0, split_sizes = tile_56, x = var_5015_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor aw_721_equation_0 = const()[name = tensor("aw_721_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_721_cast_fp16 = einsum(equation = aw_721_equation_0, values = (var_5040_cast_fp16_0, var_5018_cast_fp16_0))[name = tensor("aw_721_cast_fp16")]; + tensor aw_723_equation_0 = const()[name = tensor("aw_723_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_723_cast_fp16 = einsum(equation = aw_723_equation_0, values = (var_5040_cast_fp16_1, var_5018_cast_fp16_1))[name = tensor("aw_723_cast_fp16")]; + tensor aw_725_equation_0 = const()[name = tensor("aw_725_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_725_cast_fp16 = einsum(equation = aw_725_equation_0, values = (var_5040_cast_fp16_2, var_5018_cast_fp16_2))[name = tensor("aw_725_cast_fp16")]; + tensor aw_727_equation_0 = const()[name = tensor("aw_727_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_727_cast_fp16 = einsum(equation = aw_727_equation_0, values = (var_5040_cast_fp16_3, var_5018_cast_fp16_3))[name = tensor("aw_727_cast_fp16")]; + tensor aw_729_equation_0 = const()[name = tensor("aw_729_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_729_cast_fp16 = einsum(equation = aw_729_equation_0, values = (var_5040_cast_fp16_4, var_5018_cast_fp16_4))[name = tensor("aw_729_cast_fp16")]; + tensor aw_731_equation_0 = const()[name = tensor("aw_731_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_731_cast_fp16 = einsum(equation = aw_731_equation_0, values = (var_5040_cast_fp16_5, var_5018_cast_fp16_5))[name = tensor("aw_731_cast_fp16")]; + tensor aw_733_equation_0 = const()[name = tensor("aw_733_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_733_cast_fp16 = einsum(equation = aw_733_equation_0, values = (var_5040_cast_fp16_6, var_5018_cast_fp16_6))[name = tensor("aw_733_cast_fp16")]; + tensor aw_735_equation_0 = const()[name = tensor("aw_735_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_735_cast_fp16 = einsum(equation = aw_735_equation_0, values = (var_5040_cast_fp16_7, var_5018_cast_fp16_7))[name = tensor("aw_735_cast_fp16")]; + tensor aw_737_equation_0 = const()[name = tensor("aw_737_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_737_cast_fp16 = einsum(equation = aw_737_equation_0, values = (var_5040_cast_fp16_8, var_5018_cast_fp16_8))[name = tensor("aw_737_cast_fp16")]; + tensor aw_739_equation_0 = const()[name = tensor("aw_739_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_739_cast_fp16 = einsum(equation = aw_739_equation_0, values = (var_5040_cast_fp16_9, var_5018_cast_fp16_9))[name = tensor("aw_739_cast_fp16")]; + tensor aw_741_equation_0 = const()[name = tensor("aw_741_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_741_cast_fp16 = einsum(equation = aw_741_equation_0, values = (var_5040_cast_fp16_10, var_5018_cast_fp16_10))[name = tensor("aw_741_cast_fp16")]; + tensor aw_743_equation_0 = const()[name = tensor("aw_743_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_743_cast_fp16 = einsum(equation = aw_743_equation_0, values = (var_5040_cast_fp16_11, var_5018_cast_fp16_11))[name = tensor("aw_743_cast_fp16")]; + tensor aw_745_equation_0 = const()[name = tensor("aw_745_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_745_cast_fp16 = einsum(equation = aw_745_equation_0, values = (var_5040_cast_fp16_12, var_5018_cast_fp16_12))[name = tensor("aw_745_cast_fp16")]; + tensor aw_747_equation_0 = const()[name = tensor("aw_747_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_747_cast_fp16 = einsum(equation = aw_747_equation_0, values = (var_5040_cast_fp16_13, var_5018_cast_fp16_13))[name = tensor("aw_747_cast_fp16")]; + tensor aw_749_equation_0 = const()[name = tensor("aw_749_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_749_cast_fp16 = einsum(equation = aw_749_equation_0, values = (var_5040_cast_fp16_14, var_5018_cast_fp16_14))[name = tensor("aw_749_cast_fp16")]; + tensor aw_751_equation_0 = const()[name = tensor("aw_751_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_751_cast_fp16 = einsum(equation = aw_751_equation_0, values = (var_5040_cast_fp16_15, var_5018_cast_fp16_15))[name = tensor("aw_751_cast_fp16")]; + tensor aw_753_equation_0 = const()[name = tensor("aw_753_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_753_cast_fp16 = einsum(equation = aw_753_equation_0, values = (var_5040_cast_fp16_16, var_5018_cast_fp16_16))[name = tensor("aw_753_cast_fp16")]; + tensor aw_755_equation_0 = const()[name = tensor("aw_755_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_755_cast_fp16 = einsum(equation = aw_755_equation_0, values = (var_5040_cast_fp16_17, var_5018_cast_fp16_17))[name = tensor("aw_755_cast_fp16")]; + tensor aw_757_equation_0 = const()[name = tensor("aw_757_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_757_cast_fp16 = einsum(equation = aw_757_equation_0, values = (var_5040_cast_fp16_18, var_5018_cast_fp16_18))[name = tensor("aw_757_cast_fp16")]; + tensor aw_759_equation_0 = const()[name = tensor("aw_759_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_759_cast_fp16 = einsum(equation = aw_759_equation_0, values = (var_5040_cast_fp16_19, var_5018_cast_fp16_19))[name = tensor("aw_759_cast_fp16")]; + tensor var_5122_cast_fp16 = softmax(axis = var_4966, x = aw_721_cast_fp16)[name = tensor("op_5122_cast_fp16")]; + tensor var_5123_cast_fp16 = softmax(axis = var_4966, x = aw_723_cast_fp16)[name = tensor("op_5123_cast_fp16")]; + tensor var_5124_cast_fp16 = softmax(axis = var_4966, x = aw_725_cast_fp16)[name = tensor("op_5124_cast_fp16")]; + tensor var_5125_cast_fp16 = softmax(axis = var_4966, x = aw_727_cast_fp16)[name = tensor("op_5125_cast_fp16")]; + tensor var_5126_cast_fp16 = softmax(axis = var_4966, x = aw_729_cast_fp16)[name = tensor("op_5126_cast_fp16")]; + tensor var_5127_cast_fp16 = softmax(axis = var_4966, x = aw_731_cast_fp16)[name = tensor("op_5127_cast_fp16")]; + tensor var_5128_cast_fp16 = softmax(axis = var_4966, x = aw_733_cast_fp16)[name = tensor("op_5128_cast_fp16")]; + tensor var_5129_cast_fp16 = softmax(axis = var_4966, x = aw_735_cast_fp16)[name = tensor("op_5129_cast_fp16")]; + tensor var_5130_cast_fp16 = softmax(axis = var_4966, x = aw_737_cast_fp16)[name = tensor("op_5130_cast_fp16")]; + tensor var_5131_cast_fp16 = softmax(axis = var_4966, x = aw_739_cast_fp16)[name = tensor("op_5131_cast_fp16")]; + tensor var_5132_cast_fp16 = softmax(axis = var_4966, x = aw_741_cast_fp16)[name = tensor("op_5132_cast_fp16")]; + tensor var_5133_cast_fp16 = softmax(axis = var_4966, x = aw_743_cast_fp16)[name = tensor("op_5133_cast_fp16")]; + tensor var_5134_cast_fp16 = softmax(axis = var_4966, x = aw_745_cast_fp16)[name = tensor("op_5134_cast_fp16")]; + tensor var_5135_cast_fp16 = softmax(axis = var_4966, x = aw_747_cast_fp16)[name = tensor("op_5135_cast_fp16")]; + tensor var_5136_cast_fp16 = softmax(axis = var_4966, x = aw_749_cast_fp16)[name = tensor("op_5136_cast_fp16")]; + tensor var_5137_cast_fp16 = softmax(axis = var_4966, x = aw_751_cast_fp16)[name = tensor("op_5137_cast_fp16")]; + tensor var_5138_cast_fp16 = softmax(axis = var_4966, x = aw_753_cast_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5139_cast_fp16 = softmax(axis = var_4966, x = aw_755_cast_fp16)[name = tensor("op_5139_cast_fp16")]; + tensor var_5140_cast_fp16 = softmax(axis = var_4966, x = aw_757_cast_fp16)[name = tensor("op_5140_cast_fp16")]; + tensor var_5141_cast_fp16 = softmax(axis = var_4966, x = aw_759_cast_fp16)[name = tensor("op_5141_cast_fp16")]; + tensor var_5143_equation_0 = const()[name = tensor("op_5143_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5143_cast_fp16 = einsum(equation = var_5143_equation_0, values = (var_5061_cast_fp16_0, var_5122_cast_fp16))[name = tensor("op_5143_cast_fp16")]; + tensor var_5145_equation_0 = const()[name = tensor("op_5145_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5145_cast_fp16 = einsum(equation = var_5145_equation_0, values = (var_5061_cast_fp16_1, var_5123_cast_fp16))[name = tensor("op_5145_cast_fp16")]; + tensor var_5147_equation_0 = const()[name = tensor("op_5147_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5147_cast_fp16 = einsum(equation = var_5147_equation_0, values = (var_5061_cast_fp16_2, var_5124_cast_fp16))[name = tensor("op_5147_cast_fp16")]; + tensor var_5149_equation_0 = const()[name = tensor("op_5149_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5149_cast_fp16 = einsum(equation = var_5149_equation_0, values = (var_5061_cast_fp16_3, var_5125_cast_fp16))[name = tensor("op_5149_cast_fp16")]; + tensor var_5151_equation_0 = const()[name = tensor("op_5151_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5151_cast_fp16 = einsum(equation = var_5151_equation_0, values = (var_5061_cast_fp16_4, var_5126_cast_fp16))[name = tensor("op_5151_cast_fp16")]; + tensor var_5153_equation_0 = const()[name = tensor("op_5153_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5153_cast_fp16 = einsum(equation = var_5153_equation_0, values = (var_5061_cast_fp16_5, var_5127_cast_fp16))[name = tensor("op_5153_cast_fp16")]; + tensor var_5155_equation_0 = const()[name = tensor("op_5155_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5155_cast_fp16 = einsum(equation = var_5155_equation_0, values = (var_5061_cast_fp16_6, var_5128_cast_fp16))[name = tensor("op_5155_cast_fp16")]; + tensor var_5157_equation_0 = const()[name = tensor("op_5157_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5157_cast_fp16 = einsum(equation = var_5157_equation_0, values = (var_5061_cast_fp16_7, var_5129_cast_fp16))[name = tensor("op_5157_cast_fp16")]; + tensor var_5159_equation_0 = const()[name = tensor("op_5159_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5159_cast_fp16 = einsum(equation = var_5159_equation_0, values = (var_5061_cast_fp16_8, var_5130_cast_fp16))[name = tensor("op_5159_cast_fp16")]; + tensor var_5161_equation_0 = const()[name = tensor("op_5161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5161_cast_fp16 = einsum(equation = var_5161_equation_0, values = (var_5061_cast_fp16_9, var_5131_cast_fp16))[name = tensor("op_5161_cast_fp16")]; + tensor var_5163_equation_0 = const()[name = tensor("op_5163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5163_cast_fp16 = einsum(equation = var_5163_equation_0, values = (var_5061_cast_fp16_10, var_5132_cast_fp16))[name = tensor("op_5163_cast_fp16")]; + tensor var_5165_equation_0 = const()[name = tensor("op_5165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5165_cast_fp16 = einsum(equation = var_5165_equation_0, values = (var_5061_cast_fp16_11, var_5133_cast_fp16))[name = tensor("op_5165_cast_fp16")]; + tensor var_5167_equation_0 = const()[name = tensor("op_5167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5167_cast_fp16 = einsum(equation = var_5167_equation_0, values = (var_5061_cast_fp16_12, var_5134_cast_fp16))[name = tensor("op_5167_cast_fp16")]; + tensor var_5169_equation_0 = const()[name = tensor("op_5169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5169_cast_fp16 = einsum(equation = var_5169_equation_0, values = (var_5061_cast_fp16_13, var_5135_cast_fp16))[name = tensor("op_5169_cast_fp16")]; + tensor var_5171_equation_0 = const()[name = tensor("op_5171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5171_cast_fp16 = einsum(equation = var_5171_equation_0, values = (var_5061_cast_fp16_14, var_5136_cast_fp16))[name = tensor("op_5171_cast_fp16")]; + tensor var_5173_equation_0 = const()[name = tensor("op_5173_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5173_cast_fp16 = einsum(equation = var_5173_equation_0, values = (var_5061_cast_fp16_15, var_5137_cast_fp16))[name = tensor("op_5173_cast_fp16")]; + tensor var_5175_equation_0 = const()[name = tensor("op_5175_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5175_cast_fp16 = einsum(equation = var_5175_equation_0, values = (var_5061_cast_fp16_16, var_5138_cast_fp16))[name = tensor("op_5175_cast_fp16")]; + tensor var_5177_equation_0 = const()[name = tensor("op_5177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5177_cast_fp16 = einsum(equation = var_5177_equation_0, values = (var_5061_cast_fp16_17, var_5139_cast_fp16))[name = tensor("op_5177_cast_fp16")]; + tensor var_5179_equation_0 = const()[name = tensor("op_5179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5179_cast_fp16 = einsum(equation = var_5179_equation_0, values = (var_5061_cast_fp16_18, var_5140_cast_fp16))[name = tensor("op_5179_cast_fp16")]; + tensor var_5181_equation_0 = const()[name = tensor("op_5181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5181_cast_fp16 = einsum(equation = var_5181_equation_0, values = (var_5061_cast_fp16_19, var_5141_cast_fp16))[name = tensor("op_5181_cast_fp16")]; + tensor input_185_interleave_0 = const()[name = tensor("input_185_interleave_0"), val = tensor(false)]; + tensor input_185_cast_fp16 = concat(axis = var_4966, interleave = input_185_interleave_0, values = (var_5143_cast_fp16, var_5145_cast_fp16, var_5147_cast_fp16, var_5149_cast_fp16, var_5151_cast_fp16, var_5153_cast_fp16, var_5155_cast_fp16, var_5157_cast_fp16, var_5159_cast_fp16, var_5161_cast_fp16, var_5163_cast_fp16, var_5165_cast_fp16, var_5167_cast_fp16, var_5169_cast_fp16, var_5171_cast_fp16, var_5173_cast_fp16, var_5175_cast_fp16, var_5177_cast_fp16, var_5179_cast_fp16, var_5181_cast_fp16))[name = tensor("input_185_cast_fp16")]; + tensor var_5190_pad_type_0 = const()[name = tensor("op_5190_pad_type_0"), val = tensor("valid")]; + tensor var_5190_strides_0 = const()[name = tensor("op_5190_strides_0"), val = tensor([1, 1])]; + tensor var_5190_pad_0 = const()[name = tensor("op_5190_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5190_dilations_0 = const()[name = tensor("op_5190_dilations_0"), val = tensor([1, 1])]; + tensor var_5190_groups_0 = const()[name = tensor("op_5190_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_out_weight_to_fp16 = const()[name = tensor("blocks_18_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732490432)))]; + tensor blocks_18_attn_out_bias_to_fp16 = const()[name = tensor("blocks_18_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735767296)))]; + tensor var_5190_cast_fp16 = conv(bias = blocks_18_attn_out_bias_to_fp16, dilations = var_5190_dilations_0, groups = var_5190_groups_0, pad = var_5190_pad_0, pad_type = var_5190_pad_type_0, strides = var_5190_strides_0, weight = blocks_18_attn_out_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("op_5190_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = var_5190_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor input_187_axes_0 = const()[name = tensor("input_187_axes_0"), val = tensor([1])]; + tensor input_187_gamma_0_to_fp16 = const()[name = tensor("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735769920)))]; + tensor input_187_beta_0_to_fp16 = const()[name = tensor("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735772544)))]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = input_187_beta_0_to_fp16, epsilon = var_5200_to_fp16, gamma = input_187_gamma_0_to_fp16, x = inputs_75_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("valid")]; + tensor input_189_strides_0 = const()[name = tensor("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = tensor("input_189_dilations_0"), val = tensor([1, 1])]; + tensor input_189_groups_0 = const()[name = tensor("input_189_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735775168)))]; + tensor blocks_18_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748882432)))]; + tensor input_189_cast_fp16 = conv(bias = blocks_18_mlp_0_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = blocks_18_mlp_0_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_5226_pad_type_0 = const()[name = tensor("op_5226_pad_type_0"), val = tensor("valid")]; + tensor var_5226_strides_0 = const()[name = tensor("op_5226_strides_0"), val = tensor([1, 1])]; + tensor var_5226_pad_0 = const()[name = tensor("op_5226_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5226_dilations_0 = const()[name = tensor("op_5226_dilations_0"), val = tensor([1, 1])]; + tensor var_5226_groups_0 = const()[name = tensor("op_5226_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748892736)))]; + tensor blocks_18_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762000000)))]; + tensor var_5226_cast_fp16 = conv(bias = blocks_18_mlp_2_bias_to_fp16, dilations = var_5226_dilations_0, groups = var_5226_groups_0, pad = var_5226_pad_0, pad_type = var_5226_pad_type_0, strides = var_5226_strides_0, weight = blocks_18_mlp_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("op_5226_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = var_5226_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor(1)]; + tensor input_193_axes_0 = const()[name = tensor("input_193_axes_0"), val = tensor([1])]; + tensor input_193_gamma_0_to_fp16 = const()[name = tensor("input_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762002624)))]; + tensor input_193_beta_0_to_fp16 = const()[name = tensor("input_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762005248)))]; + tensor var_5251_to_fp16 = const()[name = tensor("op_5251_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = input_193_beta_0_to_fp16, epsilon = var_5251_to_fp16, gamma = input_193_gamma_0_to_fp16, x = inputs_77_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor q_39_pad_type_0 = const()[name = tensor("q_39_pad_type_0"), val = tensor("valid")]; + tensor q_39_strides_0 = const()[name = tensor("q_39_strides_0"), val = tensor([1, 1])]; + tensor q_39_pad_0 = const()[name = tensor("q_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_39_dilations_0 = const()[name = tensor("q_39_dilations_0"), val = tensor([1, 1])]; + tensor q_39_groups_0 = const()[name = tensor("q_39_groups_0"), val = tensor(1)]; + tensor var_5286_weight_0_to_fp16 = const()[name = tensor("op_5286_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762007872)))]; + tensor var_5286_bias_0_to_fp16 = const()[name = tensor("op_5286_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765284736)))]; + tensor var_5286_cast_fp16 = conv(bias = var_5286_bias_0_to_fp16, dilations = q_39_dilations_0, groups = q_39_groups_0, pad = q_39_pad_0, pad_type = q_39_pad_type_0, strides = q_39_strides_0, weight = var_5286_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor k_39_pad_type_0 = const()[name = tensor("k_39_pad_type_0"), val = tensor("valid")]; + tensor k_39_strides_0 = const()[name = tensor("k_39_strides_0"), val = tensor([1, 1])]; + tensor k_39_pad_0 = const()[name = tensor("k_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_39_dilations_0 = const()[name = tensor("k_39_dilations_0"), val = tensor([1, 1])]; + tensor k_39_groups_0 = const()[name = tensor("k_39_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_key_weight_to_fp16 = const()[name = tensor("blocks_19_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765287360)))]; + tensor k_39_cast_fp16 = conv(dilations = k_39_dilations_0, groups = k_39_groups_0, pad = k_39_pad_0, pad_type = k_39_pad_type_0, strides = k_39_strides_0, weight = blocks_19_attn_key_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor var_5284_pad_type_0 = const()[name = tensor("op_5284_pad_type_0"), val = tensor("valid")]; + tensor var_5284_strides_0 = const()[name = tensor("op_5284_strides_0"), val = tensor([1, 1])]; + tensor var_5284_pad_0 = const()[name = tensor("op_5284_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5284_dilations_0 = const()[name = tensor("op_5284_dilations_0"), val = tensor([1, 1])]; + tensor var_5284_groups_0 = const()[name = tensor("op_5284_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_value_weight_to_fp16 = const()[name = tensor("blocks_19_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768564224)))]; + tensor blocks_19_attn_value_bias_to_fp16 = const()[name = tensor("blocks_19_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771841088)))]; + tensor var_5284_cast_fp16 = conv(bias = blocks_19_attn_value_bias_to_fp16, dilations = var_5284_dilations_0, groups = var_5284_groups_0, pad = var_5284_pad_0, pad_type = var_5284_pad_type_0, strides = var_5284_strides_0, weight = blocks_19_attn_value_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor tile_57 = const()[name = tensor("tile_57"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5287_axis_0 = const()[name = tensor("op_5287_axis_0"), val = tensor(1)]; + tensor var_5287_cast_fp16_0, tensor var_5287_cast_fp16_1, tensor var_5287_cast_fp16_2, tensor var_5287_cast_fp16_3, tensor var_5287_cast_fp16_4, tensor var_5287_cast_fp16_5, tensor var_5287_cast_fp16_6, tensor var_5287_cast_fp16_7, tensor var_5287_cast_fp16_8, tensor var_5287_cast_fp16_9, tensor var_5287_cast_fp16_10, tensor var_5287_cast_fp16_11, tensor var_5287_cast_fp16_12, tensor var_5287_cast_fp16_13, tensor var_5287_cast_fp16_14, tensor var_5287_cast_fp16_15, tensor var_5287_cast_fp16_16, tensor var_5287_cast_fp16_17, tensor var_5287_cast_fp16_18, tensor var_5287_cast_fp16_19 = split(axis = var_5287_axis_0, split_sizes = tile_57, x = var_5286_cast_fp16)[name = tensor("op_5287_cast_fp16")]; + tensor var_5308_perm_0 = const()[name = tensor("op_5308_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_58 = const()[name = tensor("tile_58"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5309_axis_0 = const()[name = tensor("op_5309_axis_0"), val = tensor(3)]; + tensor var_5308_cast_fp16 = transpose(perm = var_5308_perm_0, x = k_39_cast_fp16)[name = tensor("transpose_13")]; + tensor var_5309_cast_fp16_0, tensor var_5309_cast_fp16_1, tensor var_5309_cast_fp16_2, tensor var_5309_cast_fp16_3, tensor var_5309_cast_fp16_4, tensor var_5309_cast_fp16_5, tensor var_5309_cast_fp16_6, tensor var_5309_cast_fp16_7, tensor var_5309_cast_fp16_8, tensor var_5309_cast_fp16_9, tensor var_5309_cast_fp16_10, tensor var_5309_cast_fp16_11, tensor var_5309_cast_fp16_12, tensor var_5309_cast_fp16_13, tensor var_5309_cast_fp16_14, tensor var_5309_cast_fp16_15, tensor var_5309_cast_fp16_16, tensor var_5309_cast_fp16_17, tensor var_5309_cast_fp16_18, tensor var_5309_cast_fp16_19 = split(axis = var_5309_axis_0, split_sizes = tile_58, x = var_5308_cast_fp16)[name = tensor("op_5309_cast_fp16")]; + tensor tile_59 = const()[name = tensor("tile_59"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5330_axis_0 = const()[name = tensor("op_5330_axis_0"), val = tensor(1)]; + tensor var_5330_cast_fp16_0, tensor var_5330_cast_fp16_1, tensor var_5330_cast_fp16_2, tensor var_5330_cast_fp16_3, tensor var_5330_cast_fp16_4, tensor var_5330_cast_fp16_5, tensor var_5330_cast_fp16_6, tensor var_5330_cast_fp16_7, tensor var_5330_cast_fp16_8, tensor var_5330_cast_fp16_9, tensor var_5330_cast_fp16_10, tensor var_5330_cast_fp16_11, tensor var_5330_cast_fp16_12, tensor var_5330_cast_fp16_13, tensor var_5330_cast_fp16_14, tensor var_5330_cast_fp16_15, tensor var_5330_cast_fp16_16, tensor var_5330_cast_fp16_17, tensor var_5330_cast_fp16_18, tensor var_5330_cast_fp16_19 = split(axis = var_5330_axis_0, split_sizes = tile_59, x = var_5284_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor aw_761_equation_0 = const()[name = tensor("aw_761_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_761_cast_fp16 = einsum(equation = aw_761_equation_0, values = (var_5309_cast_fp16_0, var_5287_cast_fp16_0))[name = tensor("aw_761_cast_fp16")]; + tensor aw_763_equation_0 = const()[name = tensor("aw_763_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_763_cast_fp16 = einsum(equation = aw_763_equation_0, values = (var_5309_cast_fp16_1, var_5287_cast_fp16_1))[name = tensor("aw_763_cast_fp16")]; + tensor aw_765_equation_0 = const()[name = tensor("aw_765_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_765_cast_fp16 = einsum(equation = aw_765_equation_0, values = (var_5309_cast_fp16_2, var_5287_cast_fp16_2))[name = tensor("aw_765_cast_fp16")]; + tensor aw_767_equation_0 = const()[name = tensor("aw_767_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_767_cast_fp16 = einsum(equation = aw_767_equation_0, values = (var_5309_cast_fp16_3, var_5287_cast_fp16_3))[name = tensor("aw_767_cast_fp16")]; + tensor aw_769_equation_0 = const()[name = tensor("aw_769_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_769_cast_fp16 = einsum(equation = aw_769_equation_0, values = (var_5309_cast_fp16_4, var_5287_cast_fp16_4))[name = tensor("aw_769_cast_fp16")]; + tensor aw_771_equation_0 = const()[name = tensor("aw_771_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_771_cast_fp16 = einsum(equation = aw_771_equation_0, values = (var_5309_cast_fp16_5, var_5287_cast_fp16_5))[name = tensor("aw_771_cast_fp16")]; + tensor aw_773_equation_0 = const()[name = tensor("aw_773_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_773_cast_fp16 = einsum(equation = aw_773_equation_0, values = (var_5309_cast_fp16_6, var_5287_cast_fp16_6))[name = tensor("aw_773_cast_fp16")]; + tensor aw_775_equation_0 = const()[name = tensor("aw_775_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_775_cast_fp16 = einsum(equation = aw_775_equation_0, values = (var_5309_cast_fp16_7, var_5287_cast_fp16_7))[name = tensor("aw_775_cast_fp16")]; + tensor aw_777_equation_0 = const()[name = tensor("aw_777_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_777_cast_fp16 = einsum(equation = aw_777_equation_0, values = (var_5309_cast_fp16_8, var_5287_cast_fp16_8))[name = tensor("aw_777_cast_fp16")]; + tensor aw_779_equation_0 = const()[name = tensor("aw_779_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_779_cast_fp16 = einsum(equation = aw_779_equation_0, values = (var_5309_cast_fp16_9, var_5287_cast_fp16_9))[name = tensor("aw_779_cast_fp16")]; + tensor aw_781_equation_0 = const()[name = tensor("aw_781_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_781_cast_fp16 = einsum(equation = aw_781_equation_0, values = (var_5309_cast_fp16_10, var_5287_cast_fp16_10))[name = tensor("aw_781_cast_fp16")]; + tensor aw_783_equation_0 = const()[name = tensor("aw_783_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_783_cast_fp16 = einsum(equation = aw_783_equation_0, values = (var_5309_cast_fp16_11, var_5287_cast_fp16_11))[name = tensor("aw_783_cast_fp16")]; + tensor aw_785_equation_0 = const()[name = tensor("aw_785_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_785_cast_fp16 = einsum(equation = aw_785_equation_0, values = (var_5309_cast_fp16_12, var_5287_cast_fp16_12))[name = tensor("aw_785_cast_fp16")]; + tensor aw_787_equation_0 = const()[name = tensor("aw_787_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_787_cast_fp16 = einsum(equation = aw_787_equation_0, values = (var_5309_cast_fp16_13, var_5287_cast_fp16_13))[name = tensor("aw_787_cast_fp16")]; + tensor aw_789_equation_0 = const()[name = tensor("aw_789_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_789_cast_fp16 = einsum(equation = aw_789_equation_0, values = (var_5309_cast_fp16_14, var_5287_cast_fp16_14))[name = tensor("aw_789_cast_fp16")]; + tensor aw_791_equation_0 = const()[name = tensor("aw_791_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_791_cast_fp16 = einsum(equation = aw_791_equation_0, values = (var_5309_cast_fp16_15, var_5287_cast_fp16_15))[name = tensor("aw_791_cast_fp16")]; + tensor aw_793_equation_0 = const()[name = tensor("aw_793_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_793_cast_fp16 = einsum(equation = aw_793_equation_0, values = (var_5309_cast_fp16_16, var_5287_cast_fp16_16))[name = tensor("aw_793_cast_fp16")]; + tensor aw_795_equation_0 = const()[name = tensor("aw_795_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_795_cast_fp16 = einsum(equation = aw_795_equation_0, values = (var_5309_cast_fp16_17, var_5287_cast_fp16_17))[name = tensor("aw_795_cast_fp16")]; + tensor aw_797_equation_0 = const()[name = tensor("aw_797_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_797_cast_fp16 = einsum(equation = aw_797_equation_0, values = (var_5309_cast_fp16_18, var_5287_cast_fp16_18))[name = tensor("aw_797_cast_fp16")]; + tensor aw_799_equation_0 = const()[name = tensor("aw_799_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_799_cast_fp16 = einsum(equation = aw_799_equation_0, values = (var_5309_cast_fp16_19, var_5287_cast_fp16_19))[name = tensor("aw_799_cast_fp16")]; + tensor var_5391_cast_fp16 = softmax(axis = var_5235, x = aw_761_cast_fp16)[name = tensor("op_5391_cast_fp16")]; + tensor var_5392_cast_fp16 = softmax(axis = var_5235, x = aw_763_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5393_cast_fp16 = softmax(axis = var_5235, x = aw_765_cast_fp16)[name = tensor("op_5393_cast_fp16")]; + tensor var_5394_cast_fp16 = softmax(axis = var_5235, x = aw_767_cast_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor var_5395_cast_fp16 = softmax(axis = var_5235, x = aw_769_cast_fp16)[name = tensor("op_5395_cast_fp16")]; + tensor var_5396_cast_fp16 = softmax(axis = var_5235, x = aw_771_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor var_5397_cast_fp16 = softmax(axis = var_5235, x = aw_773_cast_fp16)[name = tensor("op_5397_cast_fp16")]; + tensor var_5398_cast_fp16 = softmax(axis = var_5235, x = aw_775_cast_fp16)[name = tensor("op_5398_cast_fp16")]; + tensor var_5399_cast_fp16 = softmax(axis = var_5235, x = aw_777_cast_fp16)[name = tensor("op_5399_cast_fp16")]; + tensor var_5400_cast_fp16 = softmax(axis = var_5235, x = aw_779_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5401_cast_fp16 = softmax(axis = var_5235, x = aw_781_cast_fp16)[name = tensor("op_5401_cast_fp16")]; + tensor var_5402_cast_fp16 = softmax(axis = var_5235, x = aw_783_cast_fp16)[name = tensor("op_5402_cast_fp16")]; + tensor var_5403_cast_fp16 = softmax(axis = var_5235, x = aw_785_cast_fp16)[name = tensor("op_5403_cast_fp16")]; + tensor var_5404_cast_fp16 = softmax(axis = var_5235, x = aw_787_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor var_5405_cast_fp16 = softmax(axis = var_5235, x = aw_789_cast_fp16)[name = tensor("op_5405_cast_fp16")]; + tensor var_5406_cast_fp16 = softmax(axis = var_5235, x = aw_791_cast_fp16)[name = tensor("op_5406_cast_fp16")]; + tensor var_5407_cast_fp16 = softmax(axis = var_5235, x = aw_793_cast_fp16)[name = tensor("op_5407_cast_fp16")]; + tensor var_5408_cast_fp16 = softmax(axis = var_5235, x = aw_795_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor var_5409_cast_fp16 = softmax(axis = var_5235, x = aw_797_cast_fp16)[name = tensor("op_5409_cast_fp16")]; + tensor var_5410_cast_fp16 = softmax(axis = var_5235, x = aw_799_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5412_equation_0 = const()[name = tensor("op_5412_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5412_cast_fp16 = einsum(equation = var_5412_equation_0, values = (var_5330_cast_fp16_0, var_5391_cast_fp16))[name = tensor("op_5412_cast_fp16")]; + tensor var_5414_equation_0 = const()[name = tensor("op_5414_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5414_cast_fp16 = einsum(equation = var_5414_equation_0, values = (var_5330_cast_fp16_1, var_5392_cast_fp16))[name = tensor("op_5414_cast_fp16")]; + tensor var_5416_equation_0 = const()[name = tensor("op_5416_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5416_cast_fp16 = einsum(equation = var_5416_equation_0, values = (var_5330_cast_fp16_2, var_5393_cast_fp16))[name = tensor("op_5416_cast_fp16")]; + tensor var_5418_equation_0 = const()[name = tensor("op_5418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5418_cast_fp16 = einsum(equation = var_5418_equation_0, values = (var_5330_cast_fp16_3, var_5394_cast_fp16))[name = tensor("op_5418_cast_fp16")]; + tensor var_5420_equation_0 = const()[name = tensor("op_5420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5420_cast_fp16 = einsum(equation = var_5420_equation_0, values = (var_5330_cast_fp16_4, var_5395_cast_fp16))[name = tensor("op_5420_cast_fp16")]; + tensor var_5422_equation_0 = const()[name = tensor("op_5422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5422_cast_fp16 = einsum(equation = var_5422_equation_0, values = (var_5330_cast_fp16_5, var_5396_cast_fp16))[name = tensor("op_5422_cast_fp16")]; + tensor var_5424_equation_0 = const()[name = tensor("op_5424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5424_cast_fp16 = einsum(equation = var_5424_equation_0, values = (var_5330_cast_fp16_6, var_5397_cast_fp16))[name = tensor("op_5424_cast_fp16")]; + tensor var_5426_equation_0 = const()[name = tensor("op_5426_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5426_cast_fp16 = einsum(equation = var_5426_equation_0, values = (var_5330_cast_fp16_7, var_5398_cast_fp16))[name = tensor("op_5426_cast_fp16")]; + tensor var_5428_equation_0 = const()[name = tensor("op_5428_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5428_cast_fp16 = einsum(equation = var_5428_equation_0, values = (var_5330_cast_fp16_8, var_5399_cast_fp16))[name = tensor("op_5428_cast_fp16")]; + tensor var_5430_equation_0 = const()[name = tensor("op_5430_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5430_cast_fp16 = einsum(equation = var_5430_equation_0, values = (var_5330_cast_fp16_9, var_5400_cast_fp16))[name = tensor("op_5430_cast_fp16")]; + tensor var_5432_equation_0 = const()[name = tensor("op_5432_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5432_cast_fp16 = einsum(equation = var_5432_equation_0, values = (var_5330_cast_fp16_10, var_5401_cast_fp16))[name = tensor("op_5432_cast_fp16")]; + tensor var_5434_equation_0 = const()[name = tensor("op_5434_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5434_cast_fp16 = einsum(equation = var_5434_equation_0, values = (var_5330_cast_fp16_11, var_5402_cast_fp16))[name = tensor("op_5434_cast_fp16")]; + tensor var_5436_equation_0 = const()[name = tensor("op_5436_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5436_cast_fp16 = einsum(equation = var_5436_equation_0, values = (var_5330_cast_fp16_12, var_5403_cast_fp16))[name = tensor("op_5436_cast_fp16")]; + tensor var_5438_equation_0 = const()[name = tensor("op_5438_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5438_cast_fp16 = einsum(equation = var_5438_equation_0, values = (var_5330_cast_fp16_13, var_5404_cast_fp16))[name = tensor("op_5438_cast_fp16")]; + tensor var_5440_equation_0 = const()[name = tensor("op_5440_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5440_cast_fp16 = einsum(equation = var_5440_equation_0, values = (var_5330_cast_fp16_14, var_5405_cast_fp16))[name = tensor("op_5440_cast_fp16")]; + tensor var_5442_equation_0 = const()[name = tensor("op_5442_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5442_cast_fp16 = einsum(equation = var_5442_equation_0, values = (var_5330_cast_fp16_15, var_5406_cast_fp16))[name = tensor("op_5442_cast_fp16")]; + tensor var_5444_equation_0 = const()[name = tensor("op_5444_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5444_cast_fp16 = einsum(equation = var_5444_equation_0, values = (var_5330_cast_fp16_16, var_5407_cast_fp16))[name = tensor("op_5444_cast_fp16")]; + tensor var_5446_equation_0 = const()[name = tensor("op_5446_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5446_cast_fp16 = einsum(equation = var_5446_equation_0, values = (var_5330_cast_fp16_17, var_5408_cast_fp16))[name = tensor("op_5446_cast_fp16")]; + tensor var_5448_equation_0 = const()[name = tensor("op_5448_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5448_cast_fp16 = einsum(equation = var_5448_equation_0, values = (var_5330_cast_fp16_18, var_5409_cast_fp16))[name = tensor("op_5448_cast_fp16")]; + tensor var_5450_equation_0 = const()[name = tensor("op_5450_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5450_cast_fp16 = einsum(equation = var_5450_equation_0, values = (var_5330_cast_fp16_19, var_5410_cast_fp16))[name = tensor("op_5450_cast_fp16")]; + tensor input_195_interleave_0 = const()[name = tensor("input_195_interleave_0"), val = tensor(false)]; + tensor input_195_cast_fp16 = concat(axis = var_5235, interleave = input_195_interleave_0, values = (var_5412_cast_fp16, var_5414_cast_fp16, var_5416_cast_fp16, var_5418_cast_fp16, var_5420_cast_fp16, var_5422_cast_fp16, var_5424_cast_fp16, var_5426_cast_fp16, var_5428_cast_fp16, var_5430_cast_fp16, var_5432_cast_fp16, var_5434_cast_fp16, var_5436_cast_fp16, var_5438_cast_fp16, var_5440_cast_fp16, var_5442_cast_fp16, var_5444_cast_fp16, var_5446_cast_fp16, var_5448_cast_fp16, var_5450_cast_fp16))[name = tensor("input_195_cast_fp16")]; + tensor var_5459_pad_type_0 = const()[name = tensor("op_5459_pad_type_0"), val = tensor("valid")]; + tensor var_5459_strides_0 = const()[name = tensor("op_5459_strides_0"), val = tensor([1, 1])]; + tensor var_5459_pad_0 = const()[name = tensor("op_5459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5459_dilations_0 = const()[name = tensor("op_5459_dilations_0"), val = tensor([1, 1])]; + tensor var_5459_groups_0 = const()[name = tensor("op_5459_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_out_weight_to_fp16 = const()[name = tensor("blocks_19_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771843712)))]; + tensor blocks_19_attn_out_bias_to_fp16 = const()[name = tensor("blocks_19_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775120576)))]; + tensor var_5459_cast_fp16 = conv(bias = blocks_19_attn_out_bias_to_fp16, dilations = var_5459_dilations_0, groups = var_5459_groups_0, pad = var_5459_pad_0, pad_type = var_5459_pad_type_0, strides = var_5459_strides_0, weight = blocks_19_attn_out_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("op_5459_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_5459_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([1])]; + tensor input_197_gamma_0_to_fp16 = const()[name = tensor("input_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775123200)))]; + tensor input_197_beta_0_to_fp16 = const()[name = tensor("input_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775125824)))]; + tensor var_5469_to_fp16 = const()[name = tensor("op_5469_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = input_197_beta_0_to_fp16, epsilon = var_5469_to_fp16, gamma = input_197_gamma_0_to_fp16, x = inputs_79_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("valid")]; + tensor input_199_strides_0 = const()[name = tensor("input_199_strides_0"), val = tensor([1, 1])]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_199_dilations_0 = const()[name = tensor("input_199_dilations_0"), val = tensor([1, 1])]; + tensor input_199_groups_0 = const()[name = tensor("input_199_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775128448)))]; + tensor blocks_19_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788235712)))]; + tensor input_199_cast_fp16 = conv(bias = blocks_19_mlp_0_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = blocks_19_mlp_0_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_5495_pad_type_0 = const()[name = tensor("op_5495_pad_type_0"), val = tensor("valid")]; + tensor var_5495_strides_0 = const()[name = tensor("op_5495_strides_0"), val = tensor([1, 1])]; + tensor var_5495_pad_0 = const()[name = tensor("op_5495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5495_dilations_0 = const()[name = tensor("op_5495_dilations_0"), val = tensor([1, 1])]; + tensor var_5495_groups_0 = const()[name = tensor("op_5495_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788246016)))]; + tensor blocks_19_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801353280)))]; + tensor var_5495_cast_fp16 = conv(bias = blocks_19_mlp_2_bias_to_fp16, dilations = var_5495_dilations_0, groups = var_5495_groups_0, pad = var_5495_pad_0, pad_type = var_5495_pad_type_0, strides = var_5495_strides_0, weight = blocks_19_mlp_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = var_5495_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_5504 = const()[name = tensor("op_5504"), val = tensor(1)]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([1])]; + tensor input_203_gamma_0_to_fp16 = const()[name = tensor("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801355904)))]; + tensor input_203_beta_0_to_fp16 = const()[name = tensor("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801358528)))]; + tensor var_5520_to_fp16 = const()[name = tensor("op_5520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = input_203_beta_0_to_fp16, epsilon = var_5520_to_fp16, gamma = input_203_gamma_0_to_fp16, x = inputs_81_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor q_41_pad_type_0 = const()[name = tensor("q_41_pad_type_0"), val = tensor("valid")]; + tensor q_41_strides_0 = const()[name = tensor("q_41_strides_0"), val = tensor([1, 1])]; + tensor q_41_pad_0 = const()[name = tensor("q_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_41_dilations_0 = const()[name = tensor("q_41_dilations_0"), val = tensor([1, 1])]; + tensor q_41_groups_0 = const()[name = tensor("q_41_groups_0"), val = tensor(1)]; + tensor var_5555_weight_0_to_fp16 = const()[name = tensor("op_5555_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801361152)))]; + tensor var_5555_bias_0_to_fp16 = const()[name = tensor("op_5555_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804638016)))]; + tensor var_5555_cast_fp16 = conv(bias = var_5555_bias_0_to_fp16, dilations = q_41_dilations_0, groups = q_41_groups_0, pad = q_41_pad_0, pad_type = q_41_pad_type_0, strides = q_41_strides_0, weight = var_5555_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("op_5555_cast_fp16")]; + tensor k_41_pad_type_0 = const()[name = tensor("k_41_pad_type_0"), val = tensor("valid")]; + tensor k_41_strides_0 = const()[name = tensor("k_41_strides_0"), val = tensor([1, 1])]; + tensor k_41_pad_0 = const()[name = tensor("k_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_41_dilations_0 = const()[name = tensor("k_41_dilations_0"), val = tensor([1, 1])]; + tensor k_41_groups_0 = const()[name = tensor("k_41_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_key_weight_to_fp16 = const()[name = tensor("blocks_20_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804640640)))]; + tensor k_41_cast_fp16 = conv(dilations = k_41_dilations_0, groups = k_41_groups_0, pad = k_41_pad_0, pad_type = k_41_pad_type_0, strides = k_41_strides_0, weight = blocks_20_attn_key_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor var_5553_pad_type_0 = const()[name = tensor("op_5553_pad_type_0"), val = tensor("valid")]; + tensor var_5553_strides_0 = const()[name = tensor("op_5553_strides_0"), val = tensor([1, 1])]; + tensor var_5553_pad_0 = const()[name = tensor("op_5553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5553_dilations_0 = const()[name = tensor("op_5553_dilations_0"), val = tensor([1, 1])]; + tensor var_5553_groups_0 = const()[name = tensor("op_5553_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_value_weight_to_fp16 = const()[name = tensor("blocks_20_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807917504)))]; + tensor blocks_20_attn_value_bias_to_fp16 = const()[name = tensor("blocks_20_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811194368)))]; + tensor var_5553_cast_fp16 = conv(bias = blocks_20_attn_value_bias_to_fp16, dilations = var_5553_dilations_0, groups = var_5553_groups_0, pad = var_5553_pad_0, pad_type = var_5553_pad_type_0, strides = var_5553_strides_0, weight = blocks_20_attn_value_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("op_5553_cast_fp16")]; + tensor tile_60 = const()[name = tensor("tile_60"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5556_axis_0 = const()[name = tensor("op_5556_axis_0"), val = tensor(1)]; + tensor var_5556_cast_fp16_0, tensor var_5556_cast_fp16_1, tensor var_5556_cast_fp16_2, tensor var_5556_cast_fp16_3, tensor var_5556_cast_fp16_4, tensor var_5556_cast_fp16_5, tensor var_5556_cast_fp16_6, tensor var_5556_cast_fp16_7, tensor var_5556_cast_fp16_8, tensor var_5556_cast_fp16_9, tensor var_5556_cast_fp16_10, tensor var_5556_cast_fp16_11, tensor var_5556_cast_fp16_12, tensor var_5556_cast_fp16_13, tensor var_5556_cast_fp16_14, tensor var_5556_cast_fp16_15, tensor var_5556_cast_fp16_16, tensor var_5556_cast_fp16_17, tensor var_5556_cast_fp16_18, tensor var_5556_cast_fp16_19 = split(axis = var_5556_axis_0, split_sizes = tile_60, x = var_5555_cast_fp16)[name = tensor("op_5556_cast_fp16")]; + tensor var_5577_perm_0 = const()[name = tensor("op_5577_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_61 = const()[name = tensor("tile_61"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5578_axis_0 = const()[name = tensor("op_5578_axis_0"), val = tensor(3)]; + tensor var_5577_cast_fp16 = transpose(perm = var_5577_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_12")]; + tensor var_5578_cast_fp16_0, tensor var_5578_cast_fp16_1, tensor var_5578_cast_fp16_2, tensor var_5578_cast_fp16_3, tensor var_5578_cast_fp16_4, tensor var_5578_cast_fp16_5, tensor var_5578_cast_fp16_6, tensor var_5578_cast_fp16_7, tensor var_5578_cast_fp16_8, tensor var_5578_cast_fp16_9, tensor var_5578_cast_fp16_10, tensor var_5578_cast_fp16_11, tensor var_5578_cast_fp16_12, tensor var_5578_cast_fp16_13, tensor var_5578_cast_fp16_14, tensor var_5578_cast_fp16_15, tensor var_5578_cast_fp16_16, tensor var_5578_cast_fp16_17, tensor var_5578_cast_fp16_18, tensor var_5578_cast_fp16_19 = split(axis = var_5578_axis_0, split_sizes = tile_61, x = var_5577_cast_fp16)[name = tensor("op_5578_cast_fp16")]; + tensor tile_62 = const()[name = tensor("tile_62"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5599_axis_0 = const()[name = tensor("op_5599_axis_0"), val = tensor(1)]; + tensor var_5599_cast_fp16_0, tensor var_5599_cast_fp16_1, tensor var_5599_cast_fp16_2, tensor var_5599_cast_fp16_3, tensor var_5599_cast_fp16_4, tensor var_5599_cast_fp16_5, tensor var_5599_cast_fp16_6, tensor var_5599_cast_fp16_7, tensor var_5599_cast_fp16_8, tensor var_5599_cast_fp16_9, tensor var_5599_cast_fp16_10, tensor var_5599_cast_fp16_11, tensor var_5599_cast_fp16_12, tensor var_5599_cast_fp16_13, tensor var_5599_cast_fp16_14, tensor var_5599_cast_fp16_15, tensor var_5599_cast_fp16_16, tensor var_5599_cast_fp16_17, tensor var_5599_cast_fp16_18, tensor var_5599_cast_fp16_19 = split(axis = var_5599_axis_0, split_sizes = tile_62, x = var_5553_cast_fp16)[name = tensor("op_5599_cast_fp16")]; + tensor aw_801_equation_0 = const()[name = tensor("aw_801_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_801_cast_fp16 = einsum(equation = aw_801_equation_0, values = (var_5578_cast_fp16_0, var_5556_cast_fp16_0))[name = tensor("aw_801_cast_fp16")]; + tensor aw_803_equation_0 = const()[name = tensor("aw_803_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_803_cast_fp16 = einsum(equation = aw_803_equation_0, values = (var_5578_cast_fp16_1, var_5556_cast_fp16_1))[name = tensor("aw_803_cast_fp16")]; + tensor aw_805_equation_0 = const()[name = tensor("aw_805_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_805_cast_fp16 = einsum(equation = aw_805_equation_0, values = (var_5578_cast_fp16_2, var_5556_cast_fp16_2))[name = tensor("aw_805_cast_fp16")]; + tensor aw_807_equation_0 = const()[name = tensor("aw_807_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_807_cast_fp16 = einsum(equation = aw_807_equation_0, values = (var_5578_cast_fp16_3, var_5556_cast_fp16_3))[name = tensor("aw_807_cast_fp16")]; + tensor aw_809_equation_0 = const()[name = tensor("aw_809_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_809_cast_fp16 = einsum(equation = aw_809_equation_0, values = (var_5578_cast_fp16_4, var_5556_cast_fp16_4))[name = tensor("aw_809_cast_fp16")]; + tensor aw_811_equation_0 = const()[name = tensor("aw_811_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_811_cast_fp16 = einsum(equation = aw_811_equation_0, values = (var_5578_cast_fp16_5, var_5556_cast_fp16_5))[name = tensor("aw_811_cast_fp16")]; + tensor aw_813_equation_0 = const()[name = tensor("aw_813_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_813_cast_fp16 = einsum(equation = aw_813_equation_0, values = (var_5578_cast_fp16_6, var_5556_cast_fp16_6))[name = tensor("aw_813_cast_fp16")]; + tensor aw_815_equation_0 = const()[name = tensor("aw_815_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_815_cast_fp16 = einsum(equation = aw_815_equation_0, values = (var_5578_cast_fp16_7, var_5556_cast_fp16_7))[name = tensor("aw_815_cast_fp16")]; + tensor aw_817_equation_0 = const()[name = tensor("aw_817_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_817_cast_fp16 = einsum(equation = aw_817_equation_0, values = (var_5578_cast_fp16_8, var_5556_cast_fp16_8))[name = tensor("aw_817_cast_fp16")]; + tensor aw_819_equation_0 = const()[name = tensor("aw_819_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_819_cast_fp16 = einsum(equation = aw_819_equation_0, values = (var_5578_cast_fp16_9, var_5556_cast_fp16_9))[name = tensor("aw_819_cast_fp16")]; + tensor aw_821_equation_0 = const()[name = tensor("aw_821_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_821_cast_fp16 = einsum(equation = aw_821_equation_0, values = (var_5578_cast_fp16_10, var_5556_cast_fp16_10))[name = tensor("aw_821_cast_fp16")]; + tensor aw_823_equation_0 = const()[name = tensor("aw_823_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_823_cast_fp16 = einsum(equation = aw_823_equation_0, values = (var_5578_cast_fp16_11, var_5556_cast_fp16_11))[name = tensor("aw_823_cast_fp16")]; + tensor aw_825_equation_0 = const()[name = tensor("aw_825_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_825_cast_fp16 = einsum(equation = aw_825_equation_0, values = (var_5578_cast_fp16_12, var_5556_cast_fp16_12))[name = tensor("aw_825_cast_fp16")]; + tensor aw_827_equation_0 = const()[name = tensor("aw_827_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_827_cast_fp16 = einsum(equation = aw_827_equation_0, values = (var_5578_cast_fp16_13, var_5556_cast_fp16_13))[name = tensor("aw_827_cast_fp16")]; + tensor aw_829_equation_0 = const()[name = tensor("aw_829_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_829_cast_fp16 = einsum(equation = aw_829_equation_0, values = (var_5578_cast_fp16_14, var_5556_cast_fp16_14))[name = tensor("aw_829_cast_fp16")]; + tensor aw_831_equation_0 = const()[name = tensor("aw_831_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_831_cast_fp16 = einsum(equation = aw_831_equation_0, values = (var_5578_cast_fp16_15, var_5556_cast_fp16_15))[name = tensor("aw_831_cast_fp16")]; + tensor aw_833_equation_0 = const()[name = tensor("aw_833_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_833_cast_fp16 = einsum(equation = aw_833_equation_0, values = (var_5578_cast_fp16_16, var_5556_cast_fp16_16))[name = tensor("aw_833_cast_fp16")]; + tensor aw_835_equation_0 = const()[name = tensor("aw_835_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_835_cast_fp16 = einsum(equation = aw_835_equation_0, values = (var_5578_cast_fp16_17, var_5556_cast_fp16_17))[name = tensor("aw_835_cast_fp16")]; + tensor aw_837_equation_0 = const()[name = tensor("aw_837_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_837_cast_fp16 = einsum(equation = aw_837_equation_0, values = (var_5578_cast_fp16_18, var_5556_cast_fp16_18))[name = tensor("aw_837_cast_fp16")]; + tensor aw_839_equation_0 = const()[name = tensor("aw_839_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_839_cast_fp16 = einsum(equation = aw_839_equation_0, values = (var_5578_cast_fp16_19, var_5556_cast_fp16_19))[name = tensor("aw_839_cast_fp16")]; + tensor var_5660_cast_fp16 = softmax(axis = var_5504, x = aw_801_cast_fp16)[name = tensor("op_5660_cast_fp16")]; + tensor var_5661_cast_fp16 = softmax(axis = var_5504, x = aw_803_cast_fp16)[name = tensor("op_5661_cast_fp16")]; + tensor var_5662_cast_fp16 = softmax(axis = var_5504, x = aw_805_cast_fp16)[name = tensor("op_5662_cast_fp16")]; + tensor var_5663_cast_fp16 = softmax(axis = var_5504, x = aw_807_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor var_5664_cast_fp16 = softmax(axis = var_5504, x = aw_809_cast_fp16)[name = tensor("op_5664_cast_fp16")]; + tensor var_5665_cast_fp16 = softmax(axis = var_5504, x = aw_811_cast_fp16)[name = tensor("op_5665_cast_fp16")]; + tensor var_5666_cast_fp16 = softmax(axis = var_5504, x = aw_813_cast_fp16)[name = tensor("op_5666_cast_fp16")]; + tensor var_5667_cast_fp16 = softmax(axis = var_5504, x = aw_815_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = softmax(axis = var_5504, x = aw_817_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor var_5669_cast_fp16 = softmax(axis = var_5504, x = aw_819_cast_fp16)[name = tensor("op_5669_cast_fp16")]; + tensor var_5670_cast_fp16 = softmax(axis = var_5504, x = aw_821_cast_fp16)[name = tensor("op_5670_cast_fp16")]; + tensor var_5671_cast_fp16 = softmax(axis = var_5504, x = aw_823_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = softmax(axis = var_5504, x = aw_825_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor var_5673_cast_fp16 = softmax(axis = var_5504, x = aw_827_cast_fp16)[name = tensor("op_5673_cast_fp16")]; + tensor var_5674_cast_fp16 = softmax(axis = var_5504, x = aw_829_cast_fp16)[name = tensor("op_5674_cast_fp16")]; + tensor var_5675_cast_fp16 = softmax(axis = var_5504, x = aw_831_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = softmax(axis = var_5504, x = aw_833_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor var_5677_cast_fp16 = softmax(axis = var_5504, x = aw_835_cast_fp16)[name = tensor("op_5677_cast_fp16")]; + tensor var_5678_cast_fp16 = softmax(axis = var_5504, x = aw_837_cast_fp16)[name = tensor("op_5678_cast_fp16")]; + tensor var_5679_cast_fp16 = softmax(axis = var_5504, x = aw_839_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5681_equation_0 = const()[name = tensor("op_5681_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5681_cast_fp16 = einsum(equation = var_5681_equation_0, values = (var_5599_cast_fp16_0, var_5660_cast_fp16))[name = tensor("op_5681_cast_fp16")]; + tensor var_5683_equation_0 = const()[name = tensor("op_5683_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5683_cast_fp16 = einsum(equation = var_5683_equation_0, values = (var_5599_cast_fp16_1, var_5661_cast_fp16))[name = tensor("op_5683_cast_fp16")]; + tensor var_5685_equation_0 = const()[name = tensor("op_5685_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5685_cast_fp16 = einsum(equation = var_5685_equation_0, values = (var_5599_cast_fp16_2, var_5662_cast_fp16))[name = tensor("op_5685_cast_fp16")]; + tensor var_5687_equation_0 = const()[name = tensor("op_5687_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5687_cast_fp16 = einsum(equation = var_5687_equation_0, values = (var_5599_cast_fp16_3, var_5663_cast_fp16))[name = tensor("op_5687_cast_fp16")]; + tensor var_5689_equation_0 = const()[name = tensor("op_5689_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5689_cast_fp16 = einsum(equation = var_5689_equation_0, values = (var_5599_cast_fp16_4, var_5664_cast_fp16))[name = tensor("op_5689_cast_fp16")]; + tensor var_5691_equation_0 = const()[name = tensor("op_5691_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5691_cast_fp16 = einsum(equation = var_5691_equation_0, values = (var_5599_cast_fp16_5, var_5665_cast_fp16))[name = tensor("op_5691_cast_fp16")]; + tensor var_5693_equation_0 = const()[name = tensor("op_5693_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5693_cast_fp16 = einsum(equation = var_5693_equation_0, values = (var_5599_cast_fp16_6, var_5666_cast_fp16))[name = tensor("op_5693_cast_fp16")]; + tensor var_5695_equation_0 = const()[name = tensor("op_5695_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5695_cast_fp16 = einsum(equation = var_5695_equation_0, values = (var_5599_cast_fp16_7, var_5667_cast_fp16))[name = tensor("op_5695_cast_fp16")]; + tensor var_5697_equation_0 = const()[name = tensor("op_5697_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5697_cast_fp16 = einsum(equation = var_5697_equation_0, values = (var_5599_cast_fp16_8, var_5668_cast_fp16))[name = tensor("op_5697_cast_fp16")]; + tensor var_5699_equation_0 = const()[name = tensor("op_5699_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5699_cast_fp16 = einsum(equation = var_5699_equation_0, values = (var_5599_cast_fp16_9, var_5669_cast_fp16))[name = tensor("op_5699_cast_fp16")]; + tensor var_5701_equation_0 = const()[name = tensor("op_5701_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5701_cast_fp16 = einsum(equation = var_5701_equation_0, values = (var_5599_cast_fp16_10, var_5670_cast_fp16))[name = tensor("op_5701_cast_fp16")]; + tensor var_5703_equation_0 = const()[name = tensor("op_5703_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5703_cast_fp16 = einsum(equation = var_5703_equation_0, values = (var_5599_cast_fp16_11, var_5671_cast_fp16))[name = tensor("op_5703_cast_fp16")]; + tensor var_5705_equation_0 = const()[name = tensor("op_5705_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5705_cast_fp16 = einsum(equation = var_5705_equation_0, values = (var_5599_cast_fp16_12, var_5672_cast_fp16))[name = tensor("op_5705_cast_fp16")]; + tensor var_5707_equation_0 = const()[name = tensor("op_5707_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5707_cast_fp16 = einsum(equation = var_5707_equation_0, values = (var_5599_cast_fp16_13, var_5673_cast_fp16))[name = tensor("op_5707_cast_fp16")]; + tensor var_5709_equation_0 = const()[name = tensor("op_5709_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5709_cast_fp16 = einsum(equation = var_5709_equation_0, values = (var_5599_cast_fp16_14, var_5674_cast_fp16))[name = tensor("op_5709_cast_fp16")]; + tensor var_5711_equation_0 = const()[name = tensor("op_5711_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5711_cast_fp16 = einsum(equation = var_5711_equation_0, values = (var_5599_cast_fp16_15, var_5675_cast_fp16))[name = tensor("op_5711_cast_fp16")]; + tensor var_5713_equation_0 = const()[name = tensor("op_5713_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5713_cast_fp16 = einsum(equation = var_5713_equation_0, values = (var_5599_cast_fp16_16, var_5676_cast_fp16))[name = tensor("op_5713_cast_fp16")]; + tensor var_5715_equation_0 = const()[name = tensor("op_5715_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5715_cast_fp16 = einsum(equation = var_5715_equation_0, values = (var_5599_cast_fp16_17, var_5677_cast_fp16))[name = tensor("op_5715_cast_fp16")]; + tensor var_5717_equation_0 = const()[name = tensor("op_5717_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5717_cast_fp16 = einsum(equation = var_5717_equation_0, values = (var_5599_cast_fp16_18, var_5678_cast_fp16))[name = tensor("op_5717_cast_fp16")]; + tensor var_5719_equation_0 = const()[name = tensor("op_5719_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5719_cast_fp16 = einsum(equation = var_5719_equation_0, values = (var_5599_cast_fp16_19, var_5679_cast_fp16))[name = tensor("op_5719_cast_fp16")]; + tensor input_205_interleave_0 = const()[name = tensor("input_205_interleave_0"), val = tensor(false)]; + tensor input_205_cast_fp16 = concat(axis = var_5504, interleave = input_205_interleave_0, values = (var_5681_cast_fp16, var_5683_cast_fp16, var_5685_cast_fp16, var_5687_cast_fp16, var_5689_cast_fp16, var_5691_cast_fp16, var_5693_cast_fp16, var_5695_cast_fp16, var_5697_cast_fp16, var_5699_cast_fp16, var_5701_cast_fp16, var_5703_cast_fp16, var_5705_cast_fp16, var_5707_cast_fp16, var_5709_cast_fp16, var_5711_cast_fp16, var_5713_cast_fp16, var_5715_cast_fp16, var_5717_cast_fp16, var_5719_cast_fp16))[name = tensor("input_205_cast_fp16")]; + tensor var_5728_pad_type_0 = const()[name = tensor("op_5728_pad_type_0"), val = tensor("valid")]; + tensor var_5728_strides_0 = const()[name = tensor("op_5728_strides_0"), val = tensor([1, 1])]; + tensor var_5728_pad_0 = const()[name = tensor("op_5728_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5728_dilations_0 = const()[name = tensor("op_5728_dilations_0"), val = tensor([1, 1])]; + tensor var_5728_groups_0 = const()[name = tensor("op_5728_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_out_weight_to_fp16 = const()[name = tensor("blocks_20_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811196992)))]; + tensor blocks_20_attn_out_bias_to_fp16 = const()[name = tensor("blocks_20_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814473856)))]; + tensor var_5728_cast_fp16 = conv(bias = blocks_20_attn_out_bias_to_fp16, dilations = var_5728_dilations_0, groups = var_5728_groups_0, pad = var_5728_pad_0, pad_type = var_5728_pad_type_0, strides = var_5728_strides_0, weight = blocks_20_attn_out_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_5728_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([1])]; + tensor input_207_gamma_0_to_fp16 = const()[name = tensor("input_207_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814476480)))]; + tensor input_207_beta_0_to_fp16 = const()[name = tensor("input_207_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814479104)))]; + tensor var_5738_to_fp16 = const()[name = tensor("op_5738_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = input_207_beta_0_to_fp16, epsilon = var_5738_to_fp16, gamma = input_207_gamma_0_to_fp16, x = inputs_83_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("valid")]; + tensor input_209_strides_0 = const()[name = tensor("input_209_strides_0"), val = tensor([1, 1])]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_209_dilations_0 = const()[name = tensor("input_209_dilations_0"), val = tensor([1, 1])]; + tensor input_209_groups_0 = const()[name = tensor("input_209_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814481728)))]; + tensor blocks_20_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827588992)))]; + tensor input_209_cast_fp16 = conv(bias = blocks_20_mlp_0_bias_to_fp16, dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = blocks_20_mlp_0_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = input_209_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_5764_pad_type_0 = const()[name = tensor("op_5764_pad_type_0"), val = tensor("valid")]; + tensor var_5764_strides_0 = const()[name = tensor("op_5764_strides_0"), val = tensor([1, 1])]; + tensor var_5764_pad_0 = const()[name = tensor("op_5764_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5764_dilations_0 = const()[name = tensor("op_5764_dilations_0"), val = tensor([1, 1])]; + tensor var_5764_groups_0 = const()[name = tensor("op_5764_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827599296)))]; + tensor blocks_20_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840706560)))]; + tensor var_5764_cast_fp16 = conv(bias = blocks_20_mlp_2_bias_to_fp16, dilations = var_5764_dilations_0, groups = var_5764_groups_0, pad = var_5764_pad_0, pad_type = var_5764_pad_type_0, strides = var_5764_strides_0, weight = blocks_20_mlp_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("op_5764_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = var_5764_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_5773 = const()[name = tensor("op_5773"), val = tensor(1)]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([1])]; + tensor input_213_gamma_0_to_fp16 = const()[name = tensor("input_213_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840709184)))]; + tensor input_213_beta_0_to_fp16 = const()[name = tensor("input_213_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840711808)))]; + tensor var_5789_to_fp16 = const()[name = tensor("op_5789_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = input_213_beta_0_to_fp16, epsilon = var_5789_to_fp16, gamma = input_213_gamma_0_to_fp16, x = inputs_85_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor q_43_pad_type_0 = const()[name = tensor("q_43_pad_type_0"), val = tensor("valid")]; + tensor q_43_strides_0 = const()[name = tensor("q_43_strides_0"), val = tensor([1, 1])]; + tensor q_43_pad_0 = const()[name = tensor("q_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_43_dilations_0 = const()[name = tensor("q_43_dilations_0"), val = tensor([1, 1])]; + tensor q_43_groups_0 = const()[name = tensor("q_43_groups_0"), val = tensor(1)]; + tensor var_5824_weight_0_to_fp16 = const()[name = tensor("op_5824_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840714432)))]; + tensor var_5824_bias_0_to_fp16 = const()[name = tensor("op_5824_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(843991296)))]; + tensor var_5824_cast_fp16 = conv(bias = var_5824_bias_0_to_fp16, dilations = q_43_dilations_0, groups = q_43_groups_0, pad = q_43_pad_0, pad_type = q_43_pad_type_0, strides = q_43_strides_0, weight = var_5824_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5824_cast_fp16")]; + tensor k_43_pad_type_0 = const()[name = tensor("k_43_pad_type_0"), val = tensor("valid")]; + tensor k_43_strides_0 = const()[name = tensor("k_43_strides_0"), val = tensor([1, 1])]; + tensor k_43_pad_0 = const()[name = tensor("k_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_43_dilations_0 = const()[name = tensor("k_43_dilations_0"), val = tensor([1, 1])]; + tensor k_43_groups_0 = const()[name = tensor("k_43_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_key_weight_to_fp16 = const()[name = tensor("blocks_21_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(843993920)))]; + tensor k_43_cast_fp16 = conv(dilations = k_43_dilations_0, groups = k_43_groups_0, pad = k_43_pad_0, pad_type = k_43_pad_type_0, strides = k_43_strides_0, weight = blocks_21_attn_key_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor var_5822_pad_type_0 = const()[name = tensor("op_5822_pad_type_0"), val = tensor("valid")]; + tensor var_5822_strides_0 = const()[name = tensor("op_5822_strides_0"), val = tensor([1, 1])]; + tensor var_5822_pad_0 = const()[name = tensor("op_5822_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5822_dilations_0 = const()[name = tensor("op_5822_dilations_0"), val = tensor([1, 1])]; + tensor var_5822_groups_0 = const()[name = tensor("op_5822_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_value_weight_to_fp16 = const()[name = tensor("blocks_21_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847270784)))]; + tensor blocks_21_attn_value_bias_to_fp16 = const()[name = tensor("blocks_21_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850547648)))]; + tensor var_5822_cast_fp16 = conv(bias = blocks_21_attn_value_bias_to_fp16, dilations = var_5822_dilations_0, groups = var_5822_groups_0, pad = var_5822_pad_0, pad_type = var_5822_pad_type_0, strides = var_5822_strides_0, weight = blocks_21_attn_value_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5822_cast_fp16")]; + tensor tile_63 = const()[name = tensor("tile_63"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5825_axis_0 = const()[name = tensor("op_5825_axis_0"), val = tensor(1)]; + tensor var_5825_cast_fp16_0, tensor var_5825_cast_fp16_1, tensor var_5825_cast_fp16_2, tensor var_5825_cast_fp16_3, tensor var_5825_cast_fp16_4, tensor var_5825_cast_fp16_5, tensor var_5825_cast_fp16_6, tensor var_5825_cast_fp16_7, tensor var_5825_cast_fp16_8, tensor var_5825_cast_fp16_9, tensor var_5825_cast_fp16_10, tensor var_5825_cast_fp16_11, tensor var_5825_cast_fp16_12, tensor var_5825_cast_fp16_13, tensor var_5825_cast_fp16_14, tensor var_5825_cast_fp16_15, tensor var_5825_cast_fp16_16, tensor var_5825_cast_fp16_17, tensor var_5825_cast_fp16_18, tensor var_5825_cast_fp16_19 = split(axis = var_5825_axis_0, split_sizes = tile_63, x = var_5824_cast_fp16)[name = tensor("op_5825_cast_fp16")]; + tensor var_5846_perm_0 = const()[name = tensor("op_5846_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_64 = const()[name = tensor("tile_64"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5847_axis_0 = const()[name = tensor("op_5847_axis_0"), val = tensor(3)]; + tensor var_5846_cast_fp16 = transpose(perm = var_5846_perm_0, x = k_43_cast_fp16)[name = tensor("transpose_11")]; + tensor var_5847_cast_fp16_0, tensor var_5847_cast_fp16_1, tensor var_5847_cast_fp16_2, tensor var_5847_cast_fp16_3, tensor var_5847_cast_fp16_4, tensor var_5847_cast_fp16_5, tensor var_5847_cast_fp16_6, tensor var_5847_cast_fp16_7, tensor var_5847_cast_fp16_8, tensor var_5847_cast_fp16_9, tensor var_5847_cast_fp16_10, tensor var_5847_cast_fp16_11, tensor var_5847_cast_fp16_12, tensor var_5847_cast_fp16_13, tensor var_5847_cast_fp16_14, tensor var_5847_cast_fp16_15, tensor var_5847_cast_fp16_16, tensor var_5847_cast_fp16_17, tensor var_5847_cast_fp16_18, tensor var_5847_cast_fp16_19 = split(axis = var_5847_axis_0, split_sizes = tile_64, x = var_5846_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor tile_65 = const()[name = tensor("tile_65"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5868_axis_0 = const()[name = tensor("op_5868_axis_0"), val = tensor(1)]; + tensor var_5868_cast_fp16_0, tensor var_5868_cast_fp16_1, tensor var_5868_cast_fp16_2, tensor var_5868_cast_fp16_3, tensor var_5868_cast_fp16_4, tensor var_5868_cast_fp16_5, tensor var_5868_cast_fp16_6, tensor var_5868_cast_fp16_7, tensor var_5868_cast_fp16_8, tensor var_5868_cast_fp16_9, tensor var_5868_cast_fp16_10, tensor var_5868_cast_fp16_11, tensor var_5868_cast_fp16_12, tensor var_5868_cast_fp16_13, tensor var_5868_cast_fp16_14, tensor var_5868_cast_fp16_15, tensor var_5868_cast_fp16_16, tensor var_5868_cast_fp16_17, tensor var_5868_cast_fp16_18, tensor var_5868_cast_fp16_19 = split(axis = var_5868_axis_0, split_sizes = tile_65, x = var_5822_cast_fp16)[name = tensor("op_5868_cast_fp16")]; + tensor aw_841_equation_0 = const()[name = tensor("aw_841_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_841_cast_fp16 = einsum(equation = aw_841_equation_0, values = (var_5847_cast_fp16_0, var_5825_cast_fp16_0))[name = tensor("aw_841_cast_fp16")]; + tensor aw_843_equation_0 = const()[name = tensor("aw_843_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_843_cast_fp16 = einsum(equation = aw_843_equation_0, values = (var_5847_cast_fp16_1, var_5825_cast_fp16_1))[name = tensor("aw_843_cast_fp16")]; + tensor aw_845_equation_0 = const()[name = tensor("aw_845_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_845_cast_fp16 = einsum(equation = aw_845_equation_0, values = (var_5847_cast_fp16_2, var_5825_cast_fp16_2))[name = tensor("aw_845_cast_fp16")]; + tensor aw_847_equation_0 = const()[name = tensor("aw_847_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_847_cast_fp16 = einsum(equation = aw_847_equation_0, values = (var_5847_cast_fp16_3, var_5825_cast_fp16_3))[name = tensor("aw_847_cast_fp16")]; + tensor aw_849_equation_0 = const()[name = tensor("aw_849_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_849_cast_fp16 = einsum(equation = aw_849_equation_0, values = (var_5847_cast_fp16_4, var_5825_cast_fp16_4))[name = tensor("aw_849_cast_fp16")]; + tensor aw_851_equation_0 = const()[name = tensor("aw_851_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_851_cast_fp16 = einsum(equation = aw_851_equation_0, values = (var_5847_cast_fp16_5, var_5825_cast_fp16_5))[name = tensor("aw_851_cast_fp16")]; + tensor aw_853_equation_0 = const()[name = tensor("aw_853_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_853_cast_fp16 = einsum(equation = aw_853_equation_0, values = (var_5847_cast_fp16_6, var_5825_cast_fp16_6))[name = tensor("aw_853_cast_fp16")]; + tensor aw_855_equation_0 = const()[name = tensor("aw_855_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_855_cast_fp16 = einsum(equation = aw_855_equation_0, values = (var_5847_cast_fp16_7, var_5825_cast_fp16_7))[name = tensor("aw_855_cast_fp16")]; + tensor aw_857_equation_0 = const()[name = tensor("aw_857_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_857_cast_fp16 = einsum(equation = aw_857_equation_0, values = (var_5847_cast_fp16_8, var_5825_cast_fp16_8))[name = tensor("aw_857_cast_fp16")]; + tensor aw_859_equation_0 = const()[name = tensor("aw_859_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_859_cast_fp16 = einsum(equation = aw_859_equation_0, values = (var_5847_cast_fp16_9, var_5825_cast_fp16_9))[name = tensor("aw_859_cast_fp16")]; + tensor aw_861_equation_0 = const()[name = tensor("aw_861_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_861_cast_fp16 = einsum(equation = aw_861_equation_0, values = (var_5847_cast_fp16_10, var_5825_cast_fp16_10))[name = tensor("aw_861_cast_fp16")]; + tensor aw_863_equation_0 = const()[name = tensor("aw_863_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_863_cast_fp16 = einsum(equation = aw_863_equation_0, values = (var_5847_cast_fp16_11, var_5825_cast_fp16_11))[name = tensor("aw_863_cast_fp16")]; + tensor aw_865_equation_0 = const()[name = tensor("aw_865_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_865_cast_fp16 = einsum(equation = aw_865_equation_0, values = (var_5847_cast_fp16_12, var_5825_cast_fp16_12))[name = tensor("aw_865_cast_fp16")]; + tensor aw_867_equation_0 = const()[name = tensor("aw_867_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_867_cast_fp16 = einsum(equation = aw_867_equation_0, values = (var_5847_cast_fp16_13, var_5825_cast_fp16_13))[name = tensor("aw_867_cast_fp16")]; + tensor aw_869_equation_0 = const()[name = tensor("aw_869_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_869_cast_fp16 = einsum(equation = aw_869_equation_0, values = (var_5847_cast_fp16_14, var_5825_cast_fp16_14))[name = tensor("aw_869_cast_fp16")]; + tensor aw_871_equation_0 = const()[name = tensor("aw_871_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_871_cast_fp16 = einsum(equation = aw_871_equation_0, values = (var_5847_cast_fp16_15, var_5825_cast_fp16_15))[name = tensor("aw_871_cast_fp16")]; + tensor aw_873_equation_0 = const()[name = tensor("aw_873_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_873_cast_fp16 = einsum(equation = aw_873_equation_0, values = (var_5847_cast_fp16_16, var_5825_cast_fp16_16))[name = tensor("aw_873_cast_fp16")]; + tensor aw_875_equation_0 = const()[name = tensor("aw_875_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_875_cast_fp16 = einsum(equation = aw_875_equation_0, values = (var_5847_cast_fp16_17, var_5825_cast_fp16_17))[name = tensor("aw_875_cast_fp16")]; + tensor aw_877_equation_0 = const()[name = tensor("aw_877_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_877_cast_fp16 = einsum(equation = aw_877_equation_0, values = (var_5847_cast_fp16_18, var_5825_cast_fp16_18))[name = tensor("aw_877_cast_fp16")]; + tensor aw_879_equation_0 = const()[name = tensor("aw_879_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_879_cast_fp16 = einsum(equation = aw_879_equation_0, values = (var_5847_cast_fp16_19, var_5825_cast_fp16_19))[name = tensor("aw_879_cast_fp16")]; + tensor var_5929_cast_fp16 = softmax(axis = var_5773, x = aw_841_cast_fp16)[name = tensor("op_5929_cast_fp16")]; + tensor var_5930_cast_fp16 = softmax(axis = var_5773, x = aw_843_cast_fp16)[name = tensor("op_5930_cast_fp16")]; + tensor var_5931_cast_fp16 = softmax(axis = var_5773, x = aw_845_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5932_cast_fp16 = softmax(axis = var_5773, x = aw_847_cast_fp16)[name = tensor("op_5932_cast_fp16")]; + tensor var_5933_cast_fp16 = softmax(axis = var_5773, x = aw_849_cast_fp16)[name = tensor("op_5933_cast_fp16")]; + tensor var_5934_cast_fp16 = softmax(axis = var_5773, x = aw_851_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = softmax(axis = var_5773, x = aw_853_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor var_5936_cast_fp16 = softmax(axis = var_5773, x = aw_855_cast_fp16)[name = tensor("op_5936_cast_fp16")]; + tensor var_5937_cast_fp16 = softmax(axis = var_5773, x = aw_857_cast_fp16)[name = tensor("op_5937_cast_fp16")]; + tensor var_5938_cast_fp16 = softmax(axis = var_5773, x = aw_859_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = softmax(axis = var_5773, x = aw_861_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor var_5940_cast_fp16 = softmax(axis = var_5773, x = aw_863_cast_fp16)[name = tensor("op_5940_cast_fp16")]; + tensor var_5941_cast_fp16 = softmax(axis = var_5773, x = aw_865_cast_fp16)[name = tensor("op_5941_cast_fp16")]; + tensor var_5942_cast_fp16 = softmax(axis = var_5773, x = aw_867_cast_fp16)[name = tensor("op_5942_cast_fp16")]; + tensor var_5943_cast_fp16 = softmax(axis = var_5773, x = aw_869_cast_fp16)[name = tensor("op_5943_cast_fp16")]; + tensor var_5944_cast_fp16 = softmax(axis = var_5773, x = aw_871_cast_fp16)[name = tensor("op_5944_cast_fp16")]; + tensor var_5945_cast_fp16 = softmax(axis = var_5773, x = aw_873_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor var_5946_cast_fp16 = softmax(axis = var_5773, x = aw_875_cast_fp16)[name = tensor("op_5946_cast_fp16")]; + tensor var_5947_cast_fp16 = softmax(axis = var_5773, x = aw_877_cast_fp16)[name = tensor("op_5947_cast_fp16")]; + tensor var_5948_cast_fp16 = softmax(axis = var_5773, x = aw_879_cast_fp16)[name = tensor("op_5948_cast_fp16")]; + tensor var_5950_equation_0 = const()[name = tensor("op_5950_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5950_cast_fp16 = einsum(equation = var_5950_equation_0, values = (var_5868_cast_fp16_0, var_5929_cast_fp16))[name = tensor("op_5950_cast_fp16")]; + tensor var_5952_equation_0 = const()[name = tensor("op_5952_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5952_cast_fp16 = einsum(equation = var_5952_equation_0, values = (var_5868_cast_fp16_1, var_5930_cast_fp16))[name = tensor("op_5952_cast_fp16")]; + tensor var_5954_equation_0 = const()[name = tensor("op_5954_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5954_cast_fp16 = einsum(equation = var_5954_equation_0, values = (var_5868_cast_fp16_2, var_5931_cast_fp16))[name = tensor("op_5954_cast_fp16")]; + tensor var_5956_equation_0 = const()[name = tensor("op_5956_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5956_cast_fp16 = einsum(equation = var_5956_equation_0, values = (var_5868_cast_fp16_3, var_5932_cast_fp16))[name = tensor("op_5956_cast_fp16")]; + tensor var_5958_equation_0 = const()[name = tensor("op_5958_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5958_cast_fp16 = einsum(equation = var_5958_equation_0, values = (var_5868_cast_fp16_4, var_5933_cast_fp16))[name = tensor("op_5958_cast_fp16")]; + tensor var_5960_equation_0 = const()[name = tensor("op_5960_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5960_cast_fp16 = einsum(equation = var_5960_equation_0, values = (var_5868_cast_fp16_5, var_5934_cast_fp16))[name = tensor("op_5960_cast_fp16")]; + tensor var_5962_equation_0 = const()[name = tensor("op_5962_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5962_cast_fp16 = einsum(equation = var_5962_equation_0, values = (var_5868_cast_fp16_6, var_5935_cast_fp16))[name = tensor("op_5962_cast_fp16")]; + tensor var_5964_equation_0 = const()[name = tensor("op_5964_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5964_cast_fp16 = einsum(equation = var_5964_equation_0, values = (var_5868_cast_fp16_7, var_5936_cast_fp16))[name = tensor("op_5964_cast_fp16")]; + tensor var_5966_equation_0 = const()[name = tensor("op_5966_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5966_cast_fp16 = einsum(equation = var_5966_equation_0, values = (var_5868_cast_fp16_8, var_5937_cast_fp16))[name = tensor("op_5966_cast_fp16")]; + tensor var_5968_equation_0 = const()[name = tensor("op_5968_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5968_cast_fp16 = einsum(equation = var_5968_equation_0, values = (var_5868_cast_fp16_9, var_5938_cast_fp16))[name = tensor("op_5968_cast_fp16")]; + tensor var_5970_equation_0 = const()[name = tensor("op_5970_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5970_cast_fp16 = einsum(equation = var_5970_equation_0, values = (var_5868_cast_fp16_10, var_5939_cast_fp16))[name = tensor("op_5970_cast_fp16")]; + tensor var_5972_equation_0 = const()[name = tensor("op_5972_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5972_cast_fp16 = einsum(equation = var_5972_equation_0, values = (var_5868_cast_fp16_11, var_5940_cast_fp16))[name = tensor("op_5972_cast_fp16")]; + tensor var_5974_equation_0 = const()[name = tensor("op_5974_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5974_cast_fp16 = einsum(equation = var_5974_equation_0, values = (var_5868_cast_fp16_12, var_5941_cast_fp16))[name = tensor("op_5974_cast_fp16")]; + tensor var_5976_equation_0 = const()[name = tensor("op_5976_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5976_cast_fp16 = einsum(equation = var_5976_equation_0, values = (var_5868_cast_fp16_13, var_5942_cast_fp16))[name = tensor("op_5976_cast_fp16")]; + tensor var_5978_equation_0 = const()[name = tensor("op_5978_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5978_cast_fp16 = einsum(equation = var_5978_equation_0, values = (var_5868_cast_fp16_14, var_5943_cast_fp16))[name = tensor("op_5978_cast_fp16")]; + tensor var_5980_equation_0 = const()[name = tensor("op_5980_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5980_cast_fp16 = einsum(equation = var_5980_equation_0, values = (var_5868_cast_fp16_15, var_5944_cast_fp16))[name = tensor("op_5980_cast_fp16")]; + tensor var_5982_equation_0 = const()[name = tensor("op_5982_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5982_cast_fp16 = einsum(equation = var_5982_equation_0, values = (var_5868_cast_fp16_16, var_5945_cast_fp16))[name = tensor("op_5982_cast_fp16")]; + tensor var_5984_equation_0 = const()[name = tensor("op_5984_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5984_cast_fp16 = einsum(equation = var_5984_equation_0, values = (var_5868_cast_fp16_17, var_5946_cast_fp16))[name = tensor("op_5984_cast_fp16")]; + tensor var_5986_equation_0 = const()[name = tensor("op_5986_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5986_cast_fp16 = einsum(equation = var_5986_equation_0, values = (var_5868_cast_fp16_18, var_5947_cast_fp16))[name = tensor("op_5986_cast_fp16")]; + tensor var_5988_equation_0 = const()[name = tensor("op_5988_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5988_cast_fp16 = einsum(equation = var_5988_equation_0, values = (var_5868_cast_fp16_19, var_5948_cast_fp16))[name = tensor("op_5988_cast_fp16")]; + tensor input_215_interleave_0 = const()[name = tensor("input_215_interleave_0"), val = tensor(false)]; + tensor input_215_cast_fp16 = concat(axis = var_5773, interleave = input_215_interleave_0, values = (var_5950_cast_fp16, var_5952_cast_fp16, var_5954_cast_fp16, var_5956_cast_fp16, var_5958_cast_fp16, var_5960_cast_fp16, var_5962_cast_fp16, var_5964_cast_fp16, var_5966_cast_fp16, var_5968_cast_fp16, var_5970_cast_fp16, var_5972_cast_fp16, var_5974_cast_fp16, var_5976_cast_fp16, var_5978_cast_fp16, var_5980_cast_fp16, var_5982_cast_fp16, var_5984_cast_fp16, var_5986_cast_fp16, var_5988_cast_fp16))[name = tensor("input_215_cast_fp16")]; + tensor var_5997_pad_type_0 = const()[name = tensor("op_5997_pad_type_0"), val = tensor("valid")]; + tensor var_5997_strides_0 = const()[name = tensor("op_5997_strides_0"), val = tensor([1, 1])]; + tensor var_5997_pad_0 = const()[name = tensor("op_5997_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5997_dilations_0 = const()[name = tensor("op_5997_dilations_0"), val = tensor([1, 1])]; + tensor var_5997_groups_0 = const()[name = tensor("op_5997_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_out_weight_to_fp16 = const()[name = tensor("blocks_21_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850550272)))]; + tensor blocks_21_attn_out_bias_to_fp16 = const()[name = tensor("blocks_21_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853827136)))]; + tensor var_5997_cast_fp16 = conv(bias = blocks_21_attn_out_bias_to_fp16, dilations = var_5997_dilations_0, groups = var_5997_groups_0, pad = var_5997_pad_0, pad_type = var_5997_pad_type_0, strides = var_5997_strides_0, weight = blocks_21_attn_out_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("op_5997_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = var_5997_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor input_217_axes_0 = const()[name = tensor("input_217_axes_0"), val = tensor([1])]; + tensor input_217_gamma_0_to_fp16 = const()[name = tensor("input_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853829760)))]; + tensor input_217_beta_0_to_fp16 = const()[name = tensor("input_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853832384)))]; + tensor var_6007_to_fp16 = const()[name = tensor("op_6007_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_217_cast_fp16 = layer_norm(axes = input_217_axes_0, beta = input_217_beta_0_to_fp16, epsilon = var_6007_to_fp16, gamma = input_217_gamma_0_to_fp16, x = inputs_87_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("valid")]; + tensor input_219_strides_0 = const()[name = tensor("input_219_strides_0"), val = tensor([1, 1])]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_219_dilations_0 = const()[name = tensor("input_219_dilations_0"), val = tensor([1, 1])]; + tensor input_219_groups_0 = const()[name = tensor("input_219_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853835008)))]; + tensor blocks_21_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(866942272)))]; + tensor input_219_cast_fp16 = conv(bias = blocks_21_mlp_0_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = blocks_21_mlp_0_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_6033_pad_type_0 = const()[name = tensor("op_6033_pad_type_0"), val = tensor("valid")]; + tensor var_6033_strides_0 = const()[name = tensor("op_6033_strides_0"), val = tensor([1, 1])]; + tensor var_6033_pad_0 = const()[name = tensor("op_6033_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6033_dilations_0 = const()[name = tensor("op_6033_dilations_0"), val = tensor([1, 1])]; + tensor var_6033_groups_0 = const()[name = tensor("op_6033_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(866952576)))]; + tensor blocks_21_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880059840)))]; + tensor var_6033_cast_fp16 = conv(bias = blocks_21_mlp_2_bias_to_fp16, dilations = var_6033_dilations_0, groups = var_6033_groups_0, pad = var_6033_pad_0, pad_type = var_6033_pad_type_0, strides = var_6033_strides_0, weight = blocks_21_mlp_2_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("op_6033_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_6033_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(1)]; + tensor input_223_axes_0 = const()[name = tensor("input_223_axes_0"), val = tensor([1])]; + tensor input_223_gamma_0_to_fp16 = const()[name = tensor("input_223_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880062464)))]; + tensor input_223_beta_0_to_fp16 = const()[name = tensor("input_223_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880065088)))]; + tensor var_6058_to_fp16 = const()[name = tensor("op_6058_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_223_cast_fp16 = layer_norm(axes = input_223_axes_0, beta = input_223_beta_0_to_fp16, epsilon = var_6058_to_fp16, gamma = input_223_gamma_0_to_fp16, x = inputs_89_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor q_45_pad_type_0 = const()[name = tensor("q_45_pad_type_0"), val = tensor("valid")]; + tensor q_45_strides_0 = const()[name = tensor("q_45_strides_0"), val = tensor([1, 1])]; + tensor q_45_pad_0 = const()[name = tensor("q_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_45_dilations_0 = const()[name = tensor("q_45_dilations_0"), val = tensor([1, 1])]; + tensor q_45_groups_0 = const()[name = tensor("q_45_groups_0"), val = tensor(1)]; + tensor var_6093_weight_0_to_fp16 = const()[name = tensor("op_6093_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880067712)))]; + tensor var_6093_bias_0_to_fp16 = const()[name = tensor("op_6093_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883344576)))]; + tensor var_6093_cast_fp16 = conv(bias = var_6093_bias_0_to_fp16, dilations = q_45_dilations_0, groups = q_45_groups_0, pad = q_45_pad_0, pad_type = q_45_pad_type_0, strides = q_45_strides_0, weight = var_6093_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("op_6093_cast_fp16")]; + tensor k_45_pad_type_0 = const()[name = tensor("k_45_pad_type_0"), val = tensor("valid")]; + tensor k_45_strides_0 = const()[name = tensor("k_45_strides_0"), val = tensor([1, 1])]; + tensor k_45_pad_0 = const()[name = tensor("k_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_45_dilations_0 = const()[name = tensor("k_45_dilations_0"), val = tensor([1, 1])]; + tensor k_45_groups_0 = const()[name = tensor("k_45_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_key_weight_to_fp16 = const()[name = tensor("blocks_22_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883347200)))]; + tensor k_45_cast_fp16 = conv(dilations = k_45_dilations_0, groups = k_45_groups_0, pad = k_45_pad_0, pad_type = k_45_pad_type_0, strides = k_45_strides_0, weight = blocks_22_attn_key_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor var_6091_pad_type_0 = const()[name = tensor("op_6091_pad_type_0"), val = tensor("valid")]; + tensor var_6091_strides_0 = const()[name = tensor("op_6091_strides_0"), val = tensor([1, 1])]; + tensor var_6091_pad_0 = const()[name = tensor("op_6091_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6091_dilations_0 = const()[name = tensor("op_6091_dilations_0"), val = tensor([1, 1])]; + tensor var_6091_groups_0 = const()[name = tensor("op_6091_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_value_weight_to_fp16 = const()[name = tensor("blocks_22_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886624064)))]; + tensor blocks_22_attn_value_bias_to_fp16 = const()[name = tensor("blocks_22_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889900928)))]; + tensor var_6091_cast_fp16 = conv(bias = blocks_22_attn_value_bias_to_fp16, dilations = var_6091_dilations_0, groups = var_6091_groups_0, pad = var_6091_pad_0, pad_type = var_6091_pad_type_0, strides = var_6091_strides_0, weight = blocks_22_attn_value_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor tile_66 = const()[name = tensor("tile_66"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6094_axis_0 = const()[name = tensor("op_6094_axis_0"), val = tensor(1)]; + tensor var_6094_cast_fp16_0, tensor var_6094_cast_fp16_1, tensor var_6094_cast_fp16_2, tensor var_6094_cast_fp16_3, tensor var_6094_cast_fp16_4, tensor var_6094_cast_fp16_5, tensor var_6094_cast_fp16_6, tensor var_6094_cast_fp16_7, tensor var_6094_cast_fp16_8, tensor var_6094_cast_fp16_9, tensor var_6094_cast_fp16_10, tensor var_6094_cast_fp16_11, tensor var_6094_cast_fp16_12, tensor var_6094_cast_fp16_13, tensor var_6094_cast_fp16_14, tensor var_6094_cast_fp16_15, tensor var_6094_cast_fp16_16, tensor var_6094_cast_fp16_17, tensor var_6094_cast_fp16_18, tensor var_6094_cast_fp16_19 = split(axis = var_6094_axis_0, split_sizes = tile_66, x = var_6093_cast_fp16)[name = tensor("op_6094_cast_fp16")]; + tensor var_6115_perm_0 = const()[name = tensor("op_6115_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_67 = const()[name = tensor("tile_67"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6116_axis_0 = const()[name = tensor("op_6116_axis_0"), val = tensor(3)]; + tensor var_6115_cast_fp16 = transpose(perm = var_6115_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_10")]; + tensor var_6116_cast_fp16_0, tensor var_6116_cast_fp16_1, tensor var_6116_cast_fp16_2, tensor var_6116_cast_fp16_3, tensor var_6116_cast_fp16_4, tensor var_6116_cast_fp16_5, tensor var_6116_cast_fp16_6, tensor var_6116_cast_fp16_7, tensor var_6116_cast_fp16_8, tensor var_6116_cast_fp16_9, tensor var_6116_cast_fp16_10, tensor var_6116_cast_fp16_11, tensor var_6116_cast_fp16_12, tensor var_6116_cast_fp16_13, tensor var_6116_cast_fp16_14, tensor var_6116_cast_fp16_15, tensor var_6116_cast_fp16_16, tensor var_6116_cast_fp16_17, tensor var_6116_cast_fp16_18, tensor var_6116_cast_fp16_19 = split(axis = var_6116_axis_0, split_sizes = tile_67, x = var_6115_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor tile_68 = const()[name = tensor("tile_68"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6137_axis_0 = const()[name = tensor("op_6137_axis_0"), val = tensor(1)]; + tensor var_6137_cast_fp16_0, tensor var_6137_cast_fp16_1, tensor var_6137_cast_fp16_2, tensor var_6137_cast_fp16_3, tensor var_6137_cast_fp16_4, tensor var_6137_cast_fp16_5, tensor var_6137_cast_fp16_6, tensor var_6137_cast_fp16_7, tensor var_6137_cast_fp16_8, tensor var_6137_cast_fp16_9, tensor var_6137_cast_fp16_10, tensor var_6137_cast_fp16_11, tensor var_6137_cast_fp16_12, tensor var_6137_cast_fp16_13, tensor var_6137_cast_fp16_14, tensor var_6137_cast_fp16_15, tensor var_6137_cast_fp16_16, tensor var_6137_cast_fp16_17, tensor var_6137_cast_fp16_18, tensor var_6137_cast_fp16_19 = split(axis = var_6137_axis_0, split_sizes = tile_68, x = var_6091_cast_fp16)[name = tensor("op_6137_cast_fp16")]; + tensor aw_881_equation_0 = const()[name = tensor("aw_881_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_881_cast_fp16 = einsum(equation = aw_881_equation_0, values = (var_6116_cast_fp16_0, var_6094_cast_fp16_0))[name = tensor("aw_881_cast_fp16")]; + tensor aw_883_equation_0 = const()[name = tensor("aw_883_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_883_cast_fp16 = einsum(equation = aw_883_equation_0, values = (var_6116_cast_fp16_1, var_6094_cast_fp16_1))[name = tensor("aw_883_cast_fp16")]; + tensor aw_885_equation_0 = const()[name = tensor("aw_885_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_885_cast_fp16 = einsum(equation = aw_885_equation_0, values = (var_6116_cast_fp16_2, var_6094_cast_fp16_2))[name = tensor("aw_885_cast_fp16")]; + tensor aw_887_equation_0 = const()[name = tensor("aw_887_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_887_cast_fp16 = einsum(equation = aw_887_equation_0, values = (var_6116_cast_fp16_3, var_6094_cast_fp16_3))[name = tensor("aw_887_cast_fp16")]; + tensor aw_889_equation_0 = const()[name = tensor("aw_889_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_889_cast_fp16 = einsum(equation = aw_889_equation_0, values = (var_6116_cast_fp16_4, var_6094_cast_fp16_4))[name = tensor("aw_889_cast_fp16")]; + tensor aw_891_equation_0 = const()[name = tensor("aw_891_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_891_cast_fp16 = einsum(equation = aw_891_equation_0, values = (var_6116_cast_fp16_5, var_6094_cast_fp16_5))[name = tensor("aw_891_cast_fp16")]; + tensor aw_893_equation_0 = const()[name = tensor("aw_893_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_893_cast_fp16 = einsum(equation = aw_893_equation_0, values = (var_6116_cast_fp16_6, var_6094_cast_fp16_6))[name = tensor("aw_893_cast_fp16")]; + tensor aw_895_equation_0 = const()[name = tensor("aw_895_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_895_cast_fp16 = einsum(equation = aw_895_equation_0, values = (var_6116_cast_fp16_7, var_6094_cast_fp16_7))[name = tensor("aw_895_cast_fp16")]; + tensor aw_897_equation_0 = const()[name = tensor("aw_897_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_897_cast_fp16 = einsum(equation = aw_897_equation_0, values = (var_6116_cast_fp16_8, var_6094_cast_fp16_8))[name = tensor("aw_897_cast_fp16")]; + tensor aw_899_equation_0 = const()[name = tensor("aw_899_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_899_cast_fp16 = einsum(equation = aw_899_equation_0, values = (var_6116_cast_fp16_9, var_6094_cast_fp16_9))[name = tensor("aw_899_cast_fp16")]; + tensor aw_901_equation_0 = const()[name = tensor("aw_901_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_901_cast_fp16 = einsum(equation = aw_901_equation_0, values = (var_6116_cast_fp16_10, var_6094_cast_fp16_10))[name = tensor("aw_901_cast_fp16")]; + tensor aw_903_equation_0 = const()[name = tensor("aw_903_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_903_cast_fp16 = einsum(equation = aw_903_equation_0, values = (var_6116_cast_fp16_11, var_6094_cast_fp16_11))[name = tensor("aw_903_cast_fp16")]; + tensor aw_905_equation_0 = const()[name = tensor("aw_905_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_905_cast_fp16 = einsum(equation = aw_905_equation_0, values = (var_6116_cast_fp16_12, var_6094_cast_fp16_12))[name = tensor("aw_905_cast_fp16")]; + tensor aw_907_equation_0 = const()[name = tensor("aw_907_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_907_cast_fp16 = einsum(equation = aw_907_equation_0, values = (var_6116_cast_fp16_13, var_6094_cast_fp16_13))[name = tensor("aw_907_cast_fp16")]; + tensor aw_909_equation_0 = const()[name = tensor("aw_909_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_909_cast_fp16 = einsum(equation = aw_909_equation_0, values = (var_6116_cast_fp16_14, var_6094_cast_fp16_14))[name = tensor("aw_909_cast_fp16")]; + tensor aw_911_equation_0 = const()[name = tensor("aw_911_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_911_cast_fp16 = einsum(equation = aw_911_equation_0, values = (var_6116_cast_fp16_15, var_6094_cast_fp16_15))[name = tensor("aw_911_cast_fp16")]; + tensor aw_913_equation_0 = const()[name = tensor("aw_913_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_913_cast_fp16 = einsum(equation = aw_913_equation_0, values = (var_6116_cast_fp16_16, var_6094_cast_fp16_16))[name = tensor("aw_913_cast_fp16")]; + tensor aw_915_equation_0 = const()[name = tensor("aw_915_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_915_cast_fp16 = einsum(equation = aw_915_equation_0, values = (var_6116_cast_fp16_17, var_6094_cast_fp16_17))[name = tensor("aw_915_cast_fp16")]; + tensor aw_917_equation_0 = const()[name = tensor("aw_917_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_917_cast_fp16 = einsum(equation = aw_917_equation_0, values = (var_6116_cast_fp16_18, var_6094_cast_fp16_18))[name = tensor("aw_917_cast_fp16")]; + tensor aw_919_equation_0 = const()[name = tensor("aw_919_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_919_cast_fp16 = einsum(equation = aw_919_equation_0, values = (var_6116_cast_fp16_19, var_6094_cast_fp16_19))[name = tensor("aw_919_cast_fp16")]; + tensor var_6198_cast_fp16 = softmax(axis = var_6042, x = aw_881_cast_fp16)[name = tensor("op_6198_cast_fp16")]; + tensor var_6199_cast_fp16 = softmax(axis = var_6042, x = aw_883_cast_fp16)[name = tensor("op_6199_cast_fp16")]; + tensor var_6200_cast_fp16 = softmax(axis = var_6042, x = aw_885_cast_fp16)[name = tensor("op_6200_cast_fp16")]; + tensor var_6201_cast_fp16 = softmax(axis = var_6042, x = aw_887_cast_fp16)[name = tensor("op_6201_cast_fp16")]; + tensor var_6202_cast_fp16 = softmax(axis = var_6042, x = aw_889_cast_fp16)[name = tensor("op_6202_cast_fp16")]; + tensor var_6203_cast_fp16 = softmax(axis = var_6042, x = aw_891_cast_fp16)[name = tensor("op_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = softmax(axis = var_6042, x = aw_893_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_cast_fp16 = softmax(axis = var_6042, x = aw_895_cast_fp16)[name = tensor("op_6205_cast_fp16")]; + tensor var_6206_cast_fp16 = softmax(axis = var_6042, x = aw_897_cast_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor var_6207_cast_fp16 = softmax(axis = var_6042, x = aw_899_cast_fp16)[name = tensor("op_6207_cast_fp16")]; + tensor var_6208_cast_fp16 = softmax(axis = var_6042, x = aw_901_cast_fp16)[name = tensor("op_6208_cast_fp16")]; + tensor var_6209_cast_fp16 = softmax(axis = var_6042, x = aw_903_cast_fp16)[name = tensor("op_6209_cast_fp16")]; + tensor var_6210_cast_fp16 = softmax(axis = var_6042, x = aw_905_cast_fp16)[name = tensor("op_6210_cast_fp16")]; + tensor var_6211_cast_fp16 = softmax(axis = var_6042, x = aw_907_cast_fp16)[name = tensor("op_6211_cast_fp16")]; + tensor var_6212_cast_fp16 = softmax(axis = var_6042, x = aw_909_cast_fp16)[name = tensor("op_6212_cast_fp16")]; + tensor var_6213_cast_fp16 = softmax(axis = var_6042, x = aw_911_cast_fp16)[name = tensor("op_6213_cast_fp16")]; + tensor var_6214_cast_fp16 = softmax(axis = var_6042, x = aw_913_cast_fp16)[name = tensor("op_6214_cast_fp16")]; + tensor var_6215_cast_fp16 = softmax(axis = var_6042, x = aw_915_cast_fp16)[name = tensor("op_6215_cast_fp16")]; + tensor var_6216_cast_fp16 = softmax(axis = var_6042, x = aw_917_cast_fp16)[name = tensor("op_6216_cast_fp16")]; + tensor var_6217_cast_fp16 = softmax(axis = var_6042, x = aw_919_cast_fp16)[name = tensor("op_6217_cast_fp16")]; + tensor var_6219_equation_0 = const()[name = tensor("op_6219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6219_cast_fp16 = einsum(equation = var_6219_equation_0, values = (var_6137_cast_fp16_0, var_6198_cast_fp16))[name = tensor("op_6219_cast_fp16")]; + tensor var_6221_equation_0 = const()[name = tensor("op_6221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6221_cast_fp16 = einsum(equation = var_6221_equation_0, values = (var_6137_cast_fp16_1, var_6199_cast_fp16))[name = tensor("op_6221_cast_fp16")]; + tensor var_6223_equation_0 = const()[name = tensor("op_6223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6223_cast_fp16 = einsum(equation = var_6223_equation_0, values = (var_6137_cast_fp16_2, var_6200_cast_fp16))[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_equation_0 = const()[name = tensor("op_6225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6225_cast_fp16 = einsum(equation = var_6225_equation_0, values = (var_6137_cast_fp16_3, var_6201_cast_fp16))[name = tensor("op_6225_cast_fp16")]; + tensor var_6227_equation_0 = const()[name = tensor("op_6227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6227_cast_fp16 = einsum(equation = var_6227_equation_0, values = (var_6137_cast_fp16_4, var_6202_cast_fp16))[name = tensor("op_6227_cast_fp16")]; + tensor var_6229_equation_0 = const()[name = tensor("op_6229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6229_cast_fp16 = einsum(equation = var_6229_equation_0, values = (var_6137_cast_fp16_5, var_6203_cast_fp16))[name = tensor("op_6229_cast_fp16")]; + tensor var_6231_equation_0 = const()[name = tensor("op_6231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6231_cast_fp16 = einsum(equation = var_6231_equation_0, values = (var_6137_cast_fp16_6, var_6204_cast_fp16))[name = tensor("op_6231_cast_fp16")]; + tensor var_6233_equation_0 = const()[name = tensor("op_6233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6233_cast_fp16 = einsum(equation = var_6233_equation_0, values = (var_6137_cast_fp16_7, var_6205_cast_fp16))[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_equation_0 = const()[name = tensor("op_6235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6235_cast_fp16 = einsum(equation = var_6235_equation_0, values = (var_6137_cast_fp16_8, var_6206_cast_fp16))[name = tensor("op_6235_cast_fp16")]; + tensor var_6237_equation_0 = const()[name = tensor("op_6237_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6237_cast_fp16 = einsum(equation = var_6237_equation_0, values = (var_6137_cast_fp16_9, var_6207_cast_fp16))[name = tensor("op_6237_cast_fp16")]; + tensor var_6239_equation_0 = const()[name = tensor("op_6239_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6239_cast_fp16 = einsum(equation = var_6239_equation_0, values = (var_6137_cast_fp16_10, var_6208_cast_fp16))[name = tensor("op_6239_cast_fp16")]; + tensor var_6241_equation_0 = const()[name = tensor("op_6241_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6241_cast_fp16 = einsum(equation = var_6241_equation_0, values = (var_6137_cast_fp16_11, var_6209_cast_fp16))[name = tensor("op_6241_cast_fp16")]; + tensor var_6243_equation_0 = const()[name = tensor("op_6243_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6243_cast_fp16 = einsum(equation = var_6243_equation_0, values = (var_6137_cast_fp16_12, var_6210_cast_fp16))[name = tensor("op_6243_cast_fp16")]; + tensor var_6245_equation_0 = const()[name = tensor("op_6245_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6245_cast_fp16 = einsum(equation = var_6245_equation_0, values = (var_6137_cast_fp16_13, var_6211_cast_fp16))[name = tensor("op_6245_cast_fp16")]; + tensor var_6247_equation_0 = const()[name = tensor("op_6247_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6247_cast_fp16 = einsum(equation = var_6247_equation_0, values = (var_6137_cast_fp16_14, var_6212_cast_fp16))[name = tensor("op_6247_cast_fp16")]; + tensor var_6249_equation_0 = const()[name = tensor("op_6249_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6249_cast_fp16 = einsum(equation = var_6249_equation_0, values = (var_6137_cast_fp16_15, var_6213_cast_fp16))[name = tensor("op_6249_cast_fp16")]; + tensor var_6251_equation_0 = const()[name = tensor("op_6251_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6251_cast_fp16 = einsum(equation = var_6251_equation_0, values = (var_6137_cast_fp16_16, var_6214_cast_fp16))[name = tensor("op_6251_cast_fp16")]; + tensor var_6253_equation_0 = const()[name = tensor("op_6253_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6253_cast_fp16 = einsum(equation = var_6253_equation_0, values = (var_6137_cast_fp16_17, var_6215_cast_fp16))[name = tensor("op_6253_cast_fp16")]; + tensor var_6255_equation_0 = const()[name = tensor("op_6255_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6255_cast_fp16 = einsum(equation = var_6255_equation_0, values = (var_6137_cast_fp16_18, var_6216_cast_fp16))[name = tensor("op_6255_cast_fp16")]; + tensor var_6257_equation_0 = const()[name = tensor("op_6257_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6257_cast_fp16 = einsum(equation = var_6257_equation_0, values = (var_6137_cast_fp16_19, var_6217_cast_fp16))[name = tensor("op_6257_cast_fp16")]; + tensor input_225_interleave_0 = const()[name = tensor("input_225_interleave_0"), val = tensor(false)]; + tensor input_225_cast_fp16 = concat(axis = var_6042, interleave = input_225_interleave_0, values = (var_6219_cast_fp16, var_6221_cast_fp16, var_6223_cast_fp16, var_6225_cast_fp16, var_6227_cast_fp16, var_6229_cast_fp16, var_6231_cast_fp16, var_6233_cast_fp16, var_6235_cast_fp16, var_6237_cast_fp16, var_6239_cast_fp16, var_6241_cast_fp16, var_6243_cast_fp16, var_6245_cast_fp16, var_6247_cast_fp16, var_6249_cast_fp16, var_6251_cast_fp16, var_6253_cast_fp16, var_6255_cast_fp16, var_6257_cast_fp16))[name = tensor("input_225_cast_fp16")]; + tensor var_6266_pad_type_0 = const()[name = tensor("op_6266_pad_type_0"), val = tensor("valid")]; + tensor var_6266_strides_0 = const()[name = tensor("op_6266_strides_0"), val = tensor([1, 1])]; + tensor var_6266_pad_0 = const()[name = tensor("op_6266_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6266_dilations_0 = const()[name = tensor("op_6266_dilations_0"), val = tensor([1, 1])]; + tensor var_6266_groups_0 = const()[name = tensor("op_6266_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_out_weight_to_fp16 = const()[name = tensor("blocks_22_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889903552)))]; + tensor blocks_22_attn_out_bias_to_fp16 = const()[name = tensor("blocks_22_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893180416)))]; + tensor var_6266_cast_fp16 = conv(bias = blocks_22_attn_out_bias_to_fp16, dilations = var_6266_dilations_0, groups = var_6266_groups_0, pad = var_6266_pad_0, pad_type = var_6266_pad_type_0, strides = var_6266_strides_0, weight = blocks_22_attn_out_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("op_6266_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = var_6266_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([1])]; + tensor input_227_gamma_0_to_fp16 = const()[name = tensor("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893183040)))]; + tensor input_227_beta_0_to_fp16 = const()[name = tensor("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893185664)))]; + tensor var_6276_to_fp16 = const()[name = tensor("op_6276_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = input_227_beta_0_to_fp16, epsilon = var_6276_to_fp16, gamma = input_227_gamma_0_to_fp16, x = inputs_91_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("valid")]; + tensor input_229_strides_0 = const()[name = tensor("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = tensor("input_229_dilations_0"), val = tensor([1, 1])]; + tensor input_229_groups_0 = const()[name = tensor("input_229_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893188288)))]; + tensor blocks_22_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906295552)))]; + tensor input_229_cast_fp16 = conv(bias = blocks_22_mlp_0_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = blocks_22_mlp_0_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_6302_pad_type_0 = const()[name = tensor("op_6302_pad_type_0"), val = tensor("valid")]; + tensor var_6302_strides_0 = const()[name = tensor("op_6302_strides_0"), val = tensor([1, 1])]; + tensor var_6302_pad_0 = const()[name = tensor("op_6302_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6302_dilations_0 = const()[name = tensor("op_6302_dilations_0"), val = tensor([1, 1])]; + tensor var_6302_groups_0 = const()[name = tensor("op_6302_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906305856)))]; + tensor blocks_22_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919413120)))]; + tensor var_6302_cast_fp16 = conv(bias = blocks_22_mlp_2_bias_to_fp16, dilations = var_6302_dilations_0, groups = var_6302_groups_0, pad = var_6302_pad_0, pad_type = var_6302_pad_type_0, strides = var_6302_strides_0, weight = blocks_22_mlp_2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("op_6302_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_6302_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_6311 = const()[name = tensor("op_6311"), val = tensor(1)]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([1])]; + tensor input_233_gamma_0_to_fp16 = const()[name = tensor("input_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919415744)))]; + tensor input_233_beta_0_to_fp16 = const()[name = tensor("input_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919418368)))]; + tensor var_6327_to_fp16 = const()[name = tensor("op_6327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = input_233_beta_0_to_fp16, epsilon = var_6327_to_fp16, gamma = input_233_gamma_0_to_fp16, x = inputs_93_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor q_47_pad_type_0 = const()[name = tensor("q_47_pad_type_0"), val = tensor("valid")]; + tensor q_47_strides_0 = const()[name = tensor("q_47_strides_0"), val = tensor([1, 1])]; + tensor q_47_pad_0 = const()[name = tensor("q_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_47_dilations_0 = const()[name = tensor("q_47_dilations_0"), val = tensor([1, 1])]; + tensor q_47_groups_0 = const()[name = tensor("q_47_groups_0"), val = tensor(1)]; + tensor var_6362_weight_0_to_fp16 = const()[name = tensor("op_6362_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919420992)))]; + tensor var_6362_bias_0_to_fp16 = const()[name = tensor("op_6362_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922697856)))]; + tensor var_6362_cast_fp16 = conv(bias = var_6362_bias_0_to_fp16, dilations = q_47_dilations_0, groups = q_47_groups_0, pad = q_47_pad_0, pad_type = q_47_pad_type_0, strides = q_47_strides_0, weight = var_6362_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("op_6362_cast_fp16")]; + tensor k_47_pad_type_0 = const()[name = tensor("k_47_pad_type_0"), val = tensor("valid")]; + tensor k_47_strides_0 = const()[name = tensor("k_47_strides_0"), val = tensor([1, 1])]; + tensor k_47_pad_0 = const()[name = tensor("k_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_47_dilations_0 = const()[name = tensor("k_47_dilations_0"), val = tensor([1, 1])]; + tensor k_47_groups_0 = const()[name = tensor("k_47_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_key_weight_to_fp16 = const()[name = tensor("blocks_23_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922700480)))]; + tensor k_47_cast_fp16 = conv(dilations = k_47_dilations_0, groups = k_47_groups_0, pad = k_47_pad_0, pad_type = k_47_pad_type_0, strides = k_47_strides_0, weight = blocks_23_attn_key_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor var_6360_pad_type_0 = const()[name = tensor("op_6360_pad_type_0"), val = tensor("valid")]; + tensor var_6360_strides_0 = const()[name = tensor("op_6360_strides_0"), val = tensor([1, 1])]; + tensor var_6360_pad_0 = const()[name = tensor("op_6360_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6360_dilations_0 = const()[name = tensor("op_6360_dilations_0"), val = tensor([1, 1])]; + tensor var_6360_groups_0 = const()[name = tensor("op_6360_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_value_weight_to_fp16 = const()[name = tensor("blocks_23_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(925977344)))]; + tensor blocks_23_attn_value_bias_to_fp16 = const()[name = tensor("blocks_23_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929254208)))]; + tensor var_6360_cast_fp16 = conv(bias = blocks_23_attn_value_bias_to_fp16, dilations = var_6360_dilations_0, groups = var_6360_groups_0, pad = var_6360_pad_0, pad_type = var_6360_pad_type_0, strides = var_6360_strides_0, weight = blocks_23_attn_value_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("op_6360_cast_fp16")]; + tensor tile_69 = const()[name = tensor("tile_69"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6363_axis_0 = const()[name = tensor("op_6363_axis_0"), val = tensor(1)]; + tensor var_6363_cast_fp16_0, tensor var_6363_cast_fp16_1, tensor var_6363_cast_fp16_2, tensor var_6363_cast_fp16_3, tensor var_6363_cast_fp16_4, tensor var_6363_cast_fp16_5, tensor var_6363_cast_fp16_6, tensor var_6363_cast_fp16_7, tensor var_6363_cast_fp16_8, tensor var_6363_cast_fp16_9, tensor var_6363_cast_fp16_10, tensor var_6363_cast_fp16_11, tensor var_6363_cast_fp16_12, tensor var_6363_cast_fp16_13, tensor var_6363_cast_fp16_14, tensor var_6363_cast_fp16_15, tensor var_6363_cast_fp16_16, tensor var_6363_cast_fp16_17, tensor var_6363_cast_fp16_18, tensor var_6363_cast_fp16_19 = split(axis = var_6363_axis_0, split_sizes = tile_69, x = var_6362_cast_fp16)[name = tensor("op_6363_cast_fp16")]; + tensor var_6384_perm_0 = const()[name = tensor("op_6384_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_70 = const()[name = tensor("tile_70"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6385_axis_0 = const()[name = tensor("op_6385_axis_0"), val = tensor(3)]; + tensor var_6384_cast_fp16 = transpose(perm = var_6384_perm_0, x = k_47_cast_fp16)[name = tensor("transpose_9")]; + tensor var_6385_cast_fp16_0, tensor var_6385_cast_fp16_1, tensor var_6385_cast_fp16_2, tensor var_6385_cast_fp16_3, tensor var_6385_cast_fp16_4, tensor var_6385_cast_fp16_5, tensor var_6385_cast_fp16_6, tensor var_6385_cast_fp16_7, tensor var_6385_cast_fp16_8, tensor var_6385_cast_fp16_9, tensor var_6385_cast_fp16_10, tensor var_6385_cast_fp16_11, tensor var_6385_cast_fp16_12, tensor var_6385_cast_fp16_13, tensor var_6385_cast_fp16_14, tensor var_6385_cast_fp16_15, tensor var_6385_cast_fp16_16, tensor var_6385_cast_fp16_17, tensor var_6385_cast_fp16_18, tensor var_6385_cast_fp16_19 = split(axis = var_6385_axis_0, split_sizes = tile_70, x = var_6384_cast_fp16)[name = tensor("op_6385_cast_fp16")]; + tensor tile_71 = const()[name = tensor("tile_71"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6406_axis_0 = const()[name = tensor("op_6406_axis_0"), val = tensor(1)]; + tensor var_6406_cast_fp16_0, tensor var_6406_cast_fp16_1, tensor var_6406_cast_fp16_2, tensor var_6406_cast_fp16_3, tensor var_6406_cast_fp16_4, tensor var_6406_cast_fp16_5, tensor var_6406_cast_fp16_6, tensor var_6406_cast_fp16_7, tensor var_6406_cast_fp16_8, tensor var_6406_cast_fp16_9, tensor var_6406_cast_fp16_10, tensor var_6406_cast_fp16_11, tensor var_6406_cast_fp16_12, tensor var_6406_cast_fp16_13, tensor var_6406_cast_fp16_14, tensor var_6406_cast_fp16_15, tensor var_6406_cast_fp16_16, tensor var_6406_cast_fp16_17, tensor var_6406_cast_fp16_18, tensor var_6406_cast_fp16_19 = split(axis = var_6406_axis_0, split_sizes = tile_71, x = var_6360_cast_fp16)[name = tensor("op_6406_cast_fp16")]; + tensor aw_921_equation_0 = const()[name = tensor("aw_921_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_921_cast_fp16 = einsum(equation = aw_921_equation_0, values = (var_6385_cast_fp16_0, var_6363_cast_fp16_0))[name = tensor("aw_921_cast_fp16")]; + tensor aw_923_equation_0 = const()[name = tensor("aw_923_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_923_cast_fp16 = einsum(equation = aw_923_equation_0, values = (var_6385_cast_fp16_1, var_6363_cast_fp16_1))[name = tensor("aw_923_cast_fp16")]; + tensor aw_925_equation_0 = const()[name = tensor("aw_925_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_925_cast_fp16 = einsum(equation = aw_925_equation_0, values = (var_6385_cast_fp16_2, var_6363_cast_fp16_2))[name = tensor("aw_925_cast_fp16")]; + tensor aw_927_equation_0 = const()[name = tensor("aw_927_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_927_cast_fp16 = einsum(equation = aw_927_equation_0, values = (var_6385_cast_fp16_3, var_6363_cast_fp16_3))[name = tensor("aw_927_cast_fp16")]; + tensor aw_929_equation_0 = const()[name = tensor("aw_929_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_929_cast_fp16 = einsum(equation = aw_929_equation_0, values = (var_6385_cast_fp16_4, var_6363_cast_fp16_4))[name = tensor("aw_929_cast_fp16")]; + tensor aw_931_equation_0 = const()[name = tensor("aw_931_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_931_cast_fp16 = einsum(equation = aw_931_equation_0, values = (var_6385_cast_fp16_5, var_6363_cast_fp16_5))[name = tensor("aw_931_cast_fp16")]; + tensor aw_933_equation_0 = const()[name = tensor("aw_933_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_933_cast_fp16 = einsum(equation = aw_933_equation_0, values = (var_6385_cast_fp16_6, var_6363_cast_fp16_6))[name = tensor("aw_933_cast_fp16")]; + tensor aw_935_equation_0 = const()[name = tensor("aw_935_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_935_cast_fp16 = einsum(equation = aw_935_equation_0, values = (var_6385_cast_fp16_7, var_6363_cast_fp16_7))[name = tensor("aw_935_cast_fp16")]; + tensor aw_937_equation_0 = const()[name = tensor("aw_937_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_937_cast_fp16 = einsum(equation = aw_937_equation_0, values = (var_6385_cast_fp16_8, var_6363_cast_fp16_8))[name = tensor("aw_937_cast_fp16")]; + tensor aw_939_equation_0 = const()[name = tensor("aw_939_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_939_cast_fp16 = einsum(equation = aw_939_equation_0, values = (var_6385_cast_fp16_9, var_6363_cast_fp16_9))[name = tensor("aw_939_cast_fp16")]; + tensor aw_941_equation_0 = const()[name = tensor("aw_941_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_941_cast_fp16 = einsum(equation = aw_941_equation_0, values = (var_6385_cast_fp16_10, var_6363_cast_fp16_10))[name = tensor("aw_941_cast_fp16")]; + tensor aw_943_equation_0 = const()[name = tensor("aw_943_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_943_cast_fp16 = einsum(equation = aw_943_equation_0, values = (var_6385_cast_fp16_11, var_6363_cast_fp16_11))[name = tensor("aw_943_cast_fp16")]; + tensor aw_945_equation_0 = const()[name = tensor("aw_945_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_945_cast_fp16 = einsum(equation = aw_945_equation_0, values = (var_6385_cast_fp16_12, var_6363_cast_fp16_12))[name = tensor("aw_945_cast_fp16")]; + tensor aw_947_equation_0 = const()[name = tensor("aw_947_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_947_cast_fp16 = einsum(equation = aw_947_equation_0, values = (var_6385_cast_fp16_13, var_6363_cast_fp16_13))[name = tensor("aw_947_cast_fp16")]; + tensor aw_949_equation_0 = const()[name = tensor("aw_949_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_949_cast_fp16 = einsum(equation = aw_949_equation_0, values = (var_6385_cast_fp16_14, var_6363_cast_fp16_14))[name = tensor("aw_949_cast_fp16")]; + tensor aw_951_equation_0 = const()[name = tensor("aw_951_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_951_cast_fp16 = einsum(equation = aw_951_equation_0, values = (var_6385_cast_fp16_15, var_6363_cast_fp16_15))[name = tensor("aw_951_cast_fp16")]; + tensor aw_953_equation_0 = const()[name = tensor("aw_953_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_953_cast_fp16 = einsum(equation = aw_953_equation_0, values = (var_6385_cast_fp16_16, var_6363_cast_fp16_16))[name = tensor("aw_953_cast_fp16")]; + tensor aw_955_equation_0 = const()[name = tensor("aw_955_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_955_cast_fp16 = einsum(equation = aw_955_equation_0, values = (var_6385_cast_fp16_17, var_6363_cast_fp16_17))[name = tensor("aw_955_cast_fp16")]; + tensor aw_957_equation_0 = const()[name = tensor("aw_957_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_957_cast_fp16 = einsum(equation = aw_957_equation_0, values = (var_6385_cast_fp16_18, var_6363_cast_fp16_18))[name = tensor("aw_957_cast_fp16")]; + tensor aw_959_equation_0 = const()[name = tensor("aw_959_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_959_cast_fp16 = einsum(equation = aw_959_equation_0, values = (var_6385_cast_fp16_19, var_6363_cast_fp16_19))[name = tensor("aw_959_cast_fp16")]; + tensor var_6467_cast_fp16 = softmax(axis = var_6311, x = aw_921_cast_fp16)[name = tensor("op_6467_cast_fp16")]; + tensor var_6468_cast_fp16 = softmax(axis = var_6311, x = aw_923_cast_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6469_cast_fp16 = softmax(axis = var_6311, x = aw_925_cast_fp16)[name = tensor("op_6469_cast_fp16")]; + tensor var_6470_cast_fp16 = softmax(axis = var_6311, x = aw_927_cast_fp16)[name = tensor("op_6470_cast_fp16")]; + tensor var_6471_cast_fp16 = softmax(axis = var_6311, x = aw_929_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6472_cast_fp16 = softmax(axis = var_6311, x = aw_931_cast_fp16)[name = tensor("op_6472_cast_fp16")]; + tensor var_6473_cast_fp16 = softmax(axis = var_6311, x = aw_933_cast_fp16)[name = tensor("op_6473_cast_fp16")]; + tensor var_6474_cast_fp16 = softmax(axis = var_6311, x = aw_935_cast_fp16)[name = tensor("op_6474_cast_fp16")]; + tensor var_6475_cast_fp16 = softmax(axis = var_6311, x = aw_937_cast_fp16)[name = tensor("op_6475_cast_fp16")]; + tensor var_6476_cast_fp16 = softmax(axis = var_6311, x = aw_939_cast_fp16)[name = tensor("op_6476_cast_fp16")]; + tensor var_6477_cast_fp16 = softmax(axis = var_6311, x = aw_941_cast_fp16)[name = tensor("op_6477_cast_fp16")]; + tensor var_6478_cast_fp16 = softmax(axis = var_6311, x = aw_943_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6479_cast_fp16 = softmax(axis = var_6311, x = aw_945_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor var_6480_cast_fp16 = softmax(axis = var_6311, x = aw_947_cast_fp16)[name = tensor("op_6480_cast_fp16")]; + tensor var_6481_cast_fp16 = softmax(axis = var_6311, x = aw_949_cast_fp16)[name = tensor("op_6481_cast_fp16")]; + tensor var_6482_cast_fp16 = softmax(axis = var_6311, x = aw_951_cast_fp16)[name = tensor("op_6482_cast_fp16")]; + tensor var_6483_cast_fp16 = softmax(axis = var_6311, x = aw_953_cast_fp16)[name = tensor("op_6483_cast_fp16")]; + tensor var_6484_cast_fp16 = softmax(axis = var_6311, x = aw_955_cast_fp16)[name = tensor("op_6484_cast_fp16")]; + tensor var_6485_cast_fp16 = softmax(axis = var_6311, x = aw_957_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor var_6486_cast_fp16 = softmax(axis = var_6311, x = aw_959_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6488_equation_0 = const()[name = tensor("op_6488_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6488_cast_fp16 = einsum(equation = var_6488_equation_0, values = (var_6406_cast_fp16_0, var_6467_cast_fp16))[name = tensor("op_6488_cast_fp16")]; + tensor var_6490_equation_0 = const()[name = tensor("op_6490_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6490_cast_fp16 = einsum(equation = var_6490_equation_0, values = (var_6406_cast_fp16_1, var_6468_cast_fp16))[name = tensor("op_6490_cast_fp16")]; + tensor var_6492_equation_0 = const()[name = tensor("op_6492_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6492_cast_fp16 = einsum(equation = var_6492_equation_0, values = (var_6406_cast_fp16_2, var_6469_cast_fp16))[name = tensor("op_6492_cast_fp16")]; + tensor var_6494_equation_0 = const()[name = tensor("op_6494_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6494_cast_fp16 = einsum(equation = var_6494_equation_0, values = (var_6406_cast_fp16_3, var_6470_cast_fp16))[name = tensor("op_6494_cast_fp16")]; + tensor var_6496_equation_0 = const()[name = tensor("op_6496_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6496_cast_fp16 = einsum(equation = var_6496_equation_0, values = (var_6406_cast_fp16_4, var_6471_cast_fp16))[name = tensor("op_6496_cast_fp16")]; + tensor var_6498_equation_0 = const()[name = tensor("op_6498_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6498_cast_fp16 = einsum(equation = var_6498_equation_0, values = (var_6406_cast_fp16_5, var_6472_cast_fp16))[name = tensor("op_6498_cast_fp16")]; + tensor var_6500_equation_0 = const()[name = tensor("op_6500_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6500_cast_fp16 = einsum(equation = var_6500_equation_0, values = (var_6406_cast_fp16_6, var_6473_cast_fp16))[name = tensor("op_6500_cast_fp16")]; + tensor var_6502_equation_0 = const()[name = tensor("op_6502_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6502_cast_fp16 = einsum(equation = var_6502_equation_0, values = (var_6406_cast_fp16_7, var_6474_cast_fp16))[name = tensor("op_6502_cast_fp16")]; + tensor var_6504_equation_0 = const()[name = tensor("op_6504_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6504_cast_fp16 = einsum(equation = var_6504_equation_0, values = (var_6406_cast_fp16_8, var_6475_cast_fp16))[name = tensor("op_6504_cast_fp16")]; + tensor var_6506_equation_0 = const()[name = tensor("op_6506_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6506_cast_fp16 = einsum(equation = var_6506_equation_0, values = (var_6406_cast_fp16_9, var_6476_cast_fp16))[name = tensor("op_6506_cast_fp16")]; + tensor var_6508_equation_0 = const()[name = tensor("op_6508_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6508_cast_fp16 = einsum(equation = var_6508_equation_0, values = (var_6406_cast_fp16_10, var_6477_cast_fp16))[name = tensor("op_6508_cast_fp16")]; + tensor var_6510_equation_0 = const()[name = tensor("op_6510_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6510_cast_fp16 = einsum(equation = var_6510_equation_0, values = (var_6406_cast_fp16_11, var_6478_cast_fp16))[name = tensor("op_6510_cast_fp16")]; + tensor var_6512_equation_0 = const()[name = tensor("op_6512_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6512_cast_fp16 = einsum(equation = var_6512_equation_0, values = (var_6406_cast_fp16_12, var_6479_cast_fp16))[name = tensor("op_6512_cast_fp16")]; + tensor var_6514_equation_0 = const()[name = tensor("op_6514_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6514_cast_fp16 = einsum(equation = var_6514_equation_0, values = (var_6406_cast_fp16_13, var_6480_cast_fp16))[name = tensor("op_6514_cast_fp16")]; + tensor var_6516_equation_0 = const()[name = tensor("op_6516_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6516_cast_fp16 = einsum(equation = var_6516_equation_0, values = (var_6406_cast_fp16_14, var_6481_cast_fp16))[name = tensor("op_6516_cast_fp16")]; + tensor var_6518_equation_0 = const()[name = tensor("op_6518_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6518_cast_fp16 = einsum(equation = var_6518_equation_0, values = (var_6406_cast_fp16_15, var_6482_cast_fp16))[name = tensor("op_6518_cast_fp16")]; + tensor var_6520_equation_0 = const()[name = tensor("op_6520_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6520_cast_fp16 = einsum(equation = var_6520_equation_0, values = (var_6406_cast_fp16_16, var_6483_cast_fp16))[name = tensor("op_6520_cast_fp16")]; + tensor var_6522_equation_0 = const()[name = tensor("op_6522_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6522_cast_fp16 = einsum(equation = var_6522_equation_0, values = (var_6406_cast_fp16_17, var_6484_cast_fp16))[name = tensor("op_6522_cast_fp16")]; + tensor var_6524_equation_0 = const()[name = tensor("op_6524_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6524_cast_fp16 = einsum(equation = var_6524_equation_0, values = (var_6406_cast_fp16_18, var_6485_cast_fp16))[name = tensor("op_6524_cast_fp16")]; + tensor var_6526_equation_0 = const()[name = tensor("op_6526_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6526_cast_fp16 = einsum(equation = var_6526_equation_0, values = (var_6406_cast_fp16_19, var_6486_cast_fp16))[name = tensor("op_6526_cast_fp16")]; + tensor input_235_interleave_0 = const()[name = tensor("input_235_interleave_0"), val = tensor(false)]; + tensor input_235_cast_fp16 = concat(axis = var_6311, interleave = input_235_interleave_0, values = (var_6488_cast_fp16, var_6490_cast_fp16, var_6492_cast_fp16, var_6494_cast_fp16, var_6496_cast_fp16, var_6498_cast_fp16, var_6500_cast_fp16, var_6502_cast_fp16, var_6504_cast_fp16, var_6506_cast_fp16, var_6508_cast_fp16, var_6510_cast_fp16, var_6512_cast_fp16, var_6514_cast_fp16, var_6516_cast_fp16, var_6518_cast_fp16, var_6520_cast_fp16, var_6522_cast_fp16, var_6524_cast_fp16, var_6526_cast_fp16))[name = tensor("input_235_cast_fp16")]; + tensor var_6535_pad_type_0 = const()[name = tensor("op_6535_pad_type_0"), val = tensor("valid")]; + tensor var_6535_strides_0 = const()[name = tensor("op_6535_strides_0"), val = tensor([1, 1])]; + tensor var_6535_pad_0 = const()[name = tensor("op_6535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6535_dilations_0 = const()[name = tensor("op_6535_dilations_0"), val = tensor([1, 1])]; + tensor var_6535_groups_0 = const()[name = tensor("op_6535_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_out_weight_to_fp16 = const()[name = tensor("blocks_23_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929256832)))]; + tensor blocks_23_attn_out_bias_to_fp16 = const()[name = tensor("blocks_23_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932533696)))]; + tensor var_6535_cast_fp16 = conv(bias = blocks_23_attn_out_bias_to_fp16, dilations = var_6535_dilations_0, groups = var_6535_groups_0, pad = var_6535_pad_0, pad_type = var_6535_pad_type_0, strides = var_6535_strides_0, weight = blocks_23_attn_out_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("op_6535_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = var_6535_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([1])]; + tensor input_237_gamma_0_to_fp16 = const()[name = tensor("input_237_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932536320)))]; + tensor input_237_beta_0_to_fp16 = const()[name = tensor("input_237_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932538944)))]; + tensor var_6545_to_fp16 = const()[name = tensor("op_6545_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = input_237_beta_0_to_fp16, epsilon = var_6545_to_fp16, gamma = input_237_gamma_0_to_fp16, x = inputs_95_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("valid")]; + tensor input_239_strides_0 = const()[name = tensor("input_239_strides_0"), val = tensor([1, 1])]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_239_dilations_0 = const()[name = tensor("input_239_dilations_0"), val = tensor([1, 1])]; + tensor input_239_groups_0 = const()[name = tensor("input_239_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932541568)))]; + tensor blocks_23_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945648832)))]; + tensor input_239_cast_fp16 = conv(bias = blocks_23_mlp_0_bias_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = blocks_23_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = input_239_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_6571_pad_type_0 = const()[name = tensor("op_6571_pad_type_0"), val = tensor("valid")]; + tensor var_6571_strides_0 = const()[name = tensor("op_6571_strides_0"), val = tensor([1, 1])]; + tensor var_6571_pad_0 = const()[name = tensor("op_6571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6571_dilations_0 = const()[name = tensor("op_6571_dilations_0"), val = tensor([1, 1])]; + tensor var_6571_groups_0 = const()[name = tensor("op_6571_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945659136)))]; + tensor blocks_23_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958766400)))]; + tensor var_6571_cast_fp16 = conv(bias = blocks_23_mlp_2_bias_to_fp16, dilations = var_6571_dilations_0, groups = var_6571_groups_0, pad = var_6571_pad_0, pad_type = var_6571_pad_type_0, strides = var_6571_strides_0, weight = blocks_23_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("op_6571_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = var_6571_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_6580 = const()[name = tensor("op_6580"), val = tensor(1)]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([1])]; + tensor input_243_gamma_0_to_fp16 = const()[name = tensor("input_243_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958769024)))]; + tensor input_243_beta_0_to_fp16 = const()[name = tensor("input_243_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958771648)))]; + tensor var_6596_to_fp16 = const()[name = tensor("op_6596_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = input_243_beta_0_to_fp16, epsilon = var_6596_to_fp16, gamma = input_243_gamma_0_to_fp16, x = inputs_97_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor q_49_pad_type_0 = const()[name = tensor("q_49_pad_type_0"), val = tensor("valid")]; + tensor q_49_strides_0 = const()[name = tensor("q_49_strides_0"), val = tensor([1, 1])]; + tensor q_49_pad_0 = const()[name = tensor("q_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_49_dilations_0 = const()[name = tensor("q_49_dilations_0"), val = tensor([1, 1])]; + tensor q_49_groups_0 = const()[name = tensor("q_49_groups_0"), val = tensor(1)]; + tensor var_6631_weight_0_to_fp16 = const()[name = tensor("op_6631_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958774272)))]; + tensor var_6631_bias_0_to_fp16 = const()[name = tensor("op_6631_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962051136)))]; + tensor var_6631_cast_fp16 = conv(bias = var_6631_bias_0_to_fp16, dilations = q_49_dilations_0, groups = q_49_groups_0, pad = q_49_pad_0, pad_type = q_49_pad_type_0, strides = q_49_strides_0, weight = var_6631_weight_0_to_fp16, x = input_243_cast_fp16)[name = tensor("op_6631_cast_fp16")]; + tensor k_49_pad_type_0 = const()[name = tensor("k_49_pad_type_0"), val = tensor("valid")]; + tensor k_49_strides_0 = const()[name = tensor("k_49_strides_0"), val = tensor([1, 1])]; + tensor k_49_pad_0 = const()[name = tensor("k_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_49_dilations_0 = const()[name = tensor("k_49_dilations_0"), val = tensor([1, 1])]; + tensor k_49_groups_0 = const()[name = tensor("k_49_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_key_weight_to_fp16 = const()[name = tensor("blocks_24_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962053760)))]; + tensor k_49_cast_fp16 = conv(dilations = k_49_dilations_0, groups = k_49_groups_0, pad = k_49_pad_0, pad_type = k_49_pad_type_0, strides = k_49_strides_0, weight = blocks_24_attn_key_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor var_6629_pad_type_0 = const()[name = tensor("op_6629_pad_type_0"), val = tensor("valid")]; + tensor var_6629_strides_0 = const()[name = tensor("op_6629_strides_0"), val = tensor([1, 1])]; + tensor var_6629_pad_0 = const()[name = tensor("op_6629_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6629_dilations_0 = const()[name = tensor("op_6629_dilations_0"), val = tensor([1, 1])]; + tensor var_6629_groups_0 = const()[name = tensor("op_6629_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_value_weight_to_fp16 = const()[name = tensor("blocks_24_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965330624)))]; + tensor blocks_24_attn_value_bias_to_fp16 = const()[name = tensor("blocks_24_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968607488)))]; + tensor var_6629_cast_fp16 = conv(bias = blocks_24_attn_value_bias_to_fp16, dilations = var_6629_dilations_0, groups = var_6629_groups_0, pad = var_6629_pad_0, pad_type = var_6629_pad_type_0, strides = var_6629_strides_0, weight = blocks_24_attn_value_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor tile_72 = const()[name = tensor("tile_72"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6632_axis_0 = const()[name = tensor("op_6632_axis_0"), val = tensor(1)]; + tensor var_6632_cast_fp16_0, tensor var_6632_cast_fp16_1, tensor var_6632_cast_fp16_2, tensor var_6632_cast_fp16_3, tensor var_6632_cast_fp16_4, tensor var_6632_cast_fp16_5, tensor var_6632_cast_fp16_6, tensor var_6632_cast_fp16_7, tensor var_6632_cast_fp16_8, tensor var_6632_cast_fp16_9, tensor var_6632_cast_fp16_10, tensor var_6632_cast_fp16_11, tensor var_6632_cast_fp16_12, tensor var_6632_cast_fp16_13, tensor var_6632_cast_fp16_14, tensor var_6632_cast_fp16_15, tensor var_6632_cast_fp16_16, tensor var_6632_cast_fp16_17, tensor var_6632_cast_fp16_18, tensor var_6632_cast_fp16_19 = split(axis = var_6632_axis_0, split_sizes = tile_72, x = var_6631_cast_fp16)[name = tensor("op_6632_cast_fp16")]; + tensor var_6653_perm_0 = const()[name = tensor("op_6653_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_73 = const()[name = tensor("tile_73"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6654_axis_0 = const()[name = tensor("op_6654_axis_0"), val = tensor(3)]; + tensor var_6653_cast_fp16 = transpose(perm = var_6653_perm_0, x = k_49_cast_fp16)[name = tensor("transpose_8")]; + tensor var_6654_cast_fp16_0, tensor var_6654_cast_fp16_1, tensor var_6654_cast_fp16_2, tensor var_6654_cast_fp16_3, tensor var_6654_cast_fp16_4, tensor var_6654_cast_fp16_5, tensor var_6654_cast_fp16_6, tensor var_6654_cast_fp16_7, tensor var_6654_cast_fp16_8, tensor var_6654_cast_fp16_9, tensor var_6654_cast_fp16_10, tensor var_6654_cast_fp16_11, tensor var_6654_cast_fp16_12, tensor var_6654_cast_fp16_13, tensor var_6654_cast_fp16_14, tensor var_6654_cast_fp16_15, tensor var_6654_cast_fp16_16, tensor var_6654_cast_fp16_17, tensor var_6654_cast_fp16_18, tensor var_6654_cast_fp16_19 = split(axis = var_6654_axis_0, split_sizes = tile_73, x = var_6653_cast_fp16)[name = tensor("op_6654_cast_fp16")]; + tensor tile_74 = const()[name = tensor("tile_74"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6675_axis_0 = const()[name = tensor("op_6675_axis_0"), val = tensor(1)]; + tensor var_6675_cast_fp16_0, tensor var_6675_cast_fp16_1, tensor var_6675_cast_fp16_2, tensor var_6675_cast_fp16_3, tensor var_6675_cast_fp16_4, tensor var_6675_cast_fp16_5, tensor var_6675_cast_fp16_6, tensor var_6675_cast_fp16_7, tensor var_6675_cast_fp16_8, tensor var_6675_cast_fp16_9, tensor var_6675_cast_fp16_10, tensor var_6675_cast_fp16_11, tensor var_6675_cast_fp16_12, tensor var_6675_cast_fp16_13, tensor var_6675_cast_fp16_14, tensor var_6675_cast_fp16_15, tensor var_6675_cast_fp16_16, tensor var_6675_cast_fp16_17, tensor var_6675_cast_fp16_18, tensor var_6675_cast_fp16_19 = split(axis = var_6675_axis_0, split_sizes = tile_74, x = var_6629_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor aw_961_equation_0 = const()[name = tensor("aw_961_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_961_cast_fp16 = einsum(equation = aw_961_equation_0, values = (var_6654_cast_fp16_0, var_6632_cast_fp16_0))[name = tensor("aw_961_cast_fp16")]; + tensor aw_963_equation_0 = const()[name = tensor("aw_963_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_963_cast_fp16 = einsum(equation = aw_963_equation_0, values = (var_6654_cast_fp16_1, var_6632_cast_fp16_1))[name = tensor("aw_963_cast_fp16")]; + tensor aw_965_equation_0 = const()[name = tensor("aw_965_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_965_cast_fp16 = einsum(equation = aw_965_equation_0, values = (var_6654_cast_fp16_2, var_6632_cast_fp16_2))[name = tensor("aw_965_cast_fp16")]; + tensor aw_967_equation_0 = const()[name = tensor("aw_967_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_967_cast_fp16 = einsum(equation = aw_967_equation_0, values = (var_6654_cast_fp16_3, var_6632_cast_fp16_3))[name = tensor("aw_967_cast_fp16")]; + tensor aw_969_equation_0 = const()[name = tensor("aw_969_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_969_cast_fp16 = einsum(equation = aw_969_equation_0, values = (var_6654_cast_fp16_4, var_6632_cast_fp16_4))[name = tensor("aw_969_cast_fp16")]; + tensor aw_971_equation_0 = const()[name = tensor("aw_971_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_971_cast_fp16 = einsum(equation = aw_971_equation_0, values = (var_6654_cast_fp16_5, var_6632_cast_fp16_5))[name = tensor("aw_971_cast_fp16")]; + tensor aw_973_equation_0 = const()[name = tensor("aw_973_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_973_cast_fp16 = einsum(equation = aw_973_equation_0, values = (var_6654_cast_fp16_6, var_6632_cast_fp16_6))[name = tensor("aw_973_cast_fp16")]; + tensor aw_975_equation_0 = const()[name = tensor("aw_975_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_975_cast_fp16 = einsum(equation = aw_975_equation_0, values = (var_6654_cast_fp16_7, var_6632_cast_fp16_7))[name = tensor("aw_975_cast_fp16")]; + tensor aw_977_equation_0 = const()[name = tensor("aw_977_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_977_cast_fp16 = einsum(equation = aw_977_equation_0, values = (var_6654_cast_fp16_8, var_6632_cast_fp16_8))[name = tensor("aw_977_cast_fp16")]; + tensor aw_979_equation_0 = const()[name = tensor("aw_979_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_979_cast_fp16 = einsum(equation = aw_979_equation_0, values = (var_6654_cast_fp16_9, var_6632_cast_fp16_9))[name = tensor("aw_979_cast_fp16")]; + tensor aw_981_equation_0 = const()[name = tensor("aw_981_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_981_cast_fp16 = einsum(equation = aw_981_equation_0, values = (var_6654_cast_fp16_10, var_6632_cast_fp16_10))[name = tensor("aw_981_cast_fp16")]; + tensor aw_983_equation_0 = const()[name = tensor("aw_983_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_983_cast_fp16 = einsum(equation = aw_983_equation_0, values = (var_6654_cast_fp16_11, var_6632_cast_fp16_11))[name = tensor("aw_983_cast_fp16")]; + tensor aw_985_equation_0 = const()[name = tensor("aw_985_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_985_cast_fp16 = einsum(equation = aw_985_equation_0, values = (var_6654_cast_fp16_12, var_6632_cast_fp16_12))[name = tensor("aw_985_cast_fp16")]; + tensor aw_987_equation_0 = const()[name = tensor("aw_987_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_987_cast_fp16 = einsum(equation = aw_987_equation_0, values = (var_6654_cast_fp16_13, var_6632_cast_fp16_13))[name = tensor("aw_987_cast_fp16")]; + tensor aw_989_equation_0 = const()[name = tensor("aw_989_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_989_cast_fp16 = einsum(equation = aw_989_equation_0, values = (var_6654_cast_fp16_14, var_6632_cast_fp16_14))[name = tensor("aw_989_cast_fp16")]; + tensor aw_991_equation_0 = const()[name = tensor("aw_991_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_991_cast_fp16 = einsum(equation = aw_991_equation_0, values = (var_6654_cast_fp16_15, var_6632_cast_fp16_15))[name = tensor("aw_991_cast_fp16")]; + tensor aw_993_equation_0 = const()[name = tensor("aw_993_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_993_cast_fp16 = einsum(equation = aw_993_equation_0, values = (var_6654_cast_fp16_16, var_6632_cast_fp16_16))[name = tensor("aw_993_cast_fp16")]; + tensor aw_995_equation_0 = const()[name = tensor("aw_995_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_995_cast_fp16 = einsum(equation = aw_995_equation_0, values = (var_6654_cast_fp16_17, var_6632_cast_fp16_17))[name = tensor("aw_995_cast_fp16")]; + tensor aw_997_equation_0 = const()[name = tensor("aw_997_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_997_cast_fp16 = einsum(equation = aw_997_equation_0, values = (var_6654_cast_fp16_18, var_6632_cast_fp16_18))[name = tensor("aw_997_cast_fp16")]; + tensor aw_999_equation_0 = const()[name = tensor("aw_999_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_999_cast_fp16 = einsum(equation = aw_999_equation_0, values = (var_6654_cast_fp16_19, var_6632_cast_fp16_19))[name = tensor("aw_999_cast_fp16")]; + tensor var_6736_cast_fp16 = softmax(axis = var_6580, x = aw_961_cast_fp16)[name = tensor("op_6736_cast_fp16")]; + tensor var_6737_cast_fp16 = softmax(axis = var_6580, x = aw_963_cast_fp16)[name = tensor("op_6737_cast_fp16")]; + tensor var_6738_cast_fp16 = softmax(axis = var_6580, x = aw_965_cast_fp16)[name = tensor("op_6738_cast_fp16")]; + tensor var_6739_cast_fp16 = softmax(axis = var_6580, x = aw_967_cast_fp16)[name = tensor("op_6739_cast_fp16")]; + tensor var_6740_cast_fp16 = softmax(axis = var_6580, x = aw_969_cast_fp16)[name = tensor("op_6740_cast_fp16")]; + tensor var_6741_cast_fp16 = softmax(axis = var_6580, x = aw_971_cast_fp16)[name = tensor("op_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = softmax(axis = var_6580, x = aw_973_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_cast_fp16 = softmax(axis = var_6580, x = aw_975_cast_fp16)[name = tensor("op_6743_cast_fp16")]; + tensor var_6744_cast_fp16 = softmax(axis = var_6580, x = aw_977_cast_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor var_6745_cast_fp16 = softmax(axis = var_6580, x = aw_979_cast_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6746_cast_fp16 = softmax(axis = var_6580, x = aw_981_cast_fp16)[name = tensor("op_6746_cast_fp16")]; + tensor var_6747_cast_fp16 = softmax(axis = var_6580, x = aw_983_cast_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6748_cast_fp16 = softmax(axis = var_6580, x = aw_985_cast_fp16)[name = tensor("op_6748_cast_fp16")]; + tensor var_6749_cast_fp16 = softmax(axis = var_6580, x = aw_987_cast_fp16)[name = tensor("op_6749_cast_fp16")]; + tensor var_6750_cast_fp16 = softmax(axis = var_6580, x = aw_989_cast_fp16)[name = tensor("op_6750_cast_fp16")]; + tensor var_6751_cast_fp16 = softmax(axis = var_6580, x = aw_991_cast_fp16)[name = tensor("op_6751_cast_fp16")]; + tensor var_6752_cast_fp16 = softmax(axis = var_6580, x = aw_993_cast_fp16)[name = tensor("op_6752_cast_fp16")]; + tensor var_6753_cast_fp16 = softmax(axis = var_6580, x = aw_995_cast_fp16)[name = tensor("op_6753_cast_fp16")]; + tensor var_6754_cast_fp16 = softmax(axis = var_6580, x = aw_997_cast_fp16)[name = tensor("op_6754_cast_fp16")]; + tensor var_6755_cast_fp16 = softmax(axis = var_6580, x = aw_999_cast_fp16)[name = tensor("op_6755_cast_fp16")]; + tensor var_6757_equation_0 = const()[name = tensor("op_6757_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6757_cast_fp16 = einsum(equation = var_6757_equation_0, values = (var_6675_cast_fp16_0, var_6736_cast_fp16))[name = tensor("op_6757_cast_fp16")]; + tensor var_6759_equation_0 = const()[name = tensor("op_6759_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6759_cast_fp16 = einsum(equation = var_6759_equation_0, values = (var_6675_cast_fp16_1, var_6737_cast_fp16))[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_equation_0 = const()[name = tensor("op_6761_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6761_cast_fp16 = einsum(equation = var_6761_equation_0, values = (var_6675_cast_fp16_2, var_6738_cast_fp16))[name = tensor("op_6761_cast_fp16")]; + tensor var_6763_equation_0 = const()[name = tensor("op_6763_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6763_cast_fp16 = einsum(equation = var_6763_equation_0, values = (var_6675_cast_fp16_3, var_6739_cast_fp16))[name = tensor("op_6763_cast_fp16")]; + tensor var_6765_equation_0 = const()[name = tensor("op_6765_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6765_cast_fp16 = einsum(equation = var_6765_equation_0, values = (var_6675_cast_fp16_4, var_6740_cast_fp16))[name = tensor("op_6765_cast_fp16")]; + tensor var_6767_equation_0 = const()[name = tensor("op_6767_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6767_cast_fp16 = einsum(equation = var_6767_equation_0, values = (var_6675_cast_fp16_5, var_6741_cast_fp16))[name = tensor("op_6767_cast_fp16")]; + tensor var_6769_equation_0 = const()[name = tensor("op_6769_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6769_cast_fp16 = einsum(equation = var_6769_equation_0, values = (var_6675_cast_fp16_6, var_6742_cast_fp16))[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_equation_0 = const()[name = tensor("op_6771_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6771_cast_fp16 = einsum(equation = var_6771_equation_0, values = (var_6675_cast_fp16_7, var_6743_cast_fp16))[name = tensor("op_6771_cast_fp16")]; + tensor var_6773_equation_0 = const()[name = tensor("op_6773_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6773_cast_fp16 = einsum(equation = var_6773_equation_0, values = (var_6675_cast_fp16_8, var_6744_cast_fp16))[name = tensor("op_6773_cast_fp16")]; + tensor var_6775_equation_0 = const()[name = tensor("op_6775_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6775_cast_fp16 = einsum(equation = var_6775_equation_0, values = (var_6675_cast_fp16_9, var_6745_cast_fp16))[name = tensor("op_6775_cast_fp16")]; + tensor var_6777_equation_0 = const()[name = tensor("op_6777_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6777_cast_fp16 = einsum(equation = var_6777_equation_0, values = (var_6675_cast_fp16_10, var_6746_cast_fp16))[name = tensor("op_6777_cast_fp16")]; + tensor var_6779_equation_0 = const()[name = tensor("op_6779_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6779_cast_fp16 = einsum(equation = var_6779_equation_0, values = (var_6675_cast_fp16_11, var_6747_cast_fp16))[name = tensor("op_6779_cast_fp16")]; + tensor var_6781_equation_0 = const()[name = tensor("op_6781_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6781_cast_fp16 = einsum(equation = var_6781_equation_0, values = (var_6675_cast_fp16_12, var_6748_cast_fp16))[name = tensor("op_6781_cast_fp16")]; + tensor var_6783_equation_0 = const()[name = tensor("op_6783_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6783_cast_fp16 = einsum(equation = var_6783_equation_0, values = (var_6675_cast_fp16_13, var_6749_cast_fp16))[name = tensor("op_6783_cast_fp16")]; + tensor var_6785_equation_0 = const()[name = tensor("op_6785_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6785_cast_fp16 = einsum(equation = var_6785_equation_0, values = (var_6675_cast_fp16_14, var_6750_cast_fp16))[name = tensor("op_6785_cast_fp16")]; + tensor var_6787_equation_0 = const()[name = tensor("op_6787_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6787_cast_fp16 = einsum(equation = var_6787_equation_0, values = (var_6675_cast_fp16_15, var_6751_cast_fp16))[name = tensor("op_6787_cast_fp16")]; + tensor var_6789_equation_0 = const()[name = tensor("op_6789_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6789_cast_fp16 = einsum(equation = var_6789_equation_0, values = (var_6675_cast_fp16_16, var_6752_cast_fp16))[name = tensor("op_6789_cast_fp16")]; + tensor var_6791_equation_0 = const()[name = tensor("op_6791_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6791_cast_fp16 = einsum(equation = var_6791_equation_0, values = (var_6675_cast_fp16_17, var_6753_cast_fp16))[name = tensor("op_6791_cast_fp16")]; + tensor var_6793_equation_0 = const()[name = tensor("op_6793_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6793_cast_fp16 = einsum(equation = var_6793_equation_0, values = (var_6675_cast_fp16_18, var_6754_cast_fp16))[name = tensor("op_6793_cast_fp16")]; + tensor var_6795_equation_0 = const()[name = tensor("op_6795_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6795_cast_fp16 = einsum(equation = var_6795_equation_0, values = (var_6675_cast_fp16_19, var_6755_cast_fp16))[name = tensor("op_6795_cast_fp16")]; + tensor input_245_interleave_0 = const()[name = tensor("input_245_interleave_0"), val = tensor(false)]; + tensor input_245_cast_fp16 = concat(axis = var_6580, interleave = input_245_interleave_0, values = (var_6757_cast_fp16, var_6759_cast_fp16, var_6761_cast_fp16, var_6763_cast_fp16, var_6765_cast_fp16, var_6767_cast_fp16, var_6769_cast_fp16, var_6771_cast_fp16, var_6773_cast_fp16, var_6775_cast_fp16, var_6777_cast_fp16, var_6779_cast_fp16, var_6781_cast_fp16, var_6783_cast_fp16, var_6785_cast_fp16, var_6787_cast_fp16, var_6789_cast_fp16, var_6791_cast_fp16, var_6793_cast_fp16, var_6795_cast_fp16))[name = tensor("input_245_cast_fp16")]; + tensor var_6804_pad_type_0 = const()[name = tensor("op_6804_pad_type_0"), val = tensor("valid")]; + tensor var_6804_strides_0 = const()[name = tensor("op_6804_strides_0"), val = tensor([1, 1])]; + tensor var_6804_pad_0 = const()[name = tensor("op_6804_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6804_dilations_0 = const()[name = tensor("op_6804_dilations_0"), val = tensor([1, 1])]; + tensor var_6804_groups_0 = const()[name = tensor("op_6804_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_out_weight_to_fp16 = const()[name = tensor("blocks_24_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968610112)))]; + tensor blocks_24_attn_out_bias_to_fp16 = const()[name = tensor("blocks_24_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971886976)))]; + tensor var_6804_cast_fp16 = conv(bias = blocks_24_attn_out_bias_to_fp16, dilations = var_6804_dilations_0, groups = var_6804_groups_0, pad = var_6804_pad_0, pad_type = var_6804_pad_type_0, strides = var_6804_strides_0, weight = blocks_24_attn_out_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("op_6804_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_6804_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor input_247_axes_0 = const()[name = tensor("input_247_axes_0"), val = tensor([1])]; + tensor input_247_gamma_0_to_fp16 = const()[name = tensor("input_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971889600)))]; + tensor input_247_beta_0_to_fp16 = const()[name = tensor("input_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971892224)))]; + tensor var_6814_to_fp16 = const()[name = tensor("op_6814_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, beta = input_247_beta_0_to_fp16, epsilon = var_6814_to_fp16, gamma = input_247_gamma_0_to_fp16, x = inputs_99_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_pad_type_0 = const()[name = tensor("input_249_pad_type_0"), val = tensor("valid")]; + tensor input_249_strides_0 = const()[name = tensor("input_249_strides_0"), val = tensor([1, 1])]; + tensor input_249_pad_0 = const()[name = tensor("input_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_249_dilations_0 = const()[name = tensor("input_249_dilations_0"), val = tensor([1, 1])]; + tensor input_249_groups_0 = const()[name = tensor("input_249_groups_0"), val = tensor(1)]; + tensor blocks_24_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971894848)))]; + tensor blocks_24_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985002112)))]; + tensor input_249_cast_fp16 = conv(bias = blocks_24_mlp_0_bias_to_fp16, dilations = input_249_dilations_0, groups = input_249_groups_0, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = input_249_strides_0, weight = blocks_24_mlp_0_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_6840_pad_type_0 = const()[name = tensor("op_6840_pad_type_0"), val = tensor("valid")]; + tensor var_6840_strides_0 = const()[name = tensor("op_6840_strides_0"), val = tensor([1, 1])]; + tensor var_6840_pad_0 = const()[name = tensor("op_6840_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6840_dilations_0 = const()[name = tensor("op_6840_dilations_0"), val = tensor([1, 1])]; + tensor var_6840_groups_0 = const()[name = tensor("op_6840_groups_0"), val = tensor(1)]; + tensor blocks_24_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985012416)))]; + tensor blocks_24_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998119680)))]; + tensor var_6840_cast_fp16 = conv(bias = blocks_24_mlp_2_bias_to_fp16, dilations = var_6840_dilations_0, groups = var_6840_groups_0, pad = var_6840_pad_0, pad_type = var_6840_pad_type_0, strides = var_6840_strides_0, weight = blocks_24_mlp_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("op_6840_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = var_6840_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_6849 = const()[name = tensor("op_6849"), val = tensor(1)]; + tensor input_253_axes_0 = const()[name = tensor("input_253_axes_0"), val = tensor([1])]; + tensor input_253_gamma_0_to_fp16 = const()[name = tensor("input_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998122304)))]; + tensor input_253_beta_0_to_fp16 = const()[name = tensor("input_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998124928)))]; + tensor var_6865_to_fp16 = const()[name = tensor("op_6865_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_253_cast_fp16 = layer_norm(axes = input_253_axes_0, beta = input_253_beta_0_to_fp16, epsilon = var_6865_to_fp16, gamma = input_253_gamma_0_to_fp16, x = inputs_101_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor q_51_pad_type_0 = const()[name = tensor("q_51_pad_type_0"), val = tensor("valid")]; + tensor q_51_strides_0 = const()[name = tensor("q_51_strides_0"), val = tensor([1, 1])]; + tensor q_51_pad_0 = const()[name = tensor("q_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_51_dilations_0 = const()[name = tensor("q_51_dilations_0"), val = tensor([1, 1])]; + tensor q_51_groups_0 = const()[name = tensor("q_51_groups_0"), val = tensor(1)]; + tensor var_6900_weight_0_to_fp16 = const()[name = tensor("op_6900_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998127552)))]; + tensor var_6900_bias_0_to_fp16 = const()[name = tensor("op_6900_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001404416)))]; + tensor var_6900_cast_fp16 = conv(bias = var_6900_bias_0_to_fp16, dilations = q_51_dilations_0, groups = q_51_groups_0, pad = q_51_pad_0, pad_type = q_51_pad_type_0, strides = q_51_strides_0, weight = var_6900_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor k_51_pad_type_0 = const()[name = tensor("k_51_pad_type_0"), val = tensor("valid")]; + tensor k_51_strides_0 = const()[name = tensor("k_51_strides_0"), val = tensor([1, 1])]; + tensor k_51_pad_0 = const()[name = tensor("k_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_51_dilations_0 = const()[name = tensor("k_51_dilations_0"), val = tensor([1, 1])]; + tensor k_51_groups_0 = const()[name = tensor("k_51_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_key_weight_to_fp16 = const()[name = tensor("blocks_25_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001407040)))]; + tensor k_51_cast_fp16 = conv(dilations = k_51_dilations_0, groups = k_51_groups_0, pad = k_51_pad_0, pad_type = k_51_pad_type_0, strides = k_51_strides_0, weight = blocks_25_attn_key_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor var_6898_pad_type_0 = const()[name = tensor("op_6898_pad_type_0"), val = tensor("valid")]; + tensor var_6898_strides_0 = const()[name = tensor("op_6898_strides_0"), val = tensor([1, 1])]; + tensor var_6898_pad_0 = const()[name = tensor("op_6898_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6898_dilations_0 = const()[name = tensor("op_6898_dilations_0"), val = tensor([1, 1])]; + tensor var_6898_groups_0 = const()[name = tensor("op_6898_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_value_weight_to_fp16 = const()[name = tensor("blocks_25_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1004683904)))]; + tensor blocks_25_attn_value_bias_to_fp16 = const()[name = tensor("blocks_25_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1007960768)))]; + tensor var_6898_cast_fp16 = conv(bias = blocks_25_attn_value_bias_to_fp16, dilations = var_6898_dilations_0, groups = var_6898_groups_0, pad = var_6898_pad_0, pad_type = var_6898_pad_type_0, strides = var_6898_strides_0, weight = blocks_25_attn_value_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor tile_75 = const()[name = tensor("tile_75"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6901_axis_0 = const()[name = tensor("op_6901_axis_0"), val = tensor(1)]; + tensor var_6901_cast_fp16_0, tensor var_6901_cast_fp16_1, tensor var_6901_cast_fp16_2, tensor var_6901_cast_fp16_3, tensor var_6901_cast_fp16_4, tensor var_6901_cast_fp16_5, tensor var_6901_cast_fp16_6, tensor var_6901_cast_fp16_7, tensor var_6901_cast_fp16_8, tensor var_6901_cast_fp16_9, tensor var_6901_cast_fp16_10, tensor var_6901_cast_fp16_11, tensor var_6901_cast_fp16_12, tensor var_6901_cast_fp16_13, tensor var_6901_cast_fp16_14, tensor var_6901_cast_fp16_15, tensor var_6901_cast_fp16_16, tensor var_6901_cast_fp16_17, tensor var_6901_cast_fp16_18, tensor var_6901_cast_fp16_19 = split(axis = var_6901_axis_0, split_sizes = tile_75, x = var_6900_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor var_6922_perm_0 = const()[name = tensor("op_6922_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_76 = const()[name = tensor("tile_76"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6923_axis_0 = const()[name = tensor("op_6923_axis_0"), val = tensor(3)]; + tensor var_6922_cast_fp16 = transpose(perm = var_6922_perm_0, x = k_51_cast_fp16)[name = tensor("transpose_7")]; + tensor var_6923_cast_fp16_0, tensor var_6923_cast_fp16_1, tensor var_6923_cast_fp16_2, tensor var_6923_cast_fp16_3, tensor var_6923_cast_fp16_4, tensor var_6923_cast_fp16_5, tensor var_6923_cast_fp16_6, tensor var_6923_cast_fp16_7, tensor var_6923_cast_fp16_8, tensor var_6923_cast_fp16_9, tensor var_6923_cast_fp16_10, tensor var_6923_cast_fp16_11, tensor var_6923_cast_fp16_12, tensor var_6923_cast_fp16_13, tensor var_6923_cast_fp16_14, tensor var_6923_cast_fp16_15, tensor var_6923_cast_fp16_16, tensor var_6923_cast_fp16_17, tensor var_6923_cast_fp16_18, tensor var_6923_cast_fp16_19 = split(axis = var_6923_axis_0, split_sizes = tile_76, x = var_6922_cast_fp16)[name = tensor("op_6923_cast_fp16")]; + tensor tile_77 = const()[name = tensor("tile_77"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6944_axis_0 = const()[name = tensor("op_6944_axis_0"), val = tensor(1)]; + tensor var_6944_cast_fp16_0, tensor var_6944_cast_fp16_1, tensor var_6944_cast_fp16_2, tensor var_6944_cast_fp16_3, tensor var_6944_cast_fp16_4, tensor var_6944_cast_fp16_5, tensor var_6944_cast_fp16_6, tensor var_6944_cast_fp16_7, tensor var_6944_cast_fp16_8, tensor var_6944_cast_fp16_9, tensor var_6944_cast_fp16_10, tensor var_6944_cast_fp16_11, tensor var_6944_cast_fp16_12, tensor var_6944_cast_fp16_13, tensor var_6944_cast_fp16_14, tensor var_6944_cast_fp16_15, tensor var_6944_cast_fp16_16, tensor var_6944_cast_fp16_17, tensor var_6944_cast_fp16_18, tensor var_6944_cast_fp16_19 = split(axis = var_6944_axis_0, split_sizes = tile_77, x = var_6898_cast_fp16)[name = tensor("op_6944_cast_fp16")]; + tensor aw_1001_equation_0 = const()[name = tensor("aw_1001_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1001_cast_fp16 = einsum(equation = aw_1001_equation_0, values = (var_6923_cast_fp16_0, var_6901_cast_fp16_0))[name = tensor("aw_1001_cast_fp16")]; + tensor aw_1003_equation_0 = const()[name = tensor("aw_1003_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1003_cast_fp16 = einsum(equation = aw_1003_equation_0, values = (var_6923_cast_fp16_1, var_6901_cast_fp16_1))[name = tensor("aw_1003_cast_fp16")]; + tensor aw_1005_equation_0 = const()[name = tensor("aw_1005_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1005_cast_fp16 = einsum(equation = aw_1005_equation_0, values = (var_6923_cast_fp16_2, var_6901_cast_fp16_2))[name = tensor("aw_1005_cast_fp16")]; + tensor aw_1007_equation_0 = const()[name = tensor("aw_1007_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1007_cast_fp16 = einsum(equation = aw_1007_equation_0, values = (var_6923_cast_fp16_3, var_6901_cast_fp16_3))[name = tensor("aw_1007_cast_fp16")]; + tensor aw_1009_equation_0 = const()[name = tensor("aw_1009_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1009_cast_fp16 = einsum(equation = aw_1009_equation_0, values = (var_6923_cast_fp16_4, var_6901_cast_fp16_4))[name = tensor("aw_1009_cast_fp16")]; + tensor aw_1011_equation_0 = const()[name = tensor("aw_1011_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1011_cast_fp16 = einsum(equation = aw_1011_equation_0, values = (var_6923_cast_fp16_5, var_6901_cast_fp16_5))[name = tensor("aw_1011_cast_fp16")]; + tensor aw_1013_equation_0 = const()[name = tensor("aw_1013_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1013_cast_fp16 = einsum(equation = aw_1013_equation_0, values = (var_6923_cast_fp16_6, var_6901_cast_fp16_6))[name = tensor("aw_1013_cast_fp16")]; + tensor aw_1015_equation_0 = const()[name = tensor("aw_1015_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1015_cast_fp16 = einsum(equation = aw_1015_equation_0, values = (var_6923_cast_fp16_7, var_6901_cast_fp16_7))[name = tensor("aw_1015_cast_fp16")]; + tensor aw_1017_equation_0 = const()[name = tensor("aw_1017_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1017_cast_fp16 = einsum(equation = aw_1017_equation_0, values = (var_6923_cast_fp16_8, var_6901_cast_fp16_8))[name = tensor("aw_1017_cast_fp16")]; + tensor aw_1019_equation_0 = const()[name = tensor("aw_1019_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1019_cast_fp16 = einsum(equation = aw_1019_equation_0, values = (var_6923_cast_fp16_9, var_6901_cast_fp16_9))[name = tensor("aw_1019_cast_fp16")]; + tensor aw_1021_equation_0 = const()[name = tensor("aw_1021_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1021_cast_fp16 = einsum(equation = aw_1021_equation_0, values = (var_6923_cast_fp16_10, var_6901_cast_fp16_10))[name = tensor("aw_1021_cast_fp16")]; + tensor aw_1023_equation_0 = const()[name = tensor("aw_1023_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1023_cast_fp16 = einsum(equation = aw_1023_equation_0, values = (var_6923_cast_fp16_11, var_6901_cast_fp16_11))[name = tensor("aw_1023_cast_fp16")]; + tensor aw_1025_equation_0 = const()[name = tensor("aw_1025_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1025_cast_fp16 = einsum(equation = aw_1025_equation_0, values = (var_6923_cast_fp16_12, var_6901_cast_fp16_12))[name = tensor("aw_1025_cast_fp16")]; + tensor aw_1027_equation_0 = const()[name = tensor("aw_1027_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1027_cast_fp16 = einsum(equation = aw_1027_equation_0, values = (var_6923_cast_fp16_13, var_6901_cast_fp16_13))[name = tensor("aw_1027_cast_fp16")]; + tensor aw_1029_equation_0 = const()[name = tensor("aw_1029_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1029_cast_fp16 = einsum(equation = aw_1029_equation_0, values = (var_6923_cast_fp16_14, var_6901_cast_fp16_14))[name = tensor("aw_1029_cast_fp16")]; + tensor aw_1031_equation_0 = const()[name = tensor("aw_1031_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1031_cast_fp16 = einsum(equation = aw_1031_equation_0, values = (var_6923_cast_fp16_15, var_6901_cast_fp16_15))[name = tensor("aw_1031_cast_fp16")]; + tensor aw_1033_equation_0 = const()[name = tensor("aw_1033_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1033_cast_fp16 = einsum(equation = aw_1033_equation_0, values = (var_6923_cast_fp16_16, var_6901_cast_fp16_16))[name = tensor("aw_1033_cast_fp16")]; + tensor aw_1035_equation_0 = const()[name = tensor("aw_1035_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1035_cast_fp16 = einsum(equation = aw_1035_equation_0, values = (var_6923_cast_fp16_17, var_6901_cast_fp16_17))[name = tensor("aw_1035_cast_fp16")]; + tensor aw_1037_equation_0 = const()[name = tensor("aw_1037_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1037_cast_fp16 = einsum(equation = aw_1037_equation_0, values = (var_6923_cast_fp16_18, var_6901_cast_fp16_18))[name = tensor("aw_1037_cast_fp16")]; + tensor aw_1039_equation_0 = const()[name = tensor("aw_1039_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1039_cast_fp16 = einsum(equation = aw_1039_equation_0, values = (var_6923_cast_fp16_19, var_6901_cast_fp16_19))[name = tensor("aw_1039_cast_fp16")]; + tensor var_7005_cast_fp16 = softmax(axis = var_6849, x = aw_1001_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7006_cast_fp16 = softmax(axis = var_6849, x = aw_1003_cast_fp16)[name = tensor("op_7006_cast_fp16")]; + tensor var_7007_cast_fp16 = softmax(axis = var_6849, x = aw_1005_cast_fp16)[name = tensor("op_7007_cast_fp16")]; + tensor var_7008_cast_fp16 = softmax(axis = var_6849, x = aw_1007_cast_fp16)[name = tensor("op_7008_cast_fp16")]; + tensor var_7009_cast_fp16 = softmax(axis = var_6849, x = aw_1009_cast_fp16)[name = tensor("op_7009_cast_fp16")]; + tensor var_7010_cast_fp16 = softmax(axis = var_6849, x = aw_1011_cast_fp16)[name = tensor("op_7010_cast_fp16")]; + tensor var_7011_cast_fp16 = softmax(axis = var_6849, x = aw_1013_cast_fp16)[name = tensor("op_7011_cast_fp16")]; + tensor var_7012_cast_fp16 = softmax(axis = var_6849, x = aw_1015_cast_fp16)[name = tensor("op_7012_cast_fp16")]; + tensor var_7013_cast_fp16 = softmax(axis = var_6849, x = aw_1017_cast_fp16)[name = tensor("op_7013_cast_fp16")]; + tensor var_7014_cast_fp16 = softmax(axis = var_6849, x = aw_1019_cast_fp16)[name = tensor("op_7014_cast_fp16")]; + tensor var_7015_cast_fp16 = softmax(axis = var_6849, x = aw_1021_cast_fp16)[name = tensor("op_7015_cast_fp16")]; + tensor var_7016_cast_fp16 = softmax(axis = var_6849, x = aw_1023_cast_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7017_cast_fp16 = softmax(axis = var_6849, x = aw_1025_cast_fp16)[name = tensor("op_7017_cast_fp16")]; + tensor var_7018_cast_fp16 = softmax(axis = var_6849, x = aw_1027_cast_fp16)[name = tensor("op_7018_cast_fp16")]; + tensor var_7019_cast_fp16 = softmax(axis = var_6849, x = aw_1029_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor var_7020_cast_fp16 = softmax(axis = var_6849, x = aw_1031_cast_fp16)[name = tensor("op_7020_cast_fp16")]; + tensor var_7021_cast_fp16 = softmax(axis = var_6849, x = aw_1033_cast_fp16)[name = tensor("op_7021_cast_fp16")]; + tensor var_7022_cast_fp16 = softmax(axis = var_6849, x = aw_1035_cast_fp16)[name = tensor("op_7022_cast_fp16")]; + tensor var_7023_cast_fp16 = softmax(axis = var_6849, x = aw_1037_cast_fp16)[name = tensor("op_7023_cast_fp16")]; + tensor var_7024_cast_fp16 = softmax(axis = var_6849, x = aw_1039_cast_fp16)[name = tensor("op_7024_cast_fp16")]; + tensor var_7026_equation_0 = const()[name = tensor("op_7026_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7026_cast_fp16 = einsum(equation = var_7026_equation_0, values = (var_6944_cast_fp16_0, var_7005_cast_fp16))[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_equation_0 = const()[name = tensor("op_7028_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7028_cast_fp16 = einsum(equation = var_7028_equation_0, values = (var_6944_cast_fp16_1, var_7006_cast_fp16))[name = tensor("op_7028_cast_fp16")]; + tensor var_7030_equation_0 = const()[name = tensor("op_7030_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7030_cast_fp16 = einsum(equation = var_7030_equation_0, values = (var_6944_cast_fp16_2, var_7007_cast_fp16))[name = tensor("op_7030_cast_fp16")]; + tensor var_7032_equation_0 = const()[name = tensor("op_7032_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7032_cast_fp16 = einsum(equation = var_7032_equation_0, values = (var_6944_cast_fp16_3, var_7008_cast_fp16))[name = tensor("op_7032_cast_fp16")]; + tensor var_7034_equation_0 = const()[name = tensor("op_7034_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7034_cast_fp16 = einsum(equation = var_7034_equation_0, values = (var_6944_cast_fp16_4, var_7009_cast_fp16))[name = tensor("op_7034_cast_fp16")]; + tensor var_7036_equation_0 = const()[name = tensor("op_7036_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7036_cast_fp16 = einsum(equation = var_7036_equation_0, values = (var_6944_cast_fp16_5, var_7010_cast_fp16))[name = tensor("op_7036_cast_fp16")]; + tensor var_7038_equation_0 = const()[name = tensor("op_7038_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7038_cast_fp16 = einsum(equation = var_7038_equation_0, values = (var_6944_cast_fp16_6, var_7011_cast_fp16))[name = tensor("op_7038_cast_fp16")]; + tensor var_7040_equation_0 = const()[name = tensor("op_7040_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7040_cast_fp16 = einsum(equation = var_7040_equation_0, values = (var_6944_cast_fp16_7, var_7012_cast_fp16))[name = tensor("op_7040_cast_fp16")]; + tensor var_7042_equation_0 = const()[name = tensor("op_7042_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7042_cast_fp16 = einsum(equation = var_7042_equation_0, values = (var_6944_cast_fp16_8, var_7013_cast_fp16))[name = tensor("op_7042_cast_fp16")]; + tensor var_7044_equation_0 = const()[name = tensor("op_7044_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7044_cast_fp16 = einsum(equation = var_7044_equation_0, values = (var_6944_cast_fp16_9, var_7014_cast_fp16))[name = tensor("op_7044_cast_fp16")]; + tensor var_7046_equation_0 = const()[name = tensor("op_7046_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7046_cast_fp16 = einsum(equation = var_7046_equation_0, values = (var_6944_cast_fp16_10, var_7015_cast_fp16))[name = tensor("op_7046_cast_fp16")]; + tensor var_7048_equation_0 = const()[name = tensor("op_7048_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7048_cast_fp16 = einsum(equation = var_7048_equation_0, values = (var_6944_cast_fp16_11, var_7016_cast_fp16))[name = tensor("op_7048_cast_fp16")]; + tensor var_7050_equation_0 = const()[name = tensor("op_7050_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7050_cast_fp16 = einsum(equation = var_7050_equation_0, values = (var_6944_cast_fp16_12, var_7017_cast_fp16))[name = tensor("op_7050_cast_fp16")]; + tensor var_7052_equation_0 = const()[name = tensor("op_7052_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7052_cast_fp16 = einsum(equation = var_7052_equation_0, values = (var_6944_cast_fp16_13, var_7018_cast_fp16))[name = tensor("op_7052_cast_fp16")]; + tensor var_7054_equation_0 = const()[name = tensor("op_7054_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7054_cast_fp16 = einsum(equation = var_7054_equation_0, values = (var_6944_cast_fp16_14, var_7019_cast_fp16))[name = tensor("op_7054_cast_fp16")]; + tensor var_7056_equation_0 = const()[name = tensor("op_7056_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7056_cast_fp16 = einsum(equation = var_7056_equation_0, values = (var_6944_cast_fp16_15, var_7020_cast_fp16))[name = tensor("op_7056_cast_fp16")]; + tensor var_7058_equation_0 = const()[name = tensor("op_7058_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7058_cast_fp16 = einsum(equation = var_7058_equation_0, values = (var_6944_cast_fp16_16, var_7021_cast_fp16))[name = tensor("op_7058_cast_fp16")]; + tensor var_7060_equation_0 = const()[name = tensor("op_7060_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7060_cast_fp16 = einsum(equation = var_7060_equation_0, values = (var_6944_cast_fp16_17, var_7022_cast_fp16))[name = tensor("op_7060_cast_fp16")]; + tensor var_7062_equation_0 = const()[name = tensor("op_7062_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7062_cast_fp16 = einsum(equation = var_7062_equation_0, values = (var_6944_cast_fp16_18, var_7023_cast_fp16))[name = tensor("op_7062_cast_fp16")]; + tensor var_7064_equation_0 = const()[name = tensor("op_7064_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7064_cast_fp16 = einsum(equation = var_7064_equation_0, values = (var_6944_cast_fp16_19, var_7024_cast_fp16))[name = tensor("op_7064_cast_fp16")]; + tensor input_255_interleave_0 = const()[name = tensor("input_255_interleave_0"), val = tensor(false)]; + tensor input_255_cast_fp16 = concat(axis = var_6849, interleave = input_255_interleave_0, values = (var_7026_cast_fp16, var_7028_cast_fp16, var_7030_cast_fp16, var_7032_cast_fp16, var_7034_cast_fp16, var_7036_cast_fp16, var_7038_cast_fp16, var_7040_cast_fp16, var_7042_cast_fp16, var_7044_cast_fp16, var_7046_cast_fp16, var_7048_cast_fp16, var_7050_cast_fp16, var_7052_cast_fp16, var_7054_cast_fp16, var_7056_cast_fp16, var_7058_cast_fp16, var_7060_cast_fp16, var_7062_cast_fp16, var_7064_cast_fp16))[name = tensor("input_255_cast_fp16")]; + tensor var_7073_pad_type_0 = const()[name = tensor("op_7073_pad_type_0"), val = tensor("valid")]; + tensor var_7073_strides_0 = const()[name = tensor("op_7073_strides_0"), val = tensor([1, 1])]; + tensor var_7073_pad_0 = const()[name = tensor("op_7073_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7073_dilations_0 = const()[name = tensor("op_7073_dilations_0"), val = tensor([1, 1])]; + tensor var_7073_groups_0 = const()[name = tensor("op_7073_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_out_weight_to_fp16 = const()[name = tensor("blocks_25_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1007963392)))]; + tensor blocks_25_attn_out_bias_to_fp16 = const()[name = tensor("blocks_25_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011240256)))]; + tensor var_7073_cast_fp16 = conv(bias = blocks_25_attn_out_bias_to_fp16, dilations = var_7073_dilations_0, groups = var_7073_groups_0, pad = var_7073_pad_0, pad_type = var_7073_pad_type_0, strides = var_7073_strides_0, weight = blocks_25_attn_out_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("op_7073_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_7073_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([1])]; + tensor input_257_gamma_0_to_fp16 = const()[name = tensor("input_257_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011242880)))]; + tensor input_257_beta_0_to_fp16 = const()[name = tensor("input_257_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011245504)))]; + tensor var_7083_to_fp16 = const()[name = tensor("op_7083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = input_257_beta_0_to_fp16, epsilon = var_7083_to_fp16, gamma = input_257_gamma_0_to_fp16, x = inputs_103_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("valid")]; + tensor input_259_strides_0 = const()[name = tensor("input_259_strides_0"), val = tensor([1, 1])]; + tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_259_dilations_0 = const()[name = tensor("input_259_dilations_0"), val = tensor([1, 1])]; + tensor input_259_groups_0 = const()[name = tensor("input_259_groups_0"), val = tensor(1)]; + tensor blocks_25_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011248128)))]; + tensor blocks_25_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024355392)))]; + tensor input_259_cast_fp16 = conv(bias = blocks_25_mlp_0_bias_to_fp16, dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = blocks_25_mlp_0_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_mode_0 = const()[name = tensor("input_261_mode_0"), val = tensor("EXACT")]; + tensor input_261_cast_fp16 = gelu(mode = input_261_mode_0, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_7109_pad_type_0 = const()[name = tensor("op_7109_pad_type_0"), val = tensor("valid")]; + tensor var_7109_strides_0 = const()[name = tensor("op_7109_strides_0"), val = tensor([1, 1])]; + tensor var_7109_pad_0 = const()[name = tensor("op_7109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7109_dilations_0 = const()[name = tensor("op_7109_dilations_0"), val = tensor([1, 1])]; + tensor var_7109_groups_0 = const()[name = tensor("op_7109_groups_0"), val = tensor(1)]; + tensor blocks_25_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024365696)))]; + tensor blocks_25_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037472960)))]; + tensor var_7109_cast_fp16 = conv(bias = blocks_25_mlp_2_bias_to_fp16, dilations = var_7109_dilations_0, groups = var_7109_groups_0, pad = var_7109_pad_0, pad_type = var_7109_pad_type_0, strides = var_7109_strides_0, weight = blocks_25_mlp_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("op_7109_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = var_7109_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_7118 = const()[name = tensor("op_7118"), val = tensor(1)]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([1])]; + tensor input_263_gamma_0_to_fp16 = const()[name = tensor("input_263_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037475584)))]; + tensor input_263_beta_0_to_fp16 = const()[name = tensor("input_263_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037478208)))]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = input_263_beta_0_to_fp16, epsilon = var_7134_to_fp16, gamma = input_263_gamma_0_to_fp16, x = inputs_105_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor q_53_pad_type_0 = const()[name = tensor("q_53_pad_type_0"), val = tensor("valid")]; + tensor q_53_strides_0 = const()[name = tensor("q_53_strides_0"), val = tensor([1, 1])]; + tensor q_53_pad_0 = const()[name = tensor("q_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_53_dilations_0 = const()[name = tensor("q_53_dilations_0"), val = tensor([1, 1])]; + tensor q_53_groups_0 = const()[name = tensor("q_53_groups_0"), val = tensor(1)]; + tensor var_7169_weight_0_to_fp16 = const()[name = tensor("op_7169_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037480832)))]; + tensor var_7169_bias_0_to_fp16 = const()[name = tensor("op_7169_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040757696)))]; + tensor var_7169_cast_fp16 = conv(bias = var_7169_bias_0_to_fp16, dilations = q_53_dilations_0, groups = q_53_groups_0, pad = q_53_pad_0, pad_type = q_53_pad_type_0, strides = q_53_strides_0, weight = var_7169_weight_0_to_fp16, x = input_263_cast_fp16)[name = tensor("op_7169_cast_fp16")]; + tensor k_53_pad_type_0 = const()[name = tensor("k_53_pad_type_0"), val = tensor("valid")]; + tensor k_53_strides_0 = const()[name = tensor("k_53_strides_0"), val = tensor([1, 1])]; + tensor k_53_pad_0 = const()[name = tensor("k_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_53_dilations_0 = const()[name = tensor("k_53_dilations_0"), val = tensor([1, 1])]; + tensor k_53_groups_0 = const()[name = tensor("k_53_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_key_weight_to_fp16 = const()[name = tensor("blocks_26_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040760320)))]; + tensor k_53_cast_fp16 = conv(dilations = k_53_dilations_0, groups = k_53_groups_0, pad = k_53_pad_0, pad_type = k_53_pad_type_0, strides = k_53_strides_0, weight = blocks_26_attn_key_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor var_7167_pad_type_0 = const()[name = tensor("op_7167_pad_type_0"), val = tensor("valid")]; + tensor var_7167_strides_0 = const()[name = tensor("op_7167_strides_0"), val = tensor([1, 1])]; + tensor var_7167_pad_0 = const()[name = tensor("op_7167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7167_dilations_0 = const()[name = tensor("op_7167_dilations_0"), val = tensor([1, 1])]; + tensor var_7167_groups_0 = const()[name = tensor("op_7167_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_value_weight_to_fp16 = const()[name = tensor("blocks_26_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044037184)))]; + tensor blocks_26_attn_value_bias_to_fp16 = const()[name = tensor("blocks_26_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047314048)))]; + tensor var_7167_cast_fp16 = conv(bias = blocks_26_attn_value_bias_to_fp16, dilations = var_7167_dilations_0, groups = var_7167_groups_0, pad = var_7167_pad_0, pad_type = var_7167_pad_type_0, strides = var_7167_strides_0, weight = blocks_26_attn_value_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("op_7167_cast_fp16")]; + tensor tile_78 = const()[name = tensor("tile_78"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7170_axis_0 = const()[name = tensor("op_7170_axis_0"), val = tensor(1)]; + tensor var_7170_cast_fp16_0, tensor var_7170_cast_fp16_1, tensor var_7170_cast_fp16_2, tensor var_7170_cast_fp16_3, tensor var_7170_cast_fp16_4, tensor var_7170_cast_fp16_5, tensor var_7170_cast_fp16_6, tensor var_7170_cast_fp16_7, tensor var_7170_cast_fp16_8, tensor var_7170_cast_fp16_9, tensor var_7170_cast_fp16_10, tensor var_7170_cast_fp16_11, tensor var_7170_cast_fp16_12, tensor var_7170_cast_fp16_13, tensor var_7170_cast_fp16_14, tensor var_7170_cast_fp16_15, tensor var_7170_cast_fp16_16, tensor var_7170_cast_fp16_17, tensor var_7170_cast_fp16_18, tensor var_7170_cast_fp16_19 = split(axis = var_7170_axis_0, split_sizes = tile_78, x = var_7169_cast_fp16)[name = tensor("op_7170_cast_fp16")]; + tensor var_7191_perm_0 = const()[name = tensor("op_7191_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_79 = const()[name = tensor("tile_79"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7192_axis_0 = const()[name = tensor("op_7192_axis_0"), val = tensor(3)]; + tensor var_7191_cast_fp16 = transpose(perm = var_7191_perm_0, x = k_53_cast_fp16)[name = tensor("transpose_6")]; + tensor var_7192_cast_fp16_0, tensor var_7192_cast_fp16_1, tensor var_7192_cast_fp16_2, tensor var_7192_cast_fp16_3, tensor var_7192_cast_fp16_4, tensor var_7192_cast_fp16_5, tensor var_7192_cast_fp16_6, tensor var_7192_cast_fp16_7, tensor var_7192_cast_fp16_8, tensor var_7192_cast_fp16_9, tensor var_7192_cast_fp16_10, tensor var_7192_cast_fp16_11, tensor var_7192_cast_fp16_12, tensor var_7192_cast_fp16_13, tensor var_7192_cast_fp16_14, tensor var_7192_cast_fp16_15, tensor var_7192_cast_fp16_16, tensor var_7192_cast_fp16_17, tensor var_7192_cast_fp16_18, tensor var_7192_cast_fp16_19 = split(axis = var_7192_axis_0, split_sizes = tile_79, x = var_7191_cast_fp16)[name = tensor("op_7192_cast_fp16")]; + tensor tile_80 = const()[name = tensor("tile_80"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7213_axis_0 = const()[name = tensor("op_7213_axis_0"), val = tensor(1)]; + tensor var_7213_cast_fp16_0, tensor var_7213_cast_fp16_1, tensor var_7213_cast_fp16_2, tensor var_7213_cast_fp16_3, tensor var_7213_cast_fp16_4, tensor var_7213_cast_fp16_5, tensor var_7213_cast_fp16_6, tensor var_7213_cast_fp16_7, tensor var_7213_cast_fp16_8, tensor var_7213_cast_fp16_9, tensor var_7213_cast_fp16_10, tensor var_7213_cast_fp16_11, tensor var_7213_cast_fp16_12, tensor var_7213_cast_fp16_13, tensor var_7213_cast_fp16_14, tensor var_7213_cast_fp16_15, tensor var_7213_cast_fp16_16, tensor var_7213_cast_fp16_17, tensor var_7213_cast_fp16_18, tensor var_7213_cast_fp16_19 = split(axis = var_7213_axis_0, split_sizes = tile_80, x = var_7167_cast_fp16)[name = tensor("op_7213_cast_fp16")]; + tensor aw_1041_equation_0 = const()[name = tensor("aw_1041_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1041_cast_fp16 = einsum(equation = aw_1041_equation_0, values = (var_7192_cast_fp16_0, var_7170_cast_fp16_0))[name = tensor("aw_1041_cast_fp16")]; + tensor aw_1043_equation_0 = const()[name = tensor("aw_1043_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1043_cast_fp16 = einsum(equation = aw_1043_equation_0, values = (var_7192_cast_fp16_1, var_7170_cast_fp16_1))[name = tensor("aw_1043_cast_fp16")]; + tensor aw_1045_equation_0 = const()[name = tensor("aw_1045_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1045_cast_fp16 = einsum(equation = aw_1045_equation_0, values = (var_7192_cast_fp16_2, var_7170_cast_fp16_2))[name = tensor("aw_1045_cast_fp16")]; + tensor aw_1047_equation_0 = const()[name = tensor("aw_1047_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1047_cast_fp16 = einsum(equation = aw_1047_equation_0, values = (var_7192_cast_fp16_3, var_7170_cast_fp16_3))[name = tensor("aw_1047_cast_fp16")]; + tensor aw_1049_equation_0 = const()[name = tensor("aw_1049_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1049_cast_fp16 = einsum(equation = aw_1049_equation_0, values = (var_7192_cast_fp16_4, var_7170_cast_fp16_4))[name = tensor("aw_1049_cast_fp16")]; + tensor aw_1051_equation_0 = const()[name = tensor("aw_1051_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1051_cast_fp16 = einsum(equation = aw_1051_equation_0, values = (var_7192_cast_fp16_5, var_7170_cast_fp16_5))[name = tensor("aw_1051_cast_fp16")]; + tensor aw_1053_equation_0 = const()[name = tensor("aw_1053_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1053_cast_fp16 = einsum(equation = aw_1053_equation_0, values = (var_7192_cast_fp16_6, var_7170_cast_fp16_6))[name = tensor("aw_1053_cast_fp16")]; + tensor aw_1055_equation_0 = const()[name = tensor("aw_1055_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1055_cast_fp16 = einsum(equation = aw_1055_equation_0, values = (var_7192_cast_fp16_7, var_7170_cast_fp16_7))[name = tensor("aw_1055_cast_fp16")]; + tensor aw_1057_equation_0 = const()[name = tensor("aw_1057_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1057_cast_fp16 = einsum(equation = aw_1057_equation_0, values = (var_7192_cast_fp16_8, var_7170_cast_fp16_8))[name = tensor("aw_1057_cast_fp16")]; + tensor aw_1059_equation_0 = const()[name = tensor("aw_1059_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1059_cast_fp16 = einsum(equation = aw_1059_equation_0, values = (var_7192_cast_fp16_9, var_7170_cast_fp16_9))[name = tensor("aw_1059_cast_fp16")]; + tensor aw_1061_equation_0 = const()[name = tensor("aw_1061_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1061_cast_fp16 = einsum(equation = aw_1061_equation_0, values = (var_7192_cast_fp16_10, var_7170_cast_fp16_10))[name = tensor("aw_1061_cast_fp16")]; + tensor aw_1063_equation_0 = const()[name = tensor("aw_1063_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1063_cast_fp16 = einsum(equation = aw_1063_equation_0, values = (var_7192_cast_fp16_11, var_7170_cast_fp16_11))[name = tensor("aw_1063_cast_fp16")]; + tensor aw_1065_equation_0 = const()[name = tensor("aw_1065_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1065_cast_fp16 = einsum(equation = aw_1065_equation_0, values = (var_7192_cast_fp16_12, var_7170_cast_fp16_12))[name = tensor("aw_1065_cast_fp16")]; + tensor aw_1067_equation_0 = const()[name = tensor("aw_1067_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1067_cast_fp16 = einsum(equation = aw_1067_equation_0, values = (var_7192_cast_fp16_13, var_7170_cast_fp16_13))[name = tensor("aw_1067_cast_fp16")]; + tensor aw_1069_equation_0 = const()[name = tensor("aw_1069_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1069_cast_fp16 = einsum(equation = aw_1069_equation_0, values = (var_7192_cast_fp16_14, var_7170_cast_fp16_14))[name = tensor("aw_1069_cast_fp16")]; + tensor aw_1071_equation_0 = const()[name = tensor("aw_1071_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1071_cast_fp16 = einsum(equation = aw_1071_equation_0, values = (var_7192_cast_fp16_15, var_7170_cast_fp16_15))[name = tensor("aw_1071_cast_fp16")]; + tensor aw_1073_equation_0 = const()[name = tensor("aw_1073_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1073_cast_fp16 = einsum(equation = aw_1073_equation_0, values = (var_7192_cast_fp16_16, var_7170_cast_fp16_16))[name = tensor("aw_1073_cast_fp16")]; + tensor aw_1075_equation_0 = const()[name = tensor("aw_1075_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1075_cast_fp16 = einsum(equation = aw_1075_equation_0, values = (var_7192_cast_fp16_17, var_7170_cast_fp16_17))[name = tensor("aw_1075_cast_fp16")]; + tensor aw_1077_equation_0 = const()[name = tensor("aw_1077_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1077_cast_fp16 = einsum(equation = aw_1077_equation_0, values = (var_7192_cast_fp16_18, var_7170_cast_fp16_18))[name = tensor("aw_1077_cast_fp16")]; + tensor aw_1079_equation_0 = const()[name = tensor("aw_1079_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1079_cast_fp16 = einsum(equation = aw_1079_equation_0, values = (var_7192_cast_fp16_19, var_7170_cast_fp16_19))[name = tensor("aw_1079_cast_fp16")]; + tensor var_7274_cast_fp16 = softmax(axis = var_7118, x = aw_1041_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor var_7275_cast_fp16 = softmax(axis = var_7118, x = aw_1043_cast_fp16)[name = tensor("op_7275_cast_fp16")]; + tensor var_7276_cast_fp16 = softmax(axis = var_7118, x = aw_1045_cast_fp16)[name = tensor("op_7276_cast_fp16")]; + tensor var_7277_cast_fp16 = softmax(axis = var_7118, x = aw_1047_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7278_cast_fp16 = softmax(axis = var_7118, x = aw_1049_cast_fp16)[name = tensor("op_7278_cast_fp16")]; + tensor var_7279_cast_fp16 = softmax(axis = var_7118, x = aw_1051_cast_fp16)[name = tensor("op_7279_cast_fp16")]; + tensor var_7280_cast_fp16 = softmax(axis = var_7118, x = aw_1053_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor var_7281_cast_fp16 = softmax(axis = var_7118, x = aw_1055_cast_fp16)[name = tensor("op_7281_cast_fp16")]; + tensor var_7282_cast_fp16 = softmax(axis = var_7118, x = aw_1057_cast_fp16)[name = tensor("op_7282_cast_fp16")]; + tensor var_7283_cast_fp16 = softmax(axis = var_7118, x = aw_1059_cast_fp16)[name = tensor("op_7283_cast_fp16")]; + tensor var_7284_cast_fp16 = softmax(axis = var_7118, x = aw_1061_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = softmax(axis = var_7118, x = aw_1063_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7286_cast_fp16 = softmax(axis = var_7118, x = aw_1065_cast_fp16)[name = tensor("op_7286_cast_fp16")]; + tensor var_7287_cast_fp16 = softmax(axis = var_7118, x = aw_1067_cast_fp16)[name = tensor("op_7287_cast_fp16")]; + tensor var_7288_cast_fp16 = softmax(axis = var_7118, x = aw_1069_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = softmax(axis = var_7118, x = aw_1071_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor var_7290_cast_fp16 = softmax(axis = var_7118, x = aw_1073_cast_fp16)[name = tensor("op_7290_cast_fp16")]; + tensor var_7291_cast_fp16 = softmax(axis = var_7118, x = aw_1075_cast_fp16)[name = tensor("op_7291_cast_fp16")]; + tensor var_7292_cast_fp16 = softmax(axis = var_7118, x = aw_1077_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = softmax(axis = var_7118, x = aw_1079_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7295_equation_0 = const()[name = tensor("op_7295_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7295_cast_fp16 = einsum(equation = var_7295_equation_0, values = (var_7213_cast_fp16_0, var_7274_cast_fp16))[name = tensor("op_7295_cast_fp16")]; + tensor var_7297_equation_0 = const()[name = tensor("op_7297_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7297_cast_fp16 = einsum(equation = var_7297_equation_0, values = (var_7213_cast_fp16_1, var_7275_cast_fp16))[name = tensor("op_7297_cast_fp16")]; + tensor var_7299_equation_0 = const()[name = tensor("op_7299_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7299_cast_fp16 = einsum(equation = var_7299_equation_0, values = (var_7213_cast_fp16_2, var_7276_cast_fp16))[name = tensor("op_7299_cast_fp16")]; + tensor var_7301_equation_0 = const()[name = tensor("op_7301_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7301_cast_fp16 = einsum(equation = var_7301_equation_0, values = (var_7213_cast_fp16_3, var_7277_cast_fp16))[name = tensor("op_7301_cast_fp16")]; + tensor var_7303_equation_0 = const()[name = tensor("op_7303_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7303_cast_fp16 = einsum(equation = var_7303_equation_0, values = (var_7213_cast_fp16_4, var_7278_cast_fp16))[name = tensor("op_7303_cast_fp16")]; + tensor var_7305_equation_0 = const()[name = tensor("op_7305_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7305_cast_fp16 = einsum(equation = var_7305_equation_0, values = (var_7213_cast_fp16_5, var_7279_cast_fp16))[name = tensor("op_7305_cast_fp16")]; + tensor var_7307_equation_0 = const()[name = tensor("op_7307_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7307_cast_fp16 = einsum(equation = var_7307_equation_0, values = (var_7213_cast_fp16_6, var_7280_cast_fp16))[name = tensor("op_7307_cast_fp16")]; + tensor var_7309_equation_0 = const()[name = tensor("op_7309_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7309_cast_fp16 = einsum(equation = var_7309_equation_0, values = (var_7213_cast_fp16_7, var_7281_cast_fp16))[name = tensor("op_7309_cast_fp16")]; + tensor var_7311_equation_0 = const()[name = tensor("op_7311_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7311_cast_fp16 = einsum(equation = var_7311_equation_0, values = (var_7213_cast_fp16_8, var_7282_cast_fp16))[name = tensor("op_7311_cast_fp16")]; + tensor var_7313_equation_0 = const()[name = tensor("op_7313_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7313_cast_fp16 = einsum(equation = var_7313_equation_0, values = (var_7213_cast_fp16_9, var_7283_cast_fp16))[name = tensor("op_7313_cast_fp16")]; + tensor var_7315_equation_0 = const()[name = tensor("op_7315_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7315_cast_fp16 = einsum(equation = var_7315_equation_0, values = (var_7213_cast_fp16_10, var_7284_cast_fp16))[name = tensor("op_7315_cast_fp16")]; + tensor var_7317_equation_0 = const()[name = tensor("op_7317_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7317_cast_fp16 = einsum(equation = var_7317_equation_0, values = (var_7213_cast_fp16_11, var_7285_cast_fp16))[name = tensor("op_7317_cast_fp16")]; + tensor var_7319_equation_0 = const()[name = tensor("op_7319_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7319_cast_fp16 = einsum(equation = var_7319_equation_0, values = (var_7213_cast_fp16_12, var_7286_cast_fp16))[name = tensor("op_7319_cast_fp16")]; + tensor var_7321_equation_0 = const()[name = tensor("op_7321_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7321_cast_fp16 = einsum(equation = var_7321_equation_0, values = (var_7213_cast_fp16_13, var_7287_cast_fp16))[name = tensor("op_7321_cast_fp16")]; + tensor var_7323_equation_0 = const()[name = tensor("op_7323_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7323_cast_fp16 = einsum(equation = var_7323_equation_0, values = (var_7213_cast_fp16_14, var_7288_cast_fp16))[name = tensor("op_7323_cast_fp16")]; + tensor var_7325_equation_0 = const()[name = tensor("op_7325_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7325_cast_fp16 = einsum(equation = var_7325_equation_0, values = (var_7213_cast_fp16_15, var_7289_cast_fp16))[name = tensor("op_7325_cast_fp16")]; + tensor var_7327_equation_0 = const()[name = tensor("op_7327_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7327_cast_fp16 = einsum(equation = var_7327_equation_0, values = (var_7213_cast_fp16_16, var_7290_cast_fp16))[name = tensor("op_7327_cast_fp16")]; + tensor var_7329_equation_0 = const()[name = tensor("op_7329_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7329_cast_fp16 = einsum(equation = var_7329_equation_0, values = (var_7213_cast_fp16_17, var_7291_cast_fp16))[name = tensor("op_7329_cast_fp16")]; + tensor var_7331_equation_0 = const()[name = tensor("op_7331_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7331_cast_fp16 = einsum(equation = var_7331_equation_0, values = (var_7213_cast_fp16_18, var_7292_cast_fp16))[name = tensor("op_7331_cast_fp16")]; + tensor var_7333_equation_0 = const()[name = tensor("op_7333_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7333_cast_fp16 = einsum(equation = var_7333_equation_0, values = (var_7213_cast_fp16_19, var_7293_cast_fp16))[name = tensor("op_7333_cast_fp16")]; + tensor input_265_interleave_0 = const()[name = tensor("input_265_interleave_0"), val = tensor(false)]; + tensor input_265_cast_fp16 = concat(axis = var_7118, interleave = input_265_interleave_0, values = (var_7295_cast_fp16, var_7297_cast_fp16, var_7299_cast_fp16, var_7301_cast_fp16, var_7303_cast_fp16, var_7305_cast_fp16, var_7307_cast_fp16, var_7309_cast_fp16, var_7311_cast_fp16, var_7313_cast_fp16, var_7315_cast_fp16, var_7317_cast_fp16, var_7319_cast_fp16, var_7321_cast_fp16, var_7323_cast_fp16, var_7325_cast_fp16, var_7327_cast_fp16, var_7329_cast_fp16, var_7331_cast_fp16, var_7333_cast_fp16))[name = tensor("input_265_cast_fp16")]; + tensor var_7342_pad_type_0 = const()[name = tensor("op_7342_pad_type_0"), val = tensor("valid")]; + tensor var_7342_strides_0 = const()[name = tensor("op_7342_strides_0"), val = tensor([1, 1])]; + tensor var_7342_pad_0 = const()[name = tensor("op_7342_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7342_dilations_0 = const()[name = tensor("op_7342_dilations_0"), val = tensor([1, 1])]; + tensor var_7342_groups_0 = const()[name = tensor("op_7342_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_out_weight_to_fp16 = const()[name = tensor("blocks_26_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047316672)))]; + tensor blocks_26_attn_out_bias_to_fp16 = const()[name = tensor("blocks_26_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050593536)))]; + tensor var_7342_cast_fp16 = conv(bias = blocks_26_attn_out_bias_to_fp16, dilations = var_7342_dilations_0, groups = var_7342_groups_0, pad = var_7342_pad_0, pad_type = var_7342_pad_type_0, strides = var_7342_strides_0, weight = blocks_26_attn_out_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("op_7342_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = var_7342_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([1])]; + tensor input_267_gamma_0_to_fp16 = const()[name = tensor("input_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050596160)))]; + tensor input_267_beta_0_to_fp16 = const()[name = tensor("input_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050598784)))]; + tensor var_7352_to_fp16 = const()[name = tensor("op_7352_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = input_267_beta_0_to_fp16, epsilon = var_7352_to_fp16, gamma = input_267_gamma_0_to_fp16, x = inputs_107_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_pad_type_0 = const()[name = tensor("input_269_pad_type_0"), val = tensor("valid")]; + tensor input_269_strides_0 = const()[name = tensor("input_269_strides_0"), val = tensor([1, 1])]; + tensor input_269_pad_0 = const()[name = tensor("input_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_269_dilations_0 = const()[name = tensor("input_269_dilations_0"), val = tensor([1, 1])]; + tensor input_269_groups_0 = const()[name = tensor("input_269_groups_0"), val = tensor(1)]; + tensor blocks_26_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050601408)))]; + tensor blocks_26_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063708672)))]; + tensor input_269_cast_fp16 = conv(bias = blocks_26_mlp_0_bias_to_fp16, dilations = input_269_dilations_0, groups = input_269_groups_0, pad = input_269_pad_0, pad_type = input_269_pad_type_0, strides = input_269_strides_0, weight = blocks_26_mlp_0_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor input_271_mode_0 = const()[name = tensor("input_271_mode_0"), val = tensor("EXACT")]; + tensor input_271_cast_fp16 = gelu(mode = input_271_mode_0, x = input_269_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor var_7378_pad_type_0 = const()[name = tensor("op_7378_pad_type_0"), val = tensor("valid")]; + tensor var_7378_strides_0 = const()[name = tensor("op_7378_strides_0"), val = tensor([1, 1])]; + tensor var_7378_pad_0 = const()[name = tensor("op_7378_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7378_dilations_0 = const()[name = tensor("op_7378_dilations_0"), val = tensor([1, 1])]; + tensor var_7378_groups_0 = const()[name = tensor("op_7378_groups_0"), val = tensor(1)]; + tensor blocks_26_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063718976)))]; + tensor blocks_26_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076826240)))]; + tensor var_7378_cast_fp16 = conv(bias = blocks_26_mlp_2_bias_to_fp16, dilations = var_7378_dilations_0, groups = var_7378_groups_0, pad = var_7378_pad_0, pad_type = var_7378_pad_type_0, strides = var_7378_strides_0, weight = blocks_26_mlp_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("op_7378_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_7378_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_7387 = const()[name = tensor("op_7387"), val = tensor(1)]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([1])]; + tensor input_273_gamma_0_to_fp16 = const()[name = tensor("input_273_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076828864)))]; + tensor input_273_beta_0_to_fp16 = const()[name = tensor("input_273_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076831488)))]; + tensor var_7403_to_fp16 = const()[name = tensor("op_7403_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = input_273_beta_0_to_fp16, epsilon = var_7403_to_fp16, gamma = input_273_gamma_0_to_fp16, x = inputs_109_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor q_55_pad_type_0 = const()[name = tensor("q_55_pad_type_0"), val = tensor("valid")]; + tensor q_55_strides_0 = const()[name = tensor("q_55_strides_0"), val = tensor([1, 1])]; + tensor q_55_pad_0 = const()[name = tensor("q_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_55_dilations_0 = const()[name = tensor("q_55_dilations_0"), val = tensor([1, 1])]; + tensor q_55_groups_0 = const()[name = tensor("q_55_groups_0"), val = tensor(1)]; + tensor var_7438_weight_0_to_fp16 = const()[name = tensor("op_7438_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076834112)))]; + tensor var_7438_bias_0_to_fp16 = const()[name = tensor("op_7438_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080110976)))]; + tensor var_7438_cast_fp16 = conv(bias = var_7438_bias_0_to_fp16, dilations = q_55_dilations_0, groups = q_55_groups_0, pad = q_55_pad_0, pad_type = q_55_pad_type_0, strides = q_55_strides_0, weight = var_7438_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("op_7438_cast_fp16")]; + tensor k_55_pad_type_0 = const()[name = tensor("k_55_pad_type_0"), val = tensor("valid")]; + tensor k_55_strides_0 = const()[name = tensor("k_55_strides_0"), val = tensor([1, 1])]; + tensor k_55_pad_0 = const()[name = tensor("k_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_55_dilations_0 = const()[name = tensor("k_55_dilations_0"), val = tensor([1, 1])]; + tensor k_55_groups_0 = const()[name = tensor("k_55_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_key_weight_to_fp16 = const()[name = tensor("blocks_27_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080113600)))]; + tensor k_55_cast_fp16 = conv(dilations = k_55_dilations_0, groups = k_55_groups_0, pad = k_55_pad_0, pad_type = k_55_pad_type_0, strides = k_55_strides_0, weight = blocks_27_attn_key_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor var_7436_pad_type_0 = const()[name = tensor("op_7436_pad_type_0"), val = tensor("valid")]; + tensor var_7436_strides_0 = const()[name = tensor("op_7436_strides_0"), val = tensor([1, 1])]; + tensor var_7436_pad_0 = const()[name = tensor("op_7436_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7436_dilations_0 = const()[name = tensor("op_7436_dilations_0"), val = tensor([1, 1])]; + tensor var_7436_groups_0 = const()[name = tensor("op_7436_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_value_weight_to_fp16 = const()[name = tensor("blocks_27_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083390464)))]; + tensor blocks_27_attn_value_bias_to_fp16 = const()[name = tensor("blocks_27_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086667328)))]; + tensor var_7436_cast_fp16 = conv(bias = blocks_27_attn_value_bias_to_fp16, dilations = var_7436_dilations_0, groups = var_7436_groups_0, pad = var_7436_pad_0, pad_type = var_7436_pad_type_0, strides = var_7436_strides_0, weight = blocks_27_attn_value_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("op_7436_cast_fp16")]; + tensor tile_81 = const()[name = tensor("tile_81"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_cast_fp16_0, tensor var_7439_cast_fp16_1, tensor var_7439_cast_fp16_2, tensor var_7439_cast_fp16_3, tensor var_7439_cast_fp16_4, tensor var_7439_cast_fp16_5, tensor var_7439_cast_fp16_6, tensor var_7439_cast_fp16_7, tensor var_7439_cast_fp16_8, tensor var_7439_cast_fp16_9, tensor var_7439_cast_fp16_10, tensor var_7439_cast_fp16_11, tensor var_7439_cast_fp16_12, tensor var_7439_cast_fp16_13, tensor var_7439_cast_fp16_14, tensor var_7439_cast_fp16_15, tensor var_7439_cast_fp16_16, tensor var_7439_cast_fp16_17, tensor var_7439_cast_fp16_18, tensor var_7439_cast_fp16_19 = split(axis = var_7439_axis_0, split_sizes = tile_81, x = var_7438_cast_fp16)[name = tensor("op_7439_cast_fp16")]; + tensor var_7460_perm_0 = const()[name = tensor("op_7460_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_82 = const()[name = tensor("tile_82"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7461_axis_0 = const()[name = tensor("op_7461_axis_0"), val = tensor(3)]; + tensor var_7460_cast_fp16 = transpose(perm = var_7460_perm_0, x = k_55_cast_fp16)[name = tensor("transpose_5")]; + tensor var_7461_cast_fp16_0, tensor var_7461_cast_fp16_1, tensor var_7461_cast_fp16_2, tensor var_7461_cast_fp16_3, tensor var_7461_cast_fp16_4, tensor var_7461_cast_fp16_5, tensor var_7461_cast_fp16_6, tensor var_7461_cast_fp16_7, tensor var_7461_cast_fp16_8, tensor var_7461_cast_fp16_9, tensor var_7461_cast_fp16_10, tensor var_7461_cast_fp16_11, tensor var_7461_cast_fp16_12, tensor var_7461_cast_fp16_13, tensor var_7461_cast_fp16_14, tensor var_7461_cast_fp16_15, tensor var_7461_cast_fp16_16, tensor var_7461_cast_fp16_17, tensor var_7461_cast_fp16_18, tensor var_7461_cast_fp16_19 = split(axis = var_7461_axis_0, split_sizes = tile_82, x = var_7460_cast_fp16)[name = tensor("op_7461_cast_fp16")]; + tensor tile_83 = const()[name = tensor("tile_83"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7482_axis_0 = const()[name = tensor("op_7482_axis_0"), val = tensor(1)]; + tensor var_7482_cast_fp16_0, tensor var_7482_cast_fp16_1, tensor var_7482_cast_fp16_2, tensor var_7482_cast_fp16_3, tensor var_7482_cast_fp16_4, tensor var_7482_cast_fp16_5, tensor var_7482_cast_fp16_6, tensor var_7482_cast_fp16_7, tensor var_7482_cast_fp16_8, tensor var_7482_cast_fp16_9, tensor var_7482_cast_fp16_10, tensor var_7482_cast_fp16_11, tensor var_7482_cast_fp16_12, tensor var_7482_cast_fp16_13, tensor var_7482_cast_fp16_14, tensor var_7482_cast_fp16_15, tensor var_7482_cast_fp16_16, tensor var_7482_cast_fp16_17, tensor var_7482_cast_fp16_18, tensor var_7482_cast_fp16_19 = split(axis = var_7482_axis_0, split_sizes = tile_83, x = var_7436_cast_fp16)[name = tensor("op_7482_cast_fp16")]; + tensor aw_1081_equation_0 = const()[name = tensor("aw_1081_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1081_cast_fp16 = einsum(equation = aw_1081_equation_0, values = (var_7461_cast_fp16_0, var_7439_cast_fp16_0))[name = tensor("aw_1081_cast_fp16")]; + tensor aw_1083_equation_0 = const()[name = tensor("aw_1083_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1083_cast_fp16 = einsum(equation = aw_1083_equation_0, values = (var_7461_cast_fp16_1, var_7439_cast_fp16_1))[name = tensor("aw_1083_cast_fp16")]; + tensor aw_1085_equation_0 = const()[name = tensor("aw_1085_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1085_cast_fp16 = einsum(equation = aw_1085_equation_0, values = (var_7461_cast_fp16_2, var_7439_cast_fp16_2))[name = tensor("aw_1085_cast_fp16")]; + tensor aw_1087_equation_0 = const()[name = tensor("aw_1087_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1087_cast_fp16 = einsum(equation = aw_1087_equation_0, values = (var_7461_cast_fp16_3, var_7439_cast_fp16_3))[name = tensor("aw_1087_cast_fp16")]; + tensor aw_1089_equation_0 = const()[name = tensor("aw_1089_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1089_cast_fp16 = einsum(equation = aw_1089_equation_0, values = (var_7461_cast_fp16_4, var_7439_cast_fp16_4))[name = tensor("aw_1089_cast_fp16")]; + tensor aw_1091_equation_0 = const()[name = tensor("aw_1091_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1091_cast_fp16 = einsum(equation = aw_1091_equation_0, values = (var_7461_cast_fp16_5, var_7439_cast_fp16_5))[name = tensor("aw_1091_cast_fp16")]; + tensor aw_1093_equation_0 = const()[name = tensor("aw_1093_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1093_cast_fp16 = einsum(equation = aw_1093_equation_0, values = (var_7461_cast_fp16_6, var_7439_cast_fp16_6))[name = tensor("aw_1093_cast_fp16")]; + tensor aw_1095_equation_0 = const()[name = tensor("aw_1095_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1095_cast_fp16 = einsum(equation = aw_1095_equation_0, values = (var_7461_cast_fp16_7, var_7439_cast_fp16_7))[name = tensor("aw_1095_cast_fp16")]; + tensor aw_1097_equation_0 = const()[name = tensor("aw_1097_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1097_cast_fp16 = einsum(equation = aw_1097_equation_0, values = (var_7461_cast_fp16_8, var_7439_cast_fp16_8))[name = tensor("aw_1097_cast_fp16")]; + tensor aw_1099_equation_0 = const()[name = tensor("aw_1099_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1099_cast_fp16 = einsum(equation = aw_1099_equation_0, values = (var_7461_cast_fp16_9, var_7439_cast_fp16_9))[name = tensor("aw_1099_cast_fp16")]; + tensor aw_1101_equation_0 = const()[name = tensor("aw_1101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1101_cast_fp16 = einsum(equation = aw_1101_equation_0, values = (var_7461_cast_fp16_10, var_7439_cast_fp16_10))[name = tensor("aw_1101_cast_fp16")]; + tensor aw_1103_equation_0 = const()[name = tensor("aw_1103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1103_cast_fp16 = einsum(equation = aw_1103_equation_0, values = (var_7461_cast_fp16_11, var_7439_cast_fp16_11))[name = tensor("aw_1103_cast_fp16")]; + tensor aw_1105_equation_0 = const()[name = tensor("aw_1105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1105_cast_fp16 = einsum(equation = aw_1105_equation_0, values = (var_7461_cast_fp16_12, var_7439_cast_fp16_12))[name = tensor("aw_1105_cast_fp16")]; + tensor aw_1107_equation_0 = const()[name = tensor("aw_1107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1107_cast_fp16 = einsum(equation = aw_1107_equation_0, values = (var_7461_cast_fp16_13, var_7439_cast_fp16_13))[name = tensor("aw_1107_cast_fp16")]; + tensor aw_1109_equation_0 = const()[name = tensor("aw_1109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1109_cast_fp16 = einsum(equation = aw_1109_equation_0, values = (var_7461_cast_fp16_14, var_7439_cast_fp16_14))[name = tensor("aw_1109_cast_fp16")]; + tensor aw_1111_equation_0 = const()[name = tensor("aw_1111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1111_cast_fp16 = einsum(equation = aw_1111_equation_0, values = (var_7461_cast_fp16_15, var_7439_cast_fp16_15))[name = tensor("aw_1111_cast_fp16")]; + tensor aw_1113_equation_0 = const()[name = tensor("aw_1113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1113_cast_fp16 = einsum(equation = aw_1113_equation_0, values = (var_7461_cast_fp16_16, var_7439_cast_fp16_16))[name = tensor("aw_1113_cast_fp16")]; + tensor aw_1115_equation_0 = const()[name = tensor("aw_1115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1115_cast_fp16 = einsum(equation = aw_1115_equation_0, values = (var_7461_cast_fp16_17, var_7439_cast_fp16_17))[name = tensor("aw_1115_cast_fp16")]; + tensor aw_1117_equation_0 = const()[name = tensor("aw_1117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1117_cast_fp16 = einsum(equation = aw_1117_equation_0, values = (var_7461_cast_fp16_18, var_7439_cast_fp16_18))[name = tensor("aw_1117_cast_fp16")]; + tensor aw_1119_equation_0 = const()[name = tensor("aw_1119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1119_cast_fp16 = einsum(equation = aw_1119_equation_0, values = (var_7461_cast_fp16_19, var_7439_cast_fp16_19))[name = tensor("aw_1119_cast_fp16")]; + tensor var_7543_cast_fp16 = softmax(axis = var_7387, x = aw_1081_cast_fp16)[name = tensor("op_7543_cast_fp16")]; + tensor var_7544_cast_fp16 = softmax(axis = var_7387, x = aw_1083_cast_fp16)[name = tensor("op_7544_cast_fp16")]; + tensor var_7545_cast_fp16 = softmax(axis = var_7387, x = aw_1085_cast_fp16)[name = tensor("op_7545_cast_fp16")]; + tensor var_7546_cast_fp16 = softmax(axis = var_7387, x = aw_1087_cast_fp16)[name = tensor("op_7546_cast_fp16")]; + tensor var_7547_cast_fp16 = softmax(axis = var_7387, x = aw_1089_cast_fp16)[name = tensor("op_7547_cast_fp16")]; + tensor var_7548_cast_fp16 = softmax(axis = var_7387, x = aw_1091_cast_fp16)[name = tensor("op_7548_cast_fp16")]; + tensor var_7549_cast_fp16 = softmax(axis = var_7387, x = aw_1093_cast_fp16)[name = tensor("op_7549_cast_fp16")]; + tensor var_7550_cast_fp16 = softmax(axis = var_7387, x = aw_1095_cast_fp16)[name = tensor("op_7550_cast_fp16")]; + tensor var_7551_cast_fp16 = softmax(axis = var_7387, x = aw_1097_cast_fp16)[name = tensor("op_7551_cast_fp16")]; + tensor var_7552_cast_fp16 = softmax(axis = var_7387, x = aw_1099_cast_fp16)[name = tensor("op_7552_cast_fp16")]; + tensor var_7553_cast_fp16 = softmax(axis = var_7387, x = aw_1101_cast_fp16)[name = tensor("op_7553_cast_fp16")]; + tensor var_7554_cast_fp16 = softmax(axis = var_7387, x = aw_1103_cast_fp16)[name = tensor("op_7554_cast_fp16")]; + tensor var_7555_cast_fp16 = softmax(axis = var_7387, x = aw_1105_cast_fp16)[name = tensor("op_7555_cast_fp16")]; + tensor var_7556_cast_fp16 = softmax(axis = var_7387, x = aw_1107_cast_fp16)[name = tensor("op_7556_cast_fp16")]; + tensor var_7557_cast_fp16 = softmax(axis = var_7387, x = aw_1109_cast_fp16)[name = tensor("op_7557_cast_fp16")]; + tensor var_7558_cast_fp16 = softmax(axis = var_7387, x = aw_1111_cast_fp16)[name = tensor("op_7558_cast_fp16")]; + tensor var_7559_cast_fp16 = softmax(axis = var_7387, x = aw_1113_cast_fp16)[name = tensor("op_7559_cast_fp16")]; + tensor var_7560_cast_fp16 = softmax(axis = var_7387, x = aw_1115_cast_fp16)[name = tensor("op_7560_cast_fp16")]; + tensor var_7561_cast_fp16 = softmax(axis = var_7387, x = aw_1117_cast_fp16)[name = tensor("op_7561_cast_fp16")]; + tensor var_7562_cast_fp16 = softmax(axis = var_7387, x = aw_1119_cast_fp16)[name = tensor("op_7562_cast_fp16")]; + tensor var_7564_equation_0 = const()[name = tensor("op_7564_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7564_cast_fp16 = einsum(equation = var_7564_equation_0, values = (var_7482_cast_fp16_0, var_7543_cast_fp16))[name = tensor("op_7564_cast_fp16")]; + tensor var_7566_equation_0 = const()[name = tensor("op_7566_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7566_cast_fp16 = einsum(equation = var_7566_equation_0, values = (var_7482_cast_fp16_1, var_7544_cast_fp16))[name = tensor("op_7566_cast_fp16")]; + tensor var_7568_equation_0 = const()[name = tensor("op_7568_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7568_cast_fp16 = einsum(equation = var_7568_equation_0, values = (var_7482_cast_fp16_2, var_7545_cast_fp16))[name = tensor("op_7568_cast_fp16")]; + tensor var_7570_equation_0 = const()[name = tensor("op_7570_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7570_cast_fp16 = einsum(equation = var_7570_equation_0, values = (var_7482_cast_fp16_3, var_7546_cast_fp16))[name = tensor("op_7570_cast_fp16")]; + tensor var_7572_equation_0 = const()[name = tensor("op_7572_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7572_cast_fp16 = einsum(equation = var_7572_equation_0, values = (var_7482_cast_fp16_4, var_7547_cast_fp16))[name = tensor("op_7572_cast_fp16")]; + tensor var_7574_equation_0 = const()[name = tensor("op_7574_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7574_cast_fp16 = einsum(equation = var_7574_equation_0, values = (var_7482_cast_fp16_5, var_7548_cast_fp16))[name = tensor("op_7574_cast_fp16")]; + tensor var_7576_equation_0 = const()[name = tensor("op_7576_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7576_cast_fp16 = einsum(equation = var_7576_equation_0, values = (var_7482_cast_fp16_6, var_7549_cast_fp16))[name = tensor("op_7576_cast_fp16")]; + tensor var_7578_equation_0 = const()[name = tensor("op_7578_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7578_cast_fp16 = einsum(equation = var_7578_equation_0, values = (var_7482_cast_fp16_7, var_7550_cast_fp16))[name = tensor("op_7578_cast_fp16")]; + tensor var_7580_equation_0 = const()[name = tensor("op_7580_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7580_cast_fp16 = einsum(equation = var_7580_equation_0, values = (var_7482_cast_fp16_8, var_7551_cast_fp16))[name = tensor("op_7580_cast_fp16")]; + tensor var_7582_equation_0 = const()[name = tensor("op_7582_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7582_cast_fp16 = einsum(equation = var_7582_equation_0, values = (var_7482_cast_fp16_9, var_7552_cast_fp16))[name = tensor("op_7582_cast_fp16")]; + tensor var_7584_equation_0 = const()[name = tensor("op_7584_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7584_cast_fp16 = einsum(equation = var_7584_equation_0, values = (var_7482_cast_fp16_10, var_7553_cast_fp16))[name = tensor("op_7584_cast_fp16")]; + tensor var_7586_equation_0 = const()[name = tensor("op_7586_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7586_cast_fp16 = einsum(equation = var_7586_equation_0, values = (var_7482_cast_fp16_11, var_7554_cast_fp16))[name = tensor("op_7586_cast_fp16")]; + tensor var_7588_equation_0 = const()[name = tensor("op_7588_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7588_cast_fp16 = einsum(equation = var_7588_equation_0, values = (var_7482_cast_fp16_12, var_7555_cast_fp16))[name = tensor("op_7588_cast_fp16")]; + tensor var_7590_equation_0 = const()[name = tensor("op_7590_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7590_cast_fp16 = einsum(equation = var_7590_equation_0, values = (var_7482_cast_fp16_13, var_7556_cast_fp16))[name = tensor("op_7590_cast_fp16")]; + tensor var_7592_equation_0 = const()[name = tensor("op_7592_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7592_cast_fp16 = einsum(equation = var_7592_equation_0, values = (var_7482_cast_fp16_14, var_7557_cast_fp16))[name = tensor("op_7592_cast_fp16")]; + tensor var_7594_equation_0 = const()[name = tensor("op_7594_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7594_cast_fp16 = einsum(equation = var_7594_equation_0, values = (var_7482_cast_fp16_15, var_7558_cast_fp16))[name = tensor("op_7594_cast_fp16")]; + tensor var_7596_equation_0 = const()[name = tensor("op_7596_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7596_cast_fp16 = einsum(equation = var_7596_equation_0, values = (var_7482_cast_fp16_16, var_7559_cast_fp16))[name = tensor("op_7596_cast_fp16")]; + tensor var_7598_equation_0 = const()[name = tensor("op_7598_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7598_cast_fp16 = einsum(equation = var_7598_equation_0, values = (var_7482_cast_fp16_17, var_7560_cast_fp16))[name = tensor("op_7598_cast_fp16")]; + tensor var_7600_equation_0 = const()[name = tensor("op_7600_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7600_cast_fp16 = einsum(equation = var_7600_equation_0, values = (var_7482_cast_fp16_18, var_7561_cast_fp16))[name = tensor("op_7600_cast_fp16")]; + tensor var_7602_equation_0 = const()[name = tensor("op_7602_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7602_cast_fp16 = einsum(equation = var_7602_equation_0, values = (var_7482_cast_fp16_19, var_7562_cast_fp16))[name = tensor("op_7602_cast_fp16")]; + tensor input_275_interleave_0 = const()[name = tensor("input_275_interleave_0"), val = tensor(false)]; + tensor input_275_cast_fp16 = concat(axis = var_7387, interleave = input_275_interleave_0, values = (var_7564_cast_fp16, var_7566_cast_fp16, var_7568_cast_fp16, var_7570_cast_fp16, var_7572_cast_fp16, var_7574_cast_fp16, var_7576_cast_fp16, var_7578_cast_fp16, var_7580_cast_fp16, var_7582_cast_fp16, var_7584_cast_fp16, var_7586_cast_fp16, var_7588_cast_fp16, var_7590_cast_fp16, var_7592_cast_fp16, var_7594_cast_fp16, var_7596_cast_fp16, var_7598_cast_fp16, var_7600_cast_fp16, var_7602_cast_fp16))[name = tensor("input_275_cast_fp16")]; + tensor var_7611_pad_type_0 = const()[name = tensor("op_7611_pad_type_0"), val = tensor("valid")]; + tensor var_7611_strides_0 = const()[name = tensor("op_7611_strides_0"), val = tensor([1, 1])]; + tensor var_7611_pad_0 = const()[name = tensor("op_7611_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7611_dilations_0 = const()[name = tensor("op_7611_dilations_0"), val = tensor([1, 1])]; + tensor var_7611_groups_0 = const()[name = tensor("op_7611_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_out_weight_to_fp16 = const()[name = tensor("blocks_27_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086669952)))]; + tensor blocks_27_attn_out_bias_to_fp16 = const()[name = tensor("blocks_27_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089946816)))]; + tensor var_7611_cast_fp16 = conv(bias = blocks_27_attn_out_bias_to_fp16, dilations = var_7611_dilations_0, groups = var_7611_groups_0, pad = var_7611_pad_0, pad_type = var_7611_pad_type_0, strides = var_7611_strides_0, weight = blocks_27_attn_out_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("op_7611_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = var_7611_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor input_277_axes_0 = const()[name = tensor("input_277_axes_0"), val = tensor([1])]; + tensor input_277_gamma_0_to_fp16 = const()[name = tensor("input_277_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089949440)))]; + tensor input_277_beta_0_to_fp16 = const()[name = tensor("input_277_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089952064)))]; + tensor var_7621_to_fp16 = const()[name = tensor("op_7621_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_277_cast_fp16 = layer_norm(axes = input_277_axes_0, beta = input_277_beta_0_to_fp16, epsilon = var_7621_to_fp16, gamma = input_277_gamma_0_to_fp16, x = inputs_111_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_pad_type_0 = const()[name = tensor("input_279_pad_type_0"), val = tensor("valid")]; + tensor input_279_strides_0 = const()[name = tensor("input_279_strides_0"), val = tensor([1, 1])]; + tensor input_279_pad_0 = const()[name = tensor("input_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_279_dilations_0 = const()[name = tensor("input_279_dilations_0"), val = tensor([1, 1])]; + tensor input_279_groups_0 = const()[name = tensor("input_279_groups_0"), val = tensor(1)]; + tensor blocks_27_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089954688)))]; + tensor blocks_27_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103061952)))]; + tensor input_279_cast_fp16 = conv(bias = blocks_27_mlp_0_bias_to_fp16, dilations = input_279_dilations_0, groups = input_279_groups_0, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = input_279_strides_0, weight = blocks_27_mlp_0_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_mode_0 = const()[name = tensor("input_281_mode_0"), val = tensor("EXACT")]; + tensor input_281_cast_fp16 = gelu(mode = input_281_mode_0, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_7647_pad_type_0 = const()[name = tensor("op_7647_pad_type_0"), val = tensor("valid")]; + tensor var_7647_strides_0 = const()[name = tensor("op_7647_strides_0"), val = tensor([1, 1])]; + tensor var_7647_pad_0 = const()[name = tensor("op_7647_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7647_dilations_0 = const()[name = tensor("op_7647_dilations_0"), val = tensor([1, 1])]; + tensor var_7647_groups_0 = const()[name = tensor("op_7647_groups_0"), val = tensor(1)]; + tensor blocks_27_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103072256)))]; + tensor blocks_27_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116179520)))]; + tensor var_7647_cast_fp16 = conv(bias = blocks_27_mlp_2_bias_to_fp16, dilations = var_7647_dilations_0, groups = var_7647_groups_0, pad = var_7647_pad_0, pad_type = var_7647_pad_type_0, strides = var_7647_strides_0, weight = blocks_27_mlp_2_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("op_7647_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_7647_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_7656 = const()[name = tensor("op_7656"), val = tensor(1)]; + tensor input_283_axes_0 = const()[name = tensor("input_283_axes_0"), val = tensor([1])]; + tensor input_283_gamma_0_to_fp16 = const()[name = tensor("input_283_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116182144)))]; + tensor input_283_beta_0_to_fp16 = const()[name = tensor("input_283_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116184768)))]; + tensor var_7672_to_fp16 = const()[name = tensor("op_7672_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = input_283_beta_0_to_fp16, epsilon = var_7672_to_fp16, gamma = input_283_gamma_0_to_fp16, x = inputs_113_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor q_57_pad_type_0 = const()[name = tensor("q_57_pad_type_0"), val = tensor("valid")]; + tensor q_57_strides_0 = const()[name = tensor("q_57_strides_0"), val = tensor([1, 1])]; + tensor q_57_pad_0 = const()[name = tensor("q_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_57_dilations_0 = const()[name = tensor("q_57_dilations_0"), val = tensor([1, 1])]; + tensor q_57_groups_0 = const()[name = tensor("q_57_groups_0"), val = tensor(1)]; + tensor var_7707_weight_0_to_fp16 = const()[name = tensor("op_7707_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116187392)))]; + tensor var_7707_bias_0_to_fp16 = const()[name = tensor("op_7707_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119464256)))]; + tensor var_7707_cast_fp16 = conv(bias = var_7707_bias_0_to_fp16, dilations = q_57_dilations_0, groups = q_57_groups_0, pad = q_57_pad_0, pad_type = q_57_pad_type_0, strides = q_57_strides_0, weight = var_7707_weight_0_to_fp16, x = input_283_cast_fp16)[name = tensor("op_7707_cast_fp16")]; + tensor k_57_pad_type_0 = const()[name = tensor("k_57_pad_type_0"), val = tensor("valid")]; + tensor k_57_strides_0 = const()[name = tensor("k_57_strides_0"), val = tensor([1, 1])]; + tensor k_57_pad_0 = const()[name = tensor("k_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_57_dilations_0 = const()[name = tensor("k_57_dilations_0"), val = tensor([1, 1])]; + tensor k_57_groups_0 = const()[name = tensor("k_57_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_key_weight_to_fp16 = const()[name = tensor("blocks_28_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119466880)))]; + tensor k_57_cast_fp16 = conv(dilations = k_57_dilations_0, groups = k_57_groups_0, pad = k_57_pad_0, pad_type = k_57_pad_type_0, strides = k_57_strides_0, weight = blocks_28_attn_key_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor var_7705_pad_type_0 = const()[name = tensor("op_7705_pad_type_0"), val = tensor("valid")]; + tensor var_7705_strides_0 = const()[name = tensor("op_7705_strides_0"), val = tensor([1, 1])]; + tensor var_7705_pad_0 = const()[name = tensor("op_7705_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7705_dilations_0 = const()[name = tensor("op_7705_dilations_0"), val = tensor([1, 1])]; + tensor var_7705_groups_0 = const()[name = tensor("op_7705_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_value_weight_to_fp16 = const()[name = tensor("blocks_28_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1122743744)))]; + tensor blocks_28_attn_value_bias_to_fp16 = const()[name = tensor("blocks_28_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126020608)))]; + tensor var_7705_cast_fp16 = conv(bias = blocks_28_attn_value_bias_to_fp16, dilations = var_7705_dilations_0, groups = var_7705_groups_0, pad = var_7705_pad_0, pad_type = var_7705_pad_type_0, strides = var_7705_strides_0, weight = blocks_28_attn_value_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("op_7705_cast_fp16")]; + tensor tile_84 = const()[name = tensor("tile_84"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7708_axis_0 = const()[name = tensor("op_7708_axis_0"), val = tensor(1)]; + tensor var_7708_cast_fp16_0, tensor var_7708_cast_fp16_1, tensor var_7708_cast_fp16_2, tensor var_7708_cast_fp16_3, tensor var_7708_cast_fp16_4, tensor var_7708_cast_fp16_5, tensor var_7708_cast_fp16_6, tensor var_7708_cast_fp16_7, tensor var_7708_cast_fp16_8, tensor var_7708_cast_fp16_9, tensor var_7708_cast_fp16_10, tensor var_7708_cast_fp16_11, tensor var_7708_cast_fp16_12, tensor var_7708_cast_fp16_13, tensor var_7708_cast_fp16_14, tensor var_7708_cast_fp16_15, tensor var_7708_cast_fp16_16, tensor var_7708_cast_fp16_17, tensor var_7708_cast_fp16_18, tensor var_7708_cast_fp16_19 = split(axis = var_7708_axis_0, split_sizes = tile_84, x = var_7707_cast_fp16)[name = tensor("op_7708_cast_fp16")]; + tensor var_7729_perm_0 = const()[name = tensor("op_7729_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_85 = const()[name = tensor("tile_85"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7730_axis_0 = const()[name = tensor("op_7730_axis_0"), val = tensor(3)]; + tensor var_7729_cast_fp16 = transpose(perm = var_7729_perm_0, x = k_57_cast_fp16)[name = tensor("transpose_4")]; + tensor var_7730_cast_fp16_0, tensor var_7730_cast_fp16_1, tensor var_7730_cast_fp16_2, tensor var_7730_cast_fp16_3, tensor var_7730_cast_fp16_4, tensor var_7730_cast_fp16_5, tensor var_7730_cast_fp16_6, tensor var_7730_cast_fp16_7, tensor var_7730_cast_fp16_8, tensor var_7730_cast_fp16_9, tensor var_7730_cast_fp16_10, tensor var_7730_cast_fp16_11, tensor var_7730_cast_fp16_12, tensor var_7730_cast_fp16_13, tensor var_7730_cast_fp16_14, tensor var_7730_cast_fp16_15, tensor var_7730_cast_fp16_16, tensor var_7730_cast_fp16_17, tensor var_7730_cast_fp16_18, tensor var_7730_cast_fp16_19 = split(axis = var_7730_axis_0, split_sizes = tile_85, x = var_7729_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor tile_86 = const()[name = tensor("tile_86"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7751_axis_0 = const()[name = tensor("op_7751_axis_0"), val = tensor(1)]; + tensor var_7751_cast_fp16_0, tensor var_7751_cast_fp16_1, tensor var_7751_cast_fp16_2, tensor var_7751_cast_fp16_3, tensor var_7751_cast_fp16_4, tensor var_7751_cast_fp16_5, tensor var_7751_cast_fp16_6, tensor var_7751_cast_fp16_7, tensor var_7751_cast_fp16_8, tensor var_7751_cast_fp16_9, tensor var_7751_cast_fp16_10, tensor var_7751_cast_fp16_11, tensor var_7751_cast_fp16_12, tensor var_7751_cast_fp16_13, tensor var_7751_cast_fp16_14, tensor var_7751_cast_fp16_15, tensor var_7751_cast_fp16_16, tensor var_7751_cast_fp16_17, tensor var_7751_cast_fp16_18, tensor var_7751_cast_fp16_19 = split(axis = var_7751_axis_0, split_sizes = tile_86, x = var_7705_cast_fp16)[name = tensor("op_7751_cast_fp16")]; + tensor aw_1121_equation_0 = const()[name = tensor("aw_1121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1121_cast_fp16 = einsum(equation = aw_1121_equation_0, values = (var_7730_cast_fp16_0, var_7708_cast_fp16_0))[name = tensor("aw_1121_cast_fp16")]; + tensor aw_1123_equation_0 = const()[name = tensor("aw_1123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1123_cast_fp16 = einsum(equation = aw_1123_equation_0, values = (var_7730_cast_fp16_1, var_7708_cast_fp16_1))[name = tensor("aw_1123_cast_fp16")]; + tensor aw_1125_equation_0 = const()[name = tensor("aw_1125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1125_cast_fp16 = einsum(equation = aw_1125_equation_0, values = (var_7730_cast_fp16_2, var_7708_cast_fp16_2))[name = tensor("aw_1125_cast_fp16")]; + tensor aw_1127_equation_0 = const()[name = tensor("aw_1127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1127_cast_fp16 = einsum(equation = aw_1127_equation_0, values = (var_7730_cast_fp16_3, var_7708_cast_fp16_3))[name = tensor("aw_1127_cast_fp16")]; + tensor aw_1129_equation_0 = const()[name = tensor("aw_1129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1129_cast_fp16 = einsum(equation = aw_1129_equation_0, values = (var_7730_cast_fp16_4, var_7708_cast_fp16_4))[name = tensor("aw_1129_cast_fp16")]; + tensor aw_1131_equation_0 = const()[name = tensor("aw_1131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1131_cast_fp16 = einsum(equation = aw_1131_equation_0, values = (var_7730_cast_fp16_5, var_7708_cast_fp16_5))[name = tensor("aw_1131_cast_fp16")]; + tensor aw_1133_equation_0 = const()[name = tensor("aw_1133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1133_cast_fp16 = einsum(equation = aw_1133_equation_0, values = (var_7730_cast_fp16_6, var_7708_cast_fp16_6))[name = tensor("aw_1133_cast_fp16")]; + tensor aw_1135_equation_0 = const()[name = tensor("aw_1135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1135_cast_fp16 = einsum(equation = aw_1135_equation_0, values = (var_7730_cast_fp16_7, var_7708_cast_fp16_7))[name = tensor("aw_1135_cast_fp16")]; + tensor aw_1137_equation_0 = const()[name = tensor("aw_1137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1137_cast_fp16 = einsum(equation = aw_1137_equation_0, values = (var_7730_cast_fp16_8, var_7708_cast_fp16_8))[name = tensor("aw_1137_cast_fp16")]; + tensor aw_1139_equation_0 = const()[name = tensor("aw_1139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1139_cast_fp16 = einsum(equation = aw_1139_equation_0, values = (var_7730_cast_fp16_9, var_7708_cast_fp16_9))[name = tensor("aw_1139_cast_fp16")]; + tensor aw_1141_equation_0 = const()[name = tensor("aw_1141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1141_cast_fp16 = einsum(equation = aw_1141_equation_0, values = (var_7730_cast_fp16_10, var_7708_cast_fp16_10))[name = tensor("aw_1141_cast_fp16")]; + tensor aw_1143_equation_0 = const()[name = tensor("aw_1143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1143_cast_fp16 = einsum(equation = aw_1143_equation_0, values = (var_7730_cast_fp16_11, var_7708_cast_fp16_11))[name = tensor("aw_1143_cast_fp16")]; + tensor aw_1145_equation_0 = const()[name = tensor("aw_1145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1145_cast_fp16 = einsum(equation = aw_1145_equation_0, values = (var_7730_cast_fp16_12, var_7708_cast_fp16_12))[name = tensor("aw_1145_cast_fp16")]; + tensor aw_1147_equation_0 = const()[name = tensor("aw_1147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1147_cast_fp16 = einsum(equation = aw_1147_equation_0, values = (var_7730_cast_fp16_13, var_7708_cast_fp16_13))[name = tensor("aw_1147_cast_fp16")]; + tensor aw_1149_equation_0 = const()[name = tensor("aw_1149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1149_cast_fp16 = einsum(equation = aw_1149_equation_0, values = (var_7730_cast_fp16_14, var_7708_cast_fp16_14))[name = tensor("aw_1149_cast_fp16")]; + tensor aw_1151_equation_0 = const()[name = tensor("aw_1151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1151_cast_fp16 = einsum(equation = aw_1151_equation_0, values = (var_7730_cast_fp16_15, var_7708_cast_fp16_15))[name = tensor("aw_1151_cast_fp16")]; + tensor aw_1153_equation_0 = const()[name = tensor("aw_1153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1153_cast_fp16 = einsum(equation = aw_1153_equation_0, values = (var_7730_cast_fp16_16, var_7708_cast_fp16_16))[name = tensor("aw_1153_cast_fp16")]; + tensor aw_1155_equation_0 = const()[name = tensor("aw_1155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1155_cast_fp16 = einsum(equation = aw_1155_equation_0, values = (var_7730_cast_fp16_17, var_7708_cast_fp16_17))[name = tensor("aw_1155_cast_fp16")]; + tensor aw_1157_equation_0 = const()[name = tensor("aw_1157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1157_cast_fp16 = einsum(equation = aw_1157_equation_0, values = (var_7730_cast_fp16_18, var_7708_cast_fp16_18))[name = tensor("aw_1157_cast_fp16")]; + tensor aw_1159_equation_0 = const()[name = tensor("aw_1159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1159_cast_fp16 = einsum(equation = aw_1159_equation_0, values = (var_7730_cast_fp16_19, var_7708_cast_fp16_19))[name = tensor("aw_1159_cast_fp16")]; + tensor var_7812_cast_fp16 = softmax(axis = var_7656, x = aw_1121_cast_fp16)[name = tensor("op_7812_cast_fp16")]; + tensor var_7813_cast_fp16 = softmax(axis = var_7656, x = aw_1123_cast_fp16)[name = tensor("op_7813_cast_fp16")]; + tensor var_7814_cast_fp16 = softmax(axis = var_7656, x = aw_1125_cast_fp16)[name = tensor("op_7814_cast_fp16")]; + tensor var_7815_cast_fp16 = softmax(axis = var_7656, x = aw_1127_cast_fp16)[name = tensor("op_7815_cast_fp16")]; + tensor var_7816_cast_fp16 = softmax(axis = var_7656, x = aw_1129_cast_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor var_7817_cast_fp16 = softmax(axis = var_7656, x = aw_1131_cast_fp16)[name = tensor("op_7817_cast_fp16")]; + tensor var_7818_cast_fp16 = softmax(axis = var_7656, x = aw_1133_cast_fp16)[name = tensor("op_7818_cast_fp16")]; + tensor var_7819_cast_fp16 = softmax(axis = var_7656, x = aw_1135_cast_fp16)[name = tensor("op_7819_cast_fp16")]; + tensor var_7820_cast_fp16 = softmax(axis = var_7656, x = aw_1137_cast_fp16)[name = tensor("op_7820_cast_fp16")]; + tensor var_7821_cast_fp16 = softmax(axis = var_7656, x = aw_1139_cast_fp16)[name = tensor("op_7821_cast_fp16")]; + tensor var_7822_cast_fp16 = softmax(axis = var_7656, x = aw_1141_cast_fp16)[name = tensor("op_7822_cast_fp16")]; + tensor var_7823_cast_fp16 = softmax(axis = var_7656, x = aw_1143_cast_fp16)[name = tensor("op_7823_cast_fp16")]; + tensor var_7824_cast_fp16 = softmax(axis = var_7656, x = aw_1145_cast_fp16)[name = tensor("op_7824_cast_fp16")]; + tensor var_7825_cast_fp16 = softmax(axis = var_7656, x = aw_1147_cast_fp16)[name = tensor("op_7825_cast_fp16")]; + tensor var_7826_cast_fp16 = softmax(axis = var_7656, x = aw_1149_cast_fp16)[name = tensor("op_7826_cast_fp16")]; + tensor var_7827_cast_fp16 = softmax(axis = var_7656, x = aw_1151_cast_fp16)[name = tensor("op_7827_cast_fp16")]; + tensor var_7828_cast_fp16 = softmax(axis = var_7656, x = aw_1153_cast_fp16)[name = tensor("op_7828_cast_fp16")]; + tensor var_7829_cast_fp16 = softmax(axis = var_7656, x = aw_1155_cast_fp16)[name = tensor("op_7829_cast_fp16")]; + tensor var_7830_cast_fp16 = softmax(axis = var_7656, x = aw_1157_cast_fp16)[name = tensor("op_7830_cast_fp16")]; + tensor var_7831_cast_fp16 = softmax(axis = var_7656, x = aw_1159_cast_fp16)[name = tensor("op_7831_cast_fp16")]; + tensor var_7833_equation_0 = const()[name = tensor("op_7833_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7833_cast_fp16 = einsum(equation = var_7833_equation_0, values = (var_7751_cast_fp16_0, var_7812_cast_fp16))[name = tensor("op_7833_cast_fp16")]; + tensor var_7835_equation_0 = const()[name = tensor("op_7835_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7835_cast_fp16 = einsum(equation = var_7835_equation_0, values = (var_7751_cast_fp16_1, var_7813_cast_fp16))[name = tensor("op_7835_cast_fp16")]; + tensor var_7837_equation_0 = const()[name = tensor("op_7837_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7837_cast_fp16 = einsum(equation = var_7837_equation_0, values = (var_7751_cast_fp16_2, var_7814_cast_fp16))[name = tensor("op_7837_cast_fp16")]; + tensor var_7839_equation_0 = const()[name = tensor("op_7839_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7839_cast_fp16 = einsum(equation = var_7839_equation_0, values = (var_7751_cast_fp16_3, var_7815_cast_fp16))[name = tensor("op_7839_cast_fp16")]; + tensor var_7841_equation_0 = const()[name = tensor("op_7841_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7841_cast_fp16 = einsum(equation = var_7841_equation_0, values = (var_7751_cast_fp16_4, var_7816_cast_fp16))[name = tensor("op_7841_cast_fp16")]; + tensor var_7843_equation_0 = const()[name = tensor("op_7843_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7843_cast_fp16 = einsum(equation = var_7843_equation_0, values = (var_7751_cast_fp16_5, var_7817_cast_fp16))[name = tensor("op_7843_cast_fp16")]; + tensor var_7845_equation_0 = const()[name = tensor("op_7845_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7845_cast_fp16 = einsum(equation = var_7845_equation_0, values = (var_7751_cast_fp16_6, var_7818_cast_fp16))[name = tensor("op_7845_cast_fp16")]; + tensor var_7847_equation_0 = const()[name = tensor("op_7847_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7847_cast_fp16 = einsum(equation = var_7847_equation_0, values = (var_7751_cast_fp16_7, var_7819_cast_fp16))[name = tensor("op_7847_cast_fp16")]; + tensor var_7849_equation_0 = const()[name = tensor("op_7849_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7849_cast_fp16 = einsum(equation = var_7849_equation_0, values = (var_7751_cast_fp16_8, var_7820_cast_fp16))[name = tensor("op_7849_cast_fp16")]; + tensor var_7851_equation_0 = const()[name = tensor("op_7851_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7851_cast_fp16 = einsum(equation = var_7851_equation_0, values = (var_7751_cast_fp16_9, var_7821_cast_fp16))[name = tensor("op_7851_cast_fp16")]; + tensor var_7853_equation_0 = const()[name = tensor("op_7853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7853_cast_fp16 = einsum(equation = var_7853_equation_0, values = (var_7751_cast_fp16_10, var_7822_cast_fp16))[name = tensor("op_7853_cast_fp16")]; + tensor var_7855_equation_0 = const()[name = tensor("op_7855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7855_cast_fp16 = einsum(equation = var_7855_equation_0, values = (var_7751_cast_fp16_11, var_7823_cast_fp16))[name = tensor("op_7855_cast_fp16")]; + tensor var_7857_equation_0 = const()[name = tensor("op_7857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7857_cast_fp16 = einsum(equation = var_7857_equation_0, values = (var_7751_cast_fp16_12, var_7824_cast_fp16))[name = tensor("op_7857_cast_fp16")]; + tensor var_7859_equation_0 = const()[name = tensor("op_7859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7859_cast_fp16 = einsum(equation = var_7859_equation_0, values = (var_7751_cast_fp16_13, var_7825_cast_fp16))[name = tensor("op_7859_cast_fp16")]; + tensor var_7861_equation_0 = const()[name = tensor("op_7861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7861_cast_fp16 = einsum(equation = var_7861_equation_0, values = (var_7751_cast_fp16_14, var_7826_cast_fp16))[name = tensor("op_7861_cast_fp16")]; + tensor var_7863_equation_0 = const()[name = tensor("op_7863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7863_cast_fp16 = einsum(equation = var_7863_equation_0, values = (var_7751_cast_fp16_15, var_7827_cast_fp16))[name = tensor("op_7863_cast_fp16")]; + tensor var_7865_equation_0 = const()[name = tensor("op_7865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7865_cast_fp16 = einsum(equation = var_7865_equation_0, values = (var_7751_cast_fp16_16, var_7828_cast_fp16))[name = tensor("op_7865_cast_fp16")]; + tensor var_7867_equation_0 = const()[name = tensor("op_7867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7867_cast_fp16 = einsum(equation = var_7867_equation_0, values = (var_7751_cast_fp16_17, var_7829_cast_fp16))[name = tensor("op_7867_cast_fp16")]; + tensor var_7869_equation_0 = const()[name = tensor("op_7869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7869_cast_fp16 = einsum(equation = var_7869_equation_0, values = (var_7751_cast_fp16_18, var_7830_cast_fp16))[name = tensor("op_7869_cast_fp16")]; + tensor var_7871_equation_0 = const()[name = tensor("op_7871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7871_cast_fp16 = einsum(equation = var_7871_equation_0, values = (var_7751_cast_fp16_19, var_7831_cast_fp16))[name = tensor("op_7871_cast_fp16")]; + tensor input_285_interleave_0 = const()[name = tensor("input_285_interleave_0"), val = tensor(false)]; + tensor input_285_cast_fp16 = concat(axis = var_7656, interleave = input_285_interleave_0, values = (var_7833_cast_fp16, var_7835_cast_fp16, var_7837_cast_fp16, var_7839_cast_fp16, var_7841_cast_fp16, var_7843_cast_fp16, var_7845_cast_fp16, var_7847_cast_fp16, var_7849_cast_fp16, var_7851_cast_fp16, var_7853_cast_fp16, var_7855_cast_fp16, var_7857_cast_fp16, var_7859_cast_fp16, var_7861_cast_fp16, var_7863_cast_fp16, var_7865_cast_fp16, var_7867_cast_fp16, var_7869_cast_fp16, var_7871_cast_fp16))[name = tensor("input_285_cast_fp16")]; + tensor var_7880_pad_type_0 = const()[name = tensor("op_7880_pad_type_0"), val = tensor("valid")]; + tensor var_7880_strides_0 = const()[name = tensor("op_7880_strides_0"), val = tensor([1, 1])]; + tensor var_7880_pad_0 = const()[name = tensor("op_7880_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7880_dilations_0 = const()[name = tensor("op_7880_dilations_0"), val = tensor([1, 1])]; + tensor var_7880_groups_0 = const()[name = tensor("op_7880_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_out_weight_to_fp16 = const()[name = tensor("blocks_28_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126023232)))]; + tensor blocks_28_attn_out_bias_to_fp16 = const()[name = tensor("blocks_28_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129300096)))]; + tensor var_7880_cast_fp16 = conv(bias = blocks_28_attn_out_bias_to_fp16, dilations = var_7880_dilations_0, groups = var_7880_groups_0, pad = var_7880_pad_0, pad_type = var_7880_pad_type_0, strides = var_7880_strides_0, weight = blocks_28_attn_out_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("op_7880_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = var_7880_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([1])]; + tensor input_287_gamma_0_to_fp16 = const()[name = tensor("input_287_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129302720)))]; + tensor input_287_beta_0_to_fp16 = const()[name = tensor("input_287_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129305344)))]; + tensor var_7890_to_fp16 = const()[name = tensor("op_7890_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = input_287_beta_0_to_fp16, epsilon = var_7890_to_fp16, gamma = input_287_gamma_0_to_fp16, x = inputs_115_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_pad_type_0 = const()[name = tensor("input_289_pad_type_0"), val = tensor("valid")]; + tensor input_289_strides_0 = const()[name = tensor("input_289_strides_0"), val = tensor([1, 1])]; + tensor input_289_pad_0 = const()[name = tensor("input_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_289_dilations_0 = const()[name = tensor("input_289_dilations_0"), val = tensor([1, 1])]; + tensor input_289_groups_0 = const()[name = tensor("input_289_groups_0"), val = tensor(1)]; + tensor blocks_28_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129307968)))]; + tensor blocks_28_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142415232)))]; + tensor input_289_cast_fp16 = conv(bias = blocks_28_mlp_0_bias_to_fp16, dilations = input_289_dilations_0, groups = input_289_groups_0, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = input_289_strides_0, weight = blocks_28_mlp_0_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor input_291_mode_0 = const()[name = tensor("input_291_mode_0"), val = tensor("EXACT")]; + tensor input_291_cast_fp16 = gelu(mode = input_291_mode_0, x = input_289_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor var_7916_pad_type_0 = const()[name = tensor("op_7916_pad_type_0"), val = tensor("valid")]; + tensor var_7916_strides_0 = const()[name = tensor("op_7916_strides_0"), val = tensor([1, 1])]; + tensor var_7916_pad_0 = const()[name = tensor("op_7916_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7916_dilations_0 = const()[name = tensor("op_7916_dilations_0"), val = tensor([1, 1])]; + tensor var_7916_groups_0 = const()[name = tensor("op_7916_groups_0"), val = tensor(1)]; + tensor blocks_28_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142425536)))]; + tensor blocks_28_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155532800)))]; + tensor var_7916_cast_fp16 = conv(bias = blocks_28_mlp_2_bias_to_fp16, dilations = var_7916_dilations_0, groups = var_7916_groups_0, pad = var_7916_pad_0, pad_type = var_7916_pad_type_0, strides = var_7916_strides_0, weight = blocks_28_mlp_2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("op_7916_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = var_7916_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_7925 = const()[name = tensor("op_7925"), val = tensor(1)]; + tensor input_293_axes_0 = const()[name = tensor("input_293_axes_0"), val = tensor([1])]; + tensor input_293_gamma_0_to_fp16 = const()[name = tensor("input_293_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155535424)))]; + tensor input_293_beta_0_to_fp16 = const()[name = tensor("input_293_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155538048)))]; + tensor var_7941_to_fp16 = const()[name = tensor("op_7941_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = input_293_beta_0_to_fp16, epsilon = var_7941_to_fp16, gamma = input_293_gamma_0_to_fp16, x = inputs_117_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor q_59_pad_type_0 = const()[name = tensor("q_59_pad_type_0"), val = tensor("valid")]; + tensor q_59_strides_0 = const()[name = tensor("q_59_strides_0"), val = tensor([1, 1])]; + tensor q_59_pad_0 = const()[name = tensor("q_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_59_dilations_0 = const()[name = tensor("q_59_dilations_0"), val = tensor([1, 1])]; + tensor q_59_groups_0 = const()[name = tensor("q_59_groups_0"), val = tensor(1)]; + tensor var_7976_weight_0_to_fp16 = const()[name = tensor("op_7976_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155540672)))]; + tensor var_7976_bias_0_to_fp16 = const()[name = tensor("op_7976_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158817536)))]; + tensor var_7976_cast_fp16 = conv(bias = var_7976_bias_0_to_fp16, dilations = q_59_dilations_0, groups = q_59_groups_0, pad = q_59_pad_0, pad_type = q_59_pad_type_0, strides = q_59_strides_0, weight = var_7976_weight_0_to_fp16, x = input_293_cast_fp16)[name = tensor("op_7976_cast_fp16")]; + tensor k_59_pad_type_0 = const()[name = tensor("k_59_pad_type_0"), val = tensor("valid")]; + tensor k_59_strides_0 = const()[name = tensor("k_59_strides_0"), val = tensor([1, 1])]; + tensor k_59_pad_0 = const()[name = tensor("k_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_59_dilations_0 = const()[name = tensor("k_59_dilations_0"), val = tensor([1, 1])]; + tensor k_59_groups_0 = const()[name = tensor("k_59_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_key_weight_to_fp16 = const()[name = tensor("blocks_29_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158820160)))]; + tensor k_59_cast_fp16 = conv(dilations = k_59_dilations_0, groups = k_59_groups_0, pad = k_59_pad_0, pad_type = k_59_pad_type_0, strides = k_59_strides_0, weight = blocks_29_attn_key_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor var_7974_pad_type_0 = const()[name = tensor("op_7974_pad_type_0"), val = tensor("valid")]; + tensor var_7974_strides_0 = const()[name = tensor("op_7974_strides_0"), val = tensor([1, 1])]; + tensor var_7974_pad_0 = const()[name = tensor("op_7974_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7974_dilations_0 = const()[name = tensor("op_7974_dilations_0"), val = tensor([1, 1])]; + tensor var_7974_groups_0 = const()[name = tensor("op_7974_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_value_weight_to_fp16 = const()[name = tensor("blocks_29_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162097024)))]; + tensor blocks_29_attn_value_bias_to_fp16 = const()[name = tensor("blocks_29_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165373888)))]; + tensor var_7974_cast_fp16 = conv(bias = blocks_29_attn_value_bias_to_fp16, dilations = var_7974_dilations_0, groups = var_7974_groups_0, pad = var_7974_pad_0, pad_type = var_7974_pad_type_0, strides = var_7974_strides_0, weight = blocks_29_attn_value_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("op_7974_cast_fp16")]; + tensor tile_87 = const()[name = tensor("tile_87"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7977_axis_0 = const()[name = tensor("op_7977_axis_0"), val = tensor(1)]; + tensor var_7977_cast_fp16_0, tensor var_7977_cast_fp16_1, tensor var_7977_cast_fp16_2, tensor var_7977_cast_fp16_3, tensor var_7977_cast_fp16_4, tensor var_7977_cast_fp16_5, tensor var_7977_cast_fp16_6, tensor var_7977_cast_fp16_7, tensor var_7977_cast_fp16_8, tensor var_7977_cast_fp16_9, tensor var_7977_cast_fp16_10, tensor var_7977_cast_fp16_11, tensor var_7977_cast_fp16_12, tensor var_7977_cast_fp16_13, tensor var_7977_cast_fp16_14, tensor var_7977_cast_fp16_15, tensor var_7977_cast_fp16_16, tensor var_7977_cast_fp16_17, tensor var_7977_cast_fp16_18, tensor var_7977_cast_fp16_19 = split(axis = var_7977_axis_0, split_sizes = tile_87, x = var_7976_cast_fp16)[name = tensor("op_7977_cast_fp16")]; + tensor var_7998_perm_0 = const()[name = tensor("op_7998_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_88 = const()[name = tensor("tile_88"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7999_axis_0 = const()[name = tensor("op_7999_axis_0"), val = tensor(3)]; + tensor var_7998_cast_fp16 = transpose(perm = var_7998_perm_0, x = k_59_cast_fp16)[name = tensor("transpose_3")]; + tensor var_7999_cast_fp16_0, tensor var_7999_cast_fp16_1, tensor var_7999_cast_fp16_2, tensor var_7999_cast_fp16_3, tensor var_7999_cast_fp16_4, tensor var_7999_cast_fp16_5, tensor var_7999_cast_fp16_6, tensor var_7999_cast_fp16_7, tensor var_7999_cast_fp16_8, tensor var_7999_cast_fp16_9, tensor var_7999_cast_fp16_10, tensor var_7999_cast_fp16_11, tensor var_7999_cast_fp16_12, tensor var_7999_cast_fp16_13, tensor var_7999_cast_fp16_14, tensor var_7999_cast_fp16_15, tensor var_7999_cast_fp16_16, tensor var_7999_cast_fp16_17, tensor var_7999_cast_fp16_18, tensor var_7999_cast_fp16_19 = split(axis = var_7999_axis_0, split_sizes = tile_88, x = var_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor tile_89 = const()[name = tensor("tile_89"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8020_axis_0 = const()[name = tensor("op_8020_axis_0"), val = tensor(1)]; + tensor var_8020_cast_fp16_0, tensor var_8020_cast_fp16_1, tensor var_8020_cast_fp16_2, tensor var_8020_cast_fp16_3, tensor var_8020_cast_fp16_4, tensor var_8020_cast_fp16_5, tensor var_8020_cast_fp16_6, tensor var_8020_cast_fp16_7, tensor var_8020_cast_fp16_8, tensor var_8020_cast_fp16_9, tensor var_8020_cast_fp16_10, tensor var_8020_cast_fp16_11, tensor var_8020_cast_fp16_12, tensor var_8020_cast_fp16_13, tensor var_8020_cast_fp16_14, tensor var_8020_cast_fp16_15, tensor var_8020_cast_fp16_16, tensor var_8020_cast_fp16_17, tensor var_8020_cast_fp16_18, tensor var_8020_cast_fp16_19 = split(axis = var_8020_axis_0, split_sizes = tile_89, x = var_7974_cast_fp16)[name = tensor("op_8020_cast_fp16")]; + tensor aw_1161_equation_0 = const()[name = tensor("aw_1161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1161_cast_fp16 = einsum(equation = aw_1161_equation_0, values = (var_7999_cast_fp16_0, var_7977_cast_fp16_0))[name = tensor("aw_1161_cast_fp16")]; + tensor aw_1163_equation_0 = const()[name = tensor("aw_1163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1163_cast_fp16 = einsum(equation = aw_1163_equation_0, values = (var_7999_cast_fp16_1, var_7977_cast_fp16_1))[name = tensor("aw_1163_cast_fp16")]; + tensor aw_1165_equation_0 = const()[name = tensor("aw_1165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1165_cast_fp16 = einsum(equation = aw_1165_equation_0, values = (var_7999_cast_fp16_2, var_7977_cast_fp16_2))[name = tensor("aw_1165_cast_fp16")]; + tensor aw_1167_equation_0 = const()[name = tensor("aw_1167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1167_cast_fp16 = einsum(equation = aw_1167_equation_0, values = (var_7999_cast_fp16_3, var_7977_cast_fp16_3))[name = tensor("aw_1167_cast_fp16")]; + tensor aw_1169_equation_0 = const()[name = tensor("aw_1169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1169_cast_fp16 = einsum(equation = aw_1169_equation_0, values = (var_7999_cast_fp16_4, var_7977_cast_fp16_4))[name = tensor("aw_1169_cast_fp16")]; + tensor aw_1171_equation_0 = const()[name = tensor("aw_1171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1171_cast_fp16 = einsum(equation = aw_1171_equation_0, values = (var_7999_cast_fp16_5, var_7977_cast_fp16_5))[name = tensor("aw_1171_cast_fp16")]; + tensor aw_1173_equation_0 = const()[name = tensor("aw_1173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1173_cast_fp16 = einsum(equation = aw_1173_equation_0, values = (var_7999_cast_fp16_6, var_7977_cast_fp16_6))[name = tensor("aw_1173_cast_fp16")]; + tensor aw_1175_equation_0 = const()[name = tensor("aw_1175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1175_cast_fp16 = einsum(equation = aw_1175_equation_0, values = (var_7999_cast_fp16_7, var_7977_cast_fp16_7))[name = tensor("aw_1175_cast_fp16")]; + tensor aw_1177_equation_0 = const()[name = tensor("aw_1177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1177_cast_fp16 = einsum(equation = aw_1177_equation_0, values = (var_7999_cast_fp16_8, var_7977_cast_fp16_8))[name = tensor("aw_1177_cast_fp16")]; + tensor aw_1179_equation_0 = const()[name = tensor("aw_1179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1179_cast_fp16 = einsum(equation = aw_1179_equation_0, values = (var_7999_cast_fp16_9, var_7977_cast_fp16_9))[name = tensor("aw_1179_cast_fp16")]; + tensor aw_1181_equation_0 = const()[name = tensor("aw_1181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1181_cast_fp16 = einsum(equation = aw_1181_equation_0, values = (var_7999_cast_fp16_10, var_7977_cast_fp16_10))[name = tensor("aw_1181_cast_fp16")]; + tensor aw_1183_equation_0 = const()[name = tensor("aw_1183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1183_cast_fp16 = einsum(equation = aw_1183_equation_0, values = (var_7999_cast_fp16_11, var_7977_cast_fp16_11))[name = tensor("aw_1183_cast_fp16")]; + tensor aw_1185_equation_0 = const()[name = tensor("aw_1185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1185_cast_fp16 = einsum(equation = aw_1185_equation_0, values = (var_7999_cast_fp16_12, var_7977_cast_fp16_12))[name = tensor("aw_1185_cast_fp16")]; + tensor aw_1187_equation_0 = const()[name = tensor("aw_1187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1187_cast_fp16 = einsum(equation = aw_1187_equation_0, values = (var_7999_cast_fp16_13, var_7977_cast_fp16_13))[name = tensor("aw_1187_cast_fp16")]; + tensor aw_1189_equation_0 = const()[name = tensor("aw_1189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1189_cast_fp16 = einsum(equation = aw_1189_equation_0, values = (var_7999_cast_fp16_14, var_7977_cast_fp16_14))[name = tensor("aw_1189_cast_fp16")]; + tensor aw_1191_equation_0 = const()[name = tensor("aw_1191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1191_cast_fp16 = einsum(equation = aw_1191_equation_0, values = (var_7999_cast_fp16_15, var_7977_cast_fp16_15))[name = tensor("aw_1191_cast_fp16")]; + tensor aw_1193_equation_0 = const()[name = tensor("aw_1193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1193_cast_fp16 = einsum(equation = aw_1193_equation_0, values = (var_7999_cast_fp16_16, var_7977_cast_fp16_16))[name = tensor("aw_1193_cast_fp16")]; + tensor aw_1195_equation_0 = const()[name = tensor("aw_1195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1195_cast_fp16 = einsum(equation = aw_1195_equation_0, values = (var_7999_cast_fp16_17, var_7977_cast_fp16_17))[name = tensor("aw_1195_cast_fp16")]; + tensor aw_1197_equation_0 = const()[name = tensor("aw_1197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1197_cast_fp16 = einsum(equation = aw_1197_equation_0, values = (var_7999_cast_fp16_18, var_7977_cast_fp16_18))[name = tensor("aw_1197_cast_fp16")]; + tensor aw_1199_equation_0 = const()[name = tensor("aw_1199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1199_cast_fp16 = einsum(equation = aw_1199_equation_0, values = (var_7999_cast_fp16_19, var_7977_cast_fp16_19))[name = tensor("aw_1199_cast_fp16")]; + tensor var_8081_cast_fp16 = softmax(axis = var_7925, x = aw_1161_cast_fp16)[name = tensor("op_8081_cast_fp16")]; + tensor var_8082_cast_fp16 = softmax(axis = var_7925, x = aw_1163_cast_fp16)[name = tensor("op_8082_cast_fp16")]; + tensor var_8083_cast_fp16 = softmax(axis = var_7925, x = aw_1165_cast_fp16)[name = tensor("op_8083_cast_fp16")]; + tensor var_8084_cast_fp16 = softmax(axis = var_7925, x = aw_1167_cast_fp16)[name = tensor("op_8084_cast_fp16")]; + tensor var_8085_cast_fp16 = softmax(axis = var_7925, x = aw_1169_cast_fp16)[name = tensor("op_8085_cast_fp16")]; + tensor var_8086_cast_fp16 = softmax(axis = var_7925, x = aw_1171_cast_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8087_cast_fp16 = softmax(axis = var_7925, x = aw_1173_cast_fp16)[name = tensor("op_8087_cast_fp16")]; + tensor var_8088_cast_fp16 = softmax(axis = var_7925, x = aw_1175_cast_fp16)[name = tensor("op_8088_cast_fp16")]; + tensor var_8089_cast_fp16 = softmax(axis = var_7925, x = aw_1177_cast_fp16)[name = tensor("op_8089_cast_fp16")]; + tensor var_8090_cast_fp16 = softmax(axis = var_7925, x = aw_1179_cast_fp16)[name = tensor("op_8090_cast_fp16")]; + tensor var_8091_cast_fp16 = softmax(axis = var_7925, x = aw_1181_cast_fp16)[name = tensor("op_8091_cast_fp16")]; + tensor var_8092_cast_fp16 = softmax(axis = var_7925, x = aw_1183_cast_fp16)[name = tensor("op_8092_cast_fp16")]; + tensor var_8093_cast_fp16 = softmax(axis = var_7925, x = aw_1185_cast_fp16)[name = tensor("op_8093_cast_fp16")]; + tensor var_8094_cast_fp16 = softmax(axis = var_7925, x = aw_1187_cast_fp16)[name = tensor("op_8094_cast_fp16")]; + tensor var_8095_cast_fp16 = softmax(axis = var_7925, x = aw_1189_cast_fp16)[name = tensor("op_8095_cast_fp16")]; + tensor var_8096_cast_fp16 = softmax(axis = var_7925, x = aw_1191_cast_fp16)[name = tensor("op_8096_cast_fp16")]; + tensor var_8097_cast_fp16 = softmax(axis = var_7925, x = aw_1193_cast_fp16)[name = tensor("op_8097_cast_fp16")]; + tensor var_8098_cast_fp16 = softmax(axis = var_7925, x = aw_1195_cast_fp16)[name = tensor("op_8098_cast_fp16")]; + tensor var_8099_cast_fp16 = softmax(axis = var_7925, x = aw_1197_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8100_cast_fp16 = softmax(axis = var_7925, x = aw_1199_cast_fp16)[name = tensor("op_8100_cast_fp16")]; + tensor var_8102_equation_0 = const()[name = tensor("op_8102_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8102_cast_fp16 = einsum(equation = var_8102_equation_0, values = (var_8020_cast_fp16_0, var_8081_cast_fp16))[name = tensor("op_8102_cast_fp16")]; + tensor var_8104_equation_0 = const()[name = tensor("op_8104_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8104_cast_fp16 = einsum(equation = var_8104_equation_0, values = (var_8020_cast_fp16_1, var_8082_cast_fp16))[name = tensor("op_8104_cast_fp16")]; + tensor var_8106_equation_0 = const()[name = tensor("op_8106_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8106_cast_fp16 = einsum(equation = var_8106_equation_0, values = (var_8020_cast_fp16_2, var_8083_cast_fp16))[name = tensor("op_8106_cast_fp16")]; + tensor var_8108_equation_0 = const()[name = tensor("op_8108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8108_cast_fp16 = einsum(equation = var_8108_equation_0, values = (var_8020_cast_fp16_3, var_8084_cast_fp16))[name = tensor("op_8108_cast_fp16")]; + tensor var_8110_equation_0 = const()[name = tensor("op_8110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8110_cast_fp16 = einsum(equation = var_8110_equation_0, values = (var_8020_cast_fp16_4, var_8085_cast_fp16))[name = tensor("op_8110_cast_fp16")]; + tensor var_8112_equation_0 = const()[name = tensor("op_8112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8112_cast_fp16 = einsum(equation = var_8112_equation_0, values = (var_8020_cast_fp16_5, var_8086_cast_fp16))[name = tensor("op_8112_cast_fp16")]; + tensor var_8114_equation_0 = const()[name = tensor("op_8114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8114_cast_fp16 = einsum(equation = var_8114_equation_0, values = (var_8020_cast_fp16_6, var_8087_cast_fp16))[name = tensor("op_8114_cast_fp16")]; + tensor var_8116_equation_0 = const()[name = tensor("op_8116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8116_cast_fp16 = einsum(equation = var_8116_equation_0, values = (var_8020_cast_fp16_7, var_8088_cast_fp16))[name = tensor("op_8116_cast_fp16")]; + tensor var_8118_equation_0 = const()[name = tensor("op_8118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8118_cast_fp16 = einsum(equation = var_8118_equation_0, values = (var_8020_cast_fp16_8, var_8089_cast_fp16))[name = tensor("op_8118_cast_fp16")]; + tensor var_8120_equation_0 = const()[name = tensor("op_8120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8120_cast_fp16 = einsum(equation = var_8120_equation_0, values = (var_8020_cast_fp16_9, var_8090_cast_fp16))[name = tensor("op_8120_cast_fp16")]; + tensor var_8122_equation_0 = const()[name = tensor("op_8122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8122_cast_fp16 = einsum(equation = var_8122_equation_0, values = (var_8020_cast_fp16_10, var_8091_cast_fp16))[name = tensor("op_8122_cast_fp16")]; + tensor var_8124_equation_0 = const()[name = tensor("op_8124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8124_cast_fp16 = einsum(equation = var_8124_equation_0, values = (var_8020_cast_fp16_11, var_8092_cast_fp16))[name = tensor("op_8124_cast_fp16")]; + tensor var_8126_equation_0 = const()[name = tensor("op_8126_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8126_cast_fp16 = einsum(equation = var_8126_equation_0, values = (var_8020_cast_fp16_12, var_8093_cast_fp16))[name = tensor("op_8126_cast_fp16")]; + tensor var_8128_equation_0 = const()[name = tensor("op_8128_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8128_cast_fp16 = einsum(equation = var_8128_equation_0, values = (var_8020_cast_fp16_13, var_8094_cast_fp16))[name = tensor("op_8128_cast_fp16")]; + tensor var_8130_equation_0 = const()[name = tensor("op_8130_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8130_cast_fp16 = einsum(equation = var_8130_equation_0, values = (var_8020_cast_fp16_14, var_8095_cast_fp16))[name = tensor("op_8130_cast_fp16")]; + tensor var_8132_equation_0 = const()[name = tensor("op_8132_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8132_cast_fp16 = einsum(equation = var_8132_equation_0, values = (var_8020_cast_fp16_15, var_8096_cast_fp16))[name = tensor("op_8132_cast_fp16")]; + tensor var_8134_equation_0 = const()[name = tensor("op_8134_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8134_cast_fp16 = einsum(equation = var_8134_equation_0, values = (var_8020_cast_fp16_16, var_8097_cast_fp16))[name = tensor("op_8134_cast_fp16")]; + tensor var_8136_equation_0 = const()[name = tensor("op_8136_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8136_cast_fp16 = einsum(equation = var_8136_equation_0, values = (var_8020_cast_fp16_17, var_8098_cast_fp16))[name = tensor("op_8136_cast_fp16")]; + tensor var_8138_equation_0 = const()[name = tensor("op_8138_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8138_cast_fp16 = einsum(equation = var_8138_equation_0, values = (var_8020_cast_fp16_18, var_8099_cast_fp16))[name = tensor("op_8138_cast_fp16")]; + tensor var_8140_equation_0 = const()[name = tensor("op_8140_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8140_cast_fp16 = einsum(equation = var_8140_equation_0, values = (var_8020_cast_fp16_19, var_8100_cast_fp16))[name = tensor("op_8140_cast_fp16")]; + tensor input_295_interleave_0 = const()[name = tensor("input_295_interleave_0"), val = tensor(false)]; + tensor input_295_cast_fp16 = concat(axis = var_7925, interleave = input_295_interleave_0, values = (var_8102_cast_fp16, var_8104_cast_fp16, var_8106_cast_fp16, var_8108_cast_fp16, var_8110_cast_fp16, var_8112_cast_fp16, var_8114_cast_fp16, var_8116_cast_fp16, var_8118_cast_fp16, var_8120_cast_fp16, var_8122_cast_fp16, var_8124_cast_fp16, var_8126_cast_fp16, var_8128_cast_fp16, var_8130_cast_fp16, var_8132_cast_fp16, var_8134_cast_fp16, var_8136_cast_fp16, var_8138_cast_fp16, var_8140_cast_fp16))[name = tensor("input_295_cast_fp16")]; + tensor var_8149_pad_type_0 = const()[name = tensor("op_8149_pad_type_0"), val = tensor("valid")]; + tensor var_8149_strides_0 = const()[name = tensor("op_8149_strides_0"), val = tensor([1, 1])]; + tensor var_8149_pad_0 = const()[name = tensor("op_8149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8149_dilations_0 = const()[name = tensor("op_8149_dilations_0"), val = tensor([1, 1])]; + tensor var_8149_groups_0 = const()[name = tensor("op_8149_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_out_weight_to_fp16 = const()[name = tensor("blocks_29_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165376512)))]; + tensor blocks_29_attn_out_bias_to_fp16 = const()[name = tensor("blocks_29_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168653376)))]; + tensor var_8149_cast_fp16 = conv(bias = blocks_29_attn_out_bias_to_fp16, dilations = var_8149_dilations_0, groups = var_8149_groups_0, pad = var_8149_pad_0, pad_type = var_8149_pad_type_0, strides = var_8149_strides_0, weight = blocks_29_attn_out_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("op_8149_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_8149_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor input_297_axes_0 = const()[name = tensor("input_297_axes_0"), val = tensor([1])]; + tensor input_297_gamma_0_to_fp16 = const()[name = tensor("input_297_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168656000)))]; + tensor input_297_beta_0_to_fp16 = const()[name = tensor("input_297_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168658624)))]; + tensor var_8159_to_fp16 = const()[name = tensor("op_8159_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_297_cast_fp16 = layer_norm(axes = input_297_axes_0, beta = input_297_beta_0_to_fp16, epsilon = var_8159_to_fp16, gamma = input_297_gamma_0_to_fp16, x = inputs_119_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_pad_type_0 = const()[name = tensor("input_299_pad_type_0"), val = tensor("valid")]; + tensor input_299_strides_0 = const()[name = tensor("input_299_strides_0"), val = tensor([1, 1])]; + tensor input_299_pad_0 = const()[name = tensor("input_299_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_299_dilations_0 = const()[name = tensor("input_299_dilations_0"), val = tensor([1, 1])]; + tensor input_299_groups_0 = const()[name = tensor("input_299_groups_0"), val = tensor(1)]; + tensor blocks_29_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168661248)))]; + tensor blocks_29_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181768512)))]; + tensor input_299_cast_fp16 = conv(bias = blocks_29_mlp_0_bias_to_fp16, dilations = input_299_dilations_0, groups = input_299_groups_0, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = input_299_strides_0, weight = blocks_29_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor input_301_mode_0 = const()[name = tensor("input_301_mode_0"), val = tensor("EXACT")]; + tensor input_301_cast_fp16 = gelu(mode = input_301_mode_0, x = input_299_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_8185_pad_type_0 = const()[name = tensor("op_8185_pad_type_0"), val = tensor("valid")]; + tensor var_8185_strides_0 = const()[name = tensor("op_8185_strides_0"), val = tensor([1, 1])]; + tensor var_8185_pad_0 = const()[name = tensor("op_8185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8185_dilations_0 = const()[name = tensor("op_8185_dilations_0"), val = tensor([1, 1])]; + tensor var_8185_groups_0 = const()[name = tensor("op_8185_groups_0"), val = tensor(1)]; + tensor blocks_29_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181778816)))]; + tensor blocks_29_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194886080)))]; + tensor var_8185_cast_fp16 = conv(bias = blocks_29_mlp_2_bias_to_fp16, dilations = var_8185_dilations_0, groups = var_8185_groups_0, pad = var_8185_pad_0, pad_type = var_8185_pad_type_0, strides = var_8185_strides_0, weight = blocks_29_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = var_8185_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_8194 = const()[name = tensor("op_8194"), val = tensor(1)]; + tensor input_303_axes_0 = const()[name = tensor("input_303_axes_0"), val = tensor([1])]; + tensor input_303_gamma_0_to_fp16 = const()[name = tensor("input_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194888704)))]; + tensor input_303_beta_0_to_fp16 = const()[name = tensor("input_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194891328)))]; + tensor var_8210_to_fp16 = const()[name = tensor("op_8210_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_303_cast_fp16 = layer_norm(axes = input_303_axes_0, beta = input_303_beta_0_to_fp16, epsilon = var_8210_to_fp16, gamma = input_303_gamma_0_to_fp16, x = inputs_121_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor q_61_pad_type_0 = const()[name = tensor("q_61_pad_type_0"), val = tensor("valid")]; + tensor q_61_strides_0 = const()[name = tensor("q_61_strides_0"), val = tensor([1, 1])]; + tensor q_61_pad_0 = const()[name = tensor("q_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_61_dilations_0 = const()[name = tensor("q_61_dilations_0"), val = tensor([1, 1])]; + tensor q_61_groups_0 = const()[name = tensor("q_61_groups_0"), val = tensor(1)]; + tensor var_8245_weight_0_to_fp16 = const()[name = tensor("op_8245_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194893952)))]; + tensor var_8245_bias_0_to_fp16 = const()[name = tensor("op_8245_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198170816)))]; + tensor var_8245_cast_fp16 = conv(bias = var_8245_bias_0_to_fp16, dilations = q_61_dilations_0, groups = q_61_groups_0, pad = q_61_pad_0, pad_type = q_61_pad_type_0, strides = q_61_strides_0, weight = var_8245_weight_0_to_fp16, x = input_303_cast_fp16)[name = tensor("op_8245_cast_fp16")]; + tensor k_61_pad_type_0 = const()[name = tensor("k_61_pad_type_0"), val = tensor("valid")]; + tensor k_61_strides_0 = const()[name = tensor("k_61_strides_0"), val = tensor([1, 1])]; + tensor k_61_pad_0 = const()[name = tensor("k_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_61_dilations_0 = const()[name = tensor("k_61_dilations_0"), val = tensor([1, 1])]; + tensor k_61_groups_0 = const()[name = tensor("k_61_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_key_weight_to_fp16 = const()[name = tensor("blocks_30_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198173440)))]; + tensor k_61_cast_fp16 = conv(dilations = k_61_dilations_0, groups = k_61_groups_0, pad = k_61_pad_0, pad_type = k_61_pad_type_0, strides = k_61_strides_0, weight = blocks_30_attn_key_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor var_8243_pad_type_0 = const()[name = tensor("op_8243_pad_type_0"), val = tensor("valid")]; + tensor var_8243_strides_0 = const()[name = tensor("op_8243_strides_0"), val = tensor([1, 1])]; + tensor var_8243_pad_0 = const()[name = tensor("op_8243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8243_dilations_0 = const()[name = tensor("op_8243_dilations_0"), val = tensor([1, 1])]; + tensor var_8243_groups_0 = const()[name = tensor("op_8243_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_value_weight_to_fp16 = const()[name = tensor("blocks_30_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201450304)))]; + tensor blocks_30_attn_value_bias_to_fp16 = const()[name = tensor("blocks_30_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204727168)))]; + tensor var_8243_cast_fp16 = conv(bias = blocks_30_attn_value_bias_to_fp16, dilations = var_8243_dilations_0, groups = var_8243_groups_0, pad = var_8243_pad_0, pad_type = var_8243_pad_type_0, strides = var_8243_strides_0, weight = blocks_30_attn_value_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("op_8243_cast_fp16")]; + tensor tile_90 = const()[name = tensor("tile_90"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8246_axis_0 = const()[name = tensor("op_8246_axis_0"), val = tensor(1)]; + tensor var_8246_cast_fp16_0, tensor var_8246_cast_fp16_1, tensor var_8246_cast_fp16_2, tensor var_8246_cast_fp16_3, tensor var_8246_cast_fp16_4, tensor var_8246_cast_fp16_5, tensor var_8246_cast_fp16_6, tensor var_8246_cast_fp16_7, tensor var_8246_cast_fp16_8, tensor var_8246_cast_fp16_9, tensor var_8246_cast_fp16_10, tensor var_8246_cast_fp16_11, tensor var_8246_cast_fp16_12, tensor var_8246_cast_fp16_13, tensor var_8246_cast_fp16_14, tensor var_8246_cast_fp16_15, tensor var_8246_cast_fp16_16, tensor var_8246_cast_fp16_17, tensor var_8246_cast_fp16_18, tensor var_8246_cast_fp16_19 = split(axis = var_8246_axis_0, split_sizes = tile_90, x = var_8245_cast_fp16)[name = tensor("op_8246_cast_fp16")]; + tensor var_8267_perm_0 = const()[name = tensor("op_8267_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_91 = const()[name = tensor("tile_91"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8268_axis_0 = const()[name = tensor("op_8268_axis_0"), val = tensor(3)]; + tensor var_8267_cast_fp16 = transpose(perm = var_8267_perm_0, x = k_61_cast_fp16)[name = tensor("transpose_2")]; + tensor var_8268_cast_fp16_0, tensor var_8268_cast_fp16_1, tensor var_8268_cast_fp16_2, tensor var_8268_cast_fp16_3, tensor var_8268_cast_fp16_4, tensor var_8268_cast_fp16_5, tensor var_8268_cast_fp16_6, tensor var_8268_cast_fp16_7, tensor var_8268_cast_fp16_8, tensor var_8268_cast_fp16_9, tensor var_8268_cast_fp16_10, tensor var_8268_cast_fp16_11, tensor var_8268_cast_fp16_12, tensor var_8268_cast_fp16_13, tensor var_8268_cast_fp16_14, tensor var_8268_cast_fp16_15, tensor var_8268_cast_fp16_16, tensor var_8268_cast_fp16_17, tensor var_8268_cast_fp16_18, tensor var_8268_cast_fp16_19 = split(axis = var_8268_axis_0, split_sizes = tile_91, x = var_8267_cast_fp16)[name = tensor("op_8268_cast_fp16")]; + tensor tile_92 = const()[name = tensor("tile_92"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8289_axis_0 = const()[name = tensor("op_8289_axis_0"), val = tensor(1)]; + tensor var_8289_cast_fp16_0, tensor var_8289_cast_fp16_1, tensor var_8289_cast_fp16_2, tensor var_8289_cast_fp16_3, tensor var_8289_cast_fp16_4, tensor var_8289_cast_fp16_5, tensor var_8289_cast_fp16_6, tensor var_8289_cast_fp16_7, tensor var_8289_cast_fp16_8, tensor var_8289_cast_fp16_9, tensor var_8289_cast_fp16_10, tensor var_8289_cast_fp16_11, tensor var_8289_cast_fp16_12, tensor var_8289_cast_fp16_13, tensor var_8289_cast_fp16_14, tensor var_8289_cast_fp16_15, tensor var_8289_cast_fp16_16, tensor var_8289_cast_fp16_17, tensor var_8289_cast_fp16_18, tensor var_8289_cast_fp16_19 = split(axis = var_8289_axis_0, split_sizes = tile_92, x = var_8243_cast_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor aw_1201_equation_0 = const()[name = tensor("aw_1201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1201_cast_fp16 = einsum(equation = aw_1201_equation_0, values = (var_8268_cast_fp16_0, var_8246_cast_fp16_0))[name = tensor("aw_1201_cast_fp16")]; + tensor aw_1203_equation_0 = const()[name = tensor("aw_1203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1203_cast_fp16 = einsum(equation = aw_1203_equation_0, values = (var_8268_cast_fp16_1, var_8246_cast_fp16_1))[name = tensor("aw_1203_cast_fp16")]; + tensor aw_1205_equation_0 = const()[name = tensor("aw_1205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1205_cast_fp16 = einsum(equation = aw_1205_equation_0, values = (var_8268_cast_fp16_2, var_8246_cast_fp16_2))[name = tensor("aw_1205_cast_fp16")]; + tensor aw_1207_equation_0 = const()[name = tensor("aw_1207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1207_cast_fp16 = einsum(equation = aw_1207_equation_0, values = (var_8268_cast_fp16_3, var_8246_cast_fp16_3))[name = tensor("aw_1207_cast_fp16")]; + tensor aw_1209_equation_0 = const()[name = tensor("aw_1209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1209_cast_fp16 = einsum(equation = aw_1209_equation_0, values = (var_8268_cast_fp16_4, var_8246_cast_fp16_4))[name = tensor("aw_1209_cast_fp16")]; + tensor aw_1211_equation_0 = const()[name = tensor("aw_1211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1211_cast_fp16 = einsum(equation = aw_1211_equation_0, values = (var_8268_cast_fp16_5, var_8246_cast_fp16_5))[name = tensor("aw_1211_cast_fp16")]; + tensor aw_1213_equation_0 = const()[name = tensor("aw_1213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1213_cast_fp16 = einsum(equation = aw_1213_equation_0, values = (var_8268_cast_fp16_6, var_8246_cast_fp16_6))[name = tensor("aw_1213_cast_fp16")]; + tensor aw_1215_equation_0 = const()[name = tensor("aw_1215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1215_cast_fp16 = einsum(equation = aw_1215_equation_0, values = (var_8268_cast_fp16_7, var_8246_cast_fp16_7))[name = tensor("aw_1215_cast_fp16")]; + tensor aw_1217_equation_0 = const()[name = tensor("aw_1217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1217_cast_fp16 = einsum(equation = aw_1217_equation_0, values = (var_8268_cast_fp16_8, var_8246_cast_fp16_8))[name = tensor("aw_1217_cast_fp16")]; + tensor aw_1219_equation_0 = const()[name = tensor("aw_1219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1219_cast_fp16 = einsum(equation = aw_1219_equation_0, values = (var_8268_cast_fp16_9, var_8246_cast_fp16_9))[name = tensor("aw_1219_cast_fp16")]; + tensor aw_1221_equation_0 = const()[name = tensor("aw_1221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1221_cast_fp16 = einsum(equation = aw_1221_equation_0, values = (var_8268_cast_fp16_10, var_8246_cast_fp16_10))[name = tensor("aw_1221_cast_fp16")]; + tensor aw_1223_equation_0 = const()[name = tensor("aw_1223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1223_cast_fp16 = einsum(equation = aw_1223_equation_0, values = (var_8268_cast_fp16_11, var_8246_cast_fp16_11))[name = tensor("aw_1223_cast_fp16")]; + tensor aw_1225_equation_0 = const()[name = tensor("aw_1225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1225_cast_fp16 = einsum(equation = aw_1225_equation_0, values = (var_8268_cast_fp16_12, var_8246_cast_fp16_12))[name = tensor("aw_1225_cast_fp16")]; + tensor aw_1227_equation_0 = const()[name = tensor("aw_1227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1227_cast_fp16 = einsum(equation = aw_1227_equation_0, values = (var_8268_cast_fp16_13, var_8246_cast_fp16_13))[name = tensor("aw_1227_cast_fp16")]; + tensor aw_1229_equation_0 = const()[name = tensor("aw_1229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1229_cast_fp16 = einsum(equation = aw_1229_equation_0, values = (var_8268_cast_fp16_14, var_8246_cast_fp16_14))[name = tensor("aw_1229_cast_fp16")]; + tensor aw_1231_equation_0 = const()[name = tensor("aw_1231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1231_cast_fp16 = einsum(equation = aw_1231_equation_0, values = (var_8268_cast_fp16_15, var_8246_cast_fp16_15))[name = tensor("aw_1231_cast_fp16")]; + tensor aw_1233_equation_0 = const()[name = tensor("aw_1233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1233_cast_fp16 = einsum(equation = aw_1233_equation_0, values = (var_8268_cast_fp16_16, var_8246_cast_fp16_16))[name = tensor("aw_1233_cast_fp16")]; + tensor aw_1235_equation_0 = const()[name = tensor("aw_1235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1235_cast_fp16 = einsum(equation = aw_1235_equation_0, values = (var_8268_cast_fp16_17, var_8246_cast_fp16_17))[name = tensor("aw_1235_cast_fp16")]; + tensor aw_1237_equation_0 = const()[name = tensor("aw_1237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1237_cast_fp16 = einsum(equation = aw_1237_equation_0, values = (var_8268_cast_fp16_18, var_8246_cast_fp16_18))[name = tensor("aw_1237_cast_fp16")]; + tensor aw_1239_equation_0 = const()[name = tensor("aw_1239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1239_cast_fp16 = einsum(equation = aw_1239_equation_0, values = (var_8268_cast_fp16_19, var_8246_cast_fp16_19))[name = tensor("aw_1239_cast_fp16")]; + tensor var_8350_cast_fp16 = softmax(axis = var_8194, x = aw_1201_cast_fp16)[name = tensor("op_8350_cast_fp16")]; + tensor var_8351_cast_fp16 = softmax(axis = var_8194, x = aw_1203_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = softmax(axis = var_8194, x = aw_1205_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor var_8353_cast_fp16 = softmax(axis = var_8194, x = aw_1207_cast_fp16)[name = tensor("op_8353_cast_fp16")]; + tensor var_8354_cast_fp16 = softmax(axis = var_8194, x = aw_1209_cast_fp16)[name = tensor("op_8354_cast_fp16")]; + tensor var_8355_cast_fp16 = softmax(axis = var_8194, x = aw_1211_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = softmax(axis = var_8194, x = aw_1213_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor var_8357_cast_fp16 = softmax(axis = var_8194, x = aw_1215_cast_fp16)[name = tensor("op_8357_cast_fp16")]; + tensor var_8358_cast_fp16 = softmax(axis = var_8194, x = aw_1217_cast_fp16)[name = tensor("op_8358_cast_fp16")]; + tensor var_8359_cast_fp16 = softmax(axis = var_8194, x = aw_1219_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = softmax(axis = var_8194, x = aw_1221_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor var_8361_cast_fp16 = softmax(axis = var_8194, x = aw_1223_cast_fp16)[name = tensor("op_8361_cast_fp16")]; + tensor var_8362_cast_fp16 = softmax(axis = var_8194, x = aw_1225_cast_fp16)[name = tensor("op_8362_cast_fp16")]; + tensor var_8363_cast_fp16 = softmax(axis = var_8194, x = aw_1227_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = softmax(axis = var_8194, x = aw_1229_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor var_8365_cast_fp16 = softmax(axis = var_8194, x = aw_1231_cast_fp16)[name = tensor("op_8365_cast_fp16")]; + tensor var_8366_cast_fp16 = softmax(axis = var_8194, x = aw_1233_cast_fp16)[name = tensor("op_8366_cast_fp16")]; + tensor var_8367_cast_fp16 = softmax(axis = var_8194, x = aw_1235_cast_fp16)[name = tensor("op_8367_cast_fp16")]; + tensor var_8368_cast_fp16 = softmax(axis = var_8194, x = aw_1237_cast_fp16)[name = tensor("op_8368_cast_fp16")]; + tensor var_8369_cast_fp16 = softmax(axis = var_8194, x = aw_1239_cast_fp16)[name = tensor("op_8369_cast_fp16")]; + tensor var_8371_equation_0 = const()[name = tensor("op_8371_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8371_cast_fp16 = einsum(equation = var_8371_equation_0, values = (var_8289_cast_fp16_0, var_8350_cast_fp16))[name = tensor("op_8371_cast_fp16")]; + tensor var_8373_equation_0 = const()[name = tensor("op_8373_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8373_cast_fp16 = einsum(equation = var_8373_equation_0, values = (var_8289_cast_fp16_1, var_8351_cast_fp16))[name = tensor("op_8373_cast_fp16")]; + tensor var_8375_equation_0 = const()[name = tensor("op_8375_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8375_cast_fp16 = einsum(equation = var_8375_equation_0, values = (var_8289_cast_fp16_2, var_8352_cast_fp16))[name = tensor("op_8375_cast_fp16")]; + tensor var_8377_equation_0 = const()[name = tensor("op_8377_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8377_cast_fp16 = einsum(equation = var_8377_equation_0, values = (var_8289_cast_fp16_3, var_8353_cast_fp16))[name = tensor("op_8377_cast_fp16")]; + tensor var_8379_equation_0 = const()[name = tensor("op_8379_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8379_cast_fp16 = einsum(equation = var_8379_equation_0, values = (var_8289_cast_fp16_4, var_8354_cast_fp16))[name = tensor("op_8379_cast_fp16")]; + tensor var_8381_equation_0 = const()[name = tensor("op_8381_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8381_cast_fp16 = einsum(equation = var_8381_equation_0, values = (var_8289_cast_fp16_5, var_8355_cast_fp16))[name = tensor("op_8381_cast_fp16")]; + tensor var_8383_equation_0 = const()[name = tensor("op_8383_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8383_cast_fp16 = einsum(equation = var_8383_equation_0, values = (var_8289_cast_fp16_6, var_8356_cast_fp16))[name = tensor("op_8383_cast_fp16")]; + tensor var_8385_equation_0 = const()[name = tensor("op_8385_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8385_cast_fp16 = einsum(equation = var_8385_equation_0, values = (var_8289_cast_fp16_7, var_8357_cast_fp16))[name = tensor("op_8385_cast_fp16")]; + tensor var_8387_equation_0 = const()[name = tensor("op_8387_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8387_cast_fp16 = einsum(equation = var_8387_equation_0, values = (var_8289_cast_fp16_8, var_8358_cast_fp16))[name = tensor("op_8387_cast_fp16")]; + tensor var_8389_equation_0 = const()[name = tensor("op_8389_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8389_cast_fp16 = einsum(equation = var_8389_equation_0, values = (var_8289_cast_fp16_9, var_8359_cast_fp16))[name = tensor("op_8389_cast_fp16")]; + tensor var_8391_equation_0 = const()[name = tensor("op_8391_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8391_cast_fp16 = einsum(equation = var_8391_equation_0, values = (var_8289_cast_fp16_10, var_8360_cast_fp16))[name = tensor("op_8391_cast_fp16")]; + tensor var_8393_equation_0 = const()[name = tensor("op_8393_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8393_cast_fp16 = einsum(equation = var_8393_equation_0, values = (var_8289_cast_fp16_11, var_8361_cast_fp16))[name = tensor("op_8393_cast_fp16")]; + tensor var_8395_equation_0 = const()[name = tensor("op_8395_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8395_cast_fp16 = einsum(equation = var_8395_equation_0, values = (var_8289_cast_fp16_12, var_8362_cast_fp16))[name = tensor("op_8395_cast_fp16")]; + tensor var_8397_equation_0 = const()[name = tensor("op_8397_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8397_cast_fp16 = einsum(equation = var_8397_equation_0, values = (var_8289_cast_fp16_13, var_8363_cast_fp16))[name = tensor("op_8397_cast_fp16")]; + tensor var_8399_equation_0 = const()[name = tensor("op_8399_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8399_cast_fp16 = einsum(equation = var_8399_equation_0, values = (var_8289_cast_fp16_14, var_8364_cast_fp16))[name = tensor("op_8399_cast_fp16")]; + tensor var_8401_equation_0 = const()[name = tensor("op_8401_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8401_cast_fp16 = einsum(equation = var_8401_equation_0, values = (var_8289_cast_fp16_15, var_8365_cast_fp16))[name = tensor("op_8401_cast_fp16")]; + tensor var_8403_equation_0 = const()[name = tensor("op_8403_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8403_cast_fp16 = einsum(equation = var_8403_equation_0, values = (var_8289_cast_fp16_16, var_8366_cast_fp16))[name = tensor("op_8403_cast_fp16")]; + tensor var_8405_equation_0 = const()[name = tensor("op_8405_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8405_cast_fp16 = einsum(equation = var_8405_equation_0, values = (var_8289_cast_fp16_17, var_8367_cast_fp16))[name = tensor("op_8405_cast_fp16")]; + tensor var_8407_equation_0 = const()[name = tensor("op_8407_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8407_cast_fp16 = einsum(equation = var_8407_equation_0, values = (var_8289_cast_fp16_18, var_8368_cast_fp16))[name = tensor("op_8407_cast_fp16")]; + tensor var_8409_equation_0 = const()[name = tensor("op_8409_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8409_cast_fp16 = einsum(equation = var_8409_equation_0, values = (var_8289_cast_fp16_19, var_8369_cast_fp16))[name = tensor("op_8409_cast_fp16")]; + tensor input_305_interleave_0 = const()[name = tensor("input_305_interleave_0"), val = tensor(false)]; + tensor input_305_cast_fp16 = concat(axis = var_8194, interleave = input_305_interleave_0, values = (var_8371_cast_fp16, var_8373_cast_fp16, var_8375_cast_fp16, var_8377_cast_fp16, var_8379_cast_fp16, var_8381_cast_fp16, var_8383_cast_fp16, var_8385_cast_fp16, var_8387_cast_fp16, var_8389_cast_fp16, var_8391_cast_fp16, var_8393_cast_fp16, var_8395_cast_fp16, var_8397_cast_fp16, var_8399_cast_fp16, var_8401_cast_fp16, var_8403_cast_fp16, var_8405_cast_fp16, var_8407_cast_fp16, var_8409_cast_fp16))[name = tensor("input_305_cast_fp16")]; + tensor var_8418_pad_type_0 = const()[name = tensor("op_8418_pad_type_0"), val = tensor("valid")]; + tensor var_8418_strides_0 = const()[name = tensor("op_8418_strides_0"), val = tensor([1, 1])]; + tensor var_8418_pad_0 = const()[name = tensor("op_8418_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8418_dilations_0 = const()[name = tensor("op_8418_dilations_0"), val = tensor([1, 1])]; + tensor var_8418_groups_0 = const()[name = tensor("op_8418_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_out_weight_to_fp16 = const()[name = tensor("blocks_30_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204729792)))]; + tensor blocks_30_attn_out_bias_to_fp16 = const()[name = tensor("blocks_30_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208006656)))]; + tensor var_8418_cast_fp16 = conv(bias = blocks_30_attn_out_bias_to_fp16, dilations = var_8418_dilations_0, groups = var_8418_groups_0, pad = var_8418_pad_0, pad_type = var_8418_pad_type_0, strides = var_8418_strides_0, weight = blocks_30_attn_out_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_8418_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor input_307_axes_0 = const()[name = tensor("input_307_axes_0"), val = tensor([1])]; + tensor input_307_gamma_0_to_fp16 = const()[name = tensor("input_307_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208009280)))]; + tensor input_307_beta_0_to_fp16 = const()[name = tensor("input_307_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208011904)))]; + tensor var_8428_to_fp16 = const()[name = tensor("op_8428_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = input_307_beta_0_to_fp16, epsilon = var_8428_to_fp16, gamma = input_307_gamma_0_to_fp16, x = inputs_123_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor input_309_pad_type_0 = const()[name = tensor("input_309_pad_type_0"), val = tensor("valid")]; + tensor input_309_strides_0 = const()[name = tensor("input_309_strides_0"), val = tensor([1, 1])]; + tensor input_309_pad_0 = const()[name = tensor("input_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_309_dilations_0 = const()[name = tensor("input_309_dilations_0"), val = tensor([1, 1])]; + tensor input_309_groups_0 = const()[name = tensor("input_309_groups_0"), val = tensor(1)]; + tensor blocks_30_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208014528)))]; + tensor blocks_30_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221121792)))]; + tensor input_309_cast_fp16 = conv(bias = blocks_30_mlp_0_bias_to_fp16, dilations = input_309_dilations_0, groups = input_309_groups_0, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = input_309_strides_0, weight = blocks_30_mlp_0_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor input_311_mode_0 = const()[name = tensor("input_311_mode_0"), val = tensor("EXACT")]; + tensor input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = input_309_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor var_8454_pad_type_0 = const()[name = tensor("op_8454_pad_type_0"), val = tensor("valid")]; + tensor var_8454_strides_0 = const()[name = tensor("op_8454_strides_0"), val = tensor([1, 1])]; + tensor var_8454_pad_0 = const()[name = tensor("op_8454_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8454_dilations_0 = const()[name = tensor("op_8454_dilations_0"), val = tensor([1, 1])]; + tensor var_8454_groups_0 = const()[name = tensor("op_8454_groups_0"), val = tensor(1)]; + tensor blocks_30_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221132096)))]; + tensor blocks_30_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234239360)))]; + tensor var_8454_cast_fp16 = conv(bias = blocks_30_mlp_2_bias_to_fp16, dilations = var_8454_dilations_0, groups = var_8454_groups_0, pad = var_8454_pad_0, pad_type = var_8454_pad_type_0, strides = var_8454_strides_0, weight = blocks_30_mlp_2_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = var_8454_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_8463 = const()[name = tensor("op_8463"), val = tensor(1)]; + tensor input_313_axes_0 = const()[name = tensor("input_313_axes_0"), val = tensor([1])]; + tensor input_313_gamma_0_to_fp16 = const()[name = tensor("input_313_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234241984)))]; + tensor input_313_beta_0_to_fp16 = const()[name = tensor("input_313_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234244608)))]; + tensor var_8479_to_fp16 = const()[name = tensor("op_8479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_313_cast_fp16 = layer_norm(axes = input_313_axes_0, beta = input_313_beta_0_to_fp16, epsilon = var_8479_to_fp16, gamma = input_313_gamma_0_to_fp16, x = inputs_125_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_8514_weight_0_to_fp16 = const()[name = tensor("op_8514_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234247232)))]; + tensor var_8514_bias_0_to_fp16 = const()[name = tensor("op_8514_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237524096)))]; + tensor var_8514_cast_fp16 = conv(bias = var_8514_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_8514_weight_0_to_fp16, x = input_313_cast_fp16)[name = tensor("op_8514_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_key_weight_to_fp16 = const()[name = tensor("blocks_31_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237526720)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_31_attn_key_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_8512_pad_type_0 = const()[name = tensor("op_8512_pad_type_0"), val = tensor("valid")]; + tensor var_8512_strides_0 = const()[name = tensor("op_8512_strides_0"), val = tensor([1, 1])]; + tensor var_8512_pad_0 = const()[name = tensor("op_8512_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8512_dilations_0 = const()[name = tensor("op_8512_dilations_0"), val = tensor([1, 1])]; + tensor var_8512_groups_0 = const()[name = tensor("op_8512_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_value_weight_to_fp16 = const()[name = tensor("blocks_31_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1240803584)))]; + tensor blocks_31_attn_value_bias_to_fp16 = const()[name = tensor("blocks_31_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244080448)))]; + tensor var_8512_cast_fp16 = conv(bias = blocks_31_attn_value_bias_to_fp16, dilations = var_8512_dilations_0, groups = var_8512_groups_0, pad = var_8512_pad_0, pad_type = var_8512_pad_type_0, strides = var_8512_strides_0, weight = blocks_31_attn_value_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("op_8512_cast_fp16")]; + tensor tile_93 = const()[name = tensor("tile_93"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8515_axis_0 = const()[name = tensor("op_8515_axis_0"), val = tensor(1)]; + tensor var_8515_cast_fp16_0, tensor var_8515_cast_fp16_1, tensor var_8515_cast_fp16_2, tensor var_8515_cast_fp16_3, tensor var_8515_cast_fp16_4, tensor var_8515_cast_fp16_5, tensor var_8515_cast_fp16_6, tensor var_8515_cast_fp16_7, tensor var_8515_cast_fp16_8, tensor var_8515_cast_fp16_9, tensor var_8515_cast_fp16_10, tensor var_8515_cast_fp16_11, tensor var_8515_cast_fp16_12, tensor var_8515_cast_fp16_13, tensor var_8515_cast_fp16_14, tensor var_8515_cast_fp16_15, tensor var_8515_cast_fp16_16, tensor var_8515_cast_fp16_17, tensor var_8515_cast_fp16_18, tensor var_8515_cast_fp16_19 = split(axis = var_8515_axis_0, split_sizes = tile_93, x = var_8514_cast_fp16)[name = tensor("op_8515_cast_fp16")]; + tensor var_8536_perm_0 = const()[name = tensor("op_8536_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_94 = const()[name = tensor("tile_94"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8537_axis_0 = const()[name = tensor("op_8537_axis_0"), val = tensor(3)]; + tensor var_8536_cast_fp16 = transpose(perm = var_8536_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_8537_cast_fp16_0, tensor var_8537_cast_fp16_1, tensor var_8537_cast_fp16_2, tensor var_8537_cast_fp16_3, tensor var_8537_cast_fp16_4, tensor var_8537_cast_fp16_5, tensor var_8537_cast_fp16_6, tensor var_8537_cast_fp16_7, tensor var_8537_cast_fp16_8, tensor var_8537_cast_fp16_9, tensor var_8537_cast_fp16_10, tensor var_8537_cast_fp16_11, tensor var_8537_cast_fp16_12, tensor var_8537_cast_fp16_13, tensor var_8537_cast_fp16_14, tensor var_8537_cast_fp16_15, tensor var_8537_cast_fp16_16, tensor var_8537_cast_fp16_17, tensor var_8537_cast_fp16_18, tensor var_8537_cast_fp16_19 = split(axis = var_8537_axis_0, split_sizes = tile_94, x = var_8536_cast_fp16)[name = tensor("op_8537_cast_fp16")]; + tensor tile_95 = const()[name = tensor("tile_95"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8558_axis_0 = const()[name = tensor("op_8558_axis_0"), val = tensor(1)]; + tensor var_8558_cast_fp16_0, tensor var_8558_cast_fp16_1, tensor var_8558_cast_fp16_2, tensor var_8558_cast_fp16_3, tensor var_8558_cast_fp16_4, tensor var_8558_cast_fp16_5, tensor var_8558_cast_fp16_6, tensor var_8558_cast_fp16_7, tensor var_8558_cast_fp16_8, tensor var_8558_cast_fp16_9, tensor var_8558_cast_fp16_10, tensor var_8558_cast_fp16_11, tensor var_8558_cast_fp16_12, tensor var_8558_cast_fp16_13, tensor var_8558_cast_fp16_14, tensor var_8558_cast_fp16_15, tensor var_8558_cast_fp16_16, tensor var_8558_cast_fp16_17, tensor var_8558_cast_fp16_18, tensor var_8558_cast_fp16_19 = split(axis = var_8558_axis_0, split_sizes = tile_95, x = var_8512_cast_fp16)[name = tensor("op_8558_cast_fp16")]; + tensor aw_1241_equation_0 = const()[name = tensor("aw_1241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1241_cast_fp16 = einsum(equation = aw_1241_equation_0, values = (var_8537_cast_fp16_0, var_8515_cast_fp16_0))[name = tensor("aw_1241_cast_fp16")]; + tensor aw_1243_equation_0 = const()[name = tensor("aw_1243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1243_cast_fp16 = einsum(equation = aw_1243_equation_0, values = (var_8537_cast_fp16_1, var_8515_cast_fp16_1))[name = tensor("aw_1243_cast_fp16")]; + tensor aw_1245_equation_0 = const()[name = tensor("aw_1245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1245_cast_fp16 = einsum(equation = aw_1245_equation_0, values = (var_8537_cast_fp16_2, var_8515_cast_fp16_2))[name = tensor("aw_1245_cast_fp16")]; + tensor aw_1247_equation_0 = const()[name = tensor("aw_1247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1247_cast_fp16 = einsum(equation = aw_1247_equation_0, values = (var_8537_cast_fp16_3, var_8515_cast_fp16_3))[name = tensor("aw_1247_cast_fp16")]; + tensor aw_1249_equation_0 = const()[name = tensor("aw_1249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1249_cast_fp16 = einsum(equation = aw_1249_equation_0, values = (var_8537_cast_fp16_4, var_8515_cast_fp16_4))[name = tensor("aw_1249_cast_fp16")]; + tensor aw_1251_equation_0 = const()[name = tensor("aw_1251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1251_cast_fp16 = einsum(equation = aw_1251_equation_0, values = (var_8537_cast_fp16_5, var_8515_cast_fp16_5))[name = tensor("aw_1251_cast_fp16")]; + tensor aw_1253_equation_0 = const()[name = tensor("aw_1253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1253_cast_fp16 = einsum(equation = aw_1253_equation_0, values = (var_8537_cast_fp16_6, var_8515_cast_fp16_6))[name = tensor("aw_1253_cast_fp16")]; + tensor aw_1255_equation_0 = const()[name = tensor("aw_1255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1255_cast_fp16 = einsum(equation = aw_1255_equation_0, values = (var_8537_cast_fp16_7, var_8515_cast_fp16_7))[name = tensor("aw_1255_cast_fp16")]; + tensor aw_1257_equation_0 = const()[name = tensor("aw_1257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1257_cast_fp16 = einsum(equation = aw_1257_equation_0, values = (var_8537_cast_fp16_8, var_8515_cast_fp16_8))[name = tensor("aw_1257_cast_fp16")]; + tensor aw_1259_equation_0 = const()[name = tensor("aw_1259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1259_cast_fp16 = einsum(equation = aw_1259_equation_0, values = (var_8537_cast_fp16_9, var_8515_cast_fp16_9))[name = tensor("aw_1259_cast_fp16")]; + tensor aw_1261_equation_0 = const()[name = tensor("aw_1261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1261_cast_fp16 = einsum(equation = aw_1261_equation_0, values = (var_8537_cast_fp16_10, var_8515_cast_fp16_10))[name = tensor("aw_1261_cast_fp16")]; + tensor aw_1263_equation_0 = const()[name = tensor("aw_1263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1263_cast_fp16 = einsum(equation = aw_1263_equation_0, values = (var_8537_cast_fp16_11, var_8515_cast_fp16_11))[name = tensor("aw_1263_cast_fp16")]; + tensor aw_1265_equation_0 = const()[name = tensor("aw_1265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1265_cast_fp16 = einsum(equation = aw_1265_equation_0, values = (var_8537_cast_fp16_12, var_8515_cast_fp16_12))[name = tensor("aw_1265_cast_fp16")]; + tensor aw_1267_equation_0 = const()[name = tensor("aw_1267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1267_cast_fp16 = einsum(equation = aw_1267_equation_0, values = (var_8537_cast_fp16_13, var_8515_cast_fp16_13))[name = tensor("aw_1267_cast_fp16")]; + tensor aw_1269_equation_0 = const()[name = tensor("aw_1269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1269_cast_fp16 = einsum(equation = aw_1269_equation_0, values = (var_8537_cast_fp16_14, var_8515_cast_fp16_14))[name = tensor("aw_1269_cast_fp16")]; + tensor aw_1271_equation_0 = const()[name = tensor("aw_1271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1271_cast_fp16 = einsum(equation = aw_1271_equation_0, values = (var_8537_cast_fp16_15, var_8515_cast_fp16_15))[name = tensor("aw_1271_cast_fp16")]; + tensor aw_1273_equation_0 = const()[name = tensor("aw_1273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1273_cast_fp16 = einsum(equation = aw_1273_equation_0, values = (var_8537_cast_fp16_16, var_8515_cast_fp16_16))[name = tensor("aw_1273_cast_fp16")]; + tensor aw_1275_equation_0 = const()[name = tensor("aw_1275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1275_cast_fp16 = einsum(equation = aw_1275_equation_0, values = (var_8537_cast_fp16_17, var_8515_cast_fp16_17))[name = tensor("aw_1275_cast_fp16")]; + tensor aw_1277_equation_0 = const()[name = tensor("aw_1277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1277_cast_fp16 = einsum(equation = aw_1277_equation_0, values = (var_8537_cast_fp16_18, var_8515_cast_fp16_18))[name = tensor("aw_1277_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_8537_cast_fp16_19, var_8515_cast_fp16_19))[name = tensor("aw_cast_fp16")]; + tensor var_8619_cast_fp16 = softmax(axis = var_8463, x = aw_1241_cast_fp16)[name = tensor("op_8619_cast_fp16")]; + tensor var_8620_cast_fp16 = softmax(axis = var_8463, x = aw_1243_cast_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor var_8621_cast_fp16 = softmax(axis = var_8463, x = aw_1245_cast_fp16)[name = tensor("op_8621_cast_fp16")]; + tensor var_8622_cast_fp16 = softmax(axis = var_8463, x = aw_1247_cast_fp16)[name = tensor("op_8622_cast_fp16")]; + tensor var_8623_cast_fp16 = softmax(axis = var_8463, x = aw_1249_cast_fp16)[name = tensor("op_8623_cast_fp16")]; + tensor var_8624_cast_fp16 = softmax(axis = var_8463, x = aw_1251_cast_fp16)[name = tensor("op_8624_cast_fp16")]; + tensor var_8625_cast_fp16 = softmax(axis = var_8463, x = aw_1253_cast_fp16)[name = tensor("op_8625_cast_fp16")]; + tensor var_8626_cast_fp16 = softmax(axis = var_8463, x = aw_1255_cast_fp16)[name = tensor("op_8626_cast_fp16")]; + tensor var_8627_cast_fp16 = softmax(axis = var_8463, x = aw_1257_cast_fp16)[name = tensor("op_8627_cast_fp16")]; + tensor var_8628_cast_fp16 = softmax(axis = var_8463, x = aw_1259_cast_fp16)[name = tensor("op_8628_cast_fp16")]; + tensor var_8629_cast_fp16 = softmax(axis = var_8463, x = aw_1261_cast_fp16)[name = tensor("op_8629_cast_fp16")]; + tensor var_8630_cast_fp16 = softmax(axis = var_8463, x = aw_1263_cast_fp16)[name = tensor("op_8630_cast_fp16")]; + tensor var_8631_cast_fp16 = softmax(axis = var_8463, x = aw_1265_cast_fp16)[name = tensor("op_8631_cast_fp16")]; + tensor var_8632_cast_fp16 = softmax(axis = var_8463, x = aw_1267_cast_fp16)[name = tensor("op_8632_cast_fp16")]; + tensor var_8633_cast_fp16 = softmax(axis = var_8463, x = aw_1269_cast_fp16)[name = tensor("op_8633_cast_fp16")]; + tensor var_8634_cast_fp16 = softmax(axis = var_8463, x = aw_1271_cast_fp16)[name = tensor("op_8634_cast_fp16")]; + tensor var_8635_cast_fp16 = softmax(axis = var_8463, x = aw_1273_cast_fp16)[name = tensor("op_8635_cast_fp16")]; + tensor var_8636_cast_fp16 = softmax(axis = var_8463, x = aw_1275_cast_fp16)[name = tensor("op_8636_cast_fp16")]; + tensor var_8637_cast_fp16 = softmax(axis = var_8463, x = aw_1277_cast_fp16)[name = tensor("op_8637_cast_fp16")]; + tensor var_8638_cast_fp16 = softmax(axis = var_8463, x = aw_cast_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor var_8640_equation_0 = const()[name = tensor("op_8640_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8640_cast_fp16 = einsum(equation = var_8640_equation_0, values = (var_8558_cast_fp16_0, var_8619_cast_fp16))[name = tensor("op_8640_cast_fp16")]; + tensor var_8642_equation_0 = const()[name = tensor("op_8642_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8642_cast_fp16 = einsum(equation = var_8642_equation_0, values = (var_8558_cast_fp16_1, var_8620_cast_fp16))[name = tensor("op_8642_cast_fp16")]; + tensor var_8644_equation_0 = const()[name = tensor("op_8644_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8644_cast_fp16 = einsum(equation = var_8644_equation_0, values = (var_8558_cast_fp16_2, var_8621_cast_fp16))[name = tensor("op_8644_cast_fp16")]; + tensor var_8646_equation_0 = const()[name = tensor("op_8646_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8646_cast_fp16 = einsum(equation = var_8646_equation_0, values = (var_8558_cast_fp16_3, var_8622_cast_fp16))[name = tensor("op_8646_cast_fp16")]; + tensor var_8648_equation_0 = const()[name = tensor("op_8648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8648_cast_fp16 = einsum(equation = var_8648_equation_0, values = (var_8558_cast_fp16_4, var_8623_cast_fp16))[name = tensor("op_8648_cast_fp16")]; + tensor var_8650_equation_0 = const()[name = tensor("op_8650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8650_cast_fp16 = einsum(equation = var_8650_equation_0, values = (var_8558_cast_fp16_5, var_8624_cast_fp16))[name = tensor("op_8650_cast_fp16")]; + tensor var_8652_equation_0 = const()[name = tensor("op_8652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8652_cast_fp16 = einsum(equation = var_8652_equation_0, values = (var_8558_cast_fp16_6, var_8625_cast_fp16))[name = tensor("op_8652_cast_fp16")]; + tensor var_8654_equation_0 = const()[name = tensor("op_8654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8654_cast_fp16 = einsum(equation = var_8654_equation_0, values = (var_8558_cast_fp16_7, var_8626_cast_fp16))[name = tensor("op_8654_cast_fp16")]; + tensor var_8656_equation_0 = const()[name = tensor("op_8656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8656_cast_fp16 = einsum(equation = var_8656_equation_0, values = (var_8558_cast_fp16_8, var_8627_cast_fp16))[name = tensor("op_8656_cast_fp16")]; + tensor var_8658_equation_0 = const()[name = tensor("op_8658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8658_cast_fp16 = einsum(equation = var_8658_equation_0, values = (var_8558_cast_fp16_9, var_8628_cast_fp16))[name = tensor("op_8658_cast_fp16")]; + tensor var_8660_equation_0 = const()[name = tensor("op_8660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8660_cast_fp16 = einsum(equation = var_8660_equation_0, values = (var_8558_cast_fp16_10, var_8629_cast_fp16))[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_equation_0 = const()[name = tensor("op_8662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8662_cast_fp16 = einsum(equation = var_8662_equation_0, values = (var_8558_cast_fp16_11, var_8630_cast_fp16))[name = tensor("op_8662_cast_fp16")]; + tensor var_8664_equation_0 = const()[name = tensor("op_8664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8664_cast_fp16 = einsum(equation = var_8664_equation_0, values = (var_8558_cast_fp16_12, var_8631_cast_fp16))[name = tensor("op_8664_cast_fp16")]; + tensor var_8666_equation_0 = const()[name = tensor("op_8666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8666_cast_fp16 = einsum(equation = var_8666_equation_0, values = (var_8558_cast_fp16_13, var_8632_cast_fp16))[name = tensor("op_8666_cast_fp16")]; + tensor var_8668_equation_0 = const()[name = tensor("op_8668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8668_cast_fp16 = einsum(equation = var_8668_equation_0, values = (var_8558_cast_fp16_14, var_8633_cast_fp16))[name = tensor("op_8668_cast_fp16")]; + tensor var_8670_equation_0 = const()[name = tensor("op_8670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8670_cast_fp16 = einsum(equation = var_8670_equation_0, values = (var_8558_cast_fp16_15, var_8634_cast_fp16))[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_equation_0 = const()[name = tensor("op_8672_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8672_cast_fp16 = einsum(equation = var_8672_equation_0, values = (var_8558_cast_fp16_16, var_8635_cast_fp16))[name = tensor("op_8672_cast_fp16")]; + tensor var_8674_equation_0 = const()[name = tensor("op_8674_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8674_cast_fp16 = einsum(equation = var_8674_equation_0, values = (var_8558_cast_fp16_17, var_8636_cast_fp16))[name = tensor("op_8674_cast_fp16")]; + tensor var_8676_equation_0 = const()[name = tensor("op_8676_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8676_cast_fp16 = einsum(equation = var_8676_equation_0, values = (var_8558_cast_fp16_18, var_8637_cast_fp16))[name = tensor("op_8676_cast_fp16")]; + tensor var_8678_equation_0 = const()[name = tensor("op_8678_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8678_cast_fp16 = einsum(equation = var_8678_equation_0, values = (var_8558_cast_fp16_19, var_8638_cast_fp16))[name = tensor("op_8678_cast_fp16")]; + tensor input_315_interleave_0 = const()[name = tensor("input_315_interleave_0"), val = tensor(false)]; + tensor input_315_cast_fp16 = concat(axis = var_8463, interleave = input_315_interleave_0, values = (var_8640_cast_fp16, var_8642_cast_fp16, var_8644_cast_fp16, var_8646_cast_fp16, var_8648_cast_fp16, var_8650_cast_fp16, var_8652_cast_fp16, var_8654_cast_fp16, var_8656_cast_fp16, var_8658_cast_fp16, var_8660_cast_fp16, var_8662_cast_fp16, var_8664_cast_fp16, var_8666_cast_fp16, var_8668_cast_fp16, var_8670_cast_fp16, var_8672_cast_fp16, var_8674_cast_fp16, var_8676_cast_fp16, var_8678_cast_fp16))[name = tensor("input_315_cast_fp16")]; + tensor var_8687_pad_type_0 = const()[name = tensor("op_8687_pad_type_0"), val = tensor("valid")]; + tensor var_8687_strides_0 = const()[name = tensor("op_8687_strides_0"), val = tensor([1, 1])]; + tensor var_8687_pad_0 = const()[name = tensor("op_8687_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8687_dilations_0 = const()[name = tensor("op_8687_dilations_0"), val = tensor([1, 1])]; + tensor var_8687_groups_0 = const()[name = tensor("op_8687_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_out_weight_to_fp16 = const()[name = tensor("blocks_31_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244083072)))]; + tensor blocks_31_attn_out_bias_to_fp16 = const()[name = tensor("blocks_31_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247359936)))]; + tensor var_8687_cast_fp16 = conv(bias = blocks_31_attn_out_bias_to_fp16, dilations = var_8687_dilations_0, groups = var_8687_groups_0, pad = var_8687_pad_0, pad_type = var_8687_pad_type_0, strides = var_8687_strides_0, weight = blocks_31_attn_out_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("op_8687_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = var_8687_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor input_317_axes_0 = const()[name = tensor("input_317_axes_0"), val = tensor([1])]; + tensor input_317_gamma_0_to_fp16 = const()[name = tensor("input_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247362560)))]; + tensor input_317_beta_0_to_fp16 = const()[name = tensor("input_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247365184)))]; + tensor var_8697_to_fp16 = const()[name = tensor("op_8697_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = input_317_beta_0_to_fp16, epsilon = var_8697_to_fp16, gamma = input_317_gamma_0_to_fp16, x = inputs_127_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_319_pad_type_0 = const()[name = tensor("input_319_pad_type_0"), val = tensor("valid")]; + tensor input_319_strides_0 = const()[name = tensor("input_319_strides_0"), val = tensor([1, 1])]; + tensor input_319_pad_0 = const()[name = tensor("input_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_319_dilations_0 = const()[name = tensor("input_319_dilations_0"), val = tensor([1, 1])]; + tensor input_319_groups_0 = const()[name = tensor("input_319_groups_0"), val = tensor(1)]; + tensor blocks_31_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247367808)))]; + tensor blocks_31_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260475072)))]; + tensor input_319_cast_fp16 = conv(bias = blocks_31_mlp_0_bias_to_fp16, dilations = input_319_dilations_0, groups = input_319_groups_0, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = input_319_strides_0, weight = blocks_31_mlp_0_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_319_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_8723_pad_type_0 = const()[name = tensor("op_8723_pad_type_0"), val = tensor("valid")]; + tensor var_8723_strides_0 = const()[name = tensor("op_8723_strides_0"), val = tensor([1, 1])]; + tensor var_8723_pad_0 = const()[name = tensor("op_8723_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8723_dilations_0 = const()[name = tensor("op_8723_dilations_0"), val = tensor([1, 1])]; + tensor var_8723_groups_0 = const()[name = tensor("op_8723_groups_0"), val = tensor(1)]; + tensor blocks_31_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260485376)))]; + tensor blocks_31_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273592640)))]; + tensor var_8723_cast_fp16 = conv(bias = blocks_31_mlp_2_bias_to_fp16, dilations = var_8723_dilations_0, groups = var_8723_groups_0, pad = var_8723_pad_0, pad_type = var_8723_pad_type_0, strides = var_8723_strides_0, weight = blocks_31_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_8723_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_8723_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273595264)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273597888)))]; + tensor var_8737_to_fp16 = const()[name = tensor("op_8737_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_8737_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_8748_axes_0 = const()[name = tensor("op_8748_axes_0"), val = tensor([2])]; + tensor var_8748_cast_fp16 = squeeze(axes = var_8748_axes_0, x = x_cast_fp16)[name = tensor("op_8748_cast_fp16")]; + tensor var_8751_perm_0 = const()[name = tensor("op_8751_perm_0"), val = tensor([0, 2, 1])]; + tensor var_8751_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8751_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_8751_cast_fp16 = transpose(perm = var_8751_perm_0, x = var_8748_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_8751_cast_fp16_to_fp32_dtype_0, x = var_8751_cast_fp16)[name = tensor("cast_131")]; + } -> (output); +} \ No newline at end of file diff --git a/large-v2/ggml-large-v2-encoder.mlmodelc/weights/weight.bin b/large-v2/ggml-large-v2-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c39236e469b018c8f5efe3a0bbf23f40a7b4d17f --- /dev/null +++ b/large-v2/ggml-large-v2-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b0425ba64707bb4c9727963b5e2b0ee1e8ddece596a56c7795aca0d966614b2 +size 1273600512 diff --git a/large-v2/ggml-large-v2.bin b/large-v2/ggml-large-v2.bin new file mode 100644 index 0000000000000000000000000000000000000000..649aafd67e30021d0140c24342ee2ffb947f4bde --- /dev/null +++ b/large-v2/ggml-large-v2.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a423fe4d40c82774b6af34115b8b935f34152246eb19e80e376071d3f999487 +size 3094623691 diff --git a/large-v3/.DS_Store b/large-v3/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..51b786d7ea7c568805d5c86a53eea79d96e02399 Binary files /dev/null and b/large-v3/.DS_Store differ diff --git a/large-v3/ggml-large-v3-encoder.mlmodelc/analytics/coremldata.bin b/large-v3/ggml-large-v3-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fdbd526d7fb5a2e38e143d34c21be58925d0e888 --- /dev/null +++ b/large-v3/ggml-large-v3-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7cfed2b3afbbf8fa7445b0ec0424e702b83c673568ddfdaa74417d8110c109c +size 243 diff --git a/large-v3/ggml-large-v3-encoder.mlmodelc/coremldata.bin b/large-v3/ggml-large-v3-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c584a477784aac75562de2c7d88f8c7b741e78d9 --- /dev/null +++ b/large-v3/ggml-large-v3-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f9fdf4bc4270b79aa10e04c95098f9c8f171b63ca54753fbdfd1724582324db +size 321 diff --git a/large-v3/ggml-large-v3-encoder.mlmodelc/metadata.json b/large-v3/ggml-large-v3-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..cbdcd27aa4cd56870b73a4fbf017356adea727f7 --- /dev/null +++ b/large-v3/ggml-large-v3-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 1280)", + "shortDescription" : "", + "shape" : "[1, 1500, 1280]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 32, + "Gelu" : 34, + "LayerNorm" : 65, + "Transpose" : 33, + "Softmax" : 640, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 65, + "Einsum" : 1280, + "ExpandDims" : 1, + "Split" : 96, + "Conv" : 194 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 3000)", + "shortDescription" : "", + "shape" : "[1, 128, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_large_v3", + "method" : "predict" + } +] \ No newline at end of file diff --git a/large-v3/ggml-large-v3-encoder.mlmodelc/model.mil b/large-v3/ggml-large-v3-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8dd576adade7bc3f1f73e1bc769ae4fbbf686045 --- /dev/null +++ b/large-v3/ggml-large-v3-encoder.mlmodelc/model.mil @@ -0,0 +1,5643 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_84_pad_type_0 = const()[name = tensor("op_84_pad_type_0"), val = tensor("custom")]; + tensor var_84_pad_0 = const()[name = tensor("op_84_pad_0"), val = tensor([1, 1])]; + tensor var_84_strides_0 = const()[name = tensor("op_84_strides_0"), val = tensor([1])]; + tensor var_84_dilations_0 = const()[name = tensor("op_84_dilations_0"), val = tensor([1])]; + tensor var_84_groups_0 = const()[name = tensor("op_84_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983168)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_132")]; + tensor var_84_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_84_dilations_0, groups = var_84_groups_0, pad = var_84_pad_0, pad_type = var_84_pad_type_0, strides = var_84_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_84_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_102_pad_type_0 = const()[name = tensor("op_102_pad_type_0"), val = tensor("custom")]; + tensor var_102_pad_0 = const()[name = tensor("op_102_pad_0"), val = tensor([1, 1])]; + tensor var_102_strides_0 = const()[name = tensor("op_102_strides_0"), val = tensor([2])]; + tensor var_102_dilations_0 = const()[name = tensor("op_102_dilations_0"), val = tensor([1])]; + tensor var_102_groups_0 = const()[name = tensor("op_102_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985792)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10816256)))]; + tensor var_102_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_102_dilations_0, groups = var_102_groups_0, pad = var_102_pad_0, pad_type = var_102_pad_type_0, strides = var_102_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_102_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_102_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_107_to_fp16 = const()[name = tensor("op_107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10818880)))]; + tensor var_109_cast_fp16 = add(x = x_3_cast_fp16, y = var_107_to_fp16)[name = tensor("op_109_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_109_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_124 = const()[name = tensor("op_124"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14658944)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14661568)))]; + tensor var_140_to_fp16 = const()[name = tensor("op_140_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_140_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_175_weight_0_to_fp16 = const()[name = tensor("op_175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14664192)))]; + tensor var_175_bias_0_to_fp16 = const()[name = tensor("op_175_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17941056)))]; + tensor var_175_cast_fp16 = conv(bias = var_175_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_175_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_175_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17943680)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_173_pad_type_0 = const()[name = tensor("op_173_pad_type_0"), val = tensor("valid")]; + tensor var_173_strides_0 = const()[name = tensor("op_173_strides_0"), val = tensor([1, 1])]; + tensor var_173_pad_0 = const()[name = tensor("op_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_173_dilations_0 = const()[name = tensor("op_173_dilations_0"), val = tensor([1, 1])]; + tensor var_173_groups_0 = const()[name = tensor("op_173_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21220544)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24497408)))]; + tensor var_173_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_173_dilations_0, groups = var_173_groups_0, pad = var_173_pad_0, pad_type = var_173_pad_type_0, strides = var_173_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_173_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_176_axis_0 = const()[name = tensor("op_176_axis_0"), val = tensor(1)]; + tensor var_176_cast_fp16_0, tensor var_176_cast_fp16_1, tensor var_176_cast_fp16_2, tensor var_176_cast_fp16_3, tensor var_176_cast_fp16_4, tensor var_176_cast_fp16_5, tensor var_176_cast_fp16_6, tensor var_176_cast_fp16_7, tensor var_176_cast_fp16_8, tensor var_176_cast_fp16_9, tensor var_176_cast_fp16_10, tensor var_176_cast_fp16_11, tensor var_176_cast_fp16_12, tensor var_176_cast_fp16_13, tensor var_176_cast_fp16_14, tensor var_176_cast_fp16_15, tensor var_176_cast_fp16_16, tensor var_176_cast_fp16_17, tensor var_176_cast_fp16_18, tensor var_176_cast_fp16_19 = split(axis = var_176_axis_0, split_sizes = tile_0, x = var_175_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor var_197_perm_0 = const()[name = tensor("op_197_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_198_axis_0 = const()[name = tensor("op_198_axis_0"), val = tensor(3)]; + tensor var_197_cast_fp16 = transpose(perm = var_197_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_32")]; + tensor var_198_cast_fp16_0, tensor var_198_cast_fp16_1, tensor var_198_cast_fp16_2, tensor var_198_cast_fp16_3, tensor var_198_cast_fp16_4, tensor var_198_cast_fp16_5, tensor var_198_cast_fp16_6, tensor var_198_cast_fp16_7, tensor var_198_cast_fp16_8, tensor var_198_cast_fp16_9, tensor var_198_cast_fp16_10, tensor var_198_cast_fp16_11, tensor var_198_cast_fp16_12, tensor var_198_cast_fp16_13, tensor var_198_cast_fp16_14, tensor var_198_cast_fp16_15, tensor var_198_cast_fp16_16, tensor var_198_cast_fp16_17, tensor var_198_cast_fp16_18, tensor var_198_cast_fp16_19 = split(axis = var_198_axis_0, split_sizes = tile_1, x = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_219_axis_0 = const()[name = tensor("op_219_axis_0"), val = tensor(1)]; + tensor var_219_cast_fp16_0, tensor var_219_cast_fp16_1, tensor var_219_cast_fp16_2, tensor var_219_cast_fp16_3, tensor var_219_cast_fp16_4, tensor var_219_cast_fp16_5, tensor var_219_cast_fp16_6, tensor var_219_cast_fp16_7, tensor var_219_cast_fp16_8, tensor var_219_cast_fp16_9, tensor var_219_cast_fp16_10, tensor var_219_cast_fp16_11, tensor var_219_cast_fp16_12, tensor var_219_cast_fp16_13, tensor var_219_cast_fp16_14, tensor var_219_cast_fp16_15, tensor var_219_cast_fp16_16, tensor var_219_cast_fp16_17, tensor var_219_cast_fp16_18, tensor var_219_cast_fp16_19 = split(axis = var_219_axis_0, split_sizes = tile_2, x = var_173_cast_fp16)[name = tensor("op_219_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_198_cast_fp16_0, var_176_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_198_cast_fp16_1, var_176_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_198_cast_fp16_2, var_176_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_198_cast_fp16_3, var_176_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_198_cast_fp16_4, var_176_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_198_cast_fp16_5, var_176_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_198_cast_fp16_6, var_176_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_198_cast_fp16_7, var_176_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_198_cast_fp16_8, var_176_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_198_cast_fp16_9, var_176_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_198_cast_fp16_10, var_176_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_198_cast_fp16_11, var_176_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_198_cast_fp16_12, var_176_cast_fp16_12))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_198_cast_fp16_13, var_176_cast_fp16_13))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_198_cast_fp16_14, var_176_cast_fp16_14))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_198_cast_fp16_15, var_176_cast_fp16_15))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_198_cast_fp16_16, var_176_cast_fp16_16))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_198_cast_fp16_17, var_176_cast_fp16_17))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_198_cast_fp16_18, var_176_cast_fp16_18))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_198_cast_fp16_19, var_176_cast_fp16_19))[name = tensor("aw_39_cast_fp16")]; + tensor var_280_cast_fp16 = softmax(axis = var_124, x = aw_1_cast_fp16)[name = tensor("op_280_cast_fp16")]; + tensor var_281_cast_fp16 = softmax(axis = var_124, x = aw_3_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor var_282_cast_fp16 = softmax(axis = var_124, x = aw_5_cast_fp16)[name = tensor("op_282_cast_fp16")]; + tensor var_283_cast_fp16 = softmax(axis = var_124, x = aw_7_cast_fp16)[name = tensor("op_283_cast_fp16")]; + tensor var_284_cast_fp16 = softmax(axis = var_124, x = aw_9_cast_fp16)[name = tensor("op_284_cast_fp16")]; + tensor var_285_cast_fp16 = softmax(axis = var_124, x = aw_11_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_286_cast_fp16 = softmax(axis = var_124, x = aw_13_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = softmax(axis = var_124, x = aw_15_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor var_288_cast_fp16 = softmax(axis = var_124, x = aw_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_289_cast_fp16 = softmax(axis = var_124, x = aw_19_cast_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_290_cast_fp16 = softmax(axis = var_124, x = aw_21_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = softmax(axis = var_124, x = aw_23_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor var_292_cast_fp16 = softmax(axis = var_124, x = aw_25_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_293_cast_fp16 = softmax(axis = var_124, x = aw_27_cast_fp16)[name = tensor("op_293_cast_fp16")]; + tensor var_294_cast_fp16 = softmax(axis = var_124, x = aw_29_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_295_cast_fp16 = softmax(axis = var_124, x = aw_31_cast_fp16)[name = tensor("op_295_cast_fp16")]; + tensor var_296_cast_fp16 = softmax(axis = var_124, x = aw_33_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_297_cast_fp16 = softmax(axis = var_124, x = aw_35_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_298_cast_fp16 = softmax(axis = var_124, x = aw_37_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_299_cast_fp16 = softmax(axis = var_124, x = aw_39_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_equation_0 = const()[name = tensor("op_301_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_301_cast_fp16 = einsum(equation = var_301_equation_0, values = (var_219_cast_fp16_0, var_280_cast_fp16))[name = tensor("op_301_cast_fp16")]; + tensor var_303_equation_0 = const()[name = tensor("op_303_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_303_cast_fp16 = einsum(equation = var_303_equation_0, values = (var_219_cast_fp16_1, var_281_cast_fp16))[name = tensor("op_303_cast_fp16")]; + tensor var_305_equation_0 = const()[name = tensor("op_305_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_305_cast_fp16 = einsum(equation = var_305_equation_0, values = (var_219_cast_fp16_2, var_282_cast_fp16))[name = tensor("op_305_cast_fp16")]; + tensor var_307_equation_0 = const()[name = tensor("op_307_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_307_cast_fp16 = einsum(equation = var_307_equation_0, values = (var_219_cast_fp16_3, var_283_cast_fp16))[name = tensor("op_307_cast_fp16")]; + tensor var_309_equation_0 = const()[name = tensor("op_309_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_309_cast_fp16 = einsum(equation = var_309_equation_0, values = (var_219_cast_fp16_4, var_284_cast_fp16))[name = tensor("op_309_cast_fp16")]; + tensor var_311_equation_0 = const()[name = tensor("op_311_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_311_cast_fp16 = einsum(equation = var_311_equation_0, values = (var_219_cast_fp16_5, var_285_cast_fp16))[name = tensor("op_311_cast_fp16")]; + tensor var_313_equation_0 = const()[name = tensor("op_313_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_313_cast_fp16 = einsum(equation = var_313_equation_0, values = (var_219_cast_fp16_6, var_286_cast_fp16))[name = tensor("op_313_cast_fp16")]; + tensor var_315_equation_0 = const()[name = tensor("op_315_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_315_cast_fp16 = einsum(equation = var_315_equation_0, values = (var_219_cast_fp16_7, var_287_cast_fp16))[name = tensor("op_315_cast_fp16")]; + tensor var_317_equation_0 = const()[name = tensor("op_317_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_317_cast_fp16 = einsum(equation = var_317_equation_0, values = (var_219_cast_fp16_8, var_288_cast_fp16))[name = tensor("op_317_cast_fp16")]; + tensor var_319_equation_0 = const()[name = tensor("op_319_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_319_cast_fp16 = einsum(equation = var_319_equation_0, values = (var_219_cast_fp16_9, var_289_cast_fp16))[name = tensor("op_319_cast_fp16")]; + tensor var_321_equation_0 = const()[name = tensor("op_321_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_321_cast_fp16 = einsum(equation = var_321_equation_0, values = (var_219_cast_fp16_10, var_290_cast_fp16))[name = tensor("op_321_cast_fp16")]; + tensor var_323_equation_0 = const()[name = tensor("op_323_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_323_cast_fp16 = einsum(equation = var_323_equation_0, values = (var_219_cast_fp16_11, var_291_cast_fp16))[name = tensor("op_323_cast_fp16")]; + tensor var_325_equation_0 = const()[name = tensor("op_325_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_325_cast_fp16 = einsum(equation = var_325_equation_0, values = (var_219_cast_fp16_12, var_292_cast_fp16))[name = tensor("op_325_cast_fp16")]; + tensor var_327_equation_0 = const()[name = tensor("op_327_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_327_cast_fp16 = einsum(equation = var_327_equation_0, values = (var_219_cast_fp16_13, var_293_cast_fp16))[name = tensor("op_327_cast_fp16")]; + tensor var_329_equation_0 = const()[name = tensor("op_329_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_329_cast_fp16 = einsum(equation = var_329_equation_0, values = (var_219_cast_fp16_14, var_294_cast_fp16))[name = tensor("op_329_cast_fp16")]; + tensor var_331_equation_0 = const()[name = tensor("op_331_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_331_cast_fp16 = einsum(equation = var_331_equation_0, values = (var_219_cast_fp16_15, var_295_cast_fp16))[name = tensor("op_331_cast_fp16")]; + tensor var_333_equation_0 = const()[name = tensor("op_333_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_333_cast_fp16 = einsum(equation = var_333_equation_0, values = (var_219_cast_fp16_16, var_296_cast_fp16))[name = tensor("op_333_cast_fp16")]; + tensor var_335_equation_0 = const()[name = tensor("op_335_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_335_cast_fp16 = einsum(equation = var_335_equation_0, values = (var_219_cast_fp16_17, var_297_cast_fp16))[name = tensor("op_335_cast_fp16")]; + tensor var_337_equation_0 = const()[name = tensor("op_337_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_337_cast_fp16 = einsum(equation = var_337_equation_0, values = (var_219_cast_fp16_18, var_298_cast_fp16))[name = tensor("op_337_cast_fp16")]; + tensor var_339_equation_0 = const()[name = tensor("op_339_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_339_cast_fp16 = einsum(equation = var_339_equation_0, values = (var_219_cast_fp16_19, var_299_cast_fp16))[name = tensor("op_339_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_124, interleave = input_5_interleave_0, values = (var_301_cast_fp16, var_303_cast_fp16, var_305_cast_fp16, var_307_cast_fp16, var_309_cast_fp16, var_311_cast_fp16, var_313_cast_fp16, var_315_cast_fp16, var_317_cast_fp16, var_319_cast_fp16, var_321_cast_fp16, var_323_cast_fp16, var_325_cast_fp16, var_327_cast_fp16, var_329_cast_fp16, var_331_cast_fp16, var_333_cast_fp16, var_335_cast_fp16, var_337_cast_fp16, var_339_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_348_pad_type_0 = const()[name = tensor("op_348_pad_type_0"), val = tensor("valid")]; + tensor var_348_strides_0 = const()[name = tensor("op_348_strides_0"), val = tensor([1, 1])]; + tensor var_348_pad_0 = const()[name = tensor("op_348_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_348_dilations_0 = const()[name = tensor("op_348_dilations_0"), val = tensor([1, 1])]; + tensor var_348_groups_0 = const()[name = tensor("op_348_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24500032)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27776896)))]; + tensor var_348_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_348_dilations_0, groups = var_348_groups_0, pad = var_348_pad_0, pad_type = var_348_pad_type_0, strides = var_348_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_348_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27779520)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27782144)))]; + tensor var_358_to_fp16 = const()[name = tensor("op_358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_358_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27784768)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40892032)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_384_pad_type_0 = const()[name = tensor("op_384_pad_type_0"), val = tensor("valid")]; + tensor var_384_strides_0 = const()[name = tensor("op_384_strides_0"), val = tensor([1, 1])]; + tensor var_384_pad_0 = const()[name = tensor("op_384_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_384_dilations_0 = const()[name = tensor("op_384_dilations_0"), val = tensor([1, 1])]; + tensor var_384_groups_0 = const()[name = tensor("op_384_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40902336)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54009600)))]; + tensor var_384_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_384_dilations_0, groups = var_384_groups_0, pad = var_384_pad_0, pad_type = var_384_pad_type_0, strides = var_384_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_384_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54012224)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54014848)))]; + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_409_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_444_weight_0_to_fp16 = const()[name = tensor("op_444_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54017472)))]; + tensor var_444_bias_0_to_fp16 = const()[name = tensor("op_444_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57294336)))]; + tensor var_444_cast_fp16 = conv(bias = var_444_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_444_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_444_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57296960)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_442_pad_type_0 = const()[name = tensor("op_442_pad_type_0"), val = tensor("valid")]; + tensor var_442_strides_0 = const()[name = tensor("op_442_strides_0"), val = tensor([1, 1])]; + tensor var_442_pad_0 = const()[name = tensor("op_442_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_442_dilations_0 = const()[name = tensor("op_442_dilations_0"), val = tensor([1, 1])]; + tensor var_442_groups_0 = const()[name = tensor("op_442_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60573824)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63850688)))]; + tensor var_442_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_442_dilations_0, groups = var_442_groups_0, pad = var_442_pad_0, pad_type = var_442_pad_type_0, strides = var_442_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_cast_fp16_0, tensor var_445_cast_fp16_1, tensor var_445_cast_fp16_2, tensor var_445_cast_fp16_3, tensor var_445_cast_fp16_4, tensor var_445_cast_fp16_5, tensor var_445_cast_fp16_6, tensor var_445_cast_fp16_7, tensor var_445_cast_fp16_8, tensor var_445_cast_fp16_9, tensor var_445_cast_fp16_10, tensor var_445_cast_fp16_11, tensor var_445_cast_fp16_12, tensor var_445_cast_fp16_13, tensor var_445_cast_fp16_14, tensor var_445_cast_fp16_15, tensor var_445_cast_fp16_16, tensor var_445_cast_fp16_17, tensor var_445_cast_fp16_18, tensor var_445_cast_fp16_19 = split(axis = var_445_axis_0, split_sizes = tile_3, x = var_444_cast_fp16)[name = tensor("op_445_cast_fp16")]; + tensor var_466_perm_0 = const()[name = tensor("op_466_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_467_axis_0 = const()[name = tensor("op_467_axis_0"), val = tensor(3)]; + tensor var_466_cast_fp16 = transpose(perm = var_466_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_31")]; + tensor var_467_cast_fp16_0, tensor var_467_cast_fp16_1, tensor var_467_cast_fp16_2, tensor var_467_cast_fp16_3, tensor var_467_cast_fp16_4, tensor var_467_cast_fp16_5, tensor var_467_cast_fp16_6, tensor var_467_cast_fp16_7, tensor var_467_cast_fp16_8, tensor var_467_cast_fp16_9, tensor var_467_cast_fp16_10, tensor var_467_cast_fp16_11, tensor var_467_cast_fp16_12, tensor var_467_cast_fp16_13, tensor var_467_cast_fp16_14, tensor var_467_cast_fp16_15, tensor var_467_cast_fp16_16, tensor var_467_cast_fp16_17, tensor var_467_cast_fp16_18, tensor var_467_cast_fp16_19 = split(axis = var_467_axis_0, split_sizes = tile_4, x = var_466_cast_fp16)[name = tensor("op_467_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_488_axis_0 = const()[name = tensor("op_488_axis_0"), val = tensor(1)]; + tensor var_488_cast_fp16_0, tensor var_488_cast_fp16_1, tensor var_488_cast_fp16_2, tensor var_488_cast_fp16_3, tensor var_488_cast_fp16_4, tensor var_488_cast_fp16_5, tensor var_488_cast_fp16_6, tensor var_488_cast_fp16_7, tensor var_488_cast_fp16_8, tensor var_488_cast_fp16_9, tensor var_488_cast_fp16_10, tensor var_488_cast_fp16_11, tensor var_488_cast_fp16_12, tensor var_488_cast_fp16_13, tensor var_488_cast_fp16_14, tensor var_488_cast_fp16_15, tensor var_488_cast_fp16_16, tensor var_488_cast_fp16_17, tensor var_488_cast_fp16_18, tensor var_488_cast_fp16_19 = split(axis = var_488_axis_0, split_sizes = tile_5, x = var_442_cast_fp16)[name = tensor("op_488_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_467_cast_fp16_0, var_445_cast_fp16_0))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_467_cast_fp16_1, var_445_cast_fp16_1))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_467_cast_fp16_2, var_445_cast_fp16_2))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_467_cast_fp16_3, var_445_cast_fp16_3))[name = tensor("aw_47_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_467_cast_fp16_4, var_445_cast_fp16_4))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_467_cast_fp16_5, var_445_cast_fp16_5))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_467_cast_fp16_6, var_445_cast_fp16_6))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_467_cast_fp16_7, var_445_cast_fp16_7))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_467_cast_fp16_8, var_445_cast_fp16_8))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_467_cast_fp16_9, var_445_cast_fp16_9))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_467_cast_fp16_10, var_445_cast_fp16_10))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_467_cast_fp16_11, var_445_cast_fp16_11))[name = tensor("aw_63_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_467_cast_fp16_12, var_445_cast_fp16_12))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_467_cast_fp16_13, var_445_cast_fp16_13))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_467_cast_fp16_14, var_445_cast_fp16_14))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_467_cast_fp16_15, var_445_cast_fp16_15))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_467_cast_fp16_16, var_445_cast_fp16_16))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_467_cast_fp16_17, var_445_cast_fp16_17))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_467_cast_fp16_18, var_445_cast_fp16_18))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_467_cast_fp16_19, var_445_cast_fp16_19))[name = tensor("aw_79_cast_fp16")]; + tensor var_549_cast_fp16 = softmax(axis = var_393, x = aw_41_cast_fp16)[name = tensor("op_549_cast_fp16")]; + tensor var_550_cast_fp16 = softmax(axis = var_393, x = aw_43_cast_fp16)[name = tensor("op_550_cast_fp16")]; + tensor var_551_cast_fp16 = softmax(axis = var_393, x = aw_45_cast_fp16)[name = tensor("op_551_cast_fp16")]; + tensor var_552_cast_fp16 = softmax(axis = var_393, x = aw_47_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = softmax(axis = var_393, x = aw_49_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor var_554_cast_fp16 = softmax(axis = var_393, x = aw_51_cast_fp16)[name = tensor("op_554_cast_fp16")]; + tensor var_555_cast_fp16 = softmax(axis = var_393, x = aw_53_cast_fp16)[name = tensor("op_555_cast_fp16")]; + tensor var_556_cast_fp16 = softmax(axis = var_393, x = aw_55_cast_fp16)[name = tensor("op_556_cast_fp16")]; + tensor var_557_cast_fp16 = softmax(axis = var_393, x = aw_57_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_558_cast_fp16 = softmax(axis = var_393, x = aw_59_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor var_559_cast_fp16 = softmax(axis = var_393, x = aw_61_cast_fp16)[name = tensor("op_559_cast_fp16")]; + tensor var_560_cast_fp16 = softmax(axis = var_393, x = aw_63_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_561_cast_fp16 = softmax(axis = var_393, x = aw_65_cast_fp16)[name = tensor("op_561_cast_fp16")]; + tensor var_562_cast_fp16 = softmax(axis = var_393, x = aw_67_cast_fp16)[name = tensor("op_562_cast_fp16")]; + tensor var_563_cast_fp16 = softmax(axis = var_393, x = aw_69_cast_fp16)[name = tensor("op_563_cast_fp16")]; + tensor var_564_cast_fp16 = softmax(axis = var_393, x = aw_71_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor var_565_cast_fp16 = softmax(axis = var_393, x = aw_73_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_566_cast_fp16 = softmax(axis = var_393, x = aw_75_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = softmax(axis = var_393, x = aw_77_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor var_568_cast_fp16 = softmax(axis = var_393, x = aw_79_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor var_570_equation_0 = const()[name = tensor("op_570_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_570_cast_fp16 = einsum(equation = var_570_equation_0, values = (var_488_cast_fp16_0, var_549_cast_fp16))[name = tensor("op_570_cast_fp16")]; + tensor var_572_equation_0 = const()[name = tensor("op_572_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_572_cast_fp16 = einsum(equation = var_572_equation_0, values = (var_488_cast_fp16_1, var_550_cast_fp16))[name = tensor("op_572_cast_fp16")]; + tensor var_574_equation_0 = const()[name = tensor("op_574_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_574_cast_fp16 = einsum(equation = var_574_equation_0, values = (var_488_cast_fp16_2, var_551_cast_fp16))[name = tensor("op_574_cast_fp16")]; + tensor var_576_equation_0 = const()[name = tensor("op_576_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_576_cast_fp16 = einsum(equation = var_576_equation_0, values = (var_488_cast_fp16_3, var_552_cast_fp16))[name = tensor("op_576_cast_fp16")]; + tensor var_578_equation_0 = const()[name = tensor("op_578_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_578_cast_fp16 = einsum(equation = var_578_equation_0, values = (var_488_cast_fp16_4, var_553_cast_fp16))[name = tensor("op_578_cast_fp16")]; + tensor var_580_equation_0 = const()[name = tensor("op_580_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_580_cast_fp16 = einsum(equation = var_580_equation_0, values = (var_488_cast_fp16_5, var_554_cast_fp16))[name = tensor("op_580_cast_fp16")]; + tensor var_582_equation_0 = const()[name = tensor("op_582_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_582_cast_fp16 = einsum(equation = var_582_equation_0, values = (var_488_cast_fp16_6, var_555_cast_fp16))[name = tensor("op_582_cast_fp16")]; + tensor var_584_equation_0 = const()[name = tensor("op_584_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_584_cast_fp16 = einsum(equation = var_584_equation_0, values = (var_488_cast_fp16_7, var_556_cast_fp16))[name = tensor("op_584_cast_fp16")]; + tensor var_586_equation_0 = const()[name = tensor("op_586_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_586_cast_fp16 = einsum(equation = var_586_equation_0, values = (var_488_cast_fp16_8, var_557_cast_fp16))[name = tensor("op_586_cast_fp16")]; + tensor var_588_equation_0 = const()[name = tensor("op_588_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_588_cast_fp16 = einsum(equation = var_588_equation_0, values = (var_488_cast_fp16_9, var_558_cast_fp16))[name = tensor("op_588_cast_fp16")]; + tensor var_590_equation_0 = const()[name = tensor("op_590_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_590_cast_fp16 = einsum(equation = var_590_equation_0, values = (var_488_cast_fp16_10, var_559_cast_fp16))[name = tensor("op_590_cast_fp16")]; + tensor var_592_equation_0 = const()[name = tensor("op_592_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_592_cast_fp16 = einsum(equation = var_592_equation_0, values = (var_488_cast_fp16_11, var_560_cast_fp16))[name = tensor("op_592_cast_fp16")]; + tensor var_594_equation_0 = const()[name = tensor("op_594_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_594_cast_fp16 = einsum(equation = var_594_equation_0, values = (var_488_cast_fp16_12, var_561_cast_fp16))[name = tensor("op_594_cast_fp16")]; + tensor var_596_equation_0 = const()[name = tensor("op_596_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_596_cast_fp16 = einsum(equation = var_596_equation_0, values = (var_488_cast_fp16_13, var_562_cast_fp16))[name = tensor("op_596_cast_fp16")]; + tensor var_598_equation_0 = const()[name = tensor("op_598_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_598_cast_fp16 = einsum(equation = var_598_equation_0, values = (var_488_cast_fp16_14, var_563_cast_fp16))[name = tensor("op_598_cast_fp16")]; + tensor var_600_equation_0 = const()[name = tensor("op_600_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_600_cast_fp16 = einsum(equation = var_600_equation_0, values = (var_488_cast_fp16_15, var_564_cast_fp16))[name = tensor("op_600_cast_fp16")]; + tensor var_602_equation_0 = const()[name = tensor("op_602_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_602_cast_fp16 = einsum(equation = var_602_equation_0, values = (var_488_cast_fp16_16, var_565_cast_fp16))[name = tensor("op_602_cast_fp16")]; + tensor var_604_equation_0 = const()[name = tensor("op_604_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_604_cast_fp16 = einsum(equation = var_604_equation_0, values = (var_488_cast_fp16_17, var_566_cast_fp16))[name = tensor("op_604_cast_fp16")]; + tensor var_606_equation_0 = const()[name = tensor("op_606_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_606_cast_fp16 = einsum(equation = var_606_equation_0, values = (var_488_cast_fp16_18, var_567_cast_fp16))[name = tensor("op_606_cast_fp16")]; + tensor var_608_equation_0 = const()[name = tensor("op_608_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_608_cast_fp16 = einsum(equation = var_608_equation_0, values = (var_488_cast_fp16_19, var_568_cast_fp16))[name = tensor("op_608_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_393, interleave = input_15_interleave_0, values = (var_570_cast_fp16, var_572_cast_fp16, var_574_cast_fp16, var_576_cast_fp16, var_578_cast_fp16, var_580_cast_fp16, var_582_cast_fp16, var_584_cast_fp16, var_586_cast_fp16, var_588_cast_fp16, var_590_cast_fp16, var_592_cast_fp16, var_594_cast_fp16, var_596_cast_fp16, var_598_cast_fp16, var_600_cast_fp16, var_602_cast_fp16, var_604_cast_fp16, var_606_cast_fp16, var_608_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_617_pad_type_0 = const()[name = tensor("op_617_pad_type_0"), val = tensor("valid")]; + tensor var_617_strides_0 = const()[name = tensor("op_617_strides_0"), val = tensor([1, 1])]; + tensor var_617_pad_0 = const()[name = tensor("op_617_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_617_dilations_0 = const()[name = tensor("op_617_dilations_0"), val = tensor([1, 1])]; + tensor var_617_groups_0 = const()[name = tensor("op_617_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63853312)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67130176)))]; + tensor var_617_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_617_dilations_0, groups = var_617_groups_0, pad = var_617_pad_0, pad_type = var_617_pad_type_0, strides = var_617_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_617_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67132800)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67135424)))]; + tensor var_627_to_fp16 = const()[name = tensor("op_627_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_627_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67138048)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80245312)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_653_pad_type_0 = const()[name = tensor("op_653_pad_type_0"), val = tensor("valid")]; + tensor var_653_strides_0 = const()[name = tensor("op_653_strides_0"), val = tensor([1, 1])]; + tensor var_653_pad_0 = const()[name = tensor("op_653_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_653_dilations_0 = const()[name = tensor("op_653_dilations_0"), val = tensor([1, 1])]; + tensor var_653_groups_0 = const()[name = tensor("op_653_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80255616)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93362880)))]; + tensor var_653_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_653_dilations_0, groups = var_653_groups_0, pad = var_653_pad_0, pad_type = var_653_pad_type_0, strides = var_653_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_653_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_662 = const()[name = tensor("op_662"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93365504)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93368128)))]; + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_678_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_713_weight_0_to_fp16 = const()[name = tensor("op_713_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93370752)))]; + tensor var_713_bias_0_to_fp16 = const()[name = tensor("op_713_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96647616)))]; + tensor var_713_cast_fp16 = conv(bias = var_713_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_713_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96650240)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_711_pad_type_0 = const()[name = tensor("op_711_pad_type_0"), val = tensor("valid")]; + tensor var_711_strides_0 = const()[name = tensor("op_711_strides_0"), val = tensor([1, 1])]; + tensor var_711_pad_0 = const()[name = tensor("op_711_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_711_dilations_0 = const()[name = tensor("op_711_dilations_0"), val = tensor([1, 1])]; + tensor var_711_groups_0 = const()[name = tensor("op_711_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99927104)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103203968)))]; + tensor var_711_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_711_dilations_0, groups = var_711_groups_0, pad = var_711_pad_0, pad_type = var_711_pad_type_0, strides = var_711_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_711_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_714_axis_0 = const()[name = tensor("op_714_axis_0"), val = tensor(1)]; + tensor var_714_cast_fp16_0, tensor var_714_cast_fp16_1, tensor var_714_cast_fp16_2, tensor var_714_cast_fp16_3, tensor var_714_cast_fp16_4, tensor var_714_cast_fp16_5, tensor var_714_cast_fp16_6, tensor var_714_cast_fp16_7, tensor var_714_cast_fp16_8, tensor var_714_cast_fp16_9, tensor var_714_cast_fp16_10, tensor var_714_cast_fp16_11, tensor var_714_cast_fp16_12, tensor var_714_cast_fp16_13, tensor var_714_cast_fp16_14, tensor var_714_cast_fp16_15, tensor var_714_cast_fp16_16, tensor var_714_cast_fp16_17, tensor var_714_cast_fp16_18, tensor var_714_cast_fp16_19 = split(axis = var_714_axis_0, split_sizes = tile_6, x = var_713_cast_fp16)[name = tensor("op_714_cast_fp16")]; + tensor var_735_perm_0 = const()[name = tensor("op_735_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_736_axis_0 = const()[name = tensor("op_736_axis_0"), val = tensor(3)]; + tensor var_735_cast_fp16 = transpose(perm = var_735_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_30")]; + tensor var_736_cast_fp16_0, tensor var_736_cast_fp16_1, tensor var_736_cast_fp16_2, tensor var_736_cast_fp16_3, tensor var_736_cast_fp16_4, tensor var_736_cast_fp16_5, tensor var_736_cast_fp16_6, tensor var_736_cast_fp16_7, tensor var_736_cast_fp16_8, tensor var_736_cast_fp16_9, tensor var_736_cast_fp16_10, tensor var_736_cast_fp16_11, tensor var_736_cast_fp16_12, tensor var_736_cast_fp16_13, tensor var_736_cast_fp16_14, tensor var_736_cast_fp16_15, tensor var_736_cast_fp16_16, tensor var_736_cast_fp16_17, tensor var_736_cast_fp16_18, tensor var_736_cast_fp16_19 = split(axis = var_736_axis_0, split_sizes = tile_7, x = var_735_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_cast_fp16_0, tensor var_757_cast_fp16_1, tensor var_757_cast_fp16_2, tensor var_757_cast_fp16_3, tensor var_757_cast_fp16_4, tensor var_757_cast_fp16_5, tensor var_757_cast_fp16_6, tensor var_757_cast_fp16_7, tensor var_757_cast_fp16_8, tensor var_757_cast_fp16_9, tensor var_757_cast_fp16_10, tensor var_757_cast_fp16_11, tensor var_757_cast_fp16_12, tensor var_757_cast_fp16_13, tensor var_757_cast_fp16_14, tensor var_757_cast_fp16_15, tensor var_757_cast_fp16_16, tensor var_757_cast_fp16_17, tensor var_757_cast_fp16_18, tensor var_757_cast_fp16_19 = split(axis = var_757_axis_0, split_sizes = tile_8, x = var_711_cast_fp16)[name = tensor("op_757_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_736_cast_fp16_0, var_714_cast_fp16_0))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_736_cast_fp16_1, var_714_cast_fp16_1))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_736_cast_fp16_2, var_714_cast_fp16_2))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_736_cast_fp16_3, var_714_cast_fp16_3))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_736_cast_fp16_4, var_714_cast_fp16_4))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_736_cast_fp16_5, var_714_cast_fp16_5))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_736_cast_fp16_6, var_714_cast_fp16_6))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_736_cast_fp16_7, var_714_cast_fp16_7))[name = tensor("aw_95_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_736_cast_fp16_8, var_714_cast_fp16_8))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_736_cast_fp16_9, var_714_cast_fp16_9))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_736_cast_fp16_10, var_714_cast_fp16_10))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_736_cast_fp16_11, var_714_cast_fp16_11))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_736_cast_fp16_12, var_714_cast_fp16_12))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_736_cast_fp16_13, var_714_cast_fp16_13))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_736_cast_fp16_14, var_714_cast_fp16_14))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_736_cast_fp16_15, var_714_cast_fp16_15))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_736_cast_fp16_16, var_714_cast_fp16_16))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_736_cast_fp16_17, var_714_cast_fp16_17))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_736_cast_fp16_18, var_714_cast_fp16_18))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_736_cast_fp16_19, var_714_cast_fp16_19))[name = tensor("aw_119_cast_fp16")]; + tensor var_818_cast_fp16 = softmax(axis = var_662, x = aw_81_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_819_cast_fp16 = softmax(axis = var_662, x = aw_83_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820_cast_fp16 = softmax(axis = var_662, x = aw_85_cast_fp16)[name = tensor("op_820_cast_fp16")]; + tensor var_821_cast_fp16 = softmax(axis = var_662, x = aw_87_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = softmax(axis = var_662, x = aw_89_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor var_823_cast_fp16 = softmax(axis = var_662, x = aw_91_cast_fp16)[name = tensor("op_823_cast_fp16")]; + tensor var_824_cast_fp16 = softmax(axis = var_662, x = aw_93_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor var_825_cast_fp16 = softmax(axis = var_662, x = aw_95_cast_fp16)[name = tensor("op_825_cast_fp16")]; + tensor var_826_cast_fp16 = softmax(axis = var_662, x = aw_97_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_827_cast_fp16 = softmax(axis = var_662, x = aw_99_cast_fp16)[name = tensor("op_827_cast_fp16")]; + tensor var_828_cast_fp16 = softmax(axis = var_662, x = aw_101_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_cast_fp16 = softmax(axis = var_662, x = aw_103_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_830_cast_fp16 = softmax(axis = var_662, x = aw_105_cast_fp16)[name = tensor("op_830_cast_fp16")]; + tensor var_831_cast_fp16 = softmax(axis = var_662, x = aw_107_cast_fp16)[name = tensor("op_831_cast_fp16")]; + tensor var_832_cast_fp16 = softmax(axis = var_662, x = aw_109_cast_fp16)[name = tensor("op_832_cast_fp16")]; + tensor var_833_cast_fp16 = softmax(axis = var_662, x = aw_111_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_834_cast_fp16 = softmax(axis = var_662, x = aw_113_cast_fp16)[name = tensor("op_834_cast_fp16")]; + tensor var_835_cast_fp16 = softmax(axis = var_662, x = aw_115_cast_fp16)[name = tensor("op_835_cast_fp16")]; + tensor var_836_cast_fp16 = softmax(axis = var_662, x = aw_117_cast_fp16)[name = tensor("op_836_cast_fp16")]; + tensor var_837_cast_fp16 = softmax(axis = var_662, x = aw_119_cast_fp16)[name = tensor("op_837_cast_fp16")]; + tensor var_839_equation_0 = const()[name = tensor("op_839_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_839_cast_fp16 = einsum(equation = var_839_equation_0, values = (var_757_cast_fp16_0, var_818_cast_fp16))[name = tensor("op_839_cast_fp16")]; + tensor var_841_equation_0 = const()[name = tensor("op_841_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_841_cast_fp16 = einsum(equation = var_841_equation_0, values = (var_757_cast_fp16_1, var_819_cast_fp16))[name = tensor("op_841_cast_fp16")]; + tensor var_843_equation_0 = const()[name = tensor("op_843_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_843_cast_fp16 = einsum(equation = var_843_equation_0, values = (var_757_cast_fp16_2, var_820_cast_fp16))[name = tensor("op_843_cast_fp16")]; + tensor var_845_equation_0 = const()[name = tensor("op_845_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_845_cast_fp16 = einsum(equation = var_845_equation_0, values = (var_757_cast_fp16_3, var_821_cast_fp16))[name = tensor("op_845_cast_fp16")]; + tensor var_847_equation_0 = const()[name = tensor("op_847_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_847_cast_fp16 = einsum(equation = var_847_equation_0, values = (var_757_cast_fp16_4, var_822_cast_fp16))[name = tensor("op_847_cast_fp16")]; + tensor var_849_equation_0 = const()[name = tensor("op_849_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_849_cast_fp16 = einsum(equation = var_849_equation_0, values = (var_757_cast_fp16_5, var_823_cast_fp16))[name = tensor("op_849_cast_fp16")]; + tensor var_851_equation_0 = const()[name = tensor("op_851_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_851_cast_fp16 = einsum(equation = var_851_equation_0, values = (var_757_cast_fp16_6, var_824_cast_fp16))[name = tensor("op_851_cast_fp16")]; + tensor var_853_equation_0 = const()[name = tensor("op_853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_853_cast_fp16 = einsum(equation = var_853_equation_0, values = (var_757_cast_fp16_7, var_825_cast_fp16))[name = tensor("op_853_cast_fp16")]; + tensor var_855_equation_0 = const()[name = tensor("op_855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_855_cast_fp16 = einsum(equation = var_855_equation_0, values = (var_757_cast_fp16_8, var_826_cast_fp16))[name = tensor("op_855_cast_fp16")]; + tensor var_857_equation_0 = const()[name = tensor("op_857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_857_cast_fp16 = einsum(equation = var_857_equation_0, values = (var_757_cast_fp16_9, var_827_cast_fp16))[name = tensor("op_857_cast_fp16")]; + tensor var_859_equation_0 = const()[name = tensor("op_859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_859_cast_fp16 = einsum(equation = var_859_equation_0, values = (var_757_cast_fp16_10, var_828_cast_fp16))[name = tensor("op_859_cast_fp16")]; + tensor var_861_equation_0 = const()[name = tensor("op_861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_861_cast_fp16 = einsum(equation = var_861_equation_0, values = (var_757_cast_fp16_11, var_829_cast_fp16))[name = tensor("op_861_cast_fp16")]; + tensor var_863_equation_0 = const()[name = tensor("op_863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_863_cast_fp16 = einsum(equation = var_863_equation_0, values = (var_757_cast_fp16_12, var_830_cast_fp16))[name = tensor("op_863_cast_fp16")]; + tensor var_865_equation_0 = const()[name = tensor("op_865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_865_cast_fp16 = einsum(equation = var_865_equation_0, values = (var_757_cast_fp16_13, var_831_cast_fp16))[name = tensor("op_865_cast_fp16")]; + tensor var_867_equation_0 = const()[name = tensor("op_867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_867_cast_fp16 = einsum(equation = var_867_equation_0, values = (var_757_cast_fp16_14, var_832_cast_fp16))[name = tensor("op_867_cast_fp16")]; + tensor var_869_equation_0 = const()[name = tensor("op_869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_869_cast_fp16 = einsum(equation = var_869_equation_0, values = (var_757_cast_fp16_15, var_833_cast_fp16))[name = tensor("op_869_cast_fp16")]; + tensor var_871_equation_0 = const()[name = tensor("op_871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_871_cast_fp16 = einsum(equation = var_871_equation_0, values = (var_757_cast_fp16_16, var_834_cast_fp16))[name = tensor("op_871_cast_fp16")]; + tensor var_873_equation_0 = const()[name = tensor("op_873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_873_cast_fp16 = einsum(equation = var_873_equation_0, values = (var_757_cast_fp16_17, var_835_cast_fp16))[name = tensor("op_873_cast_fp16")]; + tensor var_875_equation_0 = const()[name = tensor("op_875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_875_cast_fp16 = einsum(equation = var_875_equation_0, values = (var_757_cast_fp16_18, var_836_cast_fp16))[name = tensor("op_875_cast_fp16")]; + tensor var_877_equation_0 = const()[name = tensor("op_877_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_877_cast_fp16 = einsum(equation = var_877_equation_0, values = (var_757_cast_fp16_19, var_837_cast_fp16))[name = tensor("op_877_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_662, interleave = input_25_interleave_0, values = (var_839_cast_fp16, var_841_cast_fp16, var_843_cast_fp16, var_845_cast_fp16, var_847_cast_fp16, var_849_cast_fp16, var_851_cast_fp16, var_853_cast_fp16, var_855_cast_fp16, var_857_cast_fp16, var_859_cast_fp16, var_861_cast_fp16, var_863_cast_fp16, var_865_cast_fp16, var_867_cast_fp16, var_869_cast_fp16, var_871_cast_fp16, var_873_cast_fp16, var_875_cast_fp16, var_877_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_886_pad_type_0 = const()[name = tensor("op_886_pad_type_0"), val = tensor("valid")]; + tensor var_886_strides_0 = const()[name = tensor("op_886_strides_0"), val = tensor([1, 1])]; + tensor var_886_pad_0 = const()[name = tensor("op_886_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_886_dilations_0 = const()[name = tensor("op_886_dilations_0"), val = tensor([1, 1])]; + tensor var_886_groups_0 = const()[name = tensor("op_886_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103206592)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106483456)))]; + tensor var_886_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_886_dilations_0, groups = var_886_groups_0, pad = var_886_pad_0, pad_type = var_886_pad_type_0, strides = var_886_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_886_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106486080)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106488704)))]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_896_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106491328)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119598592)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_922_pad_type_0 = const()[name = tensor("op_922_pad_type_0"), val = tensor("valid")]; + tensor var_922_strides_0 = const()[name = tensor("op_922_strides_0"), val = tensor([1, 1])]; + tensor var_922_pad_0 = const()[name = tensor("op_922_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_922_dilations_0 = const()[name = tensor("op_922_dilations_0"), val = tensor([1, 1])]; + tensor var_922_groups_0 = const()[name = tensor("op_922_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119608896)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132716160)))]; + tensor var_922_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_922_dilations_0, groups = var_922_groups_0, pad = var_922_pad_0, pad_type = var_922_pad_type_0, strides = var_922_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_922_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132718784)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132721408)))]; + tensor var_947_to_fp16 = const()[name = tensor("op_947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_947_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_982_weight_0_to_fp16 = const()[name = tensor("op_982_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132724032)))]; + tensor var_982_bias_0_to_fp16 = const()[name = tensor("op_982_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136000896)))]; + tensor var_982_cast_fp16 = conv(bias = var_982_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_982_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_982_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136003520)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_980_pad_type_0 = const()[name = tensor("op_980_pad_type_0"), val = tensor("valid")]; + tensor var_980_strides_0 = const()[name = tensor("op_980_strides_0"), val = tensor([1, 1])]; + tensor var_980_pad_0 = const()[name = tensor("op_980_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_980_dilations_0 = const()[name = tensor("op_980_dilations_0"), val = tensor([1, 1])]; + tensor var_980_groups_0 = const()[name = tensor("op_980_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139280384)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142557248)))]; + tensor var_980_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_980_dilations_0, groups = var_980_groups_0, pad = var_980_pad_0, pad_type = var_980_pad_type_0, strides = var_980_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_980_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_983_axis_0 = const()[name = tensor("op_983_axis_0"), val = tensor(1)]; + tensor var_983_cast_fp16_0, tensor var_983_cast_fp16_1, tensor var_983_cast_fp16_2, tensor var_983_cast_fp16_3, tensor var_983_cast_fp16_4, tensor var_983_cast_fp16_5, tensor var_983_cast_fp16_6, tensor var_983_cast_fp16_7, tensor var_983_cast_fp16_8, tensor var_983_cast_fp16_9, tensor var_983_cast_fp16_10, tensor var_983_cast_fp16_11, tensor var_983_cast_fp16_12, tensor var_983_cast_fp16_13, tensor var_983_cast_fp16_14, tensor var_983_cast_fp16_15, tensor var_983_cast_fp16_16, tensor var_983_cast_fp16_17, tensor var_983_cast_fp16_18, tensor var_983_cast_fp16_19 = split(axis = var_983_axis_0, split_sizes = tile_9, x = var_982_cast_fp16)[name = tensor("op_983_cast_fp16")]; + tensor var_1004_perm_0 = const()[name = tensor("op_1004_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1005_axis_0 = const()[name = tensor("op_1005_axis_0"), val = tensor(3)]; + tensor var_1004_cast_fp16 = transpose(perm = var_1004_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_29")]; + tensor var_1005_cast_fp16_0, tensor var_1005_cast_fp16_1, tensor var_1005_cast_fp16_2, tensor var_1005_cast_fp16_3, tensor var_1005_cast_fp16_4, tensor var_1005_cast_fp16_5, tensor var_1005_cast_fp16_6, tensor var_1005_cast_fp16_7, tensor var_1005_cast_fp16_8, tensor var_1005_cast_fp16_9, tensor var_1005_cast_fp16_10, tensor var_1005_cast_fp16_11, tensor var_1005_cast_fp16_12, tensor var_1005_cast_fp16_13, tensor var_1005_cast_fp16_14, tensor var_1005_cast_fp16_15, tensor var_1005_cast_fp16_16, tensor var_1005_cast_fp16_17, tensor var_1005_cast_fp16_18, tensor var_1005_cast_fp16_19 = split(axis = var_1005_axis_0, split_sizes = tile_10, x = var_1004_cast_fp16)[name = tensor("op_1005_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1026_axis_0 = const()[name = tensor("op_1026_axis_0"), val = tensor(1)]; + tensor var_1026_cast_fp16_0, tensor var_1026_cast_fp16_1, tensor var_1026_cast_fp16_2, tensor var_1026_cast_fp16_3, tensor var_1026_cast_fp16_4, tensor var_1026_cast_fp16_5, tensor var_1026_cast_fp16_6, tensor var_1026_cast_fp16_7, tensor var_1026_cast_fp16_8, tensor var_1026_cast_fp16_9, tensor var_1026_cast_fp16_10, tensor var_1026_cast_fp16_11, tensor var_1026_cast_fp16_12, tensor var_1026_cast_fp16_13, tensor var_1026_cast_fp16_14, tensor var_1026_cast_fp16_15, tensor var_1026_cast_fp16_16, tensor var_1026_cast_fp16_17, tensor var_1026_cast_fp16_18, tensor var_1026_cast_fp16_19 = split(axis = var_1026_axis_0, split_sizes = tile_11, x = var_980_cast_fp16)[name = tensor("op_1026_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_1005_cast_fp16_0, var_983_cast_fp16_0))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_1005_cast_fp16_1, var_983_cast_fp16_1))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_1005_cast_fp16_2, var_983_cast_fp16_2))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_1005_cast_fp16_3, var_983_cast_fp16_3))[name = tensor("aw_127_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1005_cast_fp16_4, var_983_cast_fp16_4))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1005_cast_fp16_5, var_983_cast_fp16_5))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1005_cast_fp16_6, var_983_cast_fp16_6))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1005_cast_fp16_7, var_983_cast_fp16_7))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1005_cast_fp16_8, var_983_cast_fp16_8))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1005_cast_fp16_9, var_983_cast_fp16_9))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1005_cast_fp16_10, var_983_cast_fp16_10))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1005_cast_fp16_11, var_983_cast_fp16_11))[name = tensor("aw_143_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1005_cast_fp16_12, var_983_cast_fp16_12))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1005_cast_fp16_13, var_983_cast_fp16_13))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1005_cast_fp16_14, var_983_cast_fp16_14))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1005_cast_fp16_15, var_983_cast_fp16_15))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1005_cast_fp16_16, var_983_cast_fp16_16))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1005_cast_fp16_17, var_983_cast_fp16_17))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1005_cast_fp16_18, var_983_cast_fp16_18))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1005_cast_fp16_19, var_983_cast_fp16_19))[name = tensor("aw_159_cast_fp16")]; + tensor var_1087_cast_fp16 = softmax(axis = var_931, x = aw_121_cast_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor var_1088_cast_fp16 = softmax(axis = var_931, x = aw_123_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = softmax(axis = var_931, x = aw_125_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor var_1090_cast_fp16 = softmax(axis = var_931, x = aw_127_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1091_cast_fp16 = softmax(axis = var_931, x = aw_129_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor var_1092_cast_fp16 = softmax(axis = var_931, x = aw_131_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = softmax(axis = var_931, x = aw_133_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor var_1094_cast_fp16 = softmax(axis = var_931, x = aw_135_cast_fp16)[name = tensor("op_1094_cast_fp16")]; + tensor var_1095_cast_fp16 = softmax(axis = var_931, x = aw_137_cast_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = softmax(axis = var_931, x = aw_139_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = softmax(axis = var_931, x = aw_141_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor var_1098_cast_fp16 = softmax(axis = var_931, x = aw_143_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = softmax(axis = var_931, x = aw_145_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor var_1100_cast_fp16 = softmax(axis = var_931, x = aw_147_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = softmax(axis = var_931, x = aw_149_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = softmax(axis = var_931, x = aw_151_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor var_1103_cast_fp16 = softmax(axis = var_931, x = aw_153_cast_fp16)[name = tensor("op_1103_cast_fp16")]; + tensor var_1104_cast_fp16 = softmax(axis = var_931, x = aw_155_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = softmax(axis = var_931, x = aw_157_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = softmax(axis = var_931, x = aw_159_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor var_1108_equation_0 = const()[name = tensor("op_1108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1108_cast_fp16 = einsum(equation = var_1108_equation_0, values = (var_1026_cast_fp16_0, var_1087_cast_fp16))[name = tensor("op_1108_cast_fp16")]; + tensor var_1110_equation_0 = const()[name = tensor("op_1110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1110_cast_fp16 = einsum(equation = var_1110_equation_0, values = (var_1026_cast_fp16_1, var_1088_cast_fp16))[name = tensor("op_1110_cast_fp16")]; + tensor var_1112_equation_0 = const()[name = tensor("op_1112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1112_cast_fp16 = einsum(equation = var_1112_equation_0, values = (var_1026_cast_fp16_2, var_1089_cast_fp16))[name = tensor("op_1112_cast_fp16")]; + tensor var_1114_equation_0 = const()[name = tensor("op_1114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1114_cast_fp16 = einsum(equation = var_1114_equation_0, values = (var_1026_cast_fp16_3, var_1090_cast_fp16))[name = tensor("op_1114_cast_fp16")]; + tensor var_1116_equation_0 = const()[name = tensor("op_1116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1116_cast_fp16 = einsum(equation = var_1116_equation_0, values = (var_1026_cast_fp16_4, var_1091_cast_fp16))[name = tensor("op_1116_cast_fp16")]; + tensor var_1118_equation_0 = const()[name = tensor("op_1118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1118_cast_fp16 = einsum(equation = var_1118_equation_0, values = (var_1026_cast_fp16_5, var_1092_cast_fp16))[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_equation_0 = const()[name = tensor("op_1120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1120_cast_fp16 = einsum(equation = var_1120_equation_0, values = (var_1026_cast_fp16_6, var_1093_cast_fp16))[name = tensor("op_1120_cast_fp16")]; + tensor var_1122_equation_0 = const()[name = tensor("op_1122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1122_cast_fp16 = einsum(equation = var_1122_equation_0, values = (var_1026_cast_fp16_7, var_1094_cast_fp16))[name = tensor("op_1122_cast_fp16")]; + tensor var_1124_equation_0 = const()[name = tensor("op_1124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1124_cast_fp16 = einsum(equation = var_1124_equation_0, values = (var_1026_cast_fp16_8, var_1095_cast_fp16))[name = tensor("op_1124_cast_fp16")]; + tensor var_1126_equation_0 = const()[name = tensor("op_1126_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1126_cast_fp16 = einsum(equation = var_1126_equation_0, values = (var_1026_cast_fp16_9, var_1096_cast_fp16))[name = tensor("op_1126_cast_fp16")]; + tensor var_1128_equation_0 = const()[name = tensor("op_1128_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1128_cast_fp16 = einsum(equation = var_1128_equation_0, values = (var_1026_cast_fp16_10, var_1097_cast_fp16))[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_equation_0 = const()[name = tensor("op_1130_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1130_cast_fp16 = einsum(equation = var_1130_equation_0, values = (var_1026_cast_fp16_11, var_1098_cast_fp16))[name = tensor("op_1130_cast_fp16")]; + tensor var_1132_equation_0 = const()[name = tensor("op_1132_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1132_cast_fp16 = einsum(equation = var_1132_equation_0, values = (var_1026_cast_fp16_12, var_1099_cast_fp16))[name = tensor("op_1132_cast_fp16")]; + tensor var_1134_equation_0 = const()[name = tensor("op_1134_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1134_cast_fp16 = einsum(equation = var_1134_equation_0, values = (var_1026_cast_fp16_13, var_1100_cast_fp16))[name = tensor("op_1134_cast_fp16")]; + tensor var_1136_equation_0 = const()[name = tensor("op_1136_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1136_cast_fp16 = einsum(equation = var_1136_equation_0, values = (var_1026_cast_fp16_14, var_1101_cast_fp16))[name = tensor("op_1136_cast_fp16")]; + tensor var_1138_equation_0 = const()[name = tensor("op_1138_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1138_cast_fp16 = einsum(equation = var_1138_equation_0, values = (var_1026_cast_fp16_15, var_1102_cast_fp16))[name = tensor("op_1138_cast_fp16")]; + tensor var_1140_equation_0 = const()[name = tensor("op_1140_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1140_cast_fp16 = einsum(equation = var_1140_equation_0, values = (var_1026_cast_fp16_16, var_1103_cast_fp16))[name = tensor("op_1140_cast_fp16")]; + tensor var_1142_equation_0 = const()[name = tensor("op_1142_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1142_cast_fp16 = einsum(equation = var_1142_equation_0, values = (var_1026_cast_fp16_17, var_1104_cast_fp16))[name = tensor("op_1142_cast_fp16")]; + tensor var_1144_equation_0 = const()[name = tensor("op_1144_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1144_cast_fp16 = einsum(equation = var_1144_equation_0, values = (var_1026_cast_fp16_18, var_1105_cast_fp16))[name = tensor("op_1144_cast_fp16")]; + tensor var_1146_equation_0 = const()[name = tensor("op_1146_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1146_cast_fp16 = einsum(equation = var_1146_equation_0, values = (var_1026_cast_fp16_19, var_1106_cast_fp16))[name = tensor("op_1146_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_931, interleave = input_35_interleave_0, values = (var_1108_cast_fp16, var_1110_cast_fp16, var_1112_cast_fp16, var_1114_cast_fp16, var_1116_cast_fp16, var_1118_cast_fp16, var_1120_cast_fp16, var_1122_cast_fp16, var_1124_cast_fp16, var_1126_cast_fp16, var_1128_cast_fp16, var_1130_cast_fp16, var_1132_cast_fp16, var_1134_cast_fp16, var_1136_cast_fp16, var_1138_cast_fp16, var_1140_cast_fp16, var_1142_cast_fp16, var_1144_cast_fp16, var_1146_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_1155_pad_type_0 = const()[name = tensor("op_1155_pad_type_0"), val = tensor("valid")]; + tensor var_1155_strides_0 = const()[name = tensor("op_1155_strides_0"), val = tensor([1, 1])]; + tensor var_1155_pad_0 = const()[name = tensor("op_1155_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1155_dilations_0 = const()[name = tensor("op_1155_dilations_0"), val = tensor([1, 1])]; + tensor var_1155_groups_0 = const()[name = tensor("op_1155_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142559872)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145836736)))]; + tensor var_1155_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_1155_dilations_0, groups = var_1155_groups_0, pad = var_1155_pad_0, pad_type = var_1155_pad_type_0, strides = var_1155_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_1155_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145839360)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145841984)))]; + tensor var_1165_to_fp16 = const()[name = tensor("op_1165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_1165_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145844608)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158951872)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1191_pad_type_0 = const()[name = tensor("op_1191_pad_type_0"), val = tensor("valid")]; + tensor var_1191_strides_0 = const()[name = tensor("op_1191_strides_0"), val = tensor([1, 1])]; + tensor var_1191_pad_0 = const()[name = tensor("op_1191_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1191_dilations_0 = const()[name = tensor("op_1191_dilations_0"), val = tensor([1, 1])]; + tensor var_1191_groups_0 = const()[name = tensor("op_1191_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158962176)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172069440)))]; + tensor var_1191_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_1191_dilations_0, groups = var_1191_groups_0, pad = var_1191_pad_0, pad_type = var_1191_pad_type_0, strides = var_1191_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_1191_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172072064)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172074688)))]; + tensor var_1216_to_fp16 = const()[name = tensor("op_1216_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_1216_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_1251_weight_0_to_fp16 = const()[name = tensor("op_1251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172077312)))]; + tensor var_1251_bias_0_to_fp16 = const()[name = tensor("op_1251_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175354176)))]; + tensor var_1251_cast_fp16 = conv(bias = var_1251_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_1251_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1251_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175356800)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_1249_pad_type_0 = const()[name = tensor("op_1249_pad_type_0"), val = tensor("valid")]; + tensor var_1249_strides_0 = const()[name = tensor("op_1249_strides_0"), val = tensor([1, 1])]; + tensor var_1249_pad_0 = const()[name = tensor("op_1249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1249_dilations_0 = const()[name = tensor("op_1249_dilations_0"), val = tensor([1, 1])]; + tensor var_1249_groups_0 = const()[name = tensor("op_1249_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178633664)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181910528)))]; + tensor var_1249_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_1249_dilations_0, groups = var_1249_groups_0, pad = var_1249_pad_0, pad_type = var_1249_pad_type_0, strides = var_1249_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1249_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1252_axis_0 = const()[name = tensor("op_1252_axis_0"), val = tensor(1)]; + tensor var_1252_cast_fp16_0, tensor var_1252_cast_fp16_1, tensor var_1252_cast_fp16_2, tensor var_1252_cast_fp16_3, tensor var_1252_cast_fp16_4, tensor var_1252_cast_fp16_5, tensor var_1252_cast_fp16_6, tensor var_1252_cast_fp16_7, tensor var_1252_cast_fp16_8, tensor var_1252_cast_fp16_9, tensor var_1252_cast_fp16_10, tensor var_1252_cast_fp16_11, tensor var_1252_cast_fp16_12, tensor var_1252_cast_fp16_13, tensor var_1252_cast_fp16_14, tensor var_1252_cast_fp16_15, tensor var_1252_cast_fp16_16, tensor var_1252_cast_fp16_17, tensor var_1252_cast_fp16_18, tensor var_1252_cast_fp16_19 = split(axis = var_1252_axis_0, split_sizes = tile_12, x = var_1251_cast_fp16)[name = tensor("op_1252_cast_fp16")]; + tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(3)]; + tensor var_1273_cast_fp16 = transpose(perm = var_1273_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_28")]; + tensor var_1274_cast_fp16_0, tensor var_1274_cast_fp16_1, tensor var_1274_cast_fp16_2, tensor var_1274_cast_fp16_3, tensor var_1274_cast_fp16_4, tensor var_1274_cast_fp16_5, tensor var_1274_cast_fp16_6, tensor var_1274_cast_fp16_7, tensor var_1274_cast_fp16_8, tensor var_1274_cast_fp16_9, tensor var_1274_cast_fp16_10, tensor var_1274_cast_fp16_11, tensor var_1274_cast_fp16_12, tensor var_1274_cast_fp16_13, tensor var_1274_cast_fp16_14, tensor var_1274_cast_fp16_15, tensor var_1274_cast_fp16_16, tensor var_1274_cast_fp16_17, tensor var_1274_cast_fp16_18, tensor var_1274_cast_fp16_19 = split(axis = var_1274_axis_0, split_sizes = tile_13, x = var_1273_cast_fp16)[name = tensor("op_1274_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1295_axis_0 = const()[name = tensor("op_1295_axis_0"), val = tensor(1)]; + tensor var_1295_cast_fp16_0, tensor var_1295_cast_fp16_1, tensor var_1295_cast_fp16_2, tensor var_1295_cast_fp16_3, tensor var_1295_cast_fp16_4, tensor var_1295_cast_fp16_5, tensor var_1295_cast_fp16_6, tensor var_1295_cast_fp16_7, tensor var_1295_cast_fp16_8, tensor var_1295_cast_fp16_9, tensor var_1295_cast_fp16_10, tensor var_1295_cast_fp16_11, tensor var_1295_cast_fp16_12, tensor var_1295_cast_fp16_13, tensor var_1295_cast_fp16_14, tensor var_1295_cast_fp16_15, tensor var_1295_cast_fp16_16, tensor var_1295_cast_fp16_17, tensor var_1295_cast_fp16_18, tensor var_1295_cast_fp16_19 = split(axis = var_1295_axis_0, split_sizes = tile_14, x = var_1249_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1274_cast_fp16_0, var_1252_cast_fp16_0))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1274_cast_fp16_1, var_1252_cast_fp16_1))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1274_cast_fp16_2, var_1252_cast_fp16_2))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1274_cast_fp16_3, var_1252_cast_fp16_3))[name = tensor("aw_167_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1274_cast_fp16_4, var_1252_cast_fp16_4))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1274_cast_fp16_5, var_1252_cast_fp16_5))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1274_cast_fp16_6, var_1252_cast_fp16_6))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1274_cast_fp16_7, var_1252_cast_fp16_7))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1274_cast_fp16_8, var_1252_cast_fp16_8))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1274_cast_fp16_9, var_1252_cast_fp16_9))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1274_cast_fp16_10, var_1252_cast_fp16_10))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1274_cast_fp16_11, var_1252_cast_fp16_11))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1274_cast_fp16_12, var_1252_cast_fp16_12))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1274_cast_fp16_13, var_1252_cast_fp16_13))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1274_cast_fp16_14, var_1252_cast_fp16_14))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1274_cast_fp16_15, var_1252_cast_fp16_15))[name = tensor("aw_191_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1274_cast_fp16_16, var_1252_cast_fp16_16))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1274_cast_fp16_17, var_1252_cast_fp16_17))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1274_cast_fp16_18, var_1252_cast_fp16_18))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1274_cast_fp16_19, var_1252_cast_fp16_19))[name = tensor("aw_199_cast_fp16")]; + tensor var_1356_cast_fp16 = softmax(axis = var_1200, x = aw_161_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor var_1357_cast_fp16 = softmax(axis = var_1200, x = aw_163_cast_fp16)[name = tensor("op_1357_cast_fp16")]; + tensor var_1358_cast_fp16 = softmax(axis = var_1200, x = aw_165_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = softmax(axis = var_1200, x = aw_167_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor var_1360_cast_fp16 = softmax(axis = var_1200, x = aw_169_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor var_1361_cast_fp16 = softmax(axis = var_1200, x = aw_171_cast_fp16)[name = tensor("op_1361_cast_fp16")]; + tensor var_1362_cast_fp16 = softmax(axis = var_1200, x = aw_173_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = softmax(axis = var_1200, x = aw_175_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = softmax(axis = var_1200, x = aw_177_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor var_1365_cast_fp16 = softmax(axis = var_1200, x = aw_179_cast_fp16)[name = tensor("op_1365_cast_fp16")]; + tensor var_1366_cast_fp16 = softmax(axis = var_1200, x = aw_181_cast_fp16)[name = tensor("op_1366_cast_fp16")]; + tensor var_1367_cast_fp16 = softmax(axis = var_1200, x = aw_183_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = softmax(axis = var_1200, x = aw_185_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor var_1369_cast_fp16 = softmax(axis = var_1200, x = aw_187_cast_fp16)[name = tensor("op_1369_cast_fp16")]; + tensor var_1370_cast_fp16 = softmax(axis = var_1200, x = aw_189_cast_fp16)[name = tensor("op_1370_cast_fp16")]; + tensor var_1371_cast_fp16 = softmax(axis = var_1200, x = aw_191_cast_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = softmax(axis = var_1200, x = aw_193_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor var_1373_cast_fp16 = softmax(axis = var_1200, x = aw_195_cast_fp16)[name = tensor("op_1373_cast_fp16")]; + tensor var_1374_cast_fp16 = softmax(axis = var_1200, x = aw_197_cast_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor var_1375_cast_fp16 = softmax(axis = var_1200, x = aw_199_cast_fp16)[name = tensor("op_1375_cast_fp16")]; + tensor var_1377_equation_0 = const()[name = tensor("op_1377_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1377_cast_fp16 = einsum(equation = var_1377_equation_0, values = (var_1295_cast_fp16_0, var_1356_cast_fp16))[name = tensor("op_1377_cast_fp16")]; + tensor var_1379_equation_0 = const()[name = tensor("op_1379_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1379_cast_fp16 = einsum(equation = var_1379_equation_0, values = (var_1295_cast_fp16_1, var_1357_cast_fp16))[name = tensor("op_1379_cast_fp16")]; + tensor var_1381_equation_0 = const()[name = tensor("op_1381_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1381_cast_fp16 = einsum(equation = var_1381_equation_0, values = (var_1295_cast_fp16_2, var_1358_cast_fp16))[name = tensor("op_1381_cast_fp16")]; + tensor var_1383_equation_0 = const()[name = tensor("op_1383_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1383_cast_fp16 = einsum(equation = var_1383_equation_0, values = (var_1295_cast_fp16_3, var_1359_cast_fp16))[name = tensor("op_1383_cast_fp16")]; + tensor var_1385_equation_0 = const()[name = tensor("op_1385_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1385_cast_fp16 = einsum(equation = var_1385_equation_0, values = (var_1295_cast_fp16_4, var_1360_cast_fp16))[name = tensor("op_1385_cast_fp16")]; + tensor var_1387_equation_0 = const()[name = tensor("op_1387_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1387_cast_fp16 = einsum(equation = var_1387_equation_0, values = (var_1295_cast_fp16_5, var_1361_cast_fp16))[name = tensor("op_1387_cast_fp16")]; + tensor var_1389_equation_0 = const()[name = tensor("op_1389_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1389_cast_fp16 = einsum(equation = var_1389_equation_0, values = (var_1295_cast_fp16_6, var_1362_cast_fp16))[name = tensor("op_1389_cast_fp16")]; + tensor var_1391_equation_0 = const()[name = tensor("op_1391_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1391_cast_fp16 = einsum(equation = var_1391_equation_0, values = (var_1295_cast_fp16_7, var_1363_cast_fp16))[name = tensor("op_1391_cast_fp16")]; + tensor var_1393_equation_0 = const()[name = tensor("op_1393_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1393_cast_fp16 = einsum(equation = var_1393_equation_0, values = (var_1295_cast_fp16_8, var_1364_cast_fp16))[name = tensor("op_1393_cast_fp16")]; + tensor var_1395_equation_0 = const()[name = tensor("op_1395_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1395_cast_fp16 = einsum(equation = var_1395_equation_0, values = (var_1295_cast_fp16_9, var_1365_cast_fp16))[name = tensor("op_1395_cast_fp16")]; + tensor var_1397_equation_0 = const()[name = tensor("op_1397_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1397_cast_fp16 = einsum(equation = var_1397_equation_0, values = (var_1295_cast_fp16_10, var_1366_cast_fp16))[name = tensor("op_1397_cast_fp16")]; + tensor var_1399_equation_0 = const()[name = tensor("op_1399_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1399_cast_fp16 = einsum(equation = var_1399_equation_0, values = (var_1295_cast_fp16_11, var_1367_cast_fp16))[name = tensor("op_1399_cast_fp16")]; + tensor var_1401_equation_0 = const()[name = tensor("op_1401_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1401_cast_fp16 = einsum(equation = var_1401_equation_0, values = (var_1295_cast_fp16_12, var_1368_cast_fp16))[name = tensor("op_1401_cast_fp16")]; + tensor var_1403_equation_0 = const()[name = tensor("op_1403_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1403_cast_fp16 = einsum(equation = var_1403_equation_0, values = (var_1295_cast_fp16_13, var_1369_cast_fp16))[name = tensor("op_1403_cast_fp16")]; + tensor var_1405_equation_0 = const()[name = tensor("op_1405_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1405_cast_fp16 = einsum(equation = var_1405_equation_0, values = (var_1295_cast_fp16_14, var_1370_cast_fp16))[name = tensor("op_1405_cast_fp16")]; + tensor var_1407_equation_0 = const()[name = tensor("op_1407_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1407_cast_fp16 = einsum(equation = var_1407_equation_0, values = (var_1295_cast_fp16_15, var_1371_cast_fp16))[name = tensor("op_1407_cast_fp16")]; + tensor var_1409_equation_0 = const()[name = tensor("op_1409_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1409_cast_fp16 = einsum(equation = var_1409_equation_0, values = (var_1295_cast_fp16_16, var_1372_cast_fp16))[name = tensor("op_1409_cast_fp16")]; + tensor var_1411_equation_0 = const()[name = tensor("op_1411_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1411_cast_fp16 = einsum(equation = var_1411_equation_0, values = (var_1295_cast_fp16_17, var_1373_cast_fp16))[name = tensor("op_1411_cast_fp16")]; + tensor var_1413_equation_0 = const()[name = tensor("op_1413_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1413_cast_fp16 = einsum(equation = var_1413_equation_0, values = (var_1295_cast_fp16_18, var_1374_cast_fp16))[name = tensor("op_1413_cast_fp16")]; + tensor var_1415_equation_0 = const()[name = tensor("op_1415_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1415_cast_fp16 = einsum(equation = var_1415_equation_0, values = (var_1295_cast_fp16_19, var_1375_cast_fp16))[name = tensor("op_1415_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_1200, interleave = input_45_interleave_0, values = (var_1377_cast_fp16, var_1379_cast_fp16, var_1381_cast_fp16, var_1383_cast_fp16, var_1385_cast_fp16, var_1387_cast_fp16, var_1389_cast_fp16, var_1391_cast_fp16, var_1393_cast_fp16, var_1395_cast_fp16, var_1397_cast_fp16, var_1399_cast_fp16, var_1401_cast_fp16, var_1403_cast_fp16, var_1405_cast_fp16, var_1407_cast_fp16, var_1409_cast_fp16, var_1411_cast_fp16, var_1413_cast_fp16, var_1415_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1424_pad_type_0 = const()[name = tensor("op_1424_pad_type_0"), val = tensor("valid")]; + tensor var_1424_strides_0 = const()[name = tensor("op_1424_strides_0"), val = tensor([1, 1])]; + tensor var_1424_pad_0 = const()[name = tensor("op_1424_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1424_dilations_0 = const()[name = tensor("op_1424_dilations_0"), val = tensor([1, 1])]; + tensor var_1424_groups_0 = const()[name = tensor("op_1424_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181913152)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185190016)))]; + tensor var_1424_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1424_dilations_0, groups = var_1424_groups_0, pad = var_1424_pad_0, pad_type = var_1424_pad_type_0, strides = var_1424_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1424_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185192640)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185195264)))]; + tensor var_1434_to_fp16 = const()[name = tensor("op_1434_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1434_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185197888)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198305152)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1460_pad_type_0 = const()[name = tensor("op_1460_pad_type_0"), val = tensor("valid")]; + tensor var_1460_strides_0 = const()[name = tensor("op_1460_strides_0"), val = tensor([1, 1])]; + tensor var_1460_pad_0 = const()[name = tensor("op_1460_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1460_dilations_0 = const()[name = tensor("op_1460_dilations_0"), val = tensor([1, 1])]; + tensor var_1460_groups_0 = const()[name = tensor("op_1460_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198315456)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211422720)))]; + tensor var_1460_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1460_dilations_0, groups = var_1460_groups_0, pad = var_1460_pad_0, pad_type = var_1460_pad_type_0, strides = var_1460_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1460_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1469 = const()[name = tensor("op_1469"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211425344)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211427968)))]; + tensor var_1485_to_fp16 = const()[name = tensor("op_1485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1485_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1520_weight_0_to_fp16 = const()[name = tensor("op_1520_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211430592)))]; + tensor var_1520_bias_0_to_fp16 = const()[name = tensor("op_1520_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214707456)))]; + tensor var_1520_cast_fp16 = conv(bias = var_1520_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1520_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214710080)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1518_pad_type_0 = const()[name = tensor("op_1518_pad_type_0"), val = tensor("valid")]; + tensor var_1518_strides_0 = const()[name = tensor("op_1518_strides_0"), val = tensor([1, 1])]; + tensor var_1518_pad_0 = const()[name = tensor("op_1518_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1518_dilations_0 = const()[name = tensor("op_1518_dilations_0"), val = tensor([1, 1])]; + tensor var_1518_groups_0 = const()[name = tensor("op_1518_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217986944)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221263808)))]; + tensor var_1518_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1518_dilations_0, groups = var_1518_groups_0, pad = var_1518_pad_0, pad_type = var_1518_pad_type_0, strides = var_1518_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1518_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1521_axis_0 = const()[name = tensor("op_1521_axis_0"), val = tensor(1)]; + tensor var_1521_cast_fp16_0, tensor var_1521_cast_fp16_1, tensor var_1521_cast_fp16_2, tensor var_1521_cast_fp16_3, tensor var_1521_cast_fp16_4, tensor var_1521_cast_fp16_5, tensor var_1521_cast_fp16_6, tensor var_1521_cast_fp16_7, tensor var_1521_cast_fp16_8, tensor var_1521_cast_fp16_9, tensor var_1521_cast_fp16_10, tensor var_1521_cast_fp16_11, tensor var_1521_cast_fp16_12, tensor var_1521_cast_fp16_13, tensor var_1521_cast_fp16_14, tensor var_1521_cast_fp16_15, tensor var_1521_cast_fp16_16, tensor var_1521_cast_fp16_17, tensor var_1521_cast_fp16_18, tensor var_1521_cast_fp16_19 = split(axis = var_1521_axis_0, split_sizes = tile_15, x = var_1520_cast_fp16)[name = tensor("op_1521_cast_fp16")]; + tensor var_1542_perm_0 = const()[name = tensor("op_1542_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1543_axis_0 = const()[name = tensor("op_1543_axis_0"), val = tensor(3)]; + tensor var_1542_cast_fp16 = transpose(perm = var_1542_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1543_cast_fp16_0, tensor var_1543_cast_fp16_1, tensor var_1543_cast_fp16_2, tensor var_1543_cast_fp16_3, tensor var_1543_cast_fp16_4, tensor var_1543_cast_fp16_5, tensor var_1543_cast_fp16_6, tensor var_1543_cast_fp16_7, tensor var_1543_cast_fp16_8, tensor var_1543_cast_fp16_9, tensor var_1543_cast_fp16_10, tensor var_1543_cast_fp16_11, tensor var_1543_cast_fp16_12, tensor var_1543_cast_fp16_13, tensor var_1543_cast_fp16_14, tensor var_1543_cast_fp16_15, tensor var_1543_cast_fp16_16, tensor var_1543_cast_fp16_17, tensor var_1543_cast_fp16_18, tensor var_1543_cast_fp16_19 = split(axis = var_1543_axis_0, split_sizes = tile_16, x = var_1542_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1564_axis_0 = const()[name = tensor("op_1564_axis_0"), val = tensor(1)]; + tensor var_1564_cast_fp16_0, tensor var_1564_cast_fp16_1, tensor var_1564_cast_fp16_2, tensor var_1564_cast_fp16_3, tensor var_1564_cast_fp16_4, tensor var_1564_cast_fp16_5, tensor var_1564_cast_fp16_6, tensor var_1564_cast_fp16_7, tensor var_1564_cast_fp16_8, tensor var_1564_cast_fp16_9, tensor var_1564_cast_fp16_10, tensor var_1564_cast_fp16_11, tensor var_1564_cast_fp16_12, tensor var_1564_cast_fp16_13, tensor var_1564_cast_fp16_14, tensor var_1564_cast_fp16_15, tensor var_1564_cast_fp16_16, tensor var_1564_cast_fp16_17, tensor var_1564_cast_fp16_18, tensor var_1564_cast_fp16_19 = split(axis = var_1564_axis_0, split_sizes = tile_17, x = var_1518_cast_fp16)[name = tensor("op_1564_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1543_cast_fp16_0, var_1521_cast_fp16_0))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1543_cast_fp16_1, var_1521_cast_fp16_1))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1543_cast_fp16_2, var_1521_cast_fp16_2))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1543_cast_fp16_3, var_1521_cast_fp16_3))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1543_cast_fp16_4, var_1521_cast_fp16_4))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1543_cast_fp16_5, var_1521_cast_fp16_5))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1543_cast_fp16_6, var_1521_cast_fp16_6))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1543_cast_fp16_7, var_1521_cast_fp16_7))[name = tensor("aw_215_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1543_cast_fp16_8, var_1521_cast_fp16_8))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1543_cast_fp16_9, var_1521_cast_fp16_9))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1543_cast_fp16_10, var_1521_cast_fp16_10))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1543_cast_fp16_11, var_1521_cast_fp16_11))[name = tensor("aw_223_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1543_cast_fp16_12, var_1521_cast_fp16_12))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1543_cast_fp16_13, var_1521_cast_fp16_13))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1543_cast_fp16_14, var_1521_cast_fp16_14))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1543_cast_fp16_15, var_1521_cast_fp16_15))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1543_cast_fp16_16, var_1521_cast_fp16_16))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1543_cast_fp16_17, var_1521_cast_fp16_17))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1543_cast_fp16_18, var_1521_cast_fp16_18))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1543_cast_fp16_19, var_1521_cast_fp16_19))[name = tensor("aw_239_cast_fp16")]; + tensor var_1625_cast_fp16 = softmax(axis = var_1469, x = aw_201_cast_fp16)[name = tensor("op_1625_cast_fp16")]; + tensor var_1626_cast_fp16 = softmax(axis = var_1469, x = aw_203_cast_fp16)[name = tensor("op_1626_cast_fp16")]; + tensor var_1627_cast_fp16 = softmax(axis = var_1469, x = aw_205_cast_fp16)[name = tensor("op_1627_cast_fp16")]; + tensor var_1628_cast_fp16 = softmax(axis = var_1469, x = aw_207_cast_fp16)[name = tensor("op_1628_cast_fp16")]; + tensor var_1629_cast_fp16 = softmax(axis = var_1469, x = aw_209_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1630_cast_fp16 = softmax(axis = var_1469, x = aw_211_cast_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = softmax(axis = var_1469, x = aw_213_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_cast_fp16 = softmax(axis = var_1469, x = aw_215_cast_fp16)[name = tensor("op_1632_cast_fp16")]; + tensor var_1633_cast_fp16 = softmax(axis = var_1469, x = aw_217_cast_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor var_1634_cast_fp16 = softmax(axis = var_1469, x = aw_219_cast_fp16)[name = tensor("op_1634_cast_fp16")]; + tensor var_1635_cast_fp16 = softmax(axis = var_1469, x = aw_221_cast_fp16)[name = tensor("op_1635_cast_fp16")]; + tensor var_1636_cast_fp16 = softmax(axis = var_1469, x = aw_223_cast_fp16)[name = tensor("op_1636_cast_fp16")]; + tensor var_1637_cast_fp16 = softmax(axis = var_1469, x = aw_225_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor var_1638_cast_fp16 = softmax(axis = var_1469, x = aw_227_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_cast_fp16 = softmax(axis = var_1469, x = aw_229_cast_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1640_cast_fp16 = softmax(axis = var_1469, x = aw_231_cast_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor var_1641_cast_fp16 = softmax(axis = var_1469, x = aw_233_cast_fp16)[name = tensor("op_1641_cast_fp16")]; + tensor var_1642_cast_fp16 = softmax(axis = var_1469, x = aw_235_cast_fp16)[name = tensor("op_1642_cast_fp16")]; + tensor var_1643_cast_fp16 = softmax(axis = var_1469, x = aw_237_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor var_1644_cast_fp16 = softmax(axis = var_1469, x = aw_239_cast_fp16)[name = tensor("op_1644_cast_fp16")]; + tensor var_1646_equation_0 = const()[name = tensor("op_1646_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1646_cast_fp16 = einsum(equation = var_1646_equation_0, values = (var_1564_cast_fp16_0, var_1625_cast_fp16))[name = tensor("op_1646_cast_fp16")]; + tensor var_1648_equation_0 = const()[name = tensor("op_1648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1648_cast_fp16 = einsum(equation = var_1648_equation_0, values = (var_1564_cast_fp16_1, var_1626_cast_fp16))[name = tensor("op_1648_cast_fp16")]; + tensor var_1650_equation_0 = const()[name = tensor("op_1650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1650_cast_fp16 = einsum(equation = var_1650_equation_0, values = (var_1564_cast_fp16_2, var_1627_cast_fp16))[name = tensor("op_1650_cast_fp16")]; + tensor var_1652_equation_0 = const()[name = tensor("op_1652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1652_cast_fp16 = einsum(equation = var_1652_equation_0, values = (var_1564_cast_fp16_3, var_1628_cast_fp16))[name = tensor("op_1652_cast_fp16")]; + tensor var_1654_equation_0 = const()[name = tensor("op_1654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1654_cast_fp16 = einsum(equation = var_1654_equation_0, values = (var_1564_cast_fp16_4, var_1629_cast_fp16))[name = tensor("op_1654_cast_fp16")]; + tensor var_1656_equation_0 = const()[name = tensor("op_1656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1656_cast_fp16 = einsum(equation = var_1656_equation_0, values = (var_1564_cast_fp16_5, var_1630_cast_fp16))[name = tensor("op_1656_cast_fp16")]; + tensor var_1658_equation_0 = const()[name = tensor("op_1658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1658_cast_fp16 = einsum(equation = var_1658_equation_0, values = (var_1564_cast_fp16_6, var_1631_cast_fp16))[name = tensor("op_1658_cast_fp16")]; + tensor var_1660_equation_0 = const()[name = tensor("op_1660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1660_cast_fp16 = einsum(equation = var_1660_equation_0, values = (var_1564_cast_fp16_7, var_1632_cast_fp16))[name = tensor("op_1660_cast_fp16")]; + tensor var_1662_equation_0 = const()[name = tensor("op_1662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1662_cast_fp16 = einsum(equation = var_1662_equation_0, values = (var_1564_cast_fp16_8, var_1633_cast_fp16))[name = tensor("op_1662_cast_fp16")]; + tensor var_1664_equation_0 = const()[name = tensor("op_1664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1664_cast_fp16 = einsum(equation = var_1664_equation_0, values = (var_1564_cast_fp16_9, var_1634_cast_fp16))[name = tensor("op_1664_cast_fp16")]; + tensor var_1666_equation_0 = const()[name = tensor("op_1666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1666_cast_fp16 = einsum(equation = var_1666_equation_0, values = (var_1564_cast_fp16_10, var_1635_cast_fp16))[name = tensor("op_1666_cast_fp16")]; + tensor var_1668_equation_0 = const()[name = tensor("op_1668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1668_cast_fp16 = einsum(equation = var_1668_equation_0, values = (var_1564_cast_fp16_11, var_1636_cast_fp16))[name = tensor("op_1668_cast_fp16")]; + tensor var_1670_equation_0 = const()[name = tensor("op_1670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1670_cast_fp16 = einsum(equation = var_1670_equation_0, values = (var_1564_cast_fp16_12, var_1637_cast_fp16))[name = tensor("op_1670_cast_fp16")]; + tensor var_1672_equation_0 = const()[name = tensor("op_1672_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1672_cast_fp16 = einsum(equation = var_1672_equation_0, values = (var_1564_cast_fp16_13, var_1638_cast_fp16))[name = tensor("op_1672_cast_fp16")]; + tensor var_1674_equation_0 = const()[name = tensor("op_1674_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1674_cast_fp16 = einsum(equation = var_1674_equation_0, values = (var_1564_cast_fp16_14, var_1639_cast_fp16))[name = tensor("op_1674_cast_fp16")]; + tensor var_1676_equation_0 = const()[name = tensor("op_1676_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1676_cast_fp16 = einsum(equation = var_1676_equation_0, values = (var_1564_cast_fp16_15, var_1640_cast_fp16))[name = tensor("op_1676_cast_fp16")]; + tensor var_1678_equation_0 = const()[name = tensor("op_1678_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1678_cast_fp16 = einsum(equation = var_1678_equation_0, values = (var_1564_cast_fp16_16, var_1641_cast_fp16))[name = tensor("op_1678_cast_fp16")]; + tensor var_1680_equation_0 = const()[name = tensor("op_1680_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1680_cast_fp16 = einsum(equation = var_1680_equation_0, values = (var_1564_cast_fp16_17, var_1642_cast_fp16))[name = tensor("op_1680_cast_fp16")]; + tensor var_1682_equation_0 = const()[name = tensor("op_1682_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1682_cast_fp16 = einsum(equation = var_1682_equation_0, values = (var_1564_cast_fp16_18, var_1643_cast_fp16))[name = tensor("op_1682_cast_fp16")]; + tensor var_1684_equation_0 = const()[name = tensor("op_1684_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1684_cast_fp16 = einsum(equation = var_1684_equation_0, values = (var_1564_cast_fp16_19, var_1644_cast_fp16))[name = tensor("op_1684_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1469, interleave = input_55_interleave_0, values = (var_1646_cast_fp16, var_1648_cast_fp16, var_1650_cast_fp16, var_1652_cast_fp16, var_1654_cast_fp16, var_1656_cast_fp16, var_1658_cast_fp16, var_1660_cast_fp16, var_1662_cast_fp16, var_1664_cast_fp16, var_1666_cast_fp16, var_1668_cast_fp16, var_1670_cast_fp16, var_1672_cast_fp16, var_1674_cast_fp16, var_1676_cast_fp16, var_1678_cast_fp16, var_1680_cast_fp16, var_1682_cast_fp16, var_1684_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1693_pad_type_0 = const()[name = tensor("op_1693_pad_type_0"), val = tensor("valid")]; + tensor var_1693_strides_0 = const()[name = tensor("op_1693_strides_0"), val = tensor([1, 1])]; + tensor var_1693_pad_0 = const()[name = tensor("op_1693_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1693_dilations_0 = const()[name = tensor("op_1693_dilations_0"), val = tensor([1, 1])]; + tensor var_1693_groups_0 = const()[name = tensor("op_1693_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221266432)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224543296)))]; + tensor var_1693_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1693_dilations_0, groups = var_1693_groups_0, pad = var_1693_pad_0, pad_type = var_1693_pad_type_0, strides = var_1693_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1693_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1693_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224545920)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224548544)))]; + tensor var_1703_to_fp16 = const()[name = tensor("op_1703_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1703_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224551168)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237658432)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1729_pad_type_0 = const()[name = tensor("op_1729_pad_type_0"), val = tensor("valid")]; + tensor var_1729_strides_0 = const()[name = tensor("op_1729_strides_0"), val = tensor([1, 1])]; + tensor var_1729_pad_0 = const()[name = tensor("op_1729_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1729_dilations_0 = const()[name = tensor("op_1729_dilations_0"), val = tensor([1, 1])]; + tensor var_1729_groups_0 = const()[name = tensor("op_1729_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237668736)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250776000)))]; + tensor var_1729_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1729_dilations_0, groups = var_1729_groups_0, pad = var_1729_pad_0, pad_type = var_1729_pad_type_0, strides = var_1729_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1729_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1729_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1738 = const()[name = tensor("op_1738"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250778624)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250781248)))]; + tensor var_1754_to_fp16 = const()[name = tensor("op_1754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1754_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1789_weight_0_to_fp16 = const()[name = tensor("op_1789_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250783872)))]; + tensor var_1789_bias_0_to_fp16 = const()[name = tensor("op_1789_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254060736)))]; + tensor var_1789_cast_fp16 = conv(bias = var_1789_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1789_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1789_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254063360)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1787_pad_type_0 = const()[name = tensor("op_1787_pad_type_0"), val = tensor("valid")]; + tensor var_1787_strides_0 = const()[name = tensor("op_1787_strides_0"), val = tensor([1, 1])]; + tensor var_1787_pad_0 = const()[name = tensor("op_1787_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1787_dilations_0 = const()[name = tensor("op_1787_dilations_0"), val = tensor([1, 1])]; + tensor var_1787_groups_0 = const()[name = tensor("op_1787_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257340224)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260617088)))]; + tensor var_1787_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1787_dilations_0, groups = var_1787_groups_0, pad = var_1787_pad_0, pad_type = var_1787_pad_type_0, strides = var_1787_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1790_axis_0 = const()[name = tensor("op_1790_axis_0"), val = tensor(1)]; + tensor var_1790_cast_fp16_0, tensor var_1790_cast_fp16_1, tensor var_1790_cast_fp16_2, tensor var_1790_cast_fp16_3, tensor var_1790_cast_fp16_4, tensor var_1790_cast_fp16_5, tensor var_1790_cast_fp16_6, tensor var_1790_cast_fp16_7, tensor var_1790_cast_fp16_8, tensor var_1790_cast_fp16_9, tensor var_1790_cast_fp16_10, tensor var_1790_cast_fp16_11, tensor var_1790_cast_fp16_12, tensor var_1790_cast_fp16_13, tensor var_1790_cast_fp16_14, tensor var_1790_cast_fp16_15, tensor var_1790_cast_fp16_16, tensor var_1790_cast_fp16_17, tensor var_1790_cast_fp16_18, tensor var_1790_cast_fp16_19 = split(axis = var_1790_axis_0, split_sizes = tile_18, x = var_1789_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1811_perm_0 = const()[name = tensor("op_1811_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1812_axis_0 = const()[name = tensor("op_1812_axis_0"), val = tensor(3)]; + tensor var_1811_cast_fp16 = transpose(perm = var_1811_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_26")]; + tensor var_1812_cast_fp16_0, tensor var_1812_cast_fp16_1, tensor var_1812_cast_fp16_2, tensor var_1812_cast_fp16_3, tensor var_1812_cast_fp16_4, tensor var_1812_cast_fp16_5, tensor var_1812_cast_fp16_6, tensor var_1812_cast_fp16_7, tensor var_1812_cast_fp16_8, tensor var_1812_cast_fp16_9, tensor var_1812_cast_fp16_10, tensor var_1812_cast_fp16_11, tensor var_1812_cast_fp16_12, tensor var_1812_cast_fp16_13, tensor var_1812_cast_fp16_14, tensor var_1812_cast_fp16_15, tensor var_1812_cast_fp16_16, tensor var_1812_cast_fp16_17, tensor var_1812_cast_fp16_18, tensor var_1812_cast_fp16_19 = split(axis = var_1812_axis_0, split_sizes = tile_19, x = var_1811_cast_fp16)[name = tensor("op_1812_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1833_axis_0 = const()[name = tensor("op_1833_axis_0"), val = tensor(1)]; + tensor var_1833_cast_fp16_0, tensor var_1833_cast_fp16_1, tensor var_1833_cast_fp16_2, tensor var_1833_cast_fp16_3, tensor var_1833_cast_fp16_4, tensor var_1833_cast_fp16_5, tensor var_1833_cast_fp16_6, tensor var_1833_cast_fp16_7, tensor var_1833_cast_fp16_8, tensor var_1833_cast_fp16_9, tensor var_1833_cast_fp16_10, tensor var_1833_cast_fp16_11, tensor var_1833_cast_fp16_12, tensor var_1833_cast_fp16_13, tensor var_1833_cast_fp16_14, tensor var_1833_cast_fp16_15, tensor var_1833_cast_fp16_16, tensor var_1833_cast_fp16_17, tensor var_1833_cast_fp16_18, tensor var_1833_cast_fp16_19 = split(axis = var_1833_axis_0, split_sizes = tile_20, x = var_1787_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_1812_cast_fp16_0, var_1790_cast_fp16_0))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_1812_cast_fp16_1, var_1790_cast_fp16_1))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_1812_cast_fp16_2, var_1790_cast_fp16_2))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_1812_cast_fp16_3, var_1790_cast_fp16_3))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_1812_cast_fp16_4, var_1790_cast_fp16_4))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_1812_cast_fp16_5, var_1790_cast_fp16_5))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_1812_cast_fp16_6, var_1790_cast_fp16_6))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_1812_cast_fp16_7, var_1790_cast_fp16_7))[name = tensor("aw_255_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_1812_cast_fp16_8, var_1790_cast_fp16_8))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_1812_cast_fp16_9, var_1790_cast_fp16_9))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_1812_cast_fp16_10, var_1790_cast_fp16_10))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_1812_cast_fp16_11, var_1790_cast_fp16_11))[name = tensor("aw_263_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_1812_cast_fp16_12, var_1790_cast_fp16_12))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_1812_cast_fp16_13, var_1790_cast_fp16_13))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_1812_cast_fp16_14, var_1790_cast_fp16_14))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_1812_cast_fp16_15, var_1790_cast_fp16_15))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_1812_cast_fp16_16, var_1790_cast_fp16_16))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_1812_cast_fp16_17, var_1790_cast_fp16_17))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_1812_cast_fp16_18, var_1790_cast_fp16_18))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_1812_cast_fp16_19, var_1790_cast_fp16_19))[name = tensor("aw_279_cast_fp16")]; + tensor var_1894_cast_fp16 = softmax(axis = var_1738, x = aw_241_cast_fp16)[name = tensor("op_1894_cast_fp16")]; + tensor var_1895_cast_fp16 = softmax(axis = var_1738, x = aw_243_cast_fp16)[name = tensor("op_1895_cast_fp16")]; + tensor var_1896_cast_fp16 = softmax(axis = var_1738, x = aw_245_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1897_cast_fp16 = softmax(axis = var_1738, x = aw_247_cast_fp16)[name = tensor("op_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = softmax(axis = var_1738, x = aw_249_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_cast_fp16 = softmax(axis = var_1738, x = aw_251_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor var_1900_cast_fp16 = softmax(axis = var_1738, x = aw_253_cast_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor var_1901_cast_fp16 = softmax(axis = var_1738, x = aw_255_cast_fp16)[name = tensor("op_1901_cast_fp16")]; + tensor var_1902_cast_fp16 = softmax(axis = var_1738, x = aw_257_cast_fp16)[name = tensor("op_1902_cast_fp16")]; + tensor var_1903_cast_fp16 = softmax(axis = var_1738, x = aw_259_cast_fp16)[name = tensor("op_1903_cast_fp16")]; + tensor var_1904_cast_fp16 = softmax(axis = var_1738, x = aw_261_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1905_cast_fp16 = softmax(axis = var_1738, x = aw_263_cast_fp16)[name = tensor("op_1905_cast_fp16")]; + tensor var_1906_cast_fp16 = softmax(axis = var_1738, x = aw_265_cast_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1907_cast_fp16 = softmax(axis = var_1738, x = aw_267_cast_fp16)[name = tensor("op_1907_cast_fp16")]; + tensor var_1908_cast_fp16 = softmax(axis = var_1738, x = aw_269_cast_fp16)[name = tensor("op_1908_cast_fp16")]; + tensor var_1909_cast_fp16 = softmax(axis = var_1738, x = aw_271_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor var_1910_cast_fp16 = softmax(axis = var_1738, x = aw_273_cast_fp16)[name = tensor("op_1910_cast_fp16")]; + tensor var_1911_cast_fp16 = softmax(axis = var_1738, x = aw_275_cast_fp16)[name = tensor("op_1911_cast_fp16")]; + tensor var_1912_cast_fp16 = softmax(axis = var_1738, x = aw_277_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1913_cast_fp16 = softmax(axis = var_1738, x = aw_279_cast_fp16)[name = tensor("op_1913_cast_fp16")]; + tensor var_1915_equation_0 = const()[name = tensor("op_1915_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1915_cast_fp16 = einsum(equation = var_1915_equation_0, values = (var_1833_cast_fp16_0, var_1894_cast_fp16))[name = tensor("op_1915_cast_fp16")]; + tensor var_1917_equation_0 = const()[name = tensor("op_1917_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1917_cast_fp16 = einsum(equation = var_1917_equation_0, values = (var_1833_cast_fp16_1, var_1895_cast_fp16))[name = tensor("op_1917_cast_fp16")]; + tensor var_1919_equation_0 = const()[name = tensor("op_1919_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1919_cast_fp16 = einsum(equation = var_1919_equation_0, values = (var_1833_cast_fp16_2, var_1896_cast_fp16))[name = tensor("op_1919_cast_fp16")]; + tensor var_1921_equation_0 = const()[name = tensor("op_1921_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1921_cast_fp16 = einsum(equation = var_1921_equation_0, values = (var_1833_cast_fp16_3, var_1897_cast_fp16))[name = tensor("op_1921_cast_fp16")]; + tensor var_1923_equation_0 = const()[name = tensor("op_1923_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1923_cast_fp16 = einsum(equation = var_1923_equation_0, values = (var_1833_cast_fp16_4, var_1898_cast_fp16))[name = tensor("op_1923_cast_fp16")]; + tensor var_1925_equation_0 = const()[name = tensor("op_1925_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1925_cast_fp16 = einsum(equation = var_1925_equation_0, values = (var_1833_cast_fp16_5, var_1899_cast_fp16))[name = tensor("op_1925_cast_fp16")]; + tensor var_1927_equation_0 = const()[name = tensor("op_1927_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1927_cast_fp16 = einsum(equation = var_1927_equation_0, values = (var_1833_cast_fp16_6, var_1900_cast_fp16))[name = tensor("op_1927_cast_fp16")]; + tensor var_1929_equation_0 = const()[name = tensor("op_1929_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1929_cast_fp16 = einsum(equation = var_1929_equation_0, values = (var_1833_cast_fp16_7, var_1901_cast_fp16))[name = tensor("op_1929_cast_fp16")]; + tensor var_1931_equation_0 = const()[name = tensor("op_1931_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1931_cast_fp16 = einsum(equation = var_1931_equation_0, values = (var_1833_cast_fp16_8, var_1902_cast_fp16))[name = tensor("op_1931_cast_fp16")]; + tensor var_1933_equation_0 = const()[name = tensor("op_1933_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1933_cast_fp16 = einsum(equation = var_1933_equation_0, values = (var_1833_cast_fp16_9, var_1903_cast_fp16))[name = tensor("op_1933_cast_fp16")]; + tensor var_1935_equation_0 = const()[name = tensor("op_1935_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1935_cast_fp16 = einsum(equation = var_1935_equation_0, values = (var_1833_cast_fp16_10, var_1904_cast_fp16))[name = tensor("op_1935_cast_fp16")]; + tensor var_1937_equation_0 = const()[name = tensor("op_1937_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1937_cast_fp16 = einsum(equation = var_1937_equation_0, values = (var_1833_cast_fp16_11, var_1905_cast_fp16))[name = tensor("op_1937_cast_fp16")]; + tensor var_1939_equation_0 = const()[name = tensor("op_1939_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1939_cast_fp16 = einsum(equation = var_1939_equation_0, values = (var_1833_cast_fp16_12, var_1906_cast_fp16))[name = tensor("op_1939_cast_fp16")]; + tensor var_1941_equation_0 = const()[name = tensor("op_1941_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1941_cast_fp16 = einsum(equation = var_1941_equation_0, values = (var_1833_cast_fp16_13, var_1907_cast_fp16))[name = tensor("op_1941_cast_fp16")]; + tensor var_1943_equation_0 = const()[name = tensor("op_1943_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1943_cast_fp16 = einsum(equation = var_1943_equation_0, values = (var_1833_cast_fp16_14, var_1908_cast_fp16))[name = tensor("op_1943_cast_fp16")]; + tensor var_1945_equation_0 = const()[name = tensor("op_1945_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1945_cast_fp16 = einsum(equation = var_1945_equation_0, values = (var_1833_cast_fp16_15, var_1909_cast_fp16))[name = tensor("op_1945_cast_fp16")]; + tensor var_1947_equation_0 = const()[name = tensor("op_1947_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1947_cast_fp16 = einsum(equation = var_1947_equation_0, values = (var_1833_cast_fp16_16, var_1910_cast_fp16))[name = tensor("op_1947_cast_fp16")]; + tensor var_1949_equation_0 = const()[name = tensor("op_1949_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1949_cast_fp16 = einsum(equation = var_1949_equation_0, values = (var_1833_cast_fp16_17, var_1911_cast_fp16))[name = tensor("op_1949_cast_fp16")]; + tensor var_1951_equation_0 = const()[name = tensor("op_1951_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1951_cast_fp16 = einsum(equation = var_1951_equation_0, values = (var_1833_cast_fp16_18, var_1912_cast_fp16))[name = tensor("op_1951_cast_fp16")]; + tensor var_1953_equation_0 = const()[name = tensor("op_1953_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1953_cast_fp16 = einsum(equation = var_1953_equation_0, values = (var_1833_cast_fp16_19, var_1913_cast_fp16))[name = tensor("op_1953_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1738, interleave = input_65_interleave_0, values = (var_1915_cast_fp16, var_1917_cast_fp16, var_1919_cast_fp16, var_1921_cast_fp16, var_1923_cast_fp16, var_1925_cast_fp16, var_1927_cast_fp16, var_1929_cast_fp16, var_1931_cast_fp16, var_1933_cast_fp16, var_1935_cast_fp16, var_1937_cast_fp16, var_1939_cast_fp16, var_1941_cast_fp16, var_1943_cast_fp16, var_1945_cast_fp16, var_1947_cast_fp16, var_1949_cast_fp16, var_1951_cast_fp16, var_1953_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1962_pad_type_0 = const()[name = tensor("op_1962_pad_type_0"), val = tensor("valid")]; + tensor var_1962_strides_0 = const()[name = tensor("op_1962_strides_0"), val = tensor([1, 1])]; + tensor var_1962_pad_0 = const()[name = tensor("op_1962_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1962_dilations_0 = const()[name = tensor("op_1962_dilations_0"), val = tensor([1, 1])]; + tensor var_1962_groups_0 = const()[name = tensor("op_1962_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260619712)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263896576)))]; + tensor var_1962_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1962_dilations_0, groups = var_1962_groups_0, pad = var_1962_pad_0, pad_type = var_1962_pad_type_0, strides = var_1962_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1962_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263899200)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263901824)))]; + tensor var_1972_to_fp16 = const()[name = tensor("op_1972_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1972_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263904448)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277011712)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1998_pad_type_0 = const()[name = tensor("op_1998_pad_type_0"), val = tensor("valid")]; + tensor var_1998_strides_0 = const()[name = tensor("op_1998_strides_0"), val = tensor([1, 1])]; + tensor var_1998_pad_0 = const()[name = tensor("op_1998_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1998_dilations_0 = const()[name = tensor("op_1998_dilations_0"), val = tensor([1, 1])]; + tensor var_1998_groups_0 = const()[name = tensor("op_1998_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277022016)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290129280)))]; + tensor var_1998_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1998_dilations_0, groups = var_1998_groups_0, pad = var_1998_pad_0, pad_type = var_1998_pad_type_0, strides = var_1998_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1998_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_2007 = const()[name = tensor("op_2007"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290131904)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290134528)))]; + tensor var_2023_to_fp16 = const()[name = tensor("op_2023_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_2023_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_2058_weight_0_to_fp16 = const()[name = tensor("op_2058_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290137152)))]; + tensor var_2058_bias_0_to_fp16 = const()[name = tensor("op_2058_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293414016)))]; + tensor var_2058_cast_fp16 = conv(bias = var_2058_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_2058_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_2058_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293416640)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_2056_pad_type_0 = const()[name = tensor("op_2056_pad_type_0"), val = tensor("valid")]; + tensor var_2056_strides_0 = const()[name = tensor("op_2056_strides_0"), val = tensor([1, 1])]; + tensor var_2056_pad_0 = const()[name = tensor("op_2056_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2056_dilations_0 = const()[name = tensor("op_2056_dilations_0"), val = tensor([1, 1])]; + tensor var_2056_groups_0 = const()[name = tensor("op_2056_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296693504)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299970368)))]; + tensor var_2056_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_2056_dilations_0, groups = var_2056_groups_0, pad = var_2056_pad_0, pad_type = var_2056_pad_type_0, strides = var_2056_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2059_axis_0 = const()[name = tensor("op_2059_axis_0"), val = tensor(1)]; + tensor var_2059_cast_fp16_0, tensor var_2059_cast_fp16_1, tensor var_2059_cast_fp16_2, tensor var_2059_cast_fp16_3, tensor var_2059_cast_fp16_4, tensor var_2059_cast_fp16_5, tensor var_2059_cast_fp16_6, tensor var_2059_cast_fp16_7, tensor var_2059_cast_fp16_8, tensor var_2059_cast_fp16_9, tensor var_2059_cast_fp16_10, tensor var_2059_cast_fp16_11, tensor var_2059_cast_fp16_12, tensor var_2059_cast_fp16_13, tensor var_2059_cast_fp16_14, tensor var_2059_cast_fp16_15, tensor var_2059_cast_fp16_16, tensor var_2059_cast_fp16_17, tensor var_2059_cast_fp16_18, tensor var_2059_cast_fp16_19 = split(axis = var_2059_axis_0, split_sizes = tile_21, x = var_2058_cast_fp16)[name = tensor("op_2059_cast_fp16")]; + tensor var_2080_perm_0 = const()[name = tensor("op_2080_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2081_axis_0 = const()[name = tensor("op_2081_axis_0"), val = tensor(3)]; + tensor var_2080_cast_fp16 = transpose(perm = var_2080_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_25")]; + tensor var_2081_cast_fp16_0, tensor var_2081_cast_fp16_1, tensor var_2081_cast_fp16_2, tensor var_2081_cast_fp16_3, tensor var_2081_cast_fp16_4, tensor var_2081_cast_fp16_5, tensor var_2081_cast_fp16_6, tensor var_2081_cast_fp16_7, tensor var_2081_cast_fp16_8, tensor var_2081_cast_fp16_9, tensor var_2081_cast_fp16_10, tensor var_2081_cast_fp16_11, tensor var_2081_cast_fp16_12, tensor var_2081_cast_fp16_13, tensor var_2081_cast_fp16_14, tensor var_2081_cast_fp16_15, tensor var_2081_cast_fp16_16, tensor var_2081_cast_fp16_17, tensor var_2081_cast_fp16_18, tensor var_2081_cast_fp16_19 = split(axis = var_2081_axis_0, split_sizes = tile_22, x = var_2080_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2102_axis_0 = const()[name = tensor("op_2102_axis_0"), val = tensor(1)]; + tensor var_2102_cast_fp16_0, tensor var_2102_cast_fp16_1, tensor var_2102_cast_fp16_2, tensor var_2102_cast_fp16_3, tensor var_2102_cast_fp16_4, tensor var_2102_cast_fp16_5, tensor var_2102_cast_fp16_6, tensor var_2102_cast_fp16_7, tensor var_2102_cast_fp16_8, tensor var_2102_cast_fp16_9, tensor var_2102_cast_fp16_10, tensor var_2102_cast_fp16_11, tensor var_2102_cast_fp16_12, tensor var_2102_cast_fp16_13, tensor var_2102_cast_fp16_14, tensor var_2102_cast_fp16_15, tensor var_2102_cast_fp16_16, tensor var_2102_cast_fp16_17, tensor var_2102_cast_fp16_18, tensor var_2102_cast_fp16_19 = split(axis = var_2102_axis_0, split_sizes = tile_23, x = var_2056_cast_fp16)[name = tensor("op_2102_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2081_cast_fp16_0, var_2059_cast_fp16_0))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2081_cast_fp16_1, var_2059_cast_fp16_1))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2081_cast_fp16_2, var_2059_cast_fp16_2))[name = tensor("aw_285_cast_fp16")]; + tensor aw_287_equation_0 = const()[name = tensor("aw_287_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_287_cast_fp16 = einsum(equation = aw_287_equation_0, values = (var_2081_cast_fp16_3, var_2059_cast_fp16_3))[name = tensor("aw_287_cast_fp16")]; + tensor aw_289_equation_0 = const()[name = tensor("aw_289_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_289_cast_fp16 = einsum(equation = aw_289_equation_0, values = (var_2081_cast_fp16_4, var_2059_cast_fp16_4))[name = tensor("aw_289_cast_fp16")]; + tensor aw_291_equation_0 = const()[name = tensor("aw_291_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_291_cast_fp16 = einsum(equation = aw_291_equation_0, values = (var_2081_cast_fp16_5, var_2059_cast_fp16_5))[name = tensor("aw_291_cast_fp16")]; + tensor aw_293_equation_0 = const()[name = tensor("aw_293_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_293_cast_fp16 = einsum(equation = aw_293_equation_0, values = (var_2081_cast_fp16_6, var_2059_cast_fp16_6))[name = tensor("aw_293_cast_fp16")]; + tensor aw_295_equation_0 = const()[name = tensor("aw_295_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_295_cast_fp16 = einsum(equation = aw_295_equation_0, values = (var_2081_cast_fp16_7, var_2059_cast_fp16_7))[name = tensor("aw_295_cast_fp16")]; + tensor aw_297_equation_0 = const()[name = tensor("aw_297_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_297_cast_fp16 = einsum(equation = aw_297_equation_0, values = (var_2081_cast_fp16_8, var_2059_cast_fp16_8))[name = tensor("aw_297_cast_fp16")]; + tensor aw_299_equation_0 = const()[name = tensor("aw_299_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_299_cast_fp16 = einsum(equation = aw_299_equation_0, values = (var_2081_cast_fp16_9, var_2059_cast_fp16_9))[name = tensor("aw_299_cast_fp16")]; + tensor aw_301_equation_0 = const()[name = tensor("aw_301_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_301_cast_fp16 = einsum(equation = aw_301_equation_0, values = (var_2081_cast_fp16_10, var_2059_cast_fp16_10))[name = tensor("aw_301_cast_fp16")]; + tensor aw_303_equation_0 = const()[name = tensor("aw_303_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_303_cast_fp16 = einsum(equation = aw_303_equation_0, values = (var_2081_cast_fp16_11, var_2059_cast_fp16_11))[name = tensor("aw_303_cast_fp16")]; + tensor aw_305_equation_0 = const()[name = tensor("aw_305_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_305_cast_fp16 = einsum(equation = aw_305_equation_0, values = (var_2081_cast_fp16_12, var_2059_cast_fp16_12))[name = tensor("aw_305_cast_fp16")]; + tensor aw_307_equation_0 = const()[name = tensor("aw_307_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_307_cast_fp16 = einsum(equation = aw_307_equation_0, values = (var_2081_cast_fp16_13, var_2059_cast_fp16_13))[name = tensor("aw_307_cast_fp16")]; + tensor aw_309_equation_0 = const()[name = tensor("aw_309_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_309_cast_fp16 = einsum(equation = aw_309_equation_0, values = (var_2081_cast_fp16_14, var_2059_cast_fp16_14))[name = tensor("aw_309_cast_fp16")]; + tensor aw_311_equation_0 = const()[name = tensor("aw_311_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_311_cast_fp16 = einsum(equation = aw_311_equation_0, values = (var_2081_cast_fp16_15, var_2059_cast_fp16_15))[name = tensor("aw_311_cast_fp16")]; + tensor aw_313_equation_0 = const()[name = tensor("aw_313_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_313_cast_fp16 = einsum(equation = aw_313_equation_0, values = (var_2081_cast_fp16_16, var_2059_cast_fp16_16))[name = tensor("aw_313_cast_fp16")]; + tensor aw_315_equation_0 = const()[name = tensor("aw_315_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_315_cast_fp16 = einsum(equation = aw_315_equation_0, values = (var_2081_cast_fp16_17, var_2059_cast_fp16_17))[name = tensor("aw_315_cast_fp16")]; + tensor aw_317_equation_0 = const()[name = tensor("aw_317_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_317_cast_fp16 = einsum(equation = aw_317_equation_0, values = (var_2081_cast_fp16_18, var_2059_cast_fp16_18))[name = tensor("aw_317_cast_fp16")]; + tensor aw_319_equation_0 = const()[name = tensor("aw_319_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_319_cast_fp16 = einsum(equation = aw_319_equation_0, values = (var_2081_cast_fp16_19, var_2059_cast_fp16_19))[name = tensor("aw_319_cast_fp16")]; + tensor var_2163_cast_fp16 = softmax(axis = var_2007, x = aw_281_cast_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = softmax(axis = var_2007, x = aw_283_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor var_2165_cast_fp16 = softmax(axis = var_2007, x = aw_285_cast_fp16)[name = tensor("op_2165_cast_fp16")]; + tensor var_2166_cast_fp16 = softmax(axis = var_2007, x = aw_287_cast_fp16)[name = tensor("op_2166_cast_fp16")]; + tensor var_2167_cast_fp16 = softmax(axis = var_2007, x = aw_289_cast_fp16)[name = tensor("op_2167_cast_fp16")]; + tensor var_2168_cast_fp16 = softmax(axis = var_2007, x = aw_291_cast_fp16)[name = tensor("op_2168_cast_fp16")]; + tensor var_2169_cast_fp16 = softmax(axis = var_2007, x = aw_293_cast_fp16)[name = tensor("op_2169_cast_fp16")]; + tensor var_2170_cast_fp16 = softmax(axis = var_2007, x = aw_295_cast_fp16)[name = tensor("op_2170_cast_fp16")]; + tensor var_2171_cast_fp16 = softmax(axis = var_2007, x = aw_297_cast_fp16)[name = tensor("op_2171_cast_fp16")]; + tensor var_2172_cast_fp16 = softmax(axis = var_2007, x = aw_299_cast_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor var_2173_cast_fp16 = softmax(axis = var_2007, x = aw_301_cast_fp16)[name = tensor("op_2173_cast_fp16")]; + tensor var_2174_cast_fp16 = softmax(axis = var_2007, x = aw_303_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor var_2175_cast_fp16 = softmax(axis = var_2007, x = aw_305_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor var_2176_cast_fp16 = softmax(axis = var_2007, x = aw_307_cast_fp16)[name = tensor("op_2176_cast_fp16")]; + tensor var_2177_cast_fp16 = softmax(axis = var_2007, x = aw_309_cast_fp16)[name = tensor("op_2177_cast_fp16")]; + tensor var_2178_cast_fp16 = softmax(axis = var_2007, x = aw_311_cast_fp16)[name = tensor("op_2178_cast_fp16")]; + tensor var_2179_cast_fp16 = softmax(axis = var_2007, x = aw_313_cast_fp16)[name = tensor("op_2179_cast_fp16")]; + tensor var_2180_cast_fp16 = softmax(axis = var_2007, x = aw_315_cast_fp16)[name = tensor("op_2180_cast_fp16")]; + tensor var_2181_cast_fp16 = softmax(axis = var_2007, x = aw_317_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2182_cast_fp16 = softmax(axis = var_2007, x = aw_319_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor var_2184_equation_0 = const()[name = tensor("op_2184_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2184_cast_fp16 = einsum(equation = var_2184_equation_0, values = (var_2102_cast_fp16_0, var_2163_cast_fp16))[name = tensor("op_2184_cast_fp16")]; + tensor var_2186_equation_0 = const()[name = tensor("op_2186_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2186_cast_fp16 = einsum(equation = var_2186_equation_0, values = (var_2102_cast_fp16_1, var_2164_cast_fp16))[name = tensor("op_2186_cast_fp16")]; + tensor var_2188_equation_0 = const()[name = tensor("op_2188_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2188_cast_fp16 = einsum(equation = var_2188_equation_0, values = (var_2102_cast_fp16_2, var_2165_cast_fp16))[name = tensor("op_2188_cast_fp16")]; + tensor var_2190_equation_0 = const()[name = tensor("op_2190_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2190_cast_fp16 = einsum(equation = var_2190_equation_0, values = (var_2102_cast_fp16_3, var_2166_cast_fp16))[name = tensor("op_2190_cast_fp16")]; + tensor var_2192_equation_0 = const()[name = tensor("op_2192_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2192_cast_fp16 = einsum(equation = var_2192_equation_0, values = (var_2102_cast_fp16_4, var_2167_cast_fp16))[name = tensor("op_2192_cast_fp16")]; + tensor var_2194_equation_0 = const()[name = tensor("op_2194_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2194_cast_fp16 = einsum(equation = var_2194_equation_0, values = (var_2102_cast_fp16_5, var_2168_cast_fp16))[name = tensor("op_2194_cast_fp16")]; + tensor var_2196_equation_0 = const()[name = tensor("op_2196_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2196_cast_fp16 = einsum(equation = var_2196_equation_0, values = (var_2102_cast_fp16_6, var_2169_cast_fp16))[name = tensor("op_2196_cast_fp16")]; + tensor var_2198_equation_0 = const()[name = tensor("op_2198_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2198_cast_fp16 = einsum(equation = var_2198_equation_0, values = (var_2102_cast_fp16_7, var_2170_cast_fp16))[name = tensor("op_2198_cast_fp16")]; + tensor var_2200_equation_0 = const()[name = tensor("op_2200_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2200_cast_fp16 = einsum(equation = var_2200_equation_0, values = (var_2102_cast_fp16_8, var_2171_cast_fp16))[name = tensor("op_2200_cast_fp16")]; + tensor var_2202_equation_0 = const()[name = tensor("op_2202_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2202_cast_fp16 = einsum(equation = var_2202_equation_0, values = (var_2102_cast_fp16_9, var_2172_cast_fp16))[name = tensor("op_2202_cast_fp16")]; + tensor var_2204_equation_0 = const()[name = tensor("op_2204_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2204_cast_fp16 = einsum(equation = var_2204_equation_0, values = (var_2102_cast_fp16_10, var_2173_cast_fp16))[name = tensor("op_2204_cast_fp16")]; + tensor var_2206_equation_0 = const()[name = tensor("op_2206_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2206_cast_fp16 = einsum(equation = var_2206_equation_0, values = (var_2102_cast_fp16_11, var_2174_cast_fp16))[name = tensor("op_2206_cast_fp16")]; + tensor var_2208_equation_0 = const()[name = tensor("op_2208_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2208_cast_fp16 = einsum(equation = var_2208_equation_0, values = (var_2102_cast_fp16_12, var_2175_cast_fp16))[name = tensor("op_2208_cast_fp16")]; + tensor var_2210_equation_0 = const()[name = tensor("op_2210_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2210_cast_fp16 = einsum(equation = var_2210_equation_0, values = (var_2102_cast_fp16_13, var_2176_cast_fp16))[name = tensor("op_2210_cast_fp16")]; + tensor var_2212_equation_0 = const()[name = tensor("op_2212_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2212_cast_fp16 = einsum(equation = var_2212_equation_0, values = (var_2102_cast_fp16_14, var_2177_cast_fp16))[name = tensor("op_2212_cast_fp16")]; + tensor var_2214_equation_0 = const()[name = tensor("op_2214_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2214_cast_fp16 = einsum(equation = var_2214_equation_0, values = (var_2102_cast_fp16_15, var_2178_cast_fp16))[name = tensor("op_2214_cast_fp16")]; + tensor var_2216_equation_0 = const()[name = tensor("op_2216_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2216_cast_fp16 = einsum(equation = var_2216_equation_0, values = (var_2102_cast_fp16_16, var_2179_cast_fp16))[name = tensor("op_2216_cast_fp16")]; + tensor var_2218_equation_0 = const()[name = tensor("op_2218_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2218_cast_fp16 = einsum(equation = var_2218_equation_0, values = (var_2102_cast_fp16_17, var_2180_cast_fp16))[name = tensor("op_2218_cast_fp16")]; + tensor var_2220_equation_0 = const()[name = tensor("op_2220_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2220_cast_fp16 = einsum(equation = var_2220_equation_0, values = (var_2102_cast_fp16_18, var_2181_cast_fp16))[name = tensor("op_2220_cast_fp16")]; + tensor var_2222_equation_0 = const()[name = tensor("op_2222_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2222_cast_fp16 = einsum(equation = var_2222_equation_0, values = (var_2102_cast_fp16_19, var_2182_cast_fp16))[name = tensor("op_2222_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_2007, interleave = input_75_interleave_0, values = (var_2184_cast_fp16, var_2186_cast_fp16, var_2188_cast_fp16, var_2190_cast_fp16, var_2192_cast_fp16, var_2194_cast_fp16, var_2196_cast_fp16, var_2198_cast_fp16, var_2200_cast_fp16, var_2202_cast_fp16, var_2204_cast_fp16, var_2206_cast_fp16, var_2208_cast_fp16, var_2210_cast_fp16, var_2212_cast_fp16, var_2214_cast_fp16, var_2216_cast_fp16, var_2218_cast_fp16, var_2220_cast_fp16, var_2222_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_2231_pad_type_0 = const()[name = tensor("op_2231_pad_type_0"), val = tensor("valid")]; + tensor var_2231_strides_0 = const()[name = tensor("op_2231_strides_0"), val = tensor([1, 1])]; + tensor var_2231_pad_0 = const()[name = tensor("op_2231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2231_dilations_0 = const()[name = tensor("op_2231_dilations_0"), val = tensor([1, 1])]; + tensor var_2231_groups_0 = const()[name = tensor("op_2231_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299972992)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303249856)))]; + tensor var_2231_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_2231_dilations_0, groups = var_2231_groups_0, pad = var_2231_pad_0, pad_type = var_2231_pad_type_0, strides = var_2231_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_2231_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303252480)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303255104)))]; + tensor var_2241_to_fp16 = const()[name = tensor("op_2241_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_2241_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303257728)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316364992)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_2267_pad_type_0 = const()[name = tensor("op_2267_pad_type_0"), val = tensor("valid")]; + tensor var_2267_strides_0 = const()[name = tensor("op_2267_strides_0"), val = tensor([1, 1])]; + tensor var_2267_pad_0 = const()[name = tensor("op_2267_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2267_dilations_0 = const()[name = tensor("op_2267_dilations_0"), val = tensor([1, 1])]; + tensor var_2267_groups_0 = const()[name = tensor("op_2267_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316375296)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329482560)))]; + tensor var_2267_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_2267_dilations_0, groups = var_2267_groups_0, pad = var_2267_pad_0, pad_type = var_2267_pad_type_0, strides = var_2267_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_2267_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_2276 = const()[name = tensor("op_2276"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329485184)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329487808)))]; + tensor var_2292_to_fp16 = const()[name = tensor("op_2292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_2292_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_2327_weight_0_to_fp16 = const()[name = tensor("op_2327_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329490432)))]; + tensor var_2327_bias_0_to_fp16 = const()[name = tensor("op_2327_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332767296)))]; + tensor var_2327_cast_fp16 = conv(bias = var_2327_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_2327_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2327_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332769920)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_2325_pad_type_0 = const()[name = tensor("op_2325_pad_type_0"), val = tensor("valid")]; + tensor var_2325_strides_0 = const()[name = tensor("op_2325_strides_0"), val = tensor([1, 1])]; + tensor var_2325_pad_0 = const()[name = tensor("op_2325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2325_dilations_0 = const()[name = tensor("op_2325_dilations_0"), val = tensor([1, 1])]; + tensor var_2325_groups_0 = const()[name = tensor("op_2325_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336046784)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339323648)))]; + tensor var_2325_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_2325_dilations_0, groups = var_2325_groups_0, pad = var_2325_pad_0, pad_type = var_2325_pad_type_0, strides = var_2325_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2328_axis_0 = const()[name = tensor("op_2328_axis_0"), val = tensor(1)]; + tensor var_2328_cast_fp16_0, tensor var_2328_cast_fp16_1, tensor var_2328_cast_fp16_2, tensor var_2328_cast_fp16_3, tensor var_2328_cast_fp16_4, tensor var_2328_cast_fp16_5, tensor var_2328_cast_fp16_6, tensor var_2328_cast_fp16_7, tensor var_2328_cast_fp16_8, tensor var_2328_cast_fp16_9, tensor var_2328_cast_fp16_10, tensor var_2328_cast_fp16_11, tensor var_2328_cast_fp16_12, tensor var_2328_cast_fp16_13, tensor var_2328_cast_fp16_14, tensor var_2328_cast_fp16_15, tensor var_2328_cast_fp16_16, tensor var_2328_cast_fp16_17, tensor var_2328_cast_fp16_18, tensor var_2328_cast_fp16_19 = split(axis = var_2328_axis_0, split_sizes = tile_24, x = var_2327_cast_fp16)[name = tensor("op_2328_cast_fp16")]; + tensor var_2349_perm_0 = const()[name = tensor("op_2349_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2350_axis_0 = const()[name = tensor("op_2350_axis_0"), val = tensor(3)]; + tensor var_2349_cast_fp16 = transpose(perm = var_2349_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_24")]; + tensor var_2350_cast_fp16_0, tensor var_2350_cast_fp16_1, tensor var_2350_cast_fp16_2, tensor var_2350_cast_fp16_3, tensor var_2350_cast_fp16_4, tensor var_2350_cast_fp16_5, tensor var_2350_cast_fp16_6, tensor var_2350_cast_fp16_7, tensor var_2350_cast_fp16_8, tensor var_2350_cast_fp16_9, tensor var_2350_cast_fp16_10, tensor var_2350_cast_fp16_11, tensor var_2350_cast_fp16_12, tensor var_2350_cast_fp16_13, tensor var_2350_cast_fp16_14, tensor var_2350_cast_fp16_15, tensor var_2350_cast_fp16_16, tensor var_2350_cast_fp16_17, tensor var_2350_cast_fp16_18, tensor var_2350_cast_fp16_19 = split(axis = var_2350_axis_0, split_sizes = tile_25, x = var_2349_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2371_axis_0 = const()[name = tensor("op_2371_axis_0"), val = tensor(1)]; + tensor var_2371_cast_fp16_0, tensor var_2371_cast_fp16_1, tensor var_2371_cast_fp16_2, tensor var_2371_cast_fp16_3, tensor var_2371_cast_fp16_4, tensor var_2371_cast_fp16_5, tensor var_2371_cast_fp16_6, tensor var_2371_cast_fp16_7, tensor var_2371_cast_fp16_8, tensor var_2371_cast_fp16_9, tensor var_2371_cast_fp16_10, tensor var_2371_cast_fp16_11, tensor var_2371_cast_fp16_12, tensor var_2371_cast_fp16_13, tensor var_2371_cast_fp16_14, tensor var_2371_cast_fp16_15, tensor var_2371_cast_fp16_16, tensor var_2371_cast_fp16_17, tensor var_2371_cast_fp16_18, tensor var_2371_cast_fp16_19 = split(axis = var_2371_axis_0, split_sizes = tile_26, x = var_2325_cast_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor aw_321_equation_0 = const()[name = tensor("aw_321_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_321_cast_fp16 = einsum(equation = aw_321_equation_0, values = (var_2350_cast_fp16_0, var_2328_cast_fp16_0))[name = tensor("aw_321_cast_fp16")]; + tensor aw_323_equation_0 = const()[name = tensor("aw_323_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_323_cast_fp16 = einsum(equation = aw_323_equation_0, values = (var_2350_cast_fp16_1, var_2328_cast_fp16_1))[name = tensor("aw_323_cast_fp16")]; + tensor aw_325_equation_0 = const()[name = tensor("aw_325_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_325_cast_fp16 = einsum(equation = aw_325_equation_0, values = (var_2350_cast_fp16_2, var_2328_cast_fp16_2))[name = tensor("aw_325_cast_fp16")]; + tensor aw_327_equation_0 = const()[name = tensor("aw_327_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_327_cast_fp16 = einsum(equation = aw_327_equation_0, values = (var_2350_cast_fp16_3, var_2328_cast_fp16_3))[name = tensor("aw_327_cast_fp16")]; + tensor aw_329_equation_0 = const()[name = tensor("aw_329_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_329_cast_fp16 = einsum(equation = aw_329_equation_0, values = (var_2350_cast_fp16_4, var_2328_cast_fp16_4))[name = tensor("aw_329_cast_fp16")]; + tensor aw_331_equation_0 = const()[name = tensor("aw_331_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_331_cast_fp16 = einsum(equation = aw_331_equation_0, values = (var_2350_cast_fp16_5, var_2328_cast_fp16_5))[name = tensor("aw_331_cast_fp16")]; + tensor aw_333_equation_0 = const()[name = tensor("aw_333_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_333_cast_fp16 = einsum(equation = aw_333_equation_0, values = (var_2350_cast_fp16_6, var_2328_cast_fp16_6))[name = tensor("aw_333_cast_fp16")]; + tensor aw_335_equation_0 = const()[name = tensor("aw_335_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_335_cast_fp16 = einsum(equation = aw_335_equation_0, values = (var_2350_cast_fp16_7, var_2328_cast_fp16_7))[name = tensor("aw_335_cast_fp16")]; + tensor aw_337_equation_0 = const()[name = tensor("aw_337_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_337_cast_fp16 = einsum(equation = aw_337_equation_0, values = (var_2350_cast_fp16_8, var_2328_cast_fp16_8))[name = tensor("aw_337_cast_fp16")]; + tensor aw_339_equation_0 = const()[name = tensor("aw_339_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_339_cast_fp16 = einsum(equation = aw_339_equation_0, values = (var_2350_cast_fp16_9, var_2328_cast_fp16_9))[name = tensor("aw_339_cast_fp16")]; + tensor aw_341_equation_0 = const()[name = tensor("aw_341_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_341_cast_fp16 = einsum(equation = aw_341_equation_0, values = (var_2350_cast_fp16_10, var_2328_cast_fp16_10))[name = tensor("aw_341_cast_fp16")]; + tensor aw_343_equation_0 = const()[name = tensor("aw_343_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_343_cast_fp16 = einsum(equation = aw_343_equation_0, values = (var_2350_cast_fp16_11, var_2328_cast_fp16_11))[name = tensor("aw_343_cast_fp16")]; + tensor aw_345_equation_0 = const()[name = tensor("aw_345_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_345_cast_fp16 = einsum(equation = aw_345_equation_0, values = (var_2350_cast_fp16_12, var_2328_cast_fp16_12))[name = tensor("aw_345_cast_fp16")]; + tensor aw_347_equation_0 = const()[name = tensor("aw_347_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_347_cast_fp16 = einsum(equation = aw_347_equation_0, values = (var_2350_cast_fp16_13, var_2328_cast_fp16_13))[name = tensor("aw_347_cast_fp16")]; + tensor aw_349_equation_0 = const()[name = tensor("aw_349_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_349_cast_fp16 = einsum(equation = aw_349_equation_0, values = (var_2350_cast_fp16_14, var_2328_cast_fp16_14))[name = tensor("aw_349_cast_fp16")]; + tensor aw_351_equation_0 = const()[name = tensor("aw_351_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_351_cast_fp16 = einsum(equation = aw_351_equation_0, values = (var_2350_cast_fp16_15, var_2328_cast_fp16_15))[name = tensor("aw_351_cast_fp16")]; + tensor aw_353_equation_0 = const()[name = tensor("aw_353_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_353_cast_fp16 = einsum(equation = aw_353_equation_0, values = (var_2350_cast_fp16_16, var_2328_cast_fp16_16))[name = tensor("aw_353_cast_fp16")]; + tensor aw_355_equation_0 = const()[name = tensor("aw_355_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_355_cast_fp16 = einsum(equation = aw_355_equation_0, values = (var_2350_cast_fp16_17, var_2328_cast_fp16_17))[name = tensor("aw_355_cast_fp16")]; + tensor aw_357_equation_0 = const()[name = tensor("aw_357_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_357_cast_fp16 = einsum(equation = aw_357_equation_0, values = (var_2350_cast_fp16_18, var_2328_cast_fp16_18))[name = tensor("aw_357_cast_fp16")]; + tensor aw_359_equation_0 = const()[name = tensor("aw_359_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_359_cast_fp16 = einsum(equation = aw_359_equation_0, values = (var_2350_cast_fp16_19, var_2328_cast_fp16_19))[name = tensor("aw_359_cast_fp16")]; + tensor var_2432_cast_fp16 = softmax(axis = var_2276, x = aw_321_cast_fp16)[name = tensor("op_2432_cast_fp16")]; + tensor var_2433_cast_fp16 = softmax(axis = var_2276, x = aw_323_cast_fp16)[name = tensor("op_2433_cast_fp16")]; + tensor var_2434_cast_fp16 = softmax(axis = var_2276, x = aw_325_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = softmax(axis = var_2276, x = aw_327_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = softmax(axis = var_2276, x = aw_329_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2437_cast_fp16 = softmax(axis = var_2276, x = aw_331_cast_fp16)[name = tensor("op_2437_cast_fp16")]; + tensor var_2438_cast_fp16 = softmax(axis = var_2276, x = aw_333_cast_fp16)[name = tensor("op_2438_cast_fp16")]; + tensor var_2439_cast_fp16 = softmax(axis = var_2276, x = aw_335_cast_fp16)[name = tensor("op_2439_cast_fp16")]; + tensor var_2440_cast_fp16 = softmax(axis = var_2276, x = aw_337_cast_fp16)[name = tensor("op_2440_cast_fp16")]; + tensor var_2441_cast_fp16 = softmax(axis = var_2276, x = aw_339_cast_fp16)[name = tensor("op_2441_cast_fp16")]; + tensor var_2442_cast_fp16 = softmax(axis = var_2276, x = aw_341_cast_fp16)[name = tensor("op_2442_cast_fp16")]; + tensor var_2443_cast_fp16 = softmax(axis = var_2276, x = aw_343_cast_fp16)[name = tensor("op_2443_cast_fp16")]; + tensor var_2444_cast_fp16 = softmax(axis = var_2276, x = aw_345_cast_fp16)[name = tensor("op_2444_cast_fp16")]; + tensor var_2445_cast_fp16 = softmax(axis = var_2276, x = aw_347_cast_fp16)[name = tensor("op_2445_cast_fp16")]; + tensor var_2446_cast_fp16 = softmax(axis = var_2276, x = aw_349_cast_fp16)[name = tensor("op_2446_cast_fp16")]; + tensor var_2447_cast_fp16 = softmax(axis = var_2276, x = aw_351_cast_fp16)[name = tensor("op_2447_cast_fp16")]; + tensor var_2448_cast_fp16 = softmax(axis = var_2276, x = aw_353_cast_fp16)[name = tensor("op_2448_cast_fp16")]; + tensor var_2449_cast_fp16 = softmax(axis = var_2276, x = aw_355_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor var_2450_cast_fp16 = softmax(axis = var_2276, x = aw_357_cast_fp16)[name = tensor("op_2450_cast_fp16")]; + tensor var_2451_cast_fp16 = softmax(axis = var_2276, x = aw_359_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2453_equation_0 = const()[name = tensor("op_2453_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2453_cast_fp16 = einsum(equation = var_2453_equation_0, values = (var_2371_cast_fp16_0, var_2432_cast_fp16))[name = tensor("op_2453_cast_fp16")]; + tensor var_2455_equation_0 = const()[name = tensor("op_2455_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2455_cast_fp16 = einsum(equation = var_2455_equation_0, values = (var_2371_cast_fp16_1, var_2433_cast_fp16))[name = tensor("op_2455_cast_fp16")]; + tensor var_2457_equation_0 = const()[name = tensor("op_2457_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2457_cast_fp16 = einsum(equation = var_2457_equation_0, values = (var_2371_cast_fp16_2, var_2434_cast_fp16))[name = tensor("op_2457_cast_fp16")]; + tensor var_2459_equation_0 = const()[name = tensor("op_2459_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2459_cast_fp16 = einsum(equation = var_2459_equation_0, values = (var_2371_cast_fp16_3, var_2435_cast_fp16))[name = tensor("op_2459_cast_fp16")]; + tensor var_2461_equation_0 = const()[name = tensor("op_2461_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2461_cast_fp16 = einsum(equation = var_2461_equation_0, values = (var_2371_cast_fp16_4, var_2436_cast_fp16))[name = tensor("op_2461_cast_fp16")]; + tensor var_2463_equation_0 = const()[name = tensor("op_2463_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2463_cast_fp16 = einsum(equation = var_2463_equation_0, values = (var_2371_cast_fp16_5, var_2437_cast_fp16))[name = tensor("op_2463_cast_fp16")]; + tensor var_2465_equation_0 = const()[name = tensor("op_2465_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2465_cast_fp16 = einsum(equation = var_2465_equation_0, values = (var_2371_cast_fp16_6, var_2438_cast_fp16))[name = tensor("op_2465_cast_fp16")]; + tensor var_2467_equation_0 = const()[name = tensor("op_2467_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2467_cast_fp16 = einsum(equation = var_2467_equation_0, values = (var_2371_cast_fp16_7, var_2439_cast_fp16))[name = tensor("op_2467_cast_fp16")]; + tensor var_2469_equation_0 = const()[name = tensor("op_2469_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2469_cast_fp16 = einsum(equation = var_2469_equation_0, values = (var_2371_cast_fp16_8, var_2440_cast_fp16))[name = tensor("op_2469_cast_fp16")]; + tensor var_2471_equation_0 = const()[name = tensor("op_2471_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2471_cast_fp16 = einsum(equation = var_2471_equation_0, values = (var_2371_cast_fp16_9, var_2441_cast_fp16))[name = tensor("op_2471_cast_fp16")]; + tensor var_2473_equation_0 = const()[name = tensor("op_2473_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2473_cast_fp16 = einsum(equation = var_2473_equation_0, values = (var_2371_cast_fp16_10, var_2442_cast_fp16))[name = tensor("op_2473_cast_fp16")]; + tensor var_2475_equation_0 = const()[name = tensor("op_2475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2475_cast_fp16 = einsum(equation = var_2475_equation_0, values = (var_2371_cast_fp16_11, var_2443_cast_fp16))[name = tensor("op_2475_cast_fp16")]; + tensor var_2477_equation_0 = const()[name = tensor("op_2477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2477_cast_fp16 = einsum(equation = var_2477_equation_0, values = (var_2371_cast_fp16_12, var_2444_cast_fp16))[name = tensor("op_2477_cast_fp16")]; + tensor var_2479_equation_0 = const()[name = tensor("op_2479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2479_cast_fp16 = einsum(equation = var_2479_equation_0, values = (var_2371_cast_fp16_13, var_2445_cast_fp16))[name = tensor("op_2479_cast_fp16")]; + tensor var_2481_equation_0 = const()[name = tensor("op_2481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2481_cast_fp16 = einsum(equation = var_2481_equation_0, values = (var_2371_cast_fp16_14, var_2446_cast_fp16))[name = tensor("op_2481_cast_fp16")]; + tensor var_2483_equation_0 = const()[name = tensor("op_2483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2483_cast_fp16 = einsum(equation = var_2483_equation_0, values = (var_2371_cast_fp16_15, var_2447_cast_fp16))[name = tensor("op_2483_cast_fp16")]; + tensor var_2485_equation_0 = const()[name = tensor("op_2485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2485_cast_fp16 = einsum(equation = var_2485_equation_0, values = (var_2371_cast_fp16_16, var_2448_cast_fp16))[name = tensor("op_2485_cast_fp16")]; + tensor var_2487_equation_0 = const()[name = tensor("op_2487_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2487_cast_fp16 = einsum(equation = var_2487_equation_0, values = (var_2371_cast_fp16_17, var_2449_cast_fp16))[name = tensor("op_2487_cast_fp16")]; + tensor var_2489_equation_0 = const()[name = tensor("op_2489_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2489_cast_fp16 = einsum(equation = var_2489_equation_0, values = (var_2371_cast_fp16_18, var_2450_cast_fp16))[name = tensor("op_2489_cast_fp16")]; + tensor var_2491_equation_0 = const()[name = tensor("op_2491_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2491_cast_fp16 = einsum(equation = var_2491_equation_0, values = (var_2371_cast_fp16_19, var_2451_cast_fp16))[name = tensor("op_2491_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_2276, interleave = input_85_interleave_0, values = (var_2453_cast_fp16, var_2455_cast_fp16, var_2457_cast_fp16, var_2459_cast_fp16, var_2461_cast_fp16, var_2463_cast_fp16, var_2465_cast_fp16, var_2467_cast_fp16, var_2469_cast_fp16, var_2471_cast_fp16, var_2473_cast_fp16, var_2475_cast_fp16, var_2477_cast_fp16, var_2479_cast_fp16, var_2481_cast_fp16, var_2483_cast_fp16, var_2485_cast_fp16, var_2487_cast_fp16, var_2489_cast_fp16, var_2491_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_2500_pad_type_0 = const()[name = tensor("op_2500_pad_type_0"), val = tensor("valid")]; + tensor var_2500_strides_0 = const()[name = tensor("op_2500_strides_0"), val = tensor([1, 1])]; + tensor var_2500_pad_0 = const()[name = tensor("op_2500_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2500_dilations_0 = const()[name = tensor("op_2500_dilations_0"), val = tensor([1, 1])]; + tensor var_2500_groups_0 = const()[name = tensor("op_2500_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339326272)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342603136)))]; + tensor var_2500_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_2500_dilations_0, groups = var_2500_groups_0, pad = var_2500_pad_0, pad_type = var_2500_pad_type_0, strides = var_2500_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_2500_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342605760)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342608384)))]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_2510_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342611008)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355718272)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2536_pad_type_0 = const()[name = tensor("op_2536_pad_type_0"), val = tensor("valid")]; + tensor var_2536_strides_0 = const()[name = tensor("op_2536_strides_0"), val = tensor([1, 1])]; + tensor var_2536_pad_0 = const()[name = tensor("op_2536_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2536_dilations_0 = const()[name = tensor("op_2536_dilations_0"), val = tensor([1, 1])]; + tensor var_2536_groups_0 = const()[name = tensor("op_2536_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355728576)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368835840)))]; + tensor var_2536_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_2536_dilations_0, groups = var_2536_groups_0, pad = var_2536_pad_0, pad_type = var_2536_pad_type_0, strides = var_2536_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_2536_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2545 = const()[name = tensor("op_2545"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368838464)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368841088)))]; + tensor var_2561_to_fp16 = const()[name = tensor("op_2561_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_2561_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_2596_weight_0_to_fp16 = const()[name = tensor("op_2596_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368843712)))]; + tensor var_2596_bias_0_to_fp16 = const()[name = tensor("op_2596_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372120576)))]; + tensor var_2596_cast_fp16 = conv(bias = var_2596_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_2596_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372123200)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_2594_pad_type_0 = const()[name = tensor("op_2594_pad_type_0"), val = tensor("valid")]; + tensor var_2594_strides_0 = const()[name = tensor("op_2594_strides_0"), val = tensor([1, 1])]; + tensor var_2594_pad_0 = const()[name = tensor("op_2594_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2594_dilations_0 = const()[name = tensor("op_2594_dilations_0"), val = tensor([1, 1])]; + tensor var_2594_groups_0 = const()[name = tensor("op_2594_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375400064)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378676928)))]; + tensor var_2594_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_2594_dilations_0, groups = var_2594_groups_0, pad = var_2594_pad_0, pad_type = var_2594_pad_type_0, strides = var_2594_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2594_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2597_axis_0 = const()[name = tensor("op_2597_axis_0"), val = tensor(1)]; + tensor var_2597_cast_fp16_0, tensor var_2597_cast_fp16_1, tensor var_2597_cast_fp16_2, tensor var_2597_cast_fp16_3, tensor var_2597_cast_fp16_4, tensor var_2597_cast_fp16_5, tensor var_2597_cast_fp16_6, tensor var_2597_cast_fp16_7, tensor var_2597_cast_fp16_8, tensor var_2597_cast_fp16_9, tensor var_2597_cast_fp16_10, tensor var_2597_cast_fp16_11, tensor var_2597_cast_fp16_12, tensor var_2597_cast_fp16_13, tensor var_2597_cast_fp16_14, tensor var_2597_cast_fp16_15, tensor var_2597_cast_fp16_16, tensor var_2597_cast_fp16_17, tensor var_2597_cast_fp16_18, tensor var_2597_cast_fp16_19 = split(axis = var_2597_axis_0, split_sizes = tile_27, x = var_2596_cast_fp16)[name = tensor("op_2597_cast_fp16")]; + tensor var_2618_perm_0 = const()[name = tensor("op_2618_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2619_axis_0 = const()[name = tensor("op_2619_axis_0"), val = tensor(3)]; + tensor var_2618_cast_fp16 = transpose(perm = var_2618_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_23")]; + tensor var_2619_cast_fp16_0, tensor var_2619_cast_fp16_1, tensor var_2619_cast_fp16_2, tensor var_2619_cast_fp16_3, tensor var_2619_cast_fp16_4, tensor var_2619_cast_fp16_5, tensor var_2619_cast_fp16_6, tensor var_2619_cast_fp16_7, tensor var_2619_cast_fp16_8, tensor var_2619_cast_fp16_9, tensor var_2619_cast_fp16_10, tensor var_2619_cast_fp16_11, tensor var_2619_cast_fp16_12, tensor var_2619_cast_fp16_13, tensor var_2619_cast_fp16_14, tensor var_2619_cast_fp16_15, tensor var_2619_cast_fp16_16, tensor var_2619_cast_fp16_17, tensor var_2619_cast_fp16_18, tensor var_2619_cast_fp16_19 = split(axis = var_2619_axis_0, split_sizes = tile_28, x = var_2618_cast_fp16)[name = tensor("op_2619_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2640_axis_0 = const()[name = tensor("op_2640_axis_0"), val = tensor(1)]; + tensor var_2640_cast_fp16_0, tensor var_2640_cast_fp16_1, tensor var_2640_cast_fp16_2, tensor var_2640_cast_fp16_3, tensor var_2640_cast_fp16_4, tensor var_2640_cast_fp16_5, tensor var_2640_cast_fp16_6, tensor var_2640_cast_fp16_7, tensor var_2640_cast_fp16_8, tensor var_2640_cast_fp16_9, tensor var_2640_cast_fp16_10, tensor var_2640_cast_fp16_11, tensor var_2640_cast_fp16_12, tensor var_2640_cast_fp16_13, tensor var_2640_cast_fp16_14, tensor var_2640_cast_fp16_15, tensor var_2640_cast_fp16_16, tensor var_2640_cast_fp16_17, tensor var_2640_cast_fp16_18, tensor var_2640_cast_fp16_19 = split(axis = var_2640_axis_0, split_sizes = tile_29, x = var_2594_cast_fp16)[name = tensor("op_2640_cast_fp16")]; + tensor aw_361_equation_0 = const()[name = tensor("aw_361_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_361_cast_fp16 = einsum(equation = aw_361_equation_0, values = (var_2619_cast_fp16_0, var_2597_cast_fp16_0))[name = tensor("aw_361_cast_fp16")]; + tensor aw_363_equation_0 = const()[name = tensor("aw_363_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_363_cast_fp16 = einsum(equation = aw_363_equation_0, values = (var_2619_cast_fp16_1, var_2597_cast_fp16_1))[name = tensor("aw_363_cast_fp16")]; + tensor aw_365_equation_0 = const()[name = tensor("aw_365_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_365_cast_fp16 = einsum(equation = aw_365_equation_0, values = (var_2619_cast_fp16_2, var_2597_cast_fp16_2))[name = tensor("aw_365_cast_fp16")]; + tensor aw_367_equation_0 = const()[name = tensor("aw_367_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_367_cast_fp16 = einsum(equation = aw_367_equation_0, values = (var_2619_cast_fp16_3, var_2597_cast_fp16_3))[name = tensor("aw_367_cast_fp16")]; + tensor aw_369_equation_0 = const()[name = tensor("aw_369_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_369_cast_fp16 = einsum(equation = aw_369_equation_0, values = (var_2619_cast_fp16_4, var_2597_cast_fp16_4))[name = tensor("aw_369_cast_fp16")]; + tensor aw_371_equation_0 = const()[name = tensor("aw_371_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_371_cast_fp16 = einsum(equation = aw_371_equation_0, values = (var_2619_cast_fp16_5, var_2597_cast_fp16_5))[name = tensor("aw_371_cast_fp16")]; + tensor aw_373_equation_0 = const()[name = tensor("aw_373_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_373_cast_fp16 = einsum(equation = aw_373_equation_0, values = (var_2619_cast_fp16_6, var_2597_cast_fp16_6))[name = tensor("aw_373_cast_fp16")]; + tensor aw_375_equation_0 = const()[name = tensor("aw_375_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_375_cast_fp16 = einsum(equation = aw_375_equation_0, values = (var_2619_cast_fp16_7, var_2597_cast_fp16_7))[name = tensor("aw_375_cast_fp16")]; + tensor aw_377_equation_0 = const()[name = tensor("aw_377_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_377_cast_fp16 = einsum(equation = aw_377_equation_0, values = (var_2619_cast_fp16_8, var_2597_cast_fp16_8))[name = tensor("aw_377_cast_fp16")]; + tensor aw_379_equation_0 = const()[name = tensor("aw_379_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_379_cast_fp16 = einsum(equation = aw_379_equation_0, values = (var_2619_cast_fp16_9, var_2597_cast_fp16_9))[name = tensor("aw_379_cast_fp16")]; + tensor aw_381_equation_0 = const()[name = tensor("aw_381_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_381_cast_fp16 = einsum(equation = aw_381_equation_0, values = (var_2619_cast_fp16_10, var_2597_cast_fp16_10))[name = tensor("aw_381_cast_fp16")]; + tensor aw_383_equation_0 = const()[name = tensor("aw_383_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_383_cast_fp16 = einsum(equation = aw_383_equation_0, values = (var_2619_cast_fp16_11, var_2597_cast_fp16_11))[name = tensor("aw_383_cast_fp16")]; + tensor aw_385_equation_0 = const()[name = tensor("aw_385_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_385_cast_fp16 = einsum(equation = aw_385_equation_0, values = (var_2619_cast_fp16_12, var_2597_cast_fp16_12))[name = tensor("aw_385_cast_fp16")]; + tensor aw_387_equation_0 = const()[name = tensor("aw_387_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_387_cast_fp16 = einsum(equation = aw_387_equation_0, values = (var_2619_cast_fp16_13, var_2597_cast_fp16_13))[name = tensor("aw_387_cast_fp16")]; + tensor aw_389_equation_0 = const()[name = tensor("aw_389_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_389_cast_fp16 = einsum(equation = aw_389_equation_0, values = (var_2619_cast_fp16_14, var_2597_cast_fp16_14))[name = tensor("aw_389_cast_fp16")]; + tensor aw_391_equation_0 = const()[name = tensor("aw_391_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_391_cast_fp16 = einsum(equation = aw_391_equation_0, values = (var_2619_cast_fp16_15, var_2597_cast_fp16_15))[name = tensor("aw_391_cast_fp16")]; + tensor aw_393_equation_0 = const()[name = tensor("aw_393_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_393_cast_fp16 = einsum(equation = aw_393_equation_0, values = (var_2619_cast_fp16_16, var_2597_cast_fp16_16))[name = tensor("aw_393_cast_fp16")]; + tensor aw_395_equation_0 = const()[name = tensor("aw_395_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_395_cast_fp16 = einsum(equation = aw_395_equation_0, values = (var_2619_cast_fp16_17, var_2597_cast_fp16_17))[name = tensor("aw_395_cast_fp16")]; + tensor aw_397_equation_0 = const()[name = tensor("aw_397_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_397_cast_fp16 = einsum(equation = aw_397_equation_0, values = (var_2619_cast_fp16_18, var_2597_cast_fp16_18))[name = tensor("aw_397_cast_fp16")]; + tensor aw_399_equation_0 = const()[name = tensor("aw_399_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_399_cast_fp16 = einsum(equation = aw_399_equation_0, values = (var_2619_cast_fp16_19, var_2597_cast_fp16_19))[name = tensor("aw_399_cast_fp16")]; + tensor var_2701_cast_fp16 = softmax(axis = var_2545, x = aw_361_cast_fp16)[name = tensor("op_2701_cast_fp16")]; + tensor var_2702_cast_fp16 = softmax(axis = var_2545, x = aw_363_cast_fp16)[name = tensor("op_2702_cast_fp16")]; + tensor var_2703_cast_fp16 = softmax(axis = var_2545, x = aw_365_cast_fp16)[name = tensor("op_2703_cast_fp16")]; + tensor var_2704_cast_fp16 = softmax(axis = var_2545, x = aw_367_cast_fp16)[name = tensor("op_2704_cast_fp16")]; + tensor var_2705_cast_fp16 = softmax(axis = var_2545, x = aw_369_cast_fp16)[name = tensor("op_2705_cast_fp16")]; + tensor var_2706_cast_fp16 = softmax(axis = var_2545, x = aw_371_cast_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = softmax(axis = var_2545, x = aw_373_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor var_2708_cast_fp16 = softmax(axis = var_2545, x = aw_375_cast_fp16)[name = tensor("op_2708_cast_fp16")]; + tensor var_2709_cast_fp16 = softmax(axis = var_2545, x = aw_377_cast_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor var_2710_cast_fp16 = softmax(axis = var_2545, x = aw_379_cast_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor var_2711_cast_fp16 = softmax(axis = var_2545, x = aw_381_cast_fp16)[name = tensor("op_2711_cast_fp16")]; + tensor var_2712_cast_fp16 = softmax(axis = var_2545, x = aw_383_cast_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = softmax(axis = var_2545, x = aw_385_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor var_2714_cast_fp16 = softmax(axis = var_2545, x = aw_387_cast_fp16)[name = tensor("op_2714_cast_fp16")]; + tensor var_2715_cast_fp16 = softmax(axis = var_2545, x = aw_389_cast_fp16)[name = tensor("op_2715_cast_fp16")]; + tensor var_2716_cast_fp16 = softmax(axis = var_2545, x = aw_391_cast_fp16)[name = tensor("op_2716_cast_fp16")]; + tensor var_2717_cast_fp16 = softmax(axis = var_2545, x = aw_393_cast_fp16)[name = tensor("op_2717_cast_fp16")]; + tensor var_2718_cast_fp16 = softmax(axis = var_2545, x = aw_395_cast_fp16)[name = tensor("op_2718_cast_fp16")]; + tensor var_2719_cast_fp16 = softmax(axis = var_2545, x = aw_397_cast_fp16)[name = tensor("op_2719_cast_fp16")]; + tensor var_2720_cast_fp16 = softmax(axis = var_2545, x = aw_399_cast_fp16)[name = tensor("op_2720_cast_fp16")]; + tensor var_2722_equation_0 = const()[name = tensor("op_2722_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2722_cast_fp16 = einsum(equation = var_2722_equation_0, values = (var_2640_cast_fp16_0, var_2701_cast_fp16))[name = tensor("op_2722_cast_fp16")]; + tensor var_2724_equation_0 = const()[name = tensor("op_2724_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2724_cast_fp16 = einsum(equation = var_2724_equation_0, values = (var_2640_cast_fp16_1, var_2702_cast_fp16))[name = tensor("op_2724_cast_fp16")]; + tensor var_2726_equation_0 = const()[name = tensor("op_2726_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2726_cast_fp16 = einsum(equation = var_2726_equation_0, values = (var_2640_cast_fp16_2, var_2703_cast_fp16))[name = tensor("op_2726_cast_fp16")]; + tensor var_2728_equation_0 = const()[name = tensor("op_2728_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2728_cast_fp16 = einsum(equation = var_2728_equation_0, values = (var_2640_cast_fp16_3, var_2704_cast_fp16))[name = tensor("op_2728_cast_fp16")]; + tensor var_2730_equation_0 = const()[name = tensor("op_2730_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2730_cast_fp16 = einsum(equation = var_2730_equation_0, values = (var_2640_cast_fp16_4, var_2705_cast_fp16))[name = tensor("op_2730_cast_fp16")]; + tensor var_2732_equation_0 = const()[name = tensor("op_2732_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2732_cast_fp16 = einsum(equation = var_2732_equation_0, values = (var_2640_cast_fp16_5, var_2706_cast_fp16))[name = tensor("op_2732_cast_fp16")]; + tensor var_2734_equation_0 = const()[name = tensor("op_2734_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2734_cast_fp16 = einsum(equation = var_2734_equation_0, values = (var_2640_cast_fp16_6, var_2707_cast_fp16))[name = tensor("op_2734_cast_fp16")]; + tensor var_2736_equation_0 = const()[name = tensor("op_2736_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2736_cast_fp16 = einsum(equation = var_2736_equation_0, values = (var_2640_cast_fp16_7, var_2708_cast_fp16))[name = tensor("op_2736_cast_fp16")]; + tensor var_2738_equation_0 = const()[name = tensor("op_2738_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2738_cast_fp16 = einsum(equation = var_2738_equation_0, values = (var_2640_cast_fp16_8, var_2709_cast_fp16))[name = tensor("op_2738_cast_fp16")]; + tensor var_2740_equation_0 = const()[name = tensor("op_2740_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2740_cast_fp16 = einsum(equation = var_2740_equation_0, values = (var_2640_cast_fp16_9, var_2710_cast_fp16))[name = tensor("op_2740_cast_fp16")]; + tensor var_2742_equation_0 = const()[name = tensor("op_2742_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2742_cast_fp16 = einsum(equation = var_2742_equation_0, values = (var_2640_cast_fp16_10, var_2711_cast_fp16))[name = tensor("op_2742_cast_fp16")]; + tensor var_2744_equation_0 = const()[name = tensor("op_2744_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2744_cast_fp16 = einsum(equation = var_2744_equation_0, values = (var_2640_cast_fp16_11, var_2712_cast_fp16))[name = tensor("op_2744_cast_fp16")]; + tensor var_2746_equation_0 = const()[name = tensor("op_2746_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2746_cast_fp16 = einsum(equation = var_2746_equation_0, values = (var_2640_cast_fp16_12, var_2713_cast_fp16))[name = tensor("op_2746_cast_fp16")]; + tensor var_2748_equation_0 = const()[name = tensor("op_2748_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2748_cast_fp16 = einsum(equation = var_2748_equation_0, values = (var_2640_cast_fp16_13, var_2714_cast_fp16))[name = tensor("op_2748_cast_fp16")]; + tensor var_2750_equation_0 = const()[name = tensor("op_2750_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2750_cast_fp16 = einsum(equation = var_2750_equation_0, values = (var_2640_cast_fp16_14, var_2715_cast_fp16))[name = tensor("op_2750_cast_fp16")]; + tensor var_2752_equation_0 = const()[name = tensor("op_2752_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2752_cast_fp16 = einsum(equation = var_2752_equation_0, values = (var_2640_cast_fp16_15, var_2716_cast_fp16))[name = tensor("op_2752_cast_fp16")]; + tensor var_2754_equation_0 = const()[name = tensor("op_2754_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2754_cast_fp16 = einsum(equation = var_2754_equation_0, values = (var_2640_cast_fp16_16, var_2717_cast_fp16))[name = tensor("op_2754_cast_fp16")]; + tensor var_2756_equation_0 = const()[name = tensor("op_2756_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2756_cast_fp16 = einsum(equation = var_2756_equation_0, values = (var_2640_cast_fp16_17, var_2718_cast_fp16))[name = tensor("op_2756_cast_fp16")]; + tensor var_2758_equation_0 = const()[name = tensor("op_2758_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2758_cast_fp16 = einsum(equation = var_2758_equation_0, values = (var_2640_cast_fp16_18, var_2719_cast_fp16))[name = tensor("op_2758_cast_fp16")]; + tensor var_2760_equation_0 = const()[name = tensor("op_2760_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2760_cast_fp16 = einsum(equation = var_2760_equation_0, values = (var_2640_cast_fp16_19, var_2720_cast_fp16))[name = tensor("op_2760_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_2545, interleave = input_95_interleave_0, values = (var_2722_cast_fp16, var_2724_cast_fp16, var_2726_cast_fp16, var_2728_cast_fp16, var_2730_cast_fp16, var_2732_cast_fp16, var_2734_cast_fp16, var_2736_cast_fp16, var_2738_cast_fp16, var_2740_cast_fp16, var_2742_cast_fp16, var_2744_cast_fp16, var_2746_cast_fp16, var_2748_cast_fp16, var_2750_cast_fp16, var_2752_cast_fp16, var_2754_cast_fp16, var_2756_cast_fp16, var_2758_cast_fp16, var_2760_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2769_pad_type_0 = const()[name = tensor("op_2769_pad_type_0"), val = tensor("valid")]; + tensor var_2769_strides_0 = const()[name = tensor("op_2769_strides_0"), val = tensor([1, 1])]; + tensor var_2769_pad_0 = const()[name = tensor("op_2769_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2769_dilations_0 = const()[name = tensor("op_2769_dilations_0"), val = tensor([1, 1])]; + tensor var_2769_groups_0 = const()[name = tensor("op_2769_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378679552)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381956416)))]; + tensor var_2769_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2769_dilations_0, groups = var_2769_groups_0, pad = var_2769_pad_0, pad_type = var_2769_pad_type_0, strides = var_2769_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2769_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2769_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381959040)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381961664)))]; + tensor var_2779_to_fp16 = const()[name = tensor("op_2779_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2779_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381964288)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395071552)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2805_pad_type_0 = const()[name = tensor("op_2805_pad_type_0"), val = tensor("valid")]; + tensor var_2805_strides_0 = const()[name = tensor("op_2805_strides_0"), val = tensor([1, 1])]; + tensor var_2805_pad_0 = const()[name = tensor("op_2805_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2805_dilations_0 = const()[name = tensor("op_2805_dilations_0"), val = tensor([1, 1])]; + tensor var_2805_groups_0 = const()[name = tensor("op_2805_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395081856)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408189120)))]; + tensor var_2805_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2805_dilations_0, groups = var_2805_groups_0, pad = var_2805_pad_0, pad_type = var_2805_pad_type_0, strides = var_2805_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2805_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2805_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408191744)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408194368)))]; + tensor var_2830_to_fp16 = const()[name = tensor("op_2830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2830_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2865_weight_0_to_fp16 = const()[name = tensor("op_2865_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408196992)))]; + tensor var_2865_bias_0_to_fp16 = const()[name = tensor("op_2865_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411473856)))]; + tensor var_2865_cast_fp16 = conv(bias = var_2865_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2865_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411476480)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2863_pad_type_0 = const()[name = tensor("op_2863_pad_type_0"), val = tensor("valid")]; + tensor var_2863_strides_0 = const()[name = tensor("op_2863_strides_0"), val = tensor([1, 1])]; + tensor var_2863_pad_0 = const()[name = tensor("op_2863_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2863_dilations_0 = const()[name = tensor("op_2863_dilations_0"), val = tensor([1, 1])]; + tensor var_2863_groups_0 = const()[name = tensor("op_2863_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414753344)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418030208)))]; + tensor var_2863_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2863_dilations_0, groups = var_2863_groups_0, pad = var_2863_pad_0, pad_type = var_2863_pad_type_0, strides = var_2863_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2863_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2866_axis_0 = const()[name = tensor("op_2866_axis_0"), val = tensor(1)]; + tensor var_2866_cast_fp16_0, tensor var_2866_cast_fp16_1, tensor var_2866_cast_fp16_2, tensor var_2866_cast_fp16_3, tensor var_2866_cast_fp16_4, tensor var_2866_cast_fp16_5, tensor var_2866_cast_fp16_6, tensor var_2866_cast_fp16_7, tensor var_2866_cast_fp16_8, tensor var_2866_cast_fp16_9, tensor var_2866_cast_fp16_10, tensor var_2866_cast_fp16_11, tensor var_2866_cast_fp16_12, tensor var_2866_cast_fp16_13, tensor var_2866_cast_fp16_14, tensor var_2866_cast_fp16_15, tensor var_2866_cast_fp16_16, tensor var_2866_cast_fp16_17, tensor var_2866_cast_fp16_18, tensor var_2866_cast_fp16_19 = split(axis = var_2866_axis_0, split_sizes = tile_30, x = var_2865_cast_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor var_2887_perm_0 = const()[name = tensor("op_2887_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2888_axis_0 = const()[name = tensor("op_2888_axis_0"), val = tensor(3)]; + tensor var_2887_cast_fp16 = transpose(perm = var_2887_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_22")]; + tensor var_2888_cast_fp16_0, tensor var_2888_cast_fp16_1, tensor var_2888_cast_fp16_2, tensor var_2888_cast_fp16_3, tensor var_2888_cast_fp16_4, tensor var_2888_cast_fp16_5, tensor var_2888_cast_fp16_6, tensor var_2888_cast_fp16_7, tensor var_2888_cast_fp16_8, tensor var_2888_cast_fp16_9, tensor var_2888_cast_fp16_10, tensor var_2888_cast_fp16_11, tensor var_2888_cast_fp16_12, tensor var_2888_cast_fp16_13, tensor var_2888_cast_fp16_14, tensor var_2888_cast_fp16_15, tensor var_2888_cast_fp16_16, tensor var_2888_cast_fp16_17, tensor var_2888_cast_fp16_18, tensor var_2888_cast_fp16_19 = split(axis = var_2888_axis_0, split_sizes = tile_31, x = var_2887_cast_fp16)[name = tensor("op_2888_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2909_axis_0 = const()[name = tensor("op_2909_axis_0"), val = tensor(1)]; + tensor var_2909_cast_fp16_0, tensor var_2909_cast_fp16_1, tensor var_2909_cast_fp16_2, tensor var_2909_cast_fp16_3, tensor var_2909_cast_fp16_4, tensor var_2909_cast_fp16_5, tensor var_2909_cast_fp16_6, tensor var_2909_cast_fp16_7, tensor var_2909_cast_fp16_8, tensor var_2909_cast_fp16_9, tensor var_2909_cast_fp16_10, tensor var_2909_cast_fp16_11, tensor var_2909_cast_fp16_12, tensor var_2909_cast_fp16_13, tensor var_2909_cast_fp16_14, tensor var_2909_cast_fp16_15, tensor var_2909_cast_fp16_16, tensor var_2909_cast_fp16_17, tensor var_2909_cast_fp16_18, tensor var_2909_cast_fp16_19 = split(axis = var_2909_axis_0, split_sizes = tile_32, x = var_2863_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor aw_401_equation_0 = const()[name = tensor("aw_401_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_401_cast_fp16 = einsum(equation = aw_401_equation_0, values = (var_2888_cast_fp16_0, var_2866_cast_fp16_0))[name = tensor("aw_401_cast_fp16")]; + tensor aw_403_equation_0 = const()[name = tensor("aw_403_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_403_cast_fp16 = einsum(equation = aw_403_equation_0, values = (var_2888_cast_fp16_1, var_2866_cast_fp16_1))[name = tensor("aw_403_cast_fp16")]; + tensor aw_405_equation_0 = const()[name = tensor("aw_405_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_405_cast_fp16 = einsum(equation = aw_405_equation_0, values = (var_2888_cast_fp16_2, var_2866_cast_fp16_2))[name = tensor("aw_405_cast_fp16")]; + tensor aw_407_equation_0 = const()[name = tensor("aw_407_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_407_cast_fp16 = einsum(equation = aw_407_equation_0, values = (var_2888_cast_fp16_3, var_2866_cast_fp16_3))[name = tensor("aw_407_cast_fp16")]; + tensor aw_409_equation_0 = const()[name = tensor("aw_409_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_409_cast_fp16 = einsum(equation = aw_409_equation_0, values = (var_2888_cast_fp16_4, var_2866_cast_fp16_4))[name = tensor("aw_409_cast_fp16")]; + tensor aw_411_equation_0 = const()[name = tensor("aw_411_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_411_cast_fp16 = einsum(equation = aw_411_equation_0, values = (var_2888_cast_fp16_5, var_2866_cast_fp16_5))[name = tensor("aw_411_cast_fp16")]; + tensor aw_413_equation_0 = const()[name = tensor("aw_413_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_413_cast_fp16 = einsum(equation = aw_413_equation_0, values = (var_2888_cast_fp16_6, var_2866_cast_fp16_6))[name = tensor("aw_413_cast_fp16")]; + tensor aw_415_equation_0 = const()[name = tensor("aw_415_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_415_cast_fp16 = einsum(equation = aw_415_equation_0, values = (var_2888_cast_fp16_7, var_2866_cast_fp16_7))[name = tensor("aw_415_cast_fp16")]; + tensor aw_417_equation_0 = const()[name = tensor("aw_417_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_417_cast_fp16 = einsum(equation = aw_417_equation_0, values = (var_2888_cast_fp16_8, var_2866_cast_fp16_8))[name = tensor("aw_417_cast_fp16")]; + tensor aw_419_equation_0 = const()[name = tensor("aw_419_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_419_cast_fp16 = einsum(equation = aw_419_equation_0, values = (var_2888_cast_fp16_9, var_2866_cast_fp16_9))[name = tensor("aw_419_cast_fp16")]; + tensor aw_421_equation_0 = const()[name = tensor("aw_421_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_421_cast_fp16 = einsum(equation = aw_421_equation_0, values = (var_2888_cast_fp16_10, var_2866_cast_fp16_10))[name = tensor("aw_421_cast_fp16")]; + tensor aw_423_equation_0 = const()[name = tensor("aw_423_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_423_cast_fp16 = einsum(equation = aw_423_equation_0, values = (var_2888_cast_fp16_11, var_2866_cast_fp16_11))[name = tensor("aw_423_cast_fp16")]; + tensor aw_425_equation_0 = const()[name = tensor("aw_425_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_425_cast_fp16 = einsum(equation = aw_425_equation_0, values = (var_2888_cast_fp16_12, var_2866_cast_fp16_12))[name = tensor("aw_425_cast_fp16")]; + tensor aw_427_equation_0 = const()[name = tensor("aw_427_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_427_cast_fp16 = einsum(equation = aw_427_equation_0, values = (var_2888_cast_fp16_13, var_2866_cast_fp16_13))[name = tensor("aw_427_cast_fp16")]; + tensor aw_429_equation_0 = const()[name = tensor("aw_429_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_429_cast_fp16 = einsum(equation = aw_429_equation_0, values = (var_2888_cast_fp16_14, var_2866_cast_fp16_14))[name = tensor("aw_429_cast_fp16")]; + tensor aw_431_equation_0 = const()[name = tensor("aw_431_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_431_cast_fp16 = einsum(equation = aw_431_equation_0, values = (var_2888_cast_fp16_15, var_2866_cast_fp16_15))[name = tensor("aw_431_cast_fp16")]; + tensor aw_433_equation_0 = const()[name = tensor("aw_433_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_433_cast_fp16 = einsum(equation = aw_433_equation_0, values = (var_2888_cast_fp16_16, var_2866_cast_fp16_16))[name = tensor("aw_433_cast_fp16")]; + tensor aw_435_equation_0 = const()[name = tensor("aw_435_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_435_cast_fp16 = einsum(equation = aw_435_equation_0, values = (var_2888_cast_fp16_17, var_2866_cast_fp16_17))[name = tensor("aw_435_cast_fp16")]; + tensor aw_437_equation_0 = const()[name = tensor("aw_437_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_437_cast_fp16 = einsum(equation = aw_437_equation_0, values = (var_2888_cast_fp16_18, var_2866_cast_fp16_18))[name = tensor("aw_437_cast_fp16")]; + tensor aw_439_equation_0 = const()[name = tensor("aw_439_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_439_cast_fp16 = einsum(equation = aw_439_equation_0, values = (var_2888_cast_fp16_19, var_2866_cast_fp16_19))[name = tensor("aw_439_cast_fp16")]; + tensor var_2970_cast_fp16 = softmax(axis = var_2814, x = aw_401_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = softmax(axis = var_2814, x = aw_403_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_cast_fp16 = softmax(axis = var_2814, x = aw_405_cast_fp16)[name = tensor("op_2972_cast_fp16")]; + tensor var_2973_cast_fp16 = softmax(axis = var_2814, x = aw_407_cast_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor var_2974_cast_fp16 = softmax(axis = var_2814, x = aw_409_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = softmax(axis = var_2814, x = aw_411_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor var_2976_cast_fp16 = softmax(axis = var_2814, x = aw_413_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor var_2977_cast_fp16 = softmax(axis = var_2814, x = aw_415_cast_fp16)[name = tensor("op_2977_cast_fp16")]; + tensor var_2978_cast_fp16 = softmax(axis = var_2814, x = aw_417_cast_fp16)[name = tensor("op_2978_cast_fp16")]; + tensor var_2979_cast_fp16 = softmax(axis = var_2814, x = aw_419_cast_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor var_2980_cast_fp16 = softmax(axis = var_2814, x = aw_421_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor var_2981_cast_fp16 = softmax(axis = var_2814, x = aw_423_cast_fp16)[name = tensor("op_2981_cast_fp16")]; + tensor var_2982_cast_fp16 = softmax(axis = var_2814, x = aw_425_cast_fp16)[name = tensor("op_2982_cast_fp16")]; + tensor var_2983_cast_fp16 = softmax(axis = var_2814, x = aw_427_cast_fp16)[name = tensor("op_2983_cast_fp16")]; + tensor var_2984_cast_fp16 = softmax(axis = var_2814, x = aw_429_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor var_2985_cast_fp16 = softmax(axis = var_2814, x = aw_431_cast_fp16)[name = tensor("op_2985_cast_fp16")]; + tensor var_2986_cast_fp16 = softmax(axis = var_2814, x = aw_433_cast_fp16)[name = tensor("op_2986_cast_fp16")]; + tensor var_2987_cast_fp16 = softmax(axis = var_2814, x = aw_435_cast_fp16)[name = tensor("op_2987_cast_fp16")]; + tensor var_2988_cast_fp16 = softmax(axis = var_2814, x = aw_437_cast_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2989_cast_fp16 = softmax(axis = var_2814, x = aw_439_cast_fp16)[name = tensor("op_2989_cast_fp16")]; + tensor var_2991_equation_0 = const()[name = tensor("op_2991_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2991_cast_fp16 = einsum(equation = var_2991_equation_0, values = (var_2909_cast_fp16_0, var_2970_cast_fp16))[name = tensor("op_2991_cast_fp16")]; + tensor var_2993_equation_0 = const()[name = tensor("op_2993_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2993_cast_fp16 = einsum(equation = var_2993_equation_0, values = (var_2909_cast_fp16_1, var_2971_cast_fp16))[name = tensor("op_2993_cast_fp16")]; + tensor var_2995_equation_0 = const()[name = tensor("op_2995_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2995_cast_fp16 = einsum(equation = var_2995_equation_0, values = (var_2909_cast_fp16_2, var_2972_cast_fp16))[name = tensor("op_2995_cast_fp16")]; + tensor var_2997_equation_0 = const()[name = tensor("op_2997_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2997_cast_fp16 = einsum(equation = var_2997_equation_0, values = (var_2909_cast_fp16_3, var_2973_cast_fp16))[name = tensor("op_2997_cast_fp16")]; + tensor var_2999_equation_0 = const()[name = tensor("op_2999_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2999_cast_fp16 = einsum(equation = var_2999_equation_0, values = (var_2909_cast_fp16_4, var_2974_cast_fp16))[name = tensor("op_2999_cast_fp16")]; + tensor var_3001_equation_0 = const()[name = tensor("op_3001_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3001_cast_fp16 = einsum(equation = var_3001_equation_0, values = (var_2909_cast_fp16_5, var_2975_cast_fp16))[name = tensor("op_3001_cast_fp16")]; + tensor var_3003_equation_0 = const()[name = tensor("op_3003_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3003_cast_fp16 = einsum(equation = var_3003_equation_0, values = (var_2909_cast_fp16_6, var_2976_cast_fp16))[name = tensor("op_3003_cast_fp16")]; + tensor var_3005_equation_0 = const()[name = tensor("op_3005_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3005_cast_fp16 = einsum(equation = var_3005_equation_0, values = (var_2909_cast_fp16_7, var_2977_cast_fp16))[name = tensor("op_3005_cast_fp16")]; + tensor var_3007_equation_0 = const()[name = tensor("op_3007_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3007_cast_fp16 = einsum(equation = var_3007_equation_0, values = (var_2909_cast_fp16_8, var_2978_cast_fp16))[name = tensor("op_3007_cast_fp16")]; + tensor var_3009_equation_0 = const()[name = tensor("op_3009_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3009_cast_fp16 = einsum(equation = var_3009_equation_0, values = (var_2909_cast_fp16_9, var_2979_cast_fp16))[name = tensor("op_3009_cast_fp16")]; + tensor var_3011_equation_0 = const()[name = tensor("op_3011_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3011_cast_fp16 = einsum(equation = var_3011_equation_0, values = (var_2909_cast_fp16_10, var_2980_cast_fp16))[name = tensor("op_3011_cast_fp16")]; + tensor var_3013_equation_0 = const()[name = tensor("op_3013_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3013_cast_fp16 = einsum(equation = var_3013_equation_0, values = (var_2909_cast_fp16_11, var_2981_cast_fp16))[name = tensor("op_3013_cast_fp16")]; + tensor var_3015_equation_0 = const()[name = tensor("op_3015_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3015_cast_fp16 = einsum(equation = var_3015_equation_0, values = (var_2909_cast_fp16_12, var_2982_cast_fp16))[name = tensor("op_3015_cast_fp16")]; + tensor var_3017_equation_0 = const()[name = tensor("op_3017_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3017_cast_fp16 = einsum(equation = var_3017_equation_0, values = (var_2909_cast_fp16_13, var_2983_cast_fp16))[name = tensor("op_3017_cast_fp16")]; + tensor var_3019_equation_0 = const()[name = tensor("op_3019_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3019_cast_fp16 = einsum(equation = var_3019_equation_0, values = (var_2909_cast_fp16_14, var_2984_cast_fp16))[name = tensor("op_3019_cast_fp16")]; + tensor var_3021_equation_0 = const()[name = tensor("op_3021_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3021_cast_fp16 = einsum(equation = var_3021_equation_0, values = (var_2909_cast_fp16_15, var_2985_cast_fp16))[name = tensor("op_3021_cast_fp16")]; + tensor var_3023_equation_0 = const()[name = tensor("op_3023_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3023_cast_fp16 = einsum(equation = var_3023_equation_0, values = (var_2909_cast_fp16_16, var_2986_cast_fp16))[name = tensor("op_3023_cast_fp16")]; + tensor var_3025_equation_0 = const()[name = tensor("op_3025_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3025_cast_fp16 = einsum(equation = var_3025_equation_0, values = (var_2909_cast_fp16_17, var_2987_cast_fp16))[name = tensor("op_3025_cast_fp16")]; + tensor var_3027_equation_0 = const()[name = tensor("op_3027_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3027_cast_fp16 = einsum(equation = var_3027_equation_0, values = (var_2909_cast_fp16_18, var_2988_cast_fp16))[name = tensor("op_3027_cast_fp16")]; + tensor var_3029_equation_0 = const()[name = tensor("op_3029_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3029_cast_fp16 = einsum(equation = var_3029_equation_0, values = (var_2909_cast_fp16_19, var_2989_cast_fp16))[name = tensor("op_3029_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2814, interleave = input_105_interleave_0, values = (var_2991_cast_fp16, var_2993_cast_fp16, var_2995_cast_fp16, var_2997_cast_fp16, var_2999_cast_fp16, var_3001_cast_fp16, var_3003_cast_fp16, var_3005_cast_fp16, var_3007_cast_fp16, var_3009_cast_fp16, var_3011_cast_fp16, var_3013_cast_fp16, var_3015_cast_fp16, var_3017_cast_fp16, var_3019_cast_fp16, var_3021_cast_fp16, var_3023_cast_fp16, var_3025_cast_fp16, var_3027_cast_fp16, var_3029_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_3038_pad_type_0 = const()[name = tensor("op_3038_pad_type_0"), val = tensor("valid")]; + tensor var_3038_strides_0 = const()[name = tensor("op_3038_strides_0"), val = tensor([1, 1])]; + tensor var_3038_pad_0 = const()[name = tensor("op_3038_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3038_dilations_0 = const()[name = tensor("op_3038_dilations_0"), val = tensor([1, 1])]; + tensor var_3038_groups_0 = const()[name = tensor("op_3038_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418032832)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421309696)))]; + tensor var_3038_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_3038_dilations_0, groups = var_3038_groups_0, pad = var_3038_pad_0, pad_type = var_3038_pad_type_0, strides = var_3038_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_3038_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421312320)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421314944)))]; + tensor var_3048_to_fp16 = const()[name = tensor("op_3048_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_3048_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421317568)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434424832)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_3074_pad_type_0 = const()[name = tensor("op_3074_pad_type_0"), val = tensor("valid")]; + tensor var_3074_strides_0 = const()[name = tensor("op_3074_strides_0"), val = tensor([1, 1])]; + tensor var_3074_pad_0 = const()[name = tensor("op_3074_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3074_dilations_0 = const()[name = tensor("op_3074_dilations_0"), val = tensor([1, 1])]; + tensor var_3074_groups_0 = const()[name = tensor("op_3074_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434435136)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447542400)))]; + tensor var_3074_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_3074_dilations_0, groups = var_3074_groups_0, pad = var_3074_pad_0, pad_type = var_3074_pad_type_0, strides = var_3074_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_3074_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_3074_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447545024)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447547648)))]; + tensor var_3099_to_fp16 = const()[name = tensor("op_3099_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_3099_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_23_pad_type_0 = const()[name = tensor("q_23_pad_type_0"), val = tensor("valid")]; + tensor q_23_strides_0 = const()[name = tensor("q_23_strides_0"), val = tensor([1, 1])]; + tensor q_23_pad_0 = const()[name = tensor("q_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_23_dilations_0 = const()[name = tensor("q_23_dilations_0"), val = tensor([1, 1])]; + tensor q_23_groups_0 = const()[name = tensor("q_23_groups_0"), val = tensor(1)]; + tensor var_3134_weight_0_to_fp16 = const()[name = tensor("op_3134_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447550272)))]; + tensor var_3134_bias_0_to_fp16 = const()[name = tensor("op_3134_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450827136)))]; + tensor var_3134_cast_fp16 = conv(bias = var_3134_bias_0_to_fp16, dilations = q_23_dilations_0, groups = q_23_groups_0, pad = q_23_pad_0, pad_type = q_23_pad_type_0, strides = q_23_strides_0, weight = var_3134_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor k_23_pad_type_0 = const()[name = tensor("k_23_pad_type_0"), val = tensor("valid")]; + tensor k_23_strides_0 = const()[name = tensor("k_23_strides_0"), val = tensor([1, 1])]; + tensor k_23_pad_0 = const()[name = tensor("k_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_23_dilations_0 = const()[name = tensor("k_23_dilations_0"), val = tensor([1, 1])]; + tensor k_23_groups_0 = const()[name = tensor("k_23_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450829760)))]; + tensor k_23_cast_fp16 = conv(dilations = k_23_dilations_0, groups = k_23_groups_0, pad = k_23_pad_0, pad_type = k_23_pad_type_0, strides = k_23_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_3132_pad_type_0 = const()[name = tensor("op_3132_pad_type_0"), val = tensor("valid")]; + tensor var_3132_strides_0 = const()[name = tensor("op_3132_strides_0"), val = tensor([1, 1])]; + tensor var_3132_pad_0 = const()[name = tensor("op_3132_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3132_dilations_0 = const()[name = tensor("op_3132_dilations_0"), val = tensor([1, 1])]; + tensor var_3132_groups_0 = const()[name = tensor("op_3132_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454106624)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457383488)))]; + tensor var_3132_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_3132_dilations_0, groups = var_3132_groups_0, pad = var_3132_pad_0, pad_type = var_3132_pad_type_0, strides = var_3132_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_3132_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3135_axis_0 = const()[name = tensor("op_3135_axis_0"), val = tensor(1)]; + tensor var_3135_cast_fp16_0, tensor var_3135_cast_fp16_1, tensor var_3135_cast_fp16_2, tensor var_3135_cast_fp16_3, tensor var_3135_cast_fp16_4, tensor var_3135_cast_fp16_5, tensor var_3135_cast_fp16_6, tensor var_3135_cast_fp16_7, tensor var_3135_cast_fp16_8, tensor var_3135_cast_fp16_9, tensor var_3135_cast_fp16_10, tensor var_3135_cast_fp16_11, tensor var_3135_cast_fp16_12, tensor var_3135_cast_fp16_13, tensor var_3135_cast_fp16_14, tensor var_3135_cast_fp16_15, tensor var_3135_cast_fp16_16, tensor var_3135_cast_fp16_17, tensor var_3135_cast_fp16_18, tensor var_3135_cast_fp16_19 = split(axis = var_3135_axis_0, split_sizes = tile_33, x = var_3134_cast_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3156_perm_0 = const()[name = tensor("op_3156_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3157_axis_0 = const()[name = tensor("op_3157_axis_0"), val = tensor(3)]; + tensor var_3156_cast_fp16 = transpose(perm = var_3156_perm_0, x = k_23_cast_fp16)[name = tensor("transpose_21")]; + tensor var_3157_cast_fp16_0, tensor var_3157_cast_fp16_1, tensor var_3157_cast_fp16_2, tensor var_3157_cast_fp16_3, tensor var_3157_cast_fp16_4, tensor var_3157_cast_fp16_5, tensor var_3157_cast_fp16_6, tensor var_3157_cast_fp16_7, tensor var_3157_cast_fp16_8, tensor var_3157_cast_fp16_9, tensor var_3157_cast_fp16_10, tensor var_3157_cast_fp16_11, tensor var_3157_cast_fp16_12, tensor var_3157_cast_fp16_13, tensor var_3157_cast_fp16_14, tensor var_3157_cast_fp16_15, tensor var_3157_cast_fp16_16, tensor var_3157_cast_fp16_17, tensor var_3157_cast_fp16_18, tensor var_3157_cast_fp16_19 = split(axis = var_3157_axis_0, split_sizes = tile_34, x = var_3156_cast_fp16)[name = tensor("op_3157_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3178_axis_0 = const()[name = tensor("op_3178_axis_0"), val = tensor(1)]; + tensor var_3178_cast_fp16_0, tensor var_3178_cast_fp16_1, tensor var_3178_cast_fp16_2, tensor var_3178_cast_fp16_3, tensor var_3178_cast_fp16_4, tensor var_3178_cast_fp16_5, tensor var_3178_cast_fp16_6, tensor var_3178_cast_fp16_7, tensor var_3178_cast_fp16_8, tensor var_3178_cast_fp16_9, tensor var_3178_cast_fp16_10, tensor var_3178_cast_fp16_11, tensor var_3178_cast_fp16_12, tensor var_3178_cast_fp16_13, tensor var_3178_cast_fp16_14, tensor var_3178_cast_fp16_15, tensor var_3178_cast_fp16_16, tensor var_3178_cast_fp16_17, tensor var_3178_cast_fp16_18, tensor var_3178_cast_fp16_19 = split(axis = var_3178_axis_0, split_sizes = tile_35, x = var_3132_cast_fp16)[name = tensor("op_3178_cast_fp16")]; + tensor aw_441_equation_0 = const()[name = tensor("aw_441_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_441_cast_fp16 = einsum(equation = aw_441_equation_0, values = (var_3157_cast_fp16_0, var_3135_cast_fp16_0))[name = tensor("aw_441_cast_fp16")]; + tensor aw_443_equation_0 = const()[name = tensor("aw_443_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_443_cast_fp16 = einsum(equation = aw_443_equation_0, values = (var_3157_cast_fp16_1, var_3135_cast_fp16_1))[name = tensor("aw_443_cast_fp16")]; + tensor aw_445_equation_0 = const()[name = tensor("aw_445_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_445_cast_fp16 = einsum(equation = aw_445_equation_0, values = (var_3157_cast_fp16_2, var_3135_cast_fp16_2))[name = tensor("aw_445_cast_fp16")]; + tensor aw_447_equation_0 = const()[name = tensor("aw_447_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_447_cast_fp16 = einsum(equation = aw_447_equation_0, values = (var_3157_cast_fp16_3, var_3135_cast_fp16_3))[name = tensor("aw_447_cast_fp16")]; + tensor aw_449_equation_0 = const()[name = tensor("aw_449_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_449_cast_fp16 = einsum(equation = aw_449_equation_0, values = (var_3157_cast_fp16_4, var_3135_cast_fp16_4))[name = tensor("aw_449_cast_fp16")]; + tensor aw_451_equation_0 = const()[name = tensor("aw_451_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_451_cast_fp16 = einsum(equation = aw_451_equation_0, values = (var_3157_cast_fp16_5, var_3135_cast_fp16_5))[name = tensor("aw_451_cast_fp16")]; + tensor aw_453_equation_0 = const()[name = tensor("aw_453_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_453_cast_fp16 = einsum(equation = aw_453_equation_0, values = (var_3157_cast_fp16_6, var_3135_cast_fp16_6))[name = tensor("aw_453_cast_fp16")]; + tensor aw_455_equation_0 = const()[name = tensor("aw_455_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_455_cast_fp16 = einsum(equation = aw_455_equation_0, values = (var_3157_cast_fp16_7, var_3135_cast_fp16_7))[name = tensor("aw_455_cast_fp16")]; + tensor aw_457_equation_0 = const()[name = tensor("aw_457_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_457_cast_fp16 = einsum(equation = aw_457_equation_0, values = (var_3157_cast_fp16_8, var_3135_cast_fp16_8))[name = tensor("aw_457_cast_fp16")]; + tensor aw_459_equation_0 = const()[name = tensor("aw_459_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_459_cast_fp16 = einsum(equation = aw_459_equation_0, values = (var_3157_cast_fp16_9, var_3135_cast_fp16_9))[name = tensor("aw_459_cast_fp16")]; + tensor aw_461_equation_0 = const()[name = tensor("aw_461_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_461_cast_fp16 = einsum(equation = aw_461_equation_0, values = (var_3157_cast_fp16_10, var_3135_cast_fp16_10))[name = tensor("aw_461_cast_fp16")]; + tensor aw_463_equation_0 = const()[name = tensor("aw_463_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_463_cast_fp16 = einsum(equation = aw_463_equation_0, values = (var_3157_cast_fp16_11, var_3135_cast_fp16_11))[name = tensor("aw_463_cast_fp16")]; + tensor aw_465_equation_0 = const()[name = tensor("aw_465_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_465_cast_fp16 = einsum(equation = aw_465_equation_0, values = (var_3157_cast_fp16_12, var_3135_cast_fp16_12))[name = tensor("aw_465_cast_fp16")]; + tensor aw_467_equation_0 = const()[name = tensor("aw_467_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_467_cast_fp16 = einsum(equation = aw_467_equation_0, values = (var_3157_cast_fp16_13, var_3135_cast_fp16_13))[name = tensor("aw_467_cast_fp16")]; + tensor aw_469_equation_0 = const()[name = tensor("aw_469_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_469_cast_fp16 = einsum(equation = aw_469_equation_0, values = (var_3157_cast_fp16_14, var_3135_cast_fp16_14))[name = tensor("aw_469_cast_fp16")]; + tensor aw_471_equation_0 = const()[name = tensor("aw_471_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_471_cast_fp16 = einsum(equation = aw_471_equation_0, values = (var_3157_cast_fp16_15, var_3135_cast_fp16_15))[name = tensor("aw_471_cast_fp16")]; + tensor aw_473_equation_0 = const()[name = tensor("aw_473_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_473_cast_fp16 = einsum(equation = aw_473_equation_0, values = (var_3157_cast_fp16_16, var_3135_cast_fp16_16))[name = tensor("aw_473_cast_fp16")]; + tensor aw_475_equation_0 = const()[name = tensor("aw_475_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_475_cast_fp16 = einsum(equation = aw_475_equation_0, values = (var_3157_cast_fp16_17, var_3135_cast_fp16_17))[name = tensor("aw_475_cast_fp16")]; + tensor aw_477_equation_0 = const()[name = tensor("aw_477_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_477_cast_fp16 = einsum(equation = aw_477_equation_0, values = (var_3157_cast_fp16_18, var_3135_cast_fp16_18))[name = tensor("aw_477_cast_fp16")]; + tensor aw_479_equation_0 = const()[name = tensor("aw_479_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_479_cast_fp16 = einsum(equation = aw_479_equation_0, values = (var_3157_cast_fp16_19, var_3135_cast_fp16_19))[name = tensor("aw_479_cast_fp16")]; + tensor var_3239_cast_fp16 = softmax(axis = var_3083, x = aw_441_cast_fp16)[name = tensor("op_3239_cast_fp16")]; + tensor var_3240_cast_fp16 = softmax(axis = var_3083, x = aw_443_cast_fp16)[name = tensor("op_3240_cast_fp16")]; + tensor var_3241_cast_fp16 = softmax(axis = var_3083, x = aw_445_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor var_3242_cast_fp16 = softmax(axis = var_3083, x = aw_447_cast_fp16)[name = tensor("op_3242_cast_fp16")]; + tensor var_3243_cast_fp16 = softmax(axis = var_3083, x = aw_449_cast_fp16)[name = tensor("op_3243_cast_fp16")]; + tensor var_3244_cast_fp16 = softmax(axis = var_3083, x = aw_451_cast_fp16)[name = tensor("op_3244_cast_fp16")]; + tensor var_3245_cast_fp16 = softmax(axis = var_3083, x = aw_453_cast_fp16)[name = tensor("op_3245_cast_fp16")]; + tensor var_3246_cast_fp16 = softmax(axis = var_3083, x = aw_455_cast_fp16)[name = tensor("op_3246_cast_fp16")]; + tensor var_3247_cast_fp16 = softmax(axis = var_3083, x = aw_457_cast_fp16)[name = tensor("op_3247_cast_fp16")]; + tensor var_3248_cast_fp16 = softmax(axis = var_3083, x = aw_459_cast_fp16)[name = tensor("op_3248_cast_fp16")]; + tensor var_3249_cast_fp16 = softmax(axis = var_3083, x = aw_461_cast_fp16)[name = tensor("op_3249_cast_fp16")]; + tensor var_3250_cast_fp16 = softmax(axis = var_3083, x = aw_463_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = softmax(axis = var_3083, x = aw_465_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor var_3252_cast_fp16 = softmax(axis = var_3083, x = aw_467_cast_fp16)[name = tensor("op_3252_cast_fp16")]; + tensor var_3253_cast_fp16 = softmax(axis = var_3083, x = aw_469_cast_fp16)[name = tensor("op_3253_cast_fp16")]; + tensor var_3254_cast_fp16 = softmax(axis = var_3083, x = aw_471_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = softmax(axis = var_3083, x = aw_473_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor var_3256_cast_fp16 = softmax(axis = var_3083, x = aw_475_cast_fp16)[name = tensor("op_3256_cast_fp16")]; + tensor var_3257_cast_fp16 = softmax(axis = var_3083, x = aw_477_cast_fp16)[name = tensor("op_3257_cast_fp16")]; + tensor var_3258_cast_fp16 = softmax(axis = var_3083, x = aw_479_cast_fp16)[name = tensor("op_3258_cast_fp16")]; + tensor var_3260_equation_0 = const()[name = tensor("op_3260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3260_cast_fp16 = einsum(equation = var_3260_equation_0, values = (var_3178_cast_fp16_0, var_3239_cast_fp16))[name = tensor("op_3260_cast_fp16")]; + tensor var_3262_equation_0 = const()[name = tensor("op_3262_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3262_cast_fp16 = einsum(equation = var_3262_equation_0, values = (var_3178_cast_fp16_1, var_3240_cast_fp16))[name = tensor("op_3262_cast_fp16")]; + tensor var_3264_equation_0 = const()[name = tensor("op_3264_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3264_cast_fp16 = einsum(equation = var_3264_equation_0, values = (var_3178_cast_fp16_2, var_3241_cast_fp16))[name = tensor("op_3264_cast_fp16")]; + tensor var_3266_equation_0 = const()[name = tensor("op_3266_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3266_cast_fp16 = einsum(equation = var_3266_equation_0, values = (var_3178_cast_fp16_3, var_3242_cast_fp16))[name = tensor("op_3266_cast_fp16")]; + tensor var_3268_equation_0 = const()[name = tensor("op_3268_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3268_cast_fp16 = einsum(equation = var_3268_equation_0, values = (var_3178_cast_fp16_4, var_3243_cast_fp16))[name = tensor("op_3268_cast_fp16")]; + tensor var_3270_equation_0 = const()[name = tensor("op_3270_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3270_cast_fp16 = einsum(equation = var_3270_equation_0, values = (var_3178_cast_fp16_5, var_3244_cast_fp16))[name = tensor("op_3270_cast_fp16")]; + tensor var_3272_equation_0 = const()[name = tensor("op_3272_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3272_cast_fp16 = einsum(equation = var_3272_equation_0, values = (var_3178_cast_fp16_6, var_3245_cast_fp16))[name = tensor("op_3272_cast_fp16")]; + tensor var_3274_equation_0 = const()[name = tensor("op_3274_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3274_cast_fp16 = einsum(equation = var_3274_equation_0, values = (var_3178_cast_fp16_7, var_3246_cast_fp16))[name = tensor("op_3274_cast_fp16")]; + tensor var_3276_equation_0 = const()[name = tensor("op_3276_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3276_cast_fp16 = einsum(equation = var_3276_equation_0, values = (var_3178_cast_fp16_8, var_3247_cast_fp16))[name = tensor("op_3276_cast_fp16")]; + tensor var_3278_equation_0 = const()[name = tensor("op_3278_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3278_cast_fp16 = einsum(equation = var_3278_equation_0, values = (var_3178_cast_fp16_9, var_3248_cast_fp16))[name = tensor("op_3278_cast_fp16")]; + tensor var_3280_equation_0 = const()[name = tensor("op_3280_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3280_cast_fp16 = einsum(equation = var_3280_equation_0, values = (var_3178_cast_fp16_10, var_3249_cast_fp16))[name = tensor("op_3280_cast_fp16")]; + tensor var_3282_equation_0 = const()[name = tensor("op_3282_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3282_cast_fp16 = einsum(equation = var_3282_equation_0, values = (var_3178_cast_fp16_11, var_3250_cast_fp16))[name = tensor("op_3282_cast_fp16")]; + tensor var_3284_equation_0 = const()[name = tensor("op_3284_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3284_cast_fp16 = einsum(equation = var_3284_equation_0, values = (var_3178_cast_fp16_12, var_3251_cast_fp16))[name = tensor("op_3284_cast_fp16")]; + tensor var_3286_equation_0 = const()[name = tensor("op_3286_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3286_cast_fp16 = einsum(equation = var_3286_equation_0, values = (var_3178_cast_fp16_13, var_3252_cast_fp16))[name = tensor("op_3286_cast_fp16")]; + tensor var_3288_equation_0 = const()[name = tensor("op_3288_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3288_cast_fp16 = einsum(equation = var_3288_equation_0, values = (var_3178_cast_fp16_14, var_3253_cast_fp16))[name = tensor("op_3288_cast_fp16")]; + tensor var_3290_equation_0 = const()[name = tensor("op_3290_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3290_cast_fp16 = einsum(equation = var_3290_equation_0, values = (var_3178_cast_fp16_15, var_3254_cast_fp16))[name = tensor("op_3290_cast_fp16")]; + tensor var_3292_equation_0 = const()[name = tensor("op_3292_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3292_cast_fp16 = einsum(equation = var_3292_equation_0, values = (var_3178_cast_fp16_16, var_3255_cast_fp16))[name = tensor("op_3292_cast_fp16")]; + tensor var_3294_equation_0 = const()[name = tensor("op_3294_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3294_cast_fp16 = einsum(equation = var_3294_equation_0, values = (var_3178_cast_fp16_17, var_3256_cast_fp16))[name = tensor("op_3294_cast_fp16")]; + tensor var_3296_equation_0 = const()[name = tensor("op_3296_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3296_cast_fp16 = einsum(equation = var_3296_equation_0, values = (var_3178_cast_fp16_18, var_3257_cast_fp16))[name = tensor("op_3296_cast_fp16")]; + tensor var_3298_equation_0 = const()[name = tensor("op_3298_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3298_cast_fp16 = einsum(equation = var_3298_equation_0, values = (var_3178_cast_fp16_19, var_3258_cast_fp16))[name = tensor("op_3298_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_3083, interleave = input_115_interleave_0, values = (var_3260_cast_fp16, var_3262_cast_fp16, var_3264_cast_fp16, var_3266_cast_fp16, var_3268_cast_fp16, var_3270_cast_fp16, var_3272_cast_fp16, var_3274_cast_fp16, var_3276_cast_fp16, var_3278_cast_fp16, var_3280_cast_fp16, var_3282_cast_fp16, var_3284_cast_fp16, var_3286_cast_fp16, var_3288_cast_fp16, var_3290_cast_fp16, var_3292_cast_fp16, var_3294_cast_fp16, var_3296_cast_fp16, var_3298_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_3307_pad_type_0 = const()[name = tensor("op_3307_pad_type_0"), val = tensor("valid")]; + tensor var_3307_strides_0 = const()[name = tensor("op_3307_strides_0"), val = tensor([1, 1])]; + tensor var_3307_pad_0 = const()[name = tensor("op_3307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3307_dilations_0 = const()[name = tensor("op_3307_dilations_0"), val = tensor([1, 1])]; + tensor var_3307_groups_0 = const()[name = tensor("op_3307_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457386112)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460662976)))]; + tensor var_3307_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_3307_dilations_0, groups = var_3307_groups_0, pad = var_3307_pad_0, pad_type = var_3307_pad_type_0, strides = var_3307_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_3307_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460665600)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460668224)))]; + tensor var_3317_to_fp16 = const()[name = tensor("op_3317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_3317_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460670848)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473778112)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_3343_pad_type_0 = const()[name = tensor("op_3343_pad_type_0"), val = tensor("valid")]; + tensor var_3343_strides_0 = const()[name = tensor("op_3343_strides_0"), val = tensor([1, 1])]; + tensor var_3343_pad_0 = const()[name = tensor("op_3343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3343_dilations_0 = const()[name = tensor("op_3343_dilations_0"), val = tensor([1, 1])]; + tensor var_3343_groups_0 = const()[name = tensor("op_3343_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473788416)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486895680)))]; + tensor var_3343_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_3343_dilations_0, groups = var_3343_groups_0, pad = var_3343_pad_0, pad_type = var_3343_pad_type_0, strides = var_3343_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("op_3343_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_3343_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor(1)]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([1])]; + tensor input_123_gamma_0_to_fp16 = const()[name = tensor("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486898304)))]; + tensor input_123_beta_0_to_fp16 = const()[name = tensor("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486900928)))]; + tensor var_3368_to_fp16 = const()[name = tensor("op_3368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = input_123_beta_0_to_fp16, epsilon = var_3368_to_fp16, gamma = input_123_gamma_0_to_fp16, x = inputs_49_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor q_25_pad_type_0 = const()[name = tensor("q_25_pad_type_0"), val = tensor("valid")]; + tensor q_25_strides_0 = const()[name = tensor("q_25_strides_0"), val = tensor([1, 1])]; + tensor q_25_pad_0 = const()[name = tensor("q_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_25_dilations_0 = const()[name = tensor("q_25_dilations_0"), val = tensor([1, 1])]; + tensor q_25_groups_0 = const()[name = tensor("q_25_groups_0"), val = tensor(1)]; + tensor var_3403_weight_0_to_fp16 = const()[name = tensor("op_3403_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486903552)))]; + tensor var_3403_bias_0_to_fp16 = const()[name = tensor("op_3403_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490180416)))]; + tensor var_3403_cast_fp16 = conv(bias = var_3403_bias_0_to_fp16, dilations = q_25_dilations_0, groups = q_25_groups_0, pad = q_25_pad_0, pad_type = q_25_pad_type_0, strides = q_25_strides_0, weight = var_3403_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3403_cast_fp16")]; + tensor k_25_pad_type_0 = const()[name = tensor("k_25_pad_type_0"), val = tensor("valid")]; + tensor k_25_strides_0 = const()[name = tensor("k_25_strides_0"), val = tensor([1, 1])]; + tensor k_25_pad_0 = const()[name = tensor("k_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_25_dilations_0 = const()[name = tensor("k_25_dilations_0"), val = tensor([1, 1])]; + tensor k_25_groups_0 = const()[name = tensor("k_25_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_key_weight_to_fp16 = const()[name = tensor("blocks_12_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490183040)))]; + tensor k_25_cast_fp16 = conv(dilations = k_25_dilations_0, groups = k_25_groups_0, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = k_25_strides_0, weight = blocks_12_attn_key_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor var_3401_pad_type_0 = const()[name = tensor("op_3401_pad_type_0"), val = tensor("valid")]; + tensor var_3401_strides_0 = const()[name = tensor("op_3401_strides_0"), val = tensor([1, 1])]; + tensor var_3401_pad_0 = const()[name = tensor("op_3401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3401_dilations_0 = const()[name = tensor("op_3401_dilations_0"), val = tensor([1, 1])]; + tensor var_3401_groups_0 = const()[name = tensor("op_3401_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_value_weight_to_fp16 = const()[name = tensor("blocks_12_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493459904)))]; + tensor blocks_12_attn_value_bias_to_fp16 = const()[name = tensor("blocks_12_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496736768)))]; + tensor var_3401_cast_fp16 = conv(bias = blocks_12_attn_value_bias_to_fp16, dilations = var_3401_dilations_0, groups = var_3401_groups_0, pad = var_3401_pad_0, pad_type = var_3401_pad_type_0, strides = var_3401_strides_0, weight = blocks_12_attn_value_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3401_cast_fp16")]; + tensor tile_36 = const()[name = tensor("tile_36"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3404_axis_0 = const()[name = tensor("op_3404_axis_0"), val = tensor(1)]; + tensor var_3404_cast_fp16_0, tensor var_3404_cast_fp16_1, tensor var_3404_cast_fp16_2, tensor var_3404_cast_fp16_3, tensor var_3404_cast_fp16_4, tensor var_3404_cast_fp16_5, tensor var_3404_cast_fp16_6, tensor var_3404_cast_fp16_7, tensor var_3404_cast_fp16_8, tensor var_3404_cast_fp16_9, tensor var_3404_cast_fp16_10, tensor var_3404_cast_fp16_11, tensor var_3404_cast_fp16_12, tensor var_3404_cast_fp16_13, tensor var_3404_cast_fp16_14, tensor var_3404_cast_fp16_15, tensor var_3404_cast_fp16_16, tensor var_3404_cast_fp16_17, tensor var_3404_cast_fp16_18, tensor var_3404_cast_fp16_19 = split(axis = var_3404_axis_0, split_sizes = tile_36, x = var_3403_cast_fp16)[name = tensor("op_3404_cast_fp16")]; + tensor var_3425_perm_0 = const()[name = tensor("op_3425_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_37 = const()[name = tensor("tile_37"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3426_axis_0 = const()[name = tensor("op_3426_axis_0"), val = tensor(3)]; + tensor var_3425_cast_fp16 = transpose(perm = var_3425_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_20")]; + tensor var_3426_cast_fp16_0, tensor var_3426_cast_fp16_1, tensor var_3426_cast_fp16_2, tensor var_3426_cast_fp16_3, tensor var_3426_cast_fp16_4, tensor var_3426_cast_fp16_5, tensor var_3426_cast_fp16_6, tensor var_3426_cast_fp16_7, tensor var_3426_cast_fp16_8, tensor var_3426_cast_fp16_9, tensor var_3426_cast_fp16_10, tensor var_3426_cast_fp16_11, tensor var_3426_cast_fp16_12, tensor var_3426_cast_fp16_13, tensor var_3426_cast_fp16_14, tensor var_3426_cast_fp16_15, tensor var_3426_cast_fp16_16, tensor var_3426_cast_fp16_17, tensor var_3426_cast_fp16_18, tensor var_3426_cast_fp16_19 = split(axis = var_3426_axis_0, split_sizes = tile_37, x = var_3425_cast_fp16)[name = tensor("op_3426_cast_fp16")]; + tensor tile_38 = const()[name = tensor("tile_38"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3447_axis_0 = const()[name = tensor("op_3447_axis_0"), val = tensor(1)]; + tensor var_3447_cast_fp16_0, tensor var_3447_cast_fp16_1, tensor var_3447_cast_fp16_2, tensor var_3447_cast_fp16_3, tensor var_3447_cast_fp16_4, tensor var_3447_cast_fp16_5, tensor var_3447_cast_fp16_6, tensor var_3447_cast_fp16_7, tensor var_3447_cast_fp16_8, tensor var_3447_cast_fp16_9, tensor var_3447_cast_fp16_10, tensor var_3447_cast_fp16_11, tensor var_3447_cast_fp16_12, tensor var_3447_cast_fp16_13, tensor var_3447_cast_fp16_14, tensor var_3447_cast_fp16_15, tensor var_3447_cast_fp16_16, tensor var_3447_cast_fp16_17, tensor var_3447_cast_fp16_18, tensor var_3447_cast_fp16_19 = split(axis = var_3447_axis_0, split_sizes = tile_38, x = var_3401_cast_fp16)[name = tensor("op_3447_cast_fp16")]; + tensor aw_481_equation_0 = const()[name = tensor("aw_481_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_481_cast_fp16 = einsum(equation = aw_481_equation_0, values = (var_3426_cast_fp16_0, var_3404_cast_fp16_0))[name = tensor("aw_481_cast_fp16")]; + tensor aw_483_equation_0 = const()[name = tensor("aw_483_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_483_cast_fp16 = einsum(equation = aw_483_equation_0, values = (var_3426_cast_fp16_1, var_3404_cast_fp16_1))[name = tensor("aw_483_cast_fp16")]; + tensor aw_485_equation_0 = const()[name = tensor("aw_485_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_485_cast_fp16 = einsum(equation = aw_485_equation_0, values = (var_3426_cast_fp16_2, var_3404_cast_fp16_2))[name = tensor("aw_485_cast_fp16")]; + tensor aw_487_equation_0 = const()[name = tensor("aw_487_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_487_cast_fp16 = einsum(equation = aw_487_equation_0, values = (var_3426_cast_fp16_3, var_3404_cast_fp16_3))[name = tensor("aw_487_cast_fp16")]; + tensor aw_489_equation_0 = const()[name = tensor("aw_489_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_489_cast_fp16 = einsum(equation = aw_489_equation_0, values = (var_3426_cast_fp16_4, var_3404_cast_fp16_4))[name = tensor("aw_489_cast_fp16")]; + tensor aw_491_equation_0 = const()[name = tensor("aw_491_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_491_cast_fp16 = einsum(equation = aw_491_equation_0, values = (var_3426_cast_fp16_5, var_3404_cast_fp16_5))[name = tensor("aw_491_cast_fp16")]; + tensor aw_493_equation_0 = const()[name = tensor("aw_493_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_493_cast_fp16 = einsum(equation = aw_493_equation_0, values = (var_3426_cast_fp16_6, var_3404_cast_fp16_6))[name = tensor("aw_493_cast_fp16")]; + tensor aw_495_equation_0 = const()[name = tensor("aw_495_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_495_cast_fp16 = einsum(equation = aw_495_equation_0, values = (var_3426_cast_fp16_7, var_3404_cast_fp16_7))[name = tensor("aw_495_cast_fp16")]; + tensor aw_497_equation_0 = const()[name = tensor("aw_497_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_497_cast_fp16 = einsum(equation = aw_497_equation_0, values = (var_3426_cast_fp16_8, var_3404_cast_fp16_8))[name = tensor("aw_497_cast_fp16")]; + tensor aw_499_equation_0 = const()[name = tensor("aw_499_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_499_cast_fp16 = einsum(equation = aw_499_equation_0, values = (var_3426_cast_fp16_9, var_3404_cast_fp16_9))[name = tensor("aw_499_cast_fp16")]; + tensor aw_501_equation_0 = const()[name = tensor("aw_501_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_501_cast_fp16 = einsum(equation = aw_501_equation_0, values = (var_3426_cast_fp16_10, var_3404_cast_fp16_10))[name = tensor("aw_501_cast_fp16")]; + tensor aw_503_equation_0 = const()[name = tensor("aw_503_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_503_cast_fp16 = einsum(equation = aw_503_equation_0, values = (var_3426_cast_fp16_11, var_3404_cast_fp16_11))[name = tensor("aw_503_cast_fp16")]; + tensor aw_505_equation_0 = const()[name = tensor("aw_505_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_505_cast_fp16 = einsum(equation = aw_505_equation_0, values = (var_3426_cast_fp16_12, var_3404_cast_fp16_12))[name = tensor("aw_505_cast_fp16")]; + tensor aw_507_equation_0 = const()[name = tensor("aw_507_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_507_cast_fp16 = einsum(equation = aw_507_equation_0, values = (var_3426_cast_fp16_13, var_3404_cast_fp16_13))[name = tensor("aw_507_cast_fp16")]; + tensor aw_509_equation_0 = const()[name = tensor("aw_509_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_509_cast_fp16 = einsum(equation = aw_509_equation_0, values = (var_3426_cast_fp16_14, var_3404_cast_fp16_14))[name = tensor("aw_509_cast_fp16")]; + tensor aw_511_equation_0 = const()[name = tensor("aw_511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_511_cast_fp16 = einsum(equation = aw_511_equation_0, values = (var_3426_cast_fp16_15, var_3404_cast_fp16_15))[name = tensor("aw_511_cast_fp16")]; + tensor aw_513_equation_0 = const()[name = tensor("aw_513_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_513_cast_fp16 = einsum(equation = aw_513_equation_0, values = (var_3426_cast_fp16_16, var_3404_cast_fp16_16))[name = tensor("aw_513_cast_fp16")]; + tensor aw_515_equation_0 = const()[name = tensor("aw_515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_515_cast_fp16 = einsum(equation = aw_515_equation_0, values = (var_3426_cast_fp16_17, var_3404_cast_fp16_17))[name = tensor("aw_515_cast_fp16")]; + tensor aw_517_equation_0 = const()[name = tensor("aw_517_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_517_cast_fp16 = einsum(equation = aw_517_equation_0, values = (var_3426_cast_fp16_18, var_3404_cast_fp16_18))[name = tensor("aw_517_cast_fp16")]; + tensor aw_519_equation_0 = const()[name = tensor("aw_519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_519_cast_fp16 = einsum(equation = aw_519_equation_0, values = (var_3426_cast_fp16_19, var_3404_cast_fp16_19))[name = tensor("aw_519_cast_fp16")]; + tensor var_3508_cast_fp16 = softmax(axis = var_3352, x = aw_481_cast_fp16)[name = tensor("op_3508_cast_fp16")]; + tensor var_3509_cast_fp16 = softmax(axis = var_3352, x = aw_483_cast_fp16)[name = tensor("op_3509_cast_fp16")]; + tensor var_3510_cast_fp16 = softmax(axis = var_3352, x = aw_485_cast_fp16)[name = tensor("op_3510_cast_fp16")]; + tensor var_3511_cast_fp16 = softmax(axis = var_3352, x = aw_487_cast_fp16)[name = tensor("op_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = softmax(axis = var_3352, x = aw_489_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_cast_fp16 = softmax(axis = var_3352, x = aw_491_cast_fp16)[name = tensor("op_3513_cast_fp16")]; + tensor var_3514_cast_fp16 = softmax(axis = var_3352, x = aw_493_cast_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor var_3515_cast_fp16 = softmax(axis = var_3352, x = aw_495_cast_fp16)[name = tensor("op_3515_cast_fp16")]; + tensor var_3516_cast_fp16 = softmax(axis = var_3352, x = aw_497_cast_fp16)[name = tensor("op_3516_cast_fp16")]; + tensor var_3517_cast_fp16 = softmax(axis = var_3352, x = aw_499_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3518_cast_fp16 = softmax(axis = var_3352, x = aw_501_cast_fp16)[name = tensor("op_3518_cast_fp16")]; + tensor var_3519_cast_fp16 = softmax(axis = var_3352, x = aw_503_cast_fp16)[name = tensor("op_3519_cast_fp16")]; + tensor var_3520_cast_fp16 = softmax(axis = var_3352, x = aw_505_cast_fp16)[name = tensor("op_3520_cast_fp16")]; + tensor var_3521_cast_fp16 = softmax(axis = var_3352, x = aw_507_cast_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3522_cast_fp16 = softmax(axis = var_3352, x = aw_509_cast_fp16)[name = tensor("op_3522_cast_fp16")]; + tensor var_3523_cast_fp16 = softmax(axis = var_3352, x = aw_511_cast_fp16)[name = tensor("op_3523_cast_fp16")]; + tensor var_3524_cast_fp16 = softmax(axis = var_3352, x = aw_513_cast_fp16)[name = tensor("op_3524_cast_fp16")]; + tensor var_3525_cast_fp16 = softmax(axis = var_3352, x = aw_515_cast_fp16)[name = tensor("op_3525_cast_fp16")]; + tensor var_3526_cast_fp16 = softmax(axis = var_3352, x = aw_517_cast_fp16)[name = tensor("op_3526_cast_fp16")]; + tensor var_3527_cast_fp16 = softmax(axis = var_3352, x = aw_519_cast_fp16)[name = tensor("op_3527_cast_fp16")]; + tensor var_3529_equation_0 = const()[name = tensor("op_3529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3529_cast_fp16 = einsum(equation = var_3529_equation_0, values = (var_3447_cast_fp16_0, var_3508_cast_fp16))[name = tensor("op_3529_cast_fp16")]; + tensor var_3531_equation_0 = const()[name = tensor("op_3531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3531_cast_fp16 = einsum(equation = var_3531_equation_0, values = (var_3447_cast_fp16_1, var_3509_cast_fp16))[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_equation_0 = const()[name = tensor("op_3533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3533_cast_fp16 = einsum(equation = var_3533_equation_0, values = (var_3447_cast_fp16_2, var_3510_cast_fp16))[name = tensor("op_3533_cast_fp16")]; + tensor var_3535_equation_0 = const()[name = tensor("op_3535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3535_cast_fp16 = einsum(equation = var_3535_equation_0, values = (var_3447_cast_fp16_3, var_3511_cast_fp16))[name = tensor("op_3535_cast_fp16")]; + tensor var_3537_equation_0 = const()[name = tensor("op_3537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3537_cast_fp16 = einsum(equation = var_3537_equation_0, values = (var_3447_cast_fp16_4, var_3512_cast_fp16))[name = tensor("op_3537_cast_fp16")]; + tensor var_3539_equation_0 = const()[name = tensor("op_3539_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3539_cast_fp16 = einsum(equation = var_3539_equation_0, values = (var_3447_cast_fp16_5, var_3513_cast_fp16))[name = tensor("op_3539_cast_fp16")]; + tensor var_3541_equation_0 = const()[name = tensor("op_3541_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3541_cast_fp16 = einsum(equation = var_3541_equation_0, values = (var_3447_cast_fp16_6, var_3514_cast_fp16))[name = tensor("op_3541_cast_fp16")]; + tensor var_3543_equation_0 = const()[name = tensor("op_3543_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3543_cast_fp16 = einsum(equation = var_3543_equation_0, values = (var_3447_cast_fp16_7, var_3515_cast_fp16))[name = tensor("op_3543_cast_fp16")]; + tensor var_3545_equation_0 = const()[name = tensor("op_3545_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3545_cast_fp16 = einsum(equation = var_3545_equation_0, values = (var_3447_cast_fp16_8, var_3516_cast_fp16))[name = tensor("op_3545_cast_fp16")]; + tensor var_3547_equation_0 = const()[name = tensor("op_3547_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3547_cast_fp16 = einsum(equation = var_3547_equation_0, values = (var_3447_cast_fp16_9, var_3517_cast_fp16))[name = tensor("op_3547_cast_fp16")]; + tensor var_3549_equation_0 = const()[name = tensor("op_3549_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3549_cast_fp16 = einsum(equation = var_3549_equation_0, values = (var_3447_cast_fp16_10, var_3518_cast_fp16))[name = tensor("op_3549_cast_fp16")]; + tensor var_3551_equation_0 = const()[name = tensor("op_3551_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3551_cast_fp16 = einsum(equation = var_3551_equation_0, values = (var_3447_cast_fp16_11, var_3519_cast_fp16))[name = tensor("op_3551_cast_fp16")]; + tensor var_3553_equation_0 = const()[name = tensor("op_3553_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3553_cast_fp16 = einsum(equation = var_3553_equation_0, values = (var_3447_cast_fp16_12, var_3520_cast_fp16))[name = tensor("op_3553_cast_fp16")]; + tensor var_3555_equation_0 = const()[name = tensor("op_3555_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3555_cast_fp16 = einsum(equation = var_3555_equation_0, values = (var_3447_cast_fp16_13, var_3521_cast_fp16))[name = tensor("op_3555_cast_fp16")]; + tensor var_3557_equation_0 = const()[name = tensor("op_3557_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3557_cast_fp16 = einsum(equation = var_3557_equation_0, values = (var_3447_cast_fp16_14, var_3522_cast_fp16))[name = tensor("op_3557_cast_fp16")]; + tensor var_3559_equation_0 = const()[name = tensor("op_3559_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3559_cast_fp16 = einsum(equation = var_3559_equation_0, values = (var_3447_cast_fp16_15, var_3523_cast_fp16))[name = tensor("op_3559_cast_fp16")]; + tensor var_3561_equation_0 = const()[name = tensor("op_3561_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3561_cast_fp16 = einsum(equation = var_3561_equation_0, values = (var_3447_cast_fp16_16, var_3524_cast_fp16))[name = tensor("op_3561_cast_fp16")]; + tensor var_3563_equation_0 = const()[name = tensor("op_3563_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3563_cast_fp16 = einsum(equation = var_3563_equation_0, values = (var_3447_cast_fp16_17, var_3525_cast_fp16))[name = tensor("op_3563_cast_fp16")]; + tensor var_3565_equation_0 = const()[name = tensor("op_3565_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3565_cast_fp16 = einsum(equation = var_3565_equation_0, values = (var_3447_cast_fp16_18, var_3526_cast_fp16))[name = tensor("op_3565_cast_fp16")]; + tensor var_3567_equation_0 = const()[name = tensor("op_3567_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3567_cast_fp16 = einsum(equation = var_3567_equation_0, values = (var_3447_cast_fp16_19, var_3527_cast_fp16))[name = tensor("op_3567_cast_fp16")]; + tensor input_125_interleave_0 = const()[name = tensor("input_125_interleave_0"), val = tensor(false)]; + tensor input_125_cast_fp16 = concat(axis = var_3352, interleave = input_125_interleave_0, values = (var_3529_cast_fp16, var_3531_cast_fp16, var_3533_cast_fp16, var_3535_cast_fp16, var_3537_cast_fp16, var_3539_cast_fp16, var_3541_cast_fp16, var_3543_cast_fp16, var_3545_cast_fp16, var_3547_cast_fp16, var_3549_cast_fp16, var_3551_cast_fp16, var_3553_cast_fp16, var_3555_cast_fp16, var_3557_cast_fp16, var_3559_cast_fp16, var_3561_cast_fp16, var_3563_cast_fp16, var_3565_cast_fp16, var_3567_cast_fp16))[name = tensor("input_125_cast_fp16")]; + tensor var_3576_pad_type_0 = const()[name = tensor("op_3576_pad_type_0"), val = tensor("valid")]; + tensor var_3576_strides_0 = const()[name = tensor("op_3576_strides_0"), val = tensor([1, 1])]; + tensor var_3576_pad_0 = const()[name = tensor("op_3576_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3576_dilations_0 = const()[name = tensor("op_3576_dilations_0"), val = tensor([1, 1])]; + tensor var_3576_groups_0 = const()[name = tensor("op_3576_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_out_weight_to_fp16 = const()[name = tensor("blocks_12_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496739392)))]; + tensor blocks_12_attn_out_bias_to_fp16 = const()[name = tensor("blocks_12_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500016256)))]; + tensor var_3576_cast_fp16 = conv(bias = blocks_12_attn_out_bias_to_fp16, dilations = var_3576_dilations_0, groups = var_3576_groups_0, pad = var_3576_pad_0, pad_type = var_3576_pad_type_0, strides = var_3576_strides_0, weight = blocks_12_attn_out_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = var_3576_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([1])]; + tensor input_127_gamma_0_to_fp16 = const()[name = tensor("input_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500018880)))]; + tensor input_127_beta_0_to_fp16 = const()[name = tensor("input_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500021504)))]; + tensor var_3586_to_fp16 = const()[name = tensor("op_3586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = input_127_beta_0_to_fp16, epsilon = var_3586_to_fp16, gamma = input_127_gamma_0_to_fp16, x = inputs_51_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("valid")]; + tensor input_129_strides_0 = const()[name = tensor("input_129_strides_0"), val = tensor([1, 1])]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_129_dilations_0 = const()[name = tensor("input_129_dilations_0"), val = tensor([1, 1])]; + tensor input_129_groups_0 = const()[name = tensor("input_129_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500024128)))]; + tensor blocks_12_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513131392)))]; + tensor input_129_cast_fp16 = conv(bias = blocks_12_mlp_0_bias_to_fp16, dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = blocks_12_mlp_0_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3612_pad_type_0 = const()[name = tensor("op_3612_pad_type_0"), val = tensor("valid")]; + tensor var_3612_strides_0 = const()[name = tensor("op_3612_strides_0"), val = tensor([1, 1])]; + tensor var_3612_pad_0 = const()[name = tensor("op_3612_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3612_dilations_0 = const()[name = tensor("op_3612_dilations_0"), val = tensor([1, 1])]; + tensor var_3612_groups_0 = const()[name = tensor("op_3612_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513141696)))]; + tensor blocks_12_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526248960)))]; + tensor var_3612_cast_fp16 = conv(bias = blocks_12_mlp_2_bias_to_fp16, dilations = var_3612_dilations_0, groups = var_3612_groups_0, pad = var_3612_pad_0, pad_type = var_3612_pad_type_0, strides = var_3612_strides_0, weight = blocks_12_mlp_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_3612_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3621 = const()[name = tensor("op_3621"), val = tensor(1)]; + tensor input_133_axes_0 = const()[name = tensor("input_133_axes_0"), val = tensor([1])]; + tensor input_133_gamma_0_to_fp16 = const()[name = tensor("input_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526251584)))]; + tensor input_133_beta_0_to_fp16 = const()[name = tensor("input_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526254208)))]; + tensor var_3637_to_fp16 = const()[name = tensor("op_3637_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_133_cast_fp16 = layer_norm(axes = input_133_axes_0, beta = input_133_beta_0_to_fp16, epsilon = var_3637_to_fp16, gamma = input_133_gamma_0_to_fp16, x = inputs_53_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor q_27_pad_type_0 = const()[name = tensor("q_27_pad_type_0"), val = tensor("valid")]; + tensor q_27_strides_0 = const()[name = tensor("q_27_strides_0"), val = tensor([1, 1])]; + tensor q_27_pad_0 = const()[name = tensor("q_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_27_dilations_0 = const()[name = tensor("q_27_dilations_0"), val = tensor([1, 1])]; + tensor q_27_groups_0 = const()[name = tensor("q_27_groups_0"), val = tensor(1)]; + tensor var_3672_weight_0_to_fp16 = const()[name = tensor("op_3672_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526256832)))]; + tensor var_3672_bias_0_to_fp16 = const()[name = tensor("op_3672_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529533696)))]; + tensor var_3672_cast_fp16 = conv(bias = var_3672_bias_0_to_fp16, dilations = q_27_dilations_0, groups = q_27_groups_0, pad = q_27_pad_0, pad_type = q_27_pad_type_0, strides = q_27_strides_0, weight = var_3672_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3672_cast_fp16")]; + tensor k_27_pad_type_0 = const()[name = tensor("k_27_pad_type_0"), val = tensor("valid")]; + tensor k_27_strides_0 = const()[name = tensor("k_27_strides_0"), val = tensor([1, 1])]; + tensor k_27_pad_0 = const()[name = tensor("k_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_27_dilations_0 = const()[name = tensor("k_27_dilations_0"), val = tensor([1, 1])]; + tensor k_27_groups_0 = const()[name = tensor("k_27_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_key_weight_to_fp16 = const()[name = tensor("blocks_13_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529536320)))]; + tensor k_27_cast_fp16 = conv(dilations = k_27_dilations_0, groups = k_27_groups_0, pad = k_27_pad_0, pad_type = k_27_pad_type_0, strides = k_27_strides_0, weight = blocks_13_attn_key_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor var_3670_pad_type_0 = const()[name = tensor("op_3670_pad_type_0"), val = tensor("valid")]; + tensor var_3670_strides_0 = const()[name = tensor("op_3670_strides_0"), val = tensor([1, 1])]; + tensor var_3670_pad_0 = const()[name = tensor("op_3670_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3670_dilations_0 = const()[name = tensor("op_3670_dilations_0"), val = tensor([1, 1])]; + tensor var_3670_groups_0 = const()[name = tensor("op_3670_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_value_weight_to_fp16 = const()[name = tensor("blocks_13_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532813184)))]; + tensor blocks_13_attn_value_bias_to_fp16 = const()[name = tensor("blocks_13_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536090048)))]; + tensor var_3670_cast_fp16 = conv(bias = blocks_13_attn_value_bias_to_fp16, dilations = var_3670_dilations_0, groups = var_3670_groups_0, pad = var_3670_pad_0, pad_type = var_3670_pad_type_0, strides = var_3670_strides_0, weight = blocks_13_attn_value_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3670_cast_fp16")]; + tensor tile_39 = const()[name = tensor("tile_39"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3673_axis_0 = const()[name = tensor("op_3673_axis_0"), val = tensor(1)]; + tensor var_3673_cast_fp16_0, tensor var_3673_cast_fp16_1, tensor var_3673_cast_fp16_2, tensor var_3673_cast_fp16_3, tensor var_3673_cast_fp16_4, tensor var_3673_cast_fp16_5, tensor var_3673_cast_fp16_6, tensor var_3673_cast_fp16_7, tensor var_3673_cast_fp16_8, tensor var_3673_cast_fp16_9, tensor var_3673_cast_fp16_10, tensor var_3673_cast_fp16_11, tensor var_3673_cast_fp16_12, tensor var_3673_cast_fp16_13, tensor var_3673_cast_fp16_14, tensor var_3673_cast_fp16_15, tensor var_3673_cast_fp16_16, tensor var_3673_cast_fp16_17, tensor var_3673_cast_fp16_18, tensor var_3673_cast_fp16_19 = split(axis = var_3673_axis_0, split_sizes = tile_39, x = var_3672_cast_fp16)[name = tensor("op_3673_cast_fp16")]; + tensor var_3694_perm_0 = const()[name = tensor("op_3694_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_40 = const()[name = tensor("tile_40"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3695_axis_0 = const()[name = tensor("op_3695_axis_0"), val = tensor(3)]; + tensor var_3694_cast_fp16 = transpose(perm = var_3694_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_19")]; + tensor var_3695_cast_fp16_0, tensor var_3695_cast_fp16_1, tensor var_3695_cast_fp16_2, tensor var_3695_cast_fp16_3, tensor var_3695_cast_fp16_4, tensor var_3695_cast_fp16_5, tensor var_3695_cast_fp16_6, tensor var_3695_cast_fp16_7, tensor var_3695_cast_fp16_8, tensor var_3695_cast_fp16_9, tensor var_3695_cast_fp16_10, tensor var_3695_cast_fp16_11, tensor var_3695_cast_fp16_12, tensor var_3695_cast_fp16_13, tensor var_3695_cast_fp16_14, tensor var_3695_cast_fp16_15, tensor var_3695_cast_fp16_16, tensor var_3695_cast_fp16_17, tensor var_3695_cast_fp16_18, tensor var_3695_cast_fp16_19 = split(axis = var_3695_axis_0, split_sizes = tile_40, x = var_3694_cast_fp16)[name = tensor("op_3695_cast_fp16")]; + tensor tile_41 = const()[name = tensor("tile_41"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3716_axis_0 = const()[name = tensor("op_3716_axis_0"), val = tensor(1)]; + tensor var_3716_cast_fp16_0, tensor var_3716_cast_fp16_1, tensor var_3716_cast_fp16_2, tensor var_3716_cast_fp16_3, tensor var_3716_cast_fp16_4, tensor var_3716_cast_fp16_5, tensor var_3716_cast_fp16_6, tensor var_3716_cast_fp16_7, tensor var_3716_cast_fp16_8, tensor var_3716_cast_fp16_9, tensor var_3716_cast_fp16_10, tensor var_3716_cast_fp16_11, tensor var_3716_cast_fp16_12, tensor var_3716_cast_fp16_13, tensor var_3716_cast_fp16_14, tensor var_3716_cast_fp16_15, tensor var_3716_cast_fp16_16, tensor var_3716_cast_fp16_17, tensor var_3716_cast_fp16_18, tensor var_3716_cast_fp16_19 = split(axis = var_3716_axis_0, split_sizes = tile_41, x = var_3670_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor aw_521_equation_0 = const()[name = tensor("aw_521_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_521_cast_fp16 = einsum(equation = aw_521_equation_0, values = (var_3695_cast_fp16_0, var_3673_cast_fp16_0))[name = tensor("aw_521_cast_fp16")]; + tensor aw_523_equation_0 = const()[name = tensor("aw_523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_523_cast_fp16 = einsum(equation = aw_523_equation_0, values = (var_3695_cast_fp16_1, var_3673_cast_fp16_1))[name = tensor("aw_523_cast_fp16")]; + tensor aw_525_equation_0 = const()[name = tensor("aw_525_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_525_cast_fp16 = einsum(equation = aw_525_equation_0, values = (var_3695_cast_fp16_2, var_3673_cast_fp16_2))[name = tensor("aw_525_cast_fp16")]; + tensor aw_527_equation_0 = const()[name = tensor("aw_527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_527_cast_fp16 = einsum(equation = aw_527_equation_0, values = (var_3695_cast_fp16_3, var_3673_cast_fp16_3))[name = tensor("aw_527_cast_fp16")]; + tensor aw_529_equation_0 = const()[name = tensor("aw_529_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_529_cast_fp16 = einsum(equation = aw_529_equation_0, values = (var_3695_cast_fp16_4, var_3673_cast_fp16_4))[name = tensor("aw_529_cast_fp16")]; + tensor aw_531_equation_0 = const()[name = tensor("aw_531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_531_cast_fp16 = einsum(equation = aw_531_equation_0, values = (var_3695_cast_fp16_5, var_3673_cast_fp16_5))[name = tensor("aw_531_cast_fp16")]; + tensor aw_533_equation_0 = const()[name = tensor("aw_533_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_533_cast_fp16 = einsum(equation = aw_533_equation_0, values = (var_3695_cast_fp16_6, var_3673_cast_fp16_6))[name = tensor("aw_533_cast_fp16")]; + tensor aw_535_equation_0 = const()[name = tensor("aw_535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_535_cast_fp16 = einsum(equation = aw_535_equation_0, values = (var_3695_cast_fp16_7, var_3673_cast_fp16_7))[name = tensor("aw_535_cast_fp16")]; + tensor aw_537_equation_0 = const()[name = tensor("aw_537_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_537_cast_fp16 = einsum(equation = aw_537_equation_0, values = (var_3695_cast_fp16_8, var_3673_cast_fp16_8))[name = tensor("aw_537_cast_fp16")]; + tensor aw_539_equation_0 = const()[name = tensor("aw_539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_539_cast_fp16 = einsum(equation = aw_539_equation_0, values = (var_3695_cast_fp16_9, var_3673_cast_fp16_9))[name = tensor("aw_539_cast_fp16")]; + tensor aw_541_equation_0 = const()[name = tensor("aw_541_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_541_cast_fp16 = einsum(equation = aw_541_equation_0, values = (var_3695_cast_fp16_10, var_3673_cast_fp16_10))[name = tensor("aw_541_cast_fp16")]; + tensor aw_543_equation_0 = const()[name = tensor("aw_543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_543_cast_fp16 = einsum(equation = aw_543_equation_0, values = (var_3695_cast_fp16_11, var_3673_cast_fp16_11))[name = tensor("aw_543_cast_fp16")]; + tensor aw_545_equation_0 = const()[name = tensor("aw_545_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_545_cast_fp16 = einsum(equation = aw_545_equation_0, values = (var_3695_cast_fp16_12, var_3673_cast_fp16_12))[name = tensor("aw_545_cast_fp16")]; + tensor aw_547_equation_0 = const()[name = tensor("aw_547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_547_cast_fp16 = einsum(equation = aw_547_equation_0, values = (var_3695_cast_fp16_13, var_3673_cast_fp16_13))[name = tensor("aw_547_cast_fp16")]; + tensor aw_549_equation_0 = const()[name = tensor("aw_549_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_549_cast_fp16 = einsum(equation = aw_549_equation_0, values = (var_3695_cast_fp16_14, var_3673_cast_fp16_14))[name = tensor("aw_549_cast_fp16")]; + tensor aw_551_equation_0 = const()[name = tensor("aw_551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_551_cast_fp16 = einsum(equation = aw_551_equation_0, values = (var_3695_cast_fp16_15, var_3673_cast_fp16_15))[name = tensor("aw_551_cast_fp16")]; + tensor aw_553_equation_0 = const()[name = tensor("aw_553_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_553_cast_fp16 = einsum(equation = aw_553_equation_0, values = (var_3695_cast_fp16_16, var_3673_cast_fp16_16))[name = tensor("aw_553_cast_fp16")]; + tensor aw_555_equation_0 = const()[name = tensor("aw_555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_555_cast_fp16 = einsum(equation = aw_555_equation_0, values = (var_3695_cast_fp16_17, var_3673_cast_fp16_17))[name = tensor("aw_555_cast_fp16")]; + tensor aw_557_equation_0 = const()[name = tensor("aw_557_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_557_cast_fp16 = einsum(equation = aw_557_equation_0, values = (var_3695_cast_fp16_18, var_3673_cast_fp16_18))[name = tensor("aw_557_cast_fp16")]; + tensor aw_559_equation_0 = const()[name = tensor("aw_559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_559_cast_fp16 = einsum(equation = aw_559_equation_0, values = (var_3695_cast_fp16_19, var_3673_cast_fp16_19))[name = tensor("aw_559_cast_fp16")]; + tensor var_3777_cast_fp16 = softmax(axis = var_3621, x = aw_521_cast_fp16)[name = tensor("op_3777_cast_fp16")]; + tensor var_3778_cast_fp16 = softmax(axis = var_3621, x = aw_523_cast_fp16)[name = tensor("op_3778_cast_fp16")]; + tensor var_3779_cast_fp16 = softmax(axis = var_3621, x = aw_525_cast_fp16)[name = tensor("op_3779_cast_fp16")]; + tensor var_3780_cast_fp16 = softmax(axis = var_3621, x = aw_527_cast_fp16)[name = tensor("op_3780_cast_fp16")]; + tensor var_3781_cast_fp16 = softmax(axis = var_3621, x = aw_529_cast_fp16)[name = tensor("op_3781_cast_fp16")]; + tensor var_3782_cast_fp16 = softmax(axis = var_3621, x = aw_531_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = softmax(axis = var_3621, x = aw_533_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor var_3784_cast_fp16 = softmax(axis = var_3621, x = aw_535_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3785_cast_fp16 = softmax(axis = var_3621, x = aw_537_cast_fp16)[name = tensor("op_3785_cast_fp16")]; + tensor var_3786_cast_fp16 = softmax(axis = var_3621, x = aw_539_cast_fp16)[name = tensor("op_3786_cast_fp16")]; + tensor var_3787_cast_fp16 = softmax(axis = var_3621, x = aw_541_cast_fp16)[name = tensor("op_3787_cast_fp16")]; + tensor var_3788_cast_fp16 = softmax(axis = var_3621, x = aw_543_cast_fp16)[name = tensor("op_3788_cast_fp16")]; + tensor var_3789_cast_fp16 = softmax(axis = var_3621, x = aw_545_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3790_cast_fp16 = softmax(axis = var_3621, x = aw_547_cast_fp16)[name = tensor("op_3790_cast_fp16")]; + tensor var_3791_cast_fp16 = softmax(axis = var_3621, x = aw_549_cast_fp16)[name = tensor("op_3791_cast_fp16")]; + tensor var_3792_cast_fp16 = softmax(axis = var_3621, x = aw_551_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3793_cast_fp16 = softmax(axis = var_3621, x = aw_553_cast_fp16)[name = tensor("op_3793_cast_fp16")]; + tensor var_3794_cast_fp16 = softmax(axis = var_3621, x = aw_555_cast_fp16)[name = tensor("op_3794_cast_fp16")]; + tensor var_3795_cast_fp16 = softmax(axis = var_3621, x = aw_557_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = softmax(axis = var_3621, x = aw_559_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor var_3798_equation_0 = const()[name = tensor("op_3798_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3798_cast_fp16 = einsum(equation = var_3798_equation_0, values = (var_3716_cast_fp16_0, var_3777_cast_fp16))[name = tensor("op_3798_cast_fp16")]; + tensor var_3800_equation_0 = const()[name = tensor("op_3800_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3800_cast_fp16 = einsum(equation = var_3800_equation_0, values = (var_3716_cast_fp16_1, var_3778_cast_fp16))[name = tensor("op_3800_cast_fp16")]; + tensor var_3802_equation_0 = const()[name = tensor("op_3802_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3802_cast_fp16 = einsum(equation = var_3802_equation_0, values = (var_3716_cast_fp16_2, var_3779_cast_fp16))[name = tensor("op_3802_cast_fp16")]; + tensor var_3804_equation_0 = const()[name = tensor("op_3804_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3804_cast_fp16 = einsum(equation = var_3804_equation_0, values = (var_3716_cast_fp16_3, var_3780_cast_fp16))[name = tensor("op_3804_cast_fp16")]; + tensor var_3806_equation_0 = const()[name = tensor("op_3806_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3806_cast_fp16 = einsum(equation = var_3806_equation_0, values = (var_3716_cast_fp16_4, var_3781_cast_fp16))[name = tensor("op_3806_cast_fp16")]; + tensor var_3808_equation_0 = const()[name = tensor("op_3808_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3808_cast_fp16 = einsum(equation = var_3808_equation_0, values = (var_3716_cast_fp16_5, var_3782_cast_fp16))[name = tensor("op_3808_cast_fp16")]; + tensor var_3810_equation_0 = const()[name = tensor("op_3810_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3810_cast_fp16 = einsum(equation = var_3810_equation_0, values = (var_3716_cast_fp16_6, var_3783_cast_fp16))[name = tensor("op_3810_cast_fp16")]; + tensor var_3812_equation_0 = const()[name = tensor("op_3812_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3812_cast_fp16 = einsum(equation = var_3812_equation_0, values = (var_3716_cast_fp16_7, var_3784_cast_fp16))[name = tensor("op_3812_cast_fp16")]; + tensor var_3814_equation_0 = const()[name = tensor("op_3814_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3814_cast_fp16 = einsum(equation = var_3814_equation_0, values = (var_3716_cast_fp16_8, var_3785_cast_fp16))[name = tensor("op_3814_cast_fp16")]; + tensor var_3816_equation_0 = const()[name = tensor("op_3816_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3816_cast_fp16 = einsum(equation = var_3816_equation_0, values = (var_3716_cast_fp16_9, var_3786_cast_fp16))[name = tensor("op_3816_cast_fp16")]; + tensor var_3818_equation_0 = const()[name = tensor("op_3818_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3818_cast_fp16 = einsum(equation = var_3818_equation_0, values = (var_3716_cast_fp16_10, var_3787_cast_fp16))[name = tensor("op_3818_cast_fp16")]; + tensor var_3820_equation_0 = const()[name = tensor("op_3820_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3820_cast_fp16 = einsum(equation = var_3820_equation_0, values = (var_3716_cast_fp16_11, var_3788_cast_fp16))[name = tensor("op_3820_cast_fp16")]; + tensor var_3822_equation_0 = const()[name = tensor("op_3822_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3822_cast_fp16 = einsum(equation = var_3822_equation_0, values = (var_3716_cast_fp16_12, var_3789_cast_fp16))[name = tensor("op_3822_cast_fp16")]; + tensor var_3824_equation_0 = const()[name = tensor("op_3824_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3824_cast_fp16 = einsum(equation = var_3824_equation_0, values = (var_3716_cast_fp16_13, var_3790_cast_fp16))[name = tensor("op_3824_cast_fp16")]; + tensor var_3826_equation_0 = const()[name = tensor("op_3826_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3826_cast_fp16 = einsum(equation = var_3826_equation_0, values = (var_3716_cast_fp16_14, var_3791_cast_fp16))[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_equation_0 = const()[name = tensor("op_3828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3828_cast_fp16 = einsum(equation = var_3828_equation_0, values = (var_3716_cast_fp16_15, var_3792_cast_fp16))[name = tensor("op_3828_cast_fp16")]; + tensor var_3830_equation_0 = const()[name = tensor("op_3830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3830_cast_fp16 = einsum(equation = var_3830_equation_0, values = (var_3716_cast_fp16_16, var_3793_cast_fp16))[name = tensor("op_3830_cast_fp16")]; + tensor var_3832_equation_0 = const()[name = tensor("op_3832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3832_cast_fp16 = einsum(equation = var_3832_equation_0, values = (var_3716_cast_fp16_17, var_3794_cast_fp16))[name = tensor("op_3832_cast_fp16")]; + tensor var_3834_equation_0 = const()[name = tensor("op_3834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3834_cast_fp16 = einsum(equation = var_3834_equation_0, values = (var_3716_cast_fp16_18, var_3795_cast_fp16))[name = tensor("op_3834_cast_fp16")]; + tensor var_3836_equation_0 = const()[name = tensor("op_3836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3836_cast_fp16 = einsum(equation = var_3836_equation_0, values = (var_3716_cast_fp16_19, var_3796_cast_fp16))[name = tensor("op_3836_cast_fp16")]; + tensor input_135_interleave_0 = const()[name = tensor("input_135_interleave_0"), val = tensor(false)]; + tensor input_135_cast_fp16 = concat(axis = var_3621, interleave = input_135_interleave_0, values = (var_3798_cast_fp16, var_3800_cast_fp16, var_3802_cast_fp16, var_3804_cast_fp16, var_3806_cast_fp16, var_3808_cast_fp16, var_3810_cast_fp16, var_3812_cast_fp16, var_3814_cast_fp16, var_3816_cast_fp16, var_3818_cast_fp16, var_3820_cast_fp16, var_3822_cast_fp16, var_3824_cast_fp16, var_3826_cast_fp16, var_3828_cast_fp16, var_3830_cast_fp16, var_3832_cast_fp16, var_3834_cast_fp16, var_3836_cast_fp16))[name = tensor("input_135_cast_fp16")]; + tensor var_3845_pad_type_0 = const()[name = tensor("op_3845_pad_type_0"), val = tensor("valid")]; + tensor var_3845_strides_0 = const()[name = tensor("op_3845_strides_0"), val = tensor([1, 1])]; + tensor var_3845_pad_0 = const()[name = tensor("op_3845_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3845_dilations_0 = const()[name = tensor("op_3845_dilations_0"), val = tensor([1, 1])]; + tensor var_3845_groups_0 = const()[name = tensor("op_3845_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_out_weight_to_fp16 = const()[name = tensor("blocks_13_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536092672)))]; + tensor blocks_13_attn_out_bias_to_fp16 = const()[name = tensor("blocks_13_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539369536)))]; + tensor var_3845_cast_fp16 = conv(bias = blocks_13_attn_out_bias_to_fp16, dilations = var_3845_dilations_0, groups = var_3845_groups_0, pad = var_3845_pad_0, pad_type = var_3845_pad_type_0, strides = var_3845_strides_0, weight = blocks_13_attn_out_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("op_3845_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = var_3845_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([1])]; + tensor input_137_gamma_0_to_fp16 = const()[name = tensor("input_137_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539372160)))]; + tensor input_137_beta_0_to_fp16 = const()[name = tensor("input_137_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539374784)))]; + tensor var_3855_to_fp16 = const()[name = tensor("op_3855_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = input_137_beta_0_to_fp16, epsilon = var_3855_to_fp16, gamma = input_137_gamma_0_to_fp16, x = inputs_55_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("valid")]; + tensor input_139_strides_0 = const()[name = tensor("input_139_strides_0"), val = tensor([1, 1])]; + tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_139_dilations_0 = const()[name = tensor("input_139_dilations_0"), val = tensor([1, 1])]; + tensor input_139_groups_0 = const()[name = tensor("input_139_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539377408)))]; + tensor blocks_13_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552484672)))]; + tensor input_139_cast_fp16 = conv(bias = blocks_13_mlp_0_bias_to_fp16, dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = blocks_13_mlp_0_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3881_pad_type_0 = const()[name = tensor("op_3881_pad_type_0"), val = tensor("valid")]; + tensor var_3881_strides_0 = const()[name = tensor("op_3881_strides_0"), val = tensor([1, 1])]; + tensor var_3881_pad_0 = const()[name = tensor("op_3881_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3881_dilations_0 = const()[name = tensor("op_3881_dilations_0"), val = tensor([1, 1])]; + tensor var_3881_groups_0 = const()[name = tensor("op_3881_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552494976)))]; + tensor blocks_13_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565602240)))]; + tensor var_3881_cast_fp16 = conv(bias = blocks_13_mlp_2_bias_to_fp16, dilations = var_3881_dilations_0, groups = var_3881_groups_0, pad = var_3881_pad_0, pad_type = var_3881_pad_type_0, strides = var_3881_strides_0, weight = blocks_13_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = var_3881_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3890 = const()[name = tensor("op_3890"), val = tensor(1)]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([1])]; + tensor input_143_gamma_0_to_fp16 = const()[name = tensor("input_143_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565604864)))]; + tensor input_143_beta_0_to_fp16 = const()[name = tensor("input_143_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565607488)))]; + tensor var_3906_to_fp16 = const()[name = tensor("op_3906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = input_143_beta_0_to_fp16, epsilon = var_3906_to_fp16, gamma = input_143_gamma_0_to_fp16, x = inputs_57_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor q_29_pad_type_0 = const()[name = tensor("q_29_pad_type_0"), val = tensor("valid")]; + tensor q_29_strides_0 = const()[name = tensor("q_29_strides_0"), val = tensor([1, 1])]; + tensor q_29_pad_0 = const()[name = tensor("q_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_29_dilations_0 = const()[name = tensor("q_29_dilations_0"), val = tensor([1, 1])]; + tensor q_29_groups_0 = const()[name = tensor("q_29_groups_0"), val = tensor(1)]; + tensor var_3941_weight_0_to_fp16 = const()[name = tensor("op_3941_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565610112)))]; + tensor var_3941_bias_0_to_fp16 = const()[name = tensor("op_3941_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568886976)))]; + tensor var_3941_cast_fp16 = conv(bias = var_3941_bias_0_to_fp16, dilations = q_29_dilations_0, groups = q_29_groups_0, pad = q_29_pad_0, pad_type = q_29_pad_type_0, strides = q_29_strides_0, weight = var_3941_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3941_cast_fp16")]; + tensor k_29_pad_type_0 = const()[name = tensor("k_29_pad_type_0"), val = tensor("valid")]; + tensor k_29_strides_0 = const()[name = tensor("k_29_strides_0"), val = tensor([1, 1])]; + tensor k_29_pad_0 = const()[name = tensor("k_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_29_dilations_0 = const()[name = tensor("k_29_dilations_0"), val = tensor([1, 1])]; + tensor k_29_groups_0 = const()[name = tensor("k_29_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_key_weight_to_fp16 = const()[name = tensor("blocks_14_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568889600)))]; + tensor k_29_cast_fp16 = conv(dilations = k_29_dilations_0, groups = k_29_groups_0, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = k_29_strides_0, weight = blocks_14_attn_key_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor var_3939_pad_type_0 = const()[name = tensor("op_3939_pad_type_0"), val = tensor("valid")]; + tensor var_3939_strides_0 = const()[name = tensor("op_3939_strides_0"), val = tensor([1, 1])]; + tensor var_3939_pad_0 = const()[name = tensor("op_3939_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3939_dilations_0 = const()[name = tensor("op_3939_dilations_0"), val = tensor([1, 1])]; + tensor var_3939_groups_0 = const()[name = tensor("op_3939_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_value_weight_to_fp16 = const()[name = tensor("blocks_14_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572166464)))]; + tensor blocks_14_attn_value_bias_to_fp16 = const()[name = tensor("blocks_14_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575443328)))]; + tensor var_3939_cast_fp16 = conv(bias = blocks_14_attn_value_bias_to_fp16, dilations = var_3939_dilations_0, groups = var_3939_groups_0, pad = var_3939_pad_0, pad_type = var_3939_pad_type_0, strides = var_3939_strides_0, weight = blocks_14_attn_value_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor tile_42 = const()[name = tensor("tile_42"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3942_axis_0 = const()[name = tensor("op_3942_axis_0"), val = tensor(1)]; + tensor var_3942_cast_fp16_0, tensor var_3942_cast_fp16_1, tensor var_3942_cast_fp16_2, tensor var_3942_cast_fp16_3, tensor var_3942_cast_fp16_4, tensor var_3942_cast_fp16_5, tensor var_3942_cast_fp16_6, tensor var_3942_cast_fp16_7, tensor var_3942_cast_fp16_8, tensor var_3942_cast_fp16_9, tensor var_3942_cast_fp16_10, tensor var_3942_cast_fp16_11, tensor var_3942_cast_fp16_12, tensor var_3942_cast_fp16_13, tensor var_3942_cast_fp16_14, tensor var_3942_cast_fp16_15, tensor var_3942_cast_fp16_16, tensor var_3942_cast_fp16_17, tensor var_3942_cast_fp16_18, tensor var_3942_cast_fp16_19 = split(axis = var_3942_axis_0, split_sizes = tile_42, x = var_3941_cast_fp16)[name = tensor("op_3942_cast_fp16")]; + tensor var_3963_perm_0 = const()[name = tensor("op_3963_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_43 = const()[name = tensor("tile_43"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3964_axis_0 = const()[name = tensor("op_3964_axis_0"), val = tensor(3)]; + tensor var_3963_cast_fp16 = transpose(perm = var_3963_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_18")]; + tensor var_3964_cast_fp16_0, tensor var_3964_cast_fp16_1, tensor var_3964_cast_fp16_2, tensor var_3964_cast_fp16_3, tensor var_3964_cast_fp16_4, tensor var_3964_cast_fp16_5, tensor var_3964_cast_fp16_6, tensor var_3964_cast_fp16_7, tensor var_3964_cast_fp16_8, tensor var_3964_cast_fp16_9, tensor var_3964_cast_fp16_10, tensor var_3964_cast_fp16_11, tensor var_3964_cast_fp16_12, tensor var_3964_cast_fp16_13, tensor var_3964_cast_fp16_14, tensor var_3964_cast_fp16_15, tensor var_3964_cast_fp16_16, tensor var_3964_cast_fp16_17, tensor var_3964_cast_fp16_18, tensor var_3964_cast_fp16_19 = split(axis = var_3964_axis_0, split_sizes = tile_43, x = var_3963_cast_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor tile_44 = const()[name = tensor("tile_44"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3985_axis_0 = const()[name = tensor("op_3985_axis_0"), val = tensor(1)]; + tensor var_3985_cast_fp16_0, tensor var_3985_cast_fp16_1, tensor var_3985_cast_fp16_2, tensor var_3985_cast_fp16_3, tensor var_3985_cast_fp16_4, tensor var_3985_cast_fp16_5, tensor var_3985_cast_fp16_6, tensor var_3985_cast_fp16_7, tensor var_3985_cast_fp16_8, tensor var_3985_cast_fp16_9, tensor var_3985_cast_fp16_10, tensor var_3985_cast_fp16_11, tensor var_3985_cast_fp16_12, tensor var_3985_cast_fp16_13, tensor var_3985_cast_fp16_14, tensor var_3985_cast_fp16_15, tensor var_3985_cast_fp16_16, tensor var_3985_cast_fp16_17, tensor var_3985_cast_fp16_18, tensor var_3985_cast_fp16_19 = split(axis = var_3985_axis_0, split_sizes = tile_44, x = var_3939_cast_fp16)[name = tensor("op_3985_cast_fp16")]; + tensor aw_561_equation_0 = const()[name = tensor("aw_561_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_561_cast_fp16 = einsum(equation = aw_561_equation_0, values = (var_3964_cast_fp16_0, var_3942_cast_fp16_0))[name = tensor("aw_561_cast_fp16")]; + tensor aw_563_equation_0 = const()[name = tensor("aw_563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_563_cast_fp16 = einsum(equation = aw_563_equation_0, values = (var_3964_cast_fp16_1, var_3942_cast_fp16_1))[name = tensor("aw_563_cast_fp16")]; + tensor aw_565_equation_0 = const()[name = tensor("aw_565_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_565_cast_fp16 = einsum(equation = aw_565_equation_0, values = (var_3964_cast_fp16_2, var_3942_cast_fp16_2))[name = tensor("aw_565_cast_fp16")]; + tensor aw_567_equation_0 = const()[name = tensor("aw_567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_567_cast_fp16 = einsum(equation = aw_567_equation_0, values = (var_3964_cast_fp16_3, var_3942_cast_fp16_3))[name = tensor("aw_567_cast_fp16")]; + tensor aw_569_equation_0 = const()[name = tensor("aw_569_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_569_cast_fp16 = einsum(equation = aw_569_equation_0, values = (var_3964_cast_fp16_4, var_3942_cast_fp16_4))[name = tensor("aw_569_cast_fp16")]; + tensor aw_571_equation_0 = const()[name = tensor("aw_571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_571_cast_fp16 = einsum(equation = aw_571_equation_0, values = (var_3964_cast_fp16_5, var_3942_cast_fp16_5))[name = tensor("aw_571_cast_fp16")]; + tensor aw_573_equation_0 = const()[name = tensor("aw_573_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_573_cast_fp16 = einsum(equation = aw_573_equation_0, values = (var_3964_cast_fp16_6, var_3942_cast_fp16_6))[name = tensor("aw_573_cast_fp16")]; + tensor aw_575_equation_0 = const()[name = tensor("aw_575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_575_cast_fp16 = einsum(equation = aw_575_equation_0, values = (var_3964_cast_fp16_7, var_3942_cast_fp16_7))[name = tensor("aw_575_cast_fp16")]; + tensor aw_577_equation_0 = const()[name = tensor("aw_577_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_577_cast_fp16 = einsum(equation = aw_577_equation_0, values = (var_3964_cast_fp16_8, var_3942_cast_fp16_8))[name = tensor("aw_577_cast_fp16")]; + tensor aw_579_equation_0 = const()[name = tensor("aw_579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_579_cast_fp16 = einsum(equation = aw_579_equation_0, values = (var_3964_cast_fp16_9, var_3942_cast_fp16_9))[name = tensor("aw_579_cast_fp16")]; + tensor aw_581_equation_0 = const()[name = tensor("aw_581_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_581_cast_fp16 = einsum(equation = aw_581_equation_0, values = (var_3964_cast_fp16_10, var_3942_cast_fp16_10))[name = tensor("aw_581_cast_fp16")]; + tensor aw_583_equation_0 = const()[name = tensor("aw_583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_583_cast_fp16 = einsum(equation = aw_583_equation_0, values = (var_3964_cast_fp16_11, var_3942_cast_fp16_11))[name = tensor("aw_583_cast_fp16")]; + tensor aw_585_equation_0 = const()[name = tensor("aw_585_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_585_cast_fp16 = einsum(equation = aw_585_equation_0, values = (var_3964_cast_fp16_12, var_3942_cast_fp16_12))[name = tensor("aw_585_cast_fp16")]; + tensor aw_587_equation_0 = const()[name = tensor("aw_587_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_587_cast_fp16 = einsum(equation = aw_587_equation_0, values = (var_3964_cast_fp16_13, var_3942_cast_fp16_13))[name = tensor("aw_587_cast_fp16")]; + tensor aw_589_equation_0 = const()[name = tensor("aw_589_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_589_cast_fp16 = einsum(equation = aw_589_equation_0, values = (var_3964_cast_fp16_14, var_3942_cast_fp16_14))[name = tensor("aw_589_cast_fp16")]; + tensor aw_591_equation_0 = const()[name = tensor("aw_591_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_591_cast_fp16 = einsum(equation = aw_591_equation_0, values = (var_3964_cast_fp16_15, var_3942_cast_fp16_15))[name = tensor("aw_591_cast_fp16")]; + tensor aw_593_equation_0 = const()[name = tensor("aw_593_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_593_cast_fp16 = einsum(equation = aw_593_equation_0, values = (var_3964_cast_fp16_16, var_3942_cast_fp16_16))[name = tensor("aw_593_cast_fp16")]; + tensor aw_595_equation_0 = const()[name = tensor("aw_595_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_595_cast_fp16 = einsum(equation = aw_595_equation_0, values = (var_3964_cast_fp16_17, var_3942_cast_fp16_17))[name = tensor("aw_595_cast_fp16")]; + tensor aw_597_equation_0 = const()[name = tensor("aw_597_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_597_cast_fp16 = einsum(equation = aw_597_equation_0, values = (var_3964_cast_fp16_18, var_3942_cast_fp16_18))[name = tensor("aw_597_cast_fp16")]; + tensor aw_599_equation_0 = const()[name = tensor("aw_599_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_599_cast_fp16 = einsum(equation = aw_599_equation_0, values = (var_3964_cast_fp16_19, var_3942_cast_fp16_19))[name = tensor("aw_599_cast_fp16")]; + tensor var_4046_cast_fp16 = softmax(axis = var_3890, x = aw_561_cast_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor var_4047_cast_fp16 = softmax(axis = var_3890, x = aw_563_cast_fp16)[name = tensor("op_4047_cast_fp16")]; + tensor var_4048_cast_fp16 = softmax(axis = var_3890, x = aw_565_cast_fp16)[name = tensor("op_4048_cast_fp16")]; + tensor var_4049_cast_fp16 = softmax(axis = var_3890, x = aw_567_cast_fp16)[name = tensor("op_4049_cast_fp16")]; + tensor var_4050_cast_fp16 = softmax(axis = var_3890, x = aw_569_cast_fp16)[name = tensor("op_4050_cast_fp16")]; + tensor var_4051_cast_fp16 = softmax(axis = var_3890, x = aw_571_cast_fp16)[name = tensor("op_4051_cast_fp16")]; + tensor var_4052_cast_fp16 = softmax(axis = var_3890, x = aw_573_cast_fp16)[name = tensor("op_4052_cast_fp16")]; + tensor var_4053_cast_fp16 = softmax(axis = var_3890, x = aw_575_cast_fp16)[name = tensor("op_4053_cast_fp16")]; + tensor var_4054_cast_fp16 = softmax(axis = var_3890, x = aw_577_cast_fp16)[name = tensor("op_4054_cast_fp16")]; + tensor var_4055_cast_fp16 = softmax(axis = var_3890, x = aw_579_cast_fp16)[name = tensor("op_4055_cast_fp16")]; + tensor var_4056_cast_fp16 = softmax(axis = var_3890, x = aw_581_cast_fp16)[name = tensor("op_4056_cast_fp16")]; + tensor var_4057_cast_fp16 = softmax(axis = var_3890, x = aw_583_cast_fp16)[name = tensor("op_4057_cast_fp16")]; + tensor var_4058_cast_fp16 = softmax(axis = var_3890, x = aw_585_cast_fp16)[name = tensor("op_4058_cast_fp16")]; + tensor var_4059_cast_fp16 = softmax(axis = var_3890, x = aw_587_cast_fp16)[name = tensor("op_4059_cast_fp16")]; + tensor var_4060_cast_fp16 = softmax(axis = var_3890, x = aw_589_cast_fp16)[name = tensor("op_4060_cast_fp16")]; + tensor var_4061_cast_fp16 = softmax(axis = var_3890, x = aw_591_cast_fp16)[name = tensor("op_4061_cast_fp16")]; + tensor var_4062_cast_fp16 = softmax(axis = var_3890, x = aw_593_cast_fp16)[name = tensor("op_4062_cast_fp16")]; + tensor var_4063_cast_fp16 = softmax(axis = var_3890, x = aw_595_cast_fp16)[name = tensor("op_4063_cast_fp16")]; + tensor var_4064_cast_fp16 = softmax(axis = var_3890, x = aw_597_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_cast_fp16 = softmax(axis = var_3890, x = aw_599_cast_fp16)[name = tensor("op_4065_cast_fp16")]; + tensor var_4067_equation_0 = const()[name = tensor("op_4067_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4067_cast_fp16 = einsum(equation = var_4067_equation_0, values = (var_3985_cast_fp16_0, var_4046_cast_fp16))[name = tensor("op_4067_cast_fp16")]; + tensor var_4069_equation_0 = const()[name = tensor("op_4069_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4069_cast_fp16 = einsum(equation = var_4069_equation_0, values = (var_3985_cast_fp16_1, var_4047_cast_fp16))[name = tensor("op_4069_cast_fp16")]; + tensor var_4071_equation_0 = const()[name = tensor("op_4071_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4071_cast_fp16 = einsum(equation = var_4071_equation_0, values = (var_3985_cast_fp16_2, var_4048_cast_fp16))[name = tensor("op_4071_cast_fp16")]; + tensor var_4073_equation_0 = const()[name = tensor("op_4073_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4073_cast_fp16 = einsum(equation = var_4073_equation_0, values = (var_3985_cast_fp16_3, var_4049_cast_fp16))[name = tensor("op_4073_cast_fp16")]; + tensor var_4075_equation_0 = const()[name = tensor("op_4075_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4075_cast_fp16 = einsum(equation = var_4075_equation_0, values = (var_3985_cast_fp16_4, var_4050_cast_fp16))[name = tensor("op_4075_cast_fp16")]; + tensor var_4077_equation_0 = const()[name = tensor("op_4077_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4077_cast_fp16 = einsum(equation = var_4077_equation_0, values = (var_3985_cast_fp16_5, var_4051_cast_fp16))[name = tensor("op_4077_cast_fp16")]; + tensor var_4079_equation_0 = const()[name = tensor("op_4079_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4079_cast_fp16 = einsum(equation = var_4079_equation_0, values = (var_3985_cast_fp16_6, var_4052_cast_fp16))[name = tensor("op_4079_cast_fp16")]; + tensor var_4081_equation_0 = const()[name = tensor("op_4081_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4081_cast_fp16 = einsum(equation = var_4081_equation_0, values = (var_3985_cast_fp16_7, var_4053_cast_fp16))[name = tensor("op_4081_cast_fp16")]; + tensor var_4083_equation_0 = const()[name = tensor("op_4083_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4083_cast_fp16 = einsum(equation = var_4083_equation_0, values = (var_3985_cast_fp16_8, var_4054_cast_fp16))[name = tensor("op_4083_cast_fp16")]; + tensor var_4085_equation_0 = const()[name = tensor("op_4085_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4085_cast_fp16 = einsum(equation = var_4085_equation_0, values = (var_3985_cast_fp16_9, var_4055_cast_fp16))[name = tensor("op_4085_cast_fp16")]; + tensor var_4087_equation_0 = const()[name = tensor("op_4087_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4087_cast_fp16 = einsum(equation = var_4087_equation_0, values = (var_3985_cast_fp16_10, var_4056_cast_fp16))[name = tensor("op_4087_cast_fp16")]; + tensor var_4089_equation_0 = const()[name = tensor("op_4089_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4089_cast_fp16 = einsum(equation = var_4089_equation_0, values = (var_3985_cast_fp16_11, var_4057_cast_fp16))[name = tensor("op_4089_cast_fp16")]; + tensor var_4091_equation_0 = const()[name = tensor("op_4091_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4091_cast_fp16 = einsum(equation = var_4091_equation_0, values = (var_3985_cast_fp16_12, var_4058_cast_fp16))[name = tensor("op_4091_cast_fp16")]; + tensor var_4093_equation_0 = const()[name = tensor("op_4093_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4093_cast_fp16 = einsum(equation = var_4093_equation_0, values = (var_3985_cast_fp16_13, var_4059_cast_fp16))[name = tensor("op_4093_cast_fp16")]; + tensor var_4095_equation_0 = const()[name = tensor("op_4095_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4095_cast_fp16 = einsum(equation = var_4095_equation_0, values = (var_3985_cast_fp16_14, var_4060_cast_fp16))[name = tensor("op_4095_cast_fp16")]; + tensor var_4097_equation_0 = const()[name = tensor("op_4097_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4097_cast_fp16 = einsum(equation = var_4097_equation_0, values = (var_3985_cast_fp16_15, var_4061_cast_fp16))[name = tensor("op_4097_cast_fp16")]; + tensor var_4099_equation_0 = const()[name = tensor("op_4099_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4099_cast_fp16 = einsum(equation = var_4099_equation_0, values = (var_3985_cast_fp16_16, var_4062_cast_fp16))[name = tensor("op_4099_cast_fp16")]; + tensor var_4101_equation_0 = const()[name = tensor("op_4101_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4101_cast_fp16 = einsum(equation = var_4101_equation_0, values = (var_3985_cast_fp16_17, var_4063_cast_fp16))[name = tensor("op_4101_cast_fp16")]; + tensor var_4103_equation_0 = const()[name = tensor("op_4103_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4103_cast_fp16 = einsum(equation = var_4103_equation_0, values = (var_3985_cast_fp16_18, var_4064_cast_fp16))[name = tensor("op_4103_cast_fp16")]; + tensor var_4105_equation_0 = const()[name = tensor("op_4105_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4105_cast_fp16 = einsum(equation = var_4105_equation_0, values = (var_3985_cast_fp16_19, var_4065_cast_fp16))[name = tensor("op_4105_cast_fp16")]; + tensor input_145_interleave_0 = const()[name = tensor("input_145_interleave_0"), val = tensor(false)]; + tensor input_145_cast_fp16 = concat(axis = var_3890, interleave = input_145_interleave_0, values = (var_4067_cast_fp16, var_4069_cast_fp16, var_4071_cast_fp16, var_4073_cast_fp16, var_4075_cast_fp16, var_4077_cast_fp16, var_4079_cast_fp16, var_4081_cast_fp16, var_4083_cast_fp16, var_4085_cast_fp16, var_4087_cast_fp16, var_4089_cast_fp16, var_4091_cast_fp16, var_4093_cast_fp16, var_4095_cast_fp16, var_4097_cast_fp16, var_4099_cast_fp16, var_4101_cast_fp16, var_4103_cast_fp16, var_4105_cast_fp16))[name = tensor("input_145_cast_fp16")]; + tensor var_4114_pad_type_0 = const()[name = tensor("op_4114_pad_type_0"), val = tensor("valid")]; + tensor var_4114_strides_0 = const()[name = tensor("op_4114_strides_0"), val = tensor([1, 1])]; + tensor var_4114_pad_0 = const()[name = tensor("op_4114_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4114_dilations_0 = const()[name = tensor("op_4114_dilations_0"), val = tensor([1, 1])]; + tensor var_4114_groups_0 = const()[name = tensor("op_4114_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_out_weight_to_fp16 = const()[name = tensor("blocks_14_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575445952)))]; + tensor blocks_14_attn_out_bias_to_fp16 = const()[name = tensor("blocks_14_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578722816)))]; + tensor var_4114_cast_fp16 = conv(bias = blocks_14_attn_out_bias_to_fp16, dilations = var_4114_dilations_0, groups = var_4114_groups_0, pad = var_4114_pad_0, pad_type = var_4114_pad_type_0, strides = var_4114_strides_0, weight = blocks_14_attn_out_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("op_4114_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_4114_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([1])]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578725440)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578728064)))]; + tensor var_4124_to_fp16 = const()[name = tensor("op_4124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = input_147_beta_0_to_fp16, epsilon = var_4124_to_fp16, gamma = input_147_gamma_0_to_fp16, x = inputs_59_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578730688)))]; + tensor blocks_14_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591837952)))]; + tensor input_149_cast_fp16 = conv(bias = blocks_14_mlp_0_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = blocks_14_mlp_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_4150_pad_type_0 = const()[name = tensor("op_4150_pad_type_0"), val = tensor("valid")]; + tensor var_4150_strides_0 = const()[name = tensor("op_4150_strides_0"), val = tensor([1, 1])]; + tensor var_4150_pad_0 = const()[name = tensor("op_4150_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4150_dilations_0 = const()[name = tensor("op_4150_dilations_0"), val = tensor([1, 1])]; + tensor var_4150_groups_0 = const()[name = tensor("op_4150_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591848256)))]; + tensor blocks_14_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604955520)))]; + tensor var_4150_cast_fp16 = conv(bias = blocks_14_mlp_2_bias_to_fp16, dilations = var_4150_dilations_0, groups = var_4150_groups_0, pad = var_4150_pad_0, pad_type = var_4150_pad_type_0, strides = var_4150_strides_0, weight = blocks_14_mlp_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = var_4150_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_4159 = const()[name = tensor("op_4159"), val = tensor(1)]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([1])]; + tensor input_153_gamma_0_to_fp16 = const()[name = tensor("input_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604958144)))]; + tensor input_153_beta_0_to_fp16 = const()[name = tensor("input_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604960768)))]; + tensor var_4175_to_fp16 = const()[name = tensor("op_4175_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = input_153_beta_0_to_fp16, epsilon = var_4175_to_fp16, gamma = input_153_gamma_0_to_fp16, x = inputs_61_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor q_31_pad_type_0 = const()[name = tensor("q_31_pad_type_0"), val = tensor("valid")]; + tensor q_31_strides_0 = const()[name = tensor("q_31_strides_0"), val = tensor([1, 1])]; + tensor q_31_pad_0 = const()[name = tensor("q_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_31_dilations_0 = const()[name = tensor("q_31_dilations_0"), val = tensor([1, 1])]; + tensor q_31_groups_0 = const()[name = tensor("q_31_groups_0"), val = tensor(1)]; + tensor var_4210_weight_0_to_fp16 = const()[name = tensor("op_4210_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604963392)))]; + tensor var_4210_bias_0_to_fp16 = const()[name = tensor("op_4210_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(608240256)))]; + tensor var_4210_cast_fp16 = conv(bias = var_4210_bias_0_to_fp16, dilations = q_31_dilations_0, groups = q_31_groups_0, pad = q_31_pad_0, pad_type = q_31_pad_type_0, strides = q_31_strides_0, weight = var_4210_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("op_4210_cast_fp16")]; + tensor k_31_pad_type_0 = const()[name = tensor("k_31_pad_type_0"), val = tensor("valid")]; + tensor k_31_strides_0 = const()[name = tensor("k_31_strides_0"), val = tensor([1, 1])]; + tensor k_31_pad_0 = const()[name = tensor("k_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_31_dilations_0 = const()[name = tensor("k_31_dilations_0"), val = tensor([1, 1])]; + tensor k_31_groups_0 = const()[name = tensor("k_31_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_key_weight_to_fp16 = const()[name = tensor("blocks_15_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(608242880)))]; + tensor k_31_cast_fp16 = conv(dilations = k_31_dilations_0, groups = k_31_groups_0, pad = k_31_pad_0, pad_type = k_31_pad_type_0, strides = k_31_strides_0, weight = blocks_15_attn_key_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor var_4208_pad_type_0 = const()[name = tensor("op_4208_pad_type_0"), val = tensor("valid")]; + tensor var_4208_strides_0 = const()[name = tensor("op_4208_strides_0"), val = tensor([1, 1])]; + tensor var_4208_pad_0 = const()[name = tensor("op_4208_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4208_dilations_0 = const()[name = tensor("op_4208_dilations_0"), val = tensor([1, 1])]; + tensor var_4208_groups_0 = const()[name = tensor("op_4208_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_value_weight_to_fp16 = const()[name = tensor("blocks_15_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611519744)))]; + tensor blocks_15_attn_value_bias_to_fp16 = const()[name = tensor("blocks_15_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614796608)))]; + tensor var_4208_cast_fp16 = conv(bias = blocks_15_attn_value_bias_to_fp16, dilations = var_4208_dilations_0, groups = var_4208_groups_0, pad = var_4208_pad_0, pad_type = var_4208_pad_type_0, strides = var_4208_strides_0, weight = blocks_15_attn_value_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor tile_45 = const()[name = tensor("tile_45"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4211_axis_0 = const()[name = tensor("op_4211_axis_0"), val = tensor(1)]; + tensor var_4211_cast_fp16_0, tensor var_4211_cast_fp16_1, tensor var_4211_cast_fp16_2, tensor var_4211_cast_fp16_3, tensor var_4211_cast_fp16_4, tensor var_4211_cast_fp16_5, tensor var_4211_cast_fp16_6, tensor var_4211_cast_fp16_7, tensor var_4211_cast_fp16_8, tensor var_4211_cast_fp16_9, tensor var_4211_cast_fp16_10, tensor var_4211_cast_fp16_11, tensor var_4211_cast_fp16_12, tensor var_4211_cast_fp16_13, tensor var_4211_cast_fp16_14, tensor var_4211_cast_fp16_15, tensor var_4211_cast_fp16_16, tensor var_4211_cast_fp16_17, tensor var_4211_cast_fp16_18, tensor var_4211_cast_fp16_19 = split(axis = var_4211_axis_0, split_sizes = tile_45, x = var_4210_cast_fp16)[name = tensor("op_4211_cast_fp16")]; + tensor var_4232_perm_0 = const()[name = tensor("op_4232_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_46 = const()[name = tensor("tile_46"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4233_axis_0 = const()[name = tensor("op_4233_axis_0"), val = tensor(3)]; + tensor var_4232_cast_fp16 = transpose(perm = var_4232_perm_0, x = k_31_cast_fp16)[name = tensor("transpose_17")]; + tensor var_4233_cast_fp16_0, tensor var_4233_cast_fp16_1, tensor var_4233_cast_fp16_2, tensor var_4233_cast_fp16_3, tensor var_4233_cast_fp16_4, tensor var_4233_cast_fp16_5, tensor var_4233_cast_fp16_6, tensor var_4233_cast_fp16_7, tensor var_4233_cast_fp16_8, tensor var_4233_cast_fp16_9, tensor var_4233_cast_fp16_10, tensor var_4233_cast_fp16_11, tensor var_4233_cast_fp16_12, tensor var_4233_cast_fp16_13, tensor var_4233_cast_fp16_14, tensor var_4233_cast_fp16_15, tensor var_4233_cast_fp16_16, tensor var_4233_cast_fp16_17, tensor var_4233_cast_fp16_18, tensor var_4233_cast_fp16_19 = split(axis = var_4233_axis_0, split_sizes = tile_46, x = var_4232_cast_fp16)[name = tensor("op_4233_cast_fp16")]; + tensor tile_47 = const()[name = tensor("tile_47"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4254_axis_0 = const()[name = tensor("op_4254_axis_0"), val = tensor(1)]; + tensor var_4254_cast_fp16_0, tensor var_4254_cast_fp16_1, tensor var_4254_cast_fp16_2, tensor var_4254_cast_fp16_3, tensor var_4254_cast_fp16_4, tensor var_4254_cast_fp16_5, tensor var_4254_cast_fp16_6, tensor var_4254_cast_fp16_7, tensor var_4254_cast_fp16_8, tensor var_4254_cast_fp16_9, tensor var_4254_cast_fp16_10, tensor var_4254_cast_fp16_11, tensor var_4254_cast_fp16_12, tensor var_4254_cast_fp16_13, tensor var_4254_cast_fp16_14, tensor var_4254_cast_fp16_15, tensor var_4254_cast_fp16_16, tensor var_4254_cast_fp16_17, tensor var_4254_cast_fp16_18, tensor var_4254_cast_fp16_19 = split(axis = var_4254_axis_0, split_sizes = tile_47, x = var_4208_cast_fp16)[name = tensor("op_4254_cast_fp16")]; + tensor aw_601_equation_0 = const()[name = tensor("aw_601_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_601_cast_fp16 = einsum(equation = aw_601_equation_0, values = (var_4233_cast_fp16_0, var_4211_cast_fp16_0))[name = tensor("aw_601_cast_fp16")]; + tensor aw_603_equation_0 = const()[name = tensor("aw_603_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_603_cast_fp16 = einsum(equation = aw_603_equation_0, values = (var_4233_cast_fp16_1, var_4211_cast_fp16_1))[name = tensor("aw_603_cast_fp16")]; + tensor aw_605_equation_0 = const()[name = tensor("aw_605_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_605_cast_fp16 = einsum(equation = aw_605_equation_0, values = (var_4233_cast_fp16_2, var_4211_cast_fp16_2))[name = tensor("aw_605_cast_fp16")]; + tensor aw_607_equation_0 = const()[name = tensor("aw_607_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_607_cast_fp16 = einsum(equation = aw_607_equation_0, values = (var_4233_cast_fp16_3, var_4211_cast_fp16_3))[name = tensor("aw_607_cast_fp16")]; + tensor aw_609_equation_0 = const()[name = tensor("aw_609_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_609_cast_fp16 = einsum(equation = aw_609_equation_0, values = (var_4233_cast_fp16_4, var_4211_cast_fp16_4))[name = tensor("aw_609_cast_fp16")]; + tensor aw_611_equation_0 = const()[name = tensor("aw_611_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_611_cast_fp16 = einsum(equation = aw_611_equation_0, values = (var_4233_cast_fp16_5, var_4211_cast_fp16_5))[name = tensor("aw_611_cast_fp16")]; + tensor aw_613_equation_0 = const()[name = tensor("aw_613_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_613_cast_fp16 = einsum(equation = aw_613_equation_0, values = (var_4233_cast_fp16_6, var_4211_cast_fp16_6))[name = tensor("aw_613_cast_fp16")]; + tensor aw_615_equation_0 = const()[name = tensor("aw_615_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_615_cast_fp16 = einsum(equation = aw_615_equation_0, values = (var_4233_cast_fp16_7, var_4211_cast_fp16_7))[name = tensor("aw_615_cast_fp16")]; + tensor aw_617_equation_0 = const()[name = tensor("aw_617_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_617_cast_fp16 = einsum(equation = aw_617_equation_0, values = (var_4233_cast_fp16_8, var_4211_cast_fp16_8))[name = tensor("aw_617_cast_fp16")]; + tensor aw_619_equation_0 = const()[name = tensor("aw_619_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_619_cast_fp16 = einsum(equation = aw_619_equation_0, values = (var_4233_cast_fp16_9, var_4211_cast_fp16_9))[name = tensor("aw_619_cast_fp16")]; + tensor aw_621_equation_0 = const()[name = tensor("aw_621_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_621_cast_fp16 = einsum(equation = aw_621_equation_0, values = (var_4233_cast_fp16_10, var_4211_cast_fp16_10))[name = tensor("aw_621_cast_fp16")]; + tensor aw_623_equation_0 = const()[name = tensor("aw_623_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_623_cast_fp16 = einsum(equation = aw_623_equation_0, values = (var_4233_cast_fp16_11, var_4211_cast_fp16_11))[name = tensor("aw_623_cast_fp16")]; + tensor aw_625_equation_0 = const()[name = tensor("aw_625_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_625_cast_fp16 = einsum(equation = aw_625_equation_0, values = (var_4233_cast_fp16_12, var_4211_cast_fp16_12))[name = tensor("aw_625_cast_fp16")]; + tensor aw_627_equation_0 = const()[name = tensor("aw_627_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_627_cast_fp16 = einsum(equation = aw_627_equation_0, values = (var_4233_cast_fp16_13, var_4211_cast_fp16_13))[name = tensor("aw_627_cast_fp16")]; + tensor aw_629_equation_0 = const()[name = tensor("aw_629_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_629_cast_fp16 = einsum(equation = aw_629_equation_0, values = (var_4233_cast_fp16_14, var_4211_cast_fp16_14))[name = tensor("aw_629_cast_fp16")]; + tensor aw_631_equation_0 = const()[name = tensor("aw_631_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_631_cast_fp16 = einsum(equation = aw_631_equation_0, values = (var_4233_cast_fp16_15, var_4211_cast_fp16_15))[name = tensor("aw_631_cast_fp16")]; + tensor aw_633_equation_0 = const()[name = tensor("aw_633_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_633_cast_fp16 = einsum(equation = aw_633_equation_0, values = (var_4233_cast_fp16_16, var_4211_cast_fp16_16))[name = tensor("aw_633_cast_fp16")]; + tensor aw_635_equation_0 = const()[name = tensor("aw_635_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_635_cast_fp16 = einsum(equation = aw_635_equation_0, values = (var_4233_cast_fp16_17, var_4211_cast_fp16_17))[name = tensor("aw_635_cast_fp16")]; + tensor aw_637_equation_0 = const()[name = tensor("aw_637_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_637_cast_fp16 = einsum(equation = aw_637_equation_0, values = (var_4233_cast_fp16_18, var_4211_cast_fp16_18))[name = tensor("aw_637_cast_fp16")]; + tensor aw_639_equation_0 = const()[name = tensor("aw_639_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_639_cast_fp16 = einsum(equation = aw_639_equation_0, values = (var_4233_cast_fp16_19, var_4211_cast_fp16_19))[name = tensor("aw_639_cast_fp16")]; + tensor var_4315_cast_fp16 = softmax(axis = var_4159, x = aw_601_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4316_cast_fp16 = softmax(axis = var_4159, x = aw_603_cast_fp16)[name = tensor("op_4316_cast_fp16")]; + tensor var_4317_cast_fp16 = softmax(axis = var_4159, x = aw_605_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor var_4318_cast_fp16 = softmax(axis = var_4159, x = aw_607_cast_fp16)[name = tensor("op_4318_cast_fp16")]; + tensor var_4319_cast_fp16 = softmax(axis = var_4159, x = aw_609_cast_fp16)[name = tensor("op_4319_cast_fp16")]; + tensor var_4320_cast_fp16 = softmax(axis = var_4159, x = aw_611_cast_fp16)[name = tensor("op_4320_cast_fp16")]; + tensor var_4321_cast_fp16 = softmax(axis = var_4159, x = aw_613_cast_fp16)[name = tensor("op_4321_cast_fp16")]; + tensor var_4322_cast_fp16 = softmax(axis = var_4159, x = aw_615_cast_fp16)[name = tensor("op_4322_cast_fp16")]; + tensor var_4323_cast_fp16 = softmax(axis = var_4159, x = aw_617_cast_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4324_cast_fp16 = softmax(axis = var_4159, x = aw_619_cast_fp16)[name = tensor("op_4324_cast_fp16")]; + tensor var_4325_cast_fp16 = softmax(axis = var_4159, x = aw_621_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor var_4326_cast_fp16 = softmax(axis = var_4159, x = aw_623_cast_fp16)[name = tensor("op_4326_cast_fp16")]; + tensor var_4327_cast_fp16 = softmax(axis = var_4159, x = aw_625_cast_fp16)[name = tensor("op_4327_cast_fp16")]; + tensor var_4328_cast_fp16 = softmax(axis = var_4159, x = aw_627_cast_fp16)[name = tensor("op_4328_cast_fp16")]; + tensor var_4329_cast_fp16 = softmax(axis = var_4159, x = aw_629_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_cast_fp16 = softmax(axis = var_4159, x = aw_631_cast_fp16)[name = tensor("op_4330_cast_fp16")]; + tensor var_4331_cast_fp16 = softmax(axis = var_4159, x = aw_633_cast_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor var_4332_cast_fp16 = softmax(axis = var_4159, x = aw_635_cast_fp16)[name = tensor("op_4332_cast_fp16")]; + tensor var_4333_cast_fp16 = softmax(axis = var_4159, x = aw_637_cast_fp16)[name = tensor("op_4333_cast_fp16")]; + tensor var_4334_cast_fp16 = softmax(axis = var_4159, x = aw_639_cast_fp16)[name = tensor("op_4334_cast_fp16")]; + tensor var_4336_equation_0 = const()[name = tensor("op_4336_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4336_cast_fp16 = einsum(equation = var_4336_equation_0, values = (var_4254_cast_fp16_0, var_4315_cast_fp16))[name = tensor("op_4336_cast_fp16")]; + tensor var_4338_equation_0 = const()[name = tensor("op_4338_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4338_cast_fp16 = einsum(equation = var_4338_equation_0, values = (var_4254_cast_fp16_1, var_4316_cast_fp16))[name = tensor("op_4338_cast_fp16")]; + tensor var_4340_equation_0 = const()[name = tensor("op_4340_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4340_cast_fp16 = einsum(equation = var_4340_equation_0, values = (var_4254_cast_fp16_2, var_4317_cast_fp16))[name = tensor("op_4340_cast_fp16")]; + tensor var_4342_equation_0 = const()[name = tensor("op_4342_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4342_cast_fp16 = einsum(equation = var_4342_equation_0, values = (var_4254_cast_fp16_3, var_4318_cast_fp16))[name = tensor("op_4342_cast_fp16")]; + tensor var_4344_equation_0 = const()[name = tensor("op_4344_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4344_cast_fp16 = einsum(equation = var_4344_equation_0, values = (var_4254_cast_fp16_4, var_4319_cast_fp16))[name = tensor("op_4344_cast_fp16")]; + tensor var_4346_equation_0 = const()[name = tensor("op_4346_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4346_cast_fp16 = einsum(equation = var_4346_equation_0, values = (var_4254_cast_fp16_5, var_4320_cast_fp16))[name = tensor("op_4346_cast_fp16")]; + tensor var_4348_equation_0 = const()[name = tensor("op_4348_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4348_cast_fp16 = einsum(equation = var_4348_equation_0, values = (var_4254_cast_fp16_6, var_4321_cast_fp16))[name = tensor("op_4348_cast_fp16")]; + tensor var_4350_equation_0 = const()[name = tensor("op_4350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4350_cast_fp16 = einsum(equation = var_4350_equation_0, values = (var_4254_cast_fp16_7, var_4322_cast_fp16))[name = tensor("op_4350_cast_fp16")]; + tensor var_4352_equation_0 = const()[name = tensor("op_4352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4352_cast_fp16 = einsum(equation = var_4352_equation_0, values = (var_4254_cast_fp16_8, var_4323_cast_fp16))[name = tensor("op_4352_cast_fp16")]; + tensor var_4354_equation_0 = const()[name = tensor("op_4354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4354_cast_fp16 = einsum(equation = var_4354_equation_0, values = (var_4254_cast_fp16_9, var_4324_cast_fp16))[name = tensor("op_4354_cast_fp16")]; + tensor var_4356_equation_0 = const()[name = tensor("op_4356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4356_cast_fp16 = einsum(equation = var_4356_equation_0, values = (var_4254_cast_fp16_10, var_4325_cast_fp16))[name = tensor("op_4356_cast_fp16")]; + tensor var_4358_equation_0 = const()[name = tensor("op_4358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4358_cast_fp16 = einsum(equation = var_4358_equation_0, values = (var_4254_cast_fp16_11, var_4326_cast_fp16))[name = tensor("op_4358_cast_fp16")]; + tensor var_4360_equation_0 = const()[name = tensor("op_4360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4360_cast_fp16 = einsum(equation = var_4360_equation_0, values = (var_4254_cast_fp16_12, var_4327_cast_fp16))[name = tensor("op_4360_cast_fp16")]; + tensor var_4362_equation_0 = const()[name = tensor("op_4362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4362_cast_fp16 = einsum(equation = var_4362_equation_0, values = (var_4254_cast_fp16_13, var_4328_cast_fp16))[name = tensor("op_4362_cast_fp16")]; + tensor var_4364_equation_0 = const()[name = tensor("op_4364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4364_cast_fp16 = einsum(equation = var_4364_equation_0, values = (var_4254_cast_fp16_14, var_4329_cast_fp16))[name = tensor("op_4364_cast_fp16")]; + tensor var_4366_equation_0 = const()[name = tensor("op_4366_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4366_cast_fp16 = einsum(equation = var_4366_equation_0, values = (var_4254_cast_fp16_15, var_4330_cast_fp16))[name = tensor("op_4366_cast_fp16")]; + tensor var_4368_equation_0 = const()[name = tensor("op_4368_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4368_cast_fp16 = einsum(equation = var_4368_equation_0, values = (var_4254_cast_fp16_16, var_4331_cast_fp16))[name = tensor("op_4368_cast_fp16")]; + tensor var_4370_equation_0 = const()[name = tensor("op_4370_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4370_cast_fp16 = einsum(equation = var_4370_equation_0, values = (var_4254_cast_fp16_17, var_4332_cast_fp16))[name = tensor("op_4370_cast_fp16")]; + tensor var_4372_equation_0 = const()[name = tensor("op_4372_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4372_cast_fp16 = einsum(equation = var_4372_equation_0, values = (var_4254_cast_fp16_18, var_4333_cast_fp16))[name = tensor("op_4372_cast_fp16")]; + tensor var_4374_equation_0 = const()[name = tensor("op_4374_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4374_cast_fp16 = einsum(equation = var_4374_equation_0, values = (var_4254_cast_fp16_19, var_4334_cast_fp16))[name = tensor("op_4374_cast_fp16")]; + tensor input_155_interleave_0 = const()[name = tensor("input_155_interleave_0"), val = tensor(false)]; + tensor input_155_cast_fp16 = concat(axis = var_4159, interleave = input_155_interleave_0, values = (var_4336_cast_fp16, var_4338_cast_fp16, var_4340_cast_fp16, var_4342_cast_fp16, var_4344_cast_fp16, var_4346_cast_fp16, var_4348_cast_fp16, var_4350_cast_fp16, var_4352_cast_fp16, var_4354_cast_fp16, var_4356_cast_fp16, var_4358_cast_fp16, var_4360_cast_fp16, var_4362_cast_fp16, var_4364_cast_fp16, var_4366_cast_fp16, var_4368_cast_fp16, var_4370_cast_fp16, var_4372_cast_fp16, var_4374_cast_fp16))[name = tensor("input_155_cast_fp16")]; + tensor var_4383_pad_type_0 = const()[name = tensor("op_4383_pad_type_0"), val = tensor("valid")]; + tensor var_4383_strides_0 = const()[name = tensor("op_4383_strides_0"), val = tensor([1, 1])]; + tensor var_4383_pad_0 = const()[name = tensor("op_4383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4383_dilations_0 = const()[name = tensor("op_4383_dilations_0"), val = tensor([1, 1])]; + tensor var_4383_groups_0 = const()[name = tensor("op_4383_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_out_weight_to_fp16 = const()[name = tensor("blocks_15_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614799232)))]; + tensor blocks_15_attn_out_bias_to_fp16 = const()[name = tensor("blocks_15_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618076096)))]; + tensor var_4383_cast_fp16 = conv(bias = blocks_15_attn_out_bias_to_fp16, dilations = var_4383_dilations_0, groups = var_4383_groups_0, pad = var_4383_pad_0, pad_type = var_4383_pad_type_0, strides = var_4383_strides_0, weight = blocks_15_attn_out_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("op_4383_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_4383_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor input_157_axes_0 = const()[name = tensor("input_157_axes_0"), val = tensor([1])]; + tensor input_157_gamma_0_to_fp16 = const()[name = tensor("input_157_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618078720)))]; + tensor input_157_beta_0_to_fp16 = const()[name = tensor("input_157_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618081344)))]; + tensor var_4393_to_fp16 = const()[name = tensor("op_4393_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = input_157_beta_0_to_fp16, epsilon = var_4393_to_fp16, gamma = input_157_gamma_0_to_fp16, x = inputs_63_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("valid")]; + tensor input_159_strides_0 = const()[name = tensor("input_159_strides_0"), val = tensor([1, 1])]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_159_dilations_0 = const()[name = tensor("input_159_dilations_0"), val = tensor([1, 1])]; + tensor input_159_groups_0 = const()[name = tensor("input_159_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618083968)))]; + tensor blocks_15_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(631191232)))]; + tensor input_159_cast_fp16 = conv(bias = blocks_15_mlp_0_bias_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = blocks_15_mlp_0_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_4419_pad_type_0 = const()[name = tensor("op_4419_pad_type_0"), val = tensor("valid")]; + tensor var_4419_strides_0 = const()[name = tensor("op_4419_strides_0"), val = tensor([1, 1])]; + tensor var_4419_pad_0 = const()[name = tensor("op_4419_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4419_dilations_0 = const()[name = tensor("op_4419_dilations_0"), val = tensor([1, 1])]; + tensor var_4419_groups_0 = const()[name = tensor("op_4419_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(631201536)))]; + tensor blocks_15_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644308800)))]; + tensor var_4419_cast_fp16 = conv(bias = blocks_15_mlp_2_bias_to_fp16, dilations = var_4419_dilations_0, groups = var_4419_groups_0, pad = var_4419_pad_0, pad_type = var_4419_pad_type_0, strides = var_4419_strides_0, weight = blocks_15_mlp_2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("op_4419_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = var_4419_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_4428 = const()[name = tensor("op_4428"), val = tensor(1)]; + tensor input_163_axes_0 = const()[name = tensor("input_163_axes_0"), val = tensor([1])]; + tensor input_163_gamma_0_to_fp16 = const()[name = tensor("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644311424)))]; + tensor input_163_beta_0_to_fp16 = const()[name = tensor("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644314048)))]; + tensor var_4444_to_fp16 = const()[name = tensor("op_4444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = input_163_beta_0_to_fp16, epsilon = var_4444_to_fp16, gamma = input_163_gamma_0_to_fp16, x = inputs_65_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor q_33_pad_type_0 = const()[name = tensor("q_33_pad_type_0"), val = tensor("valid")]; + tensor q_33_strides_0 = const()[name = tensor("q_33_strides_0"), val = tensor([1, 1])]; + tensor q_33_pad_0 = const()[name = tensor("q_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_33_dilations_0 = const()[name = tensor("q_33_dilations_0"), val = tensor([1, 1])]; + tensor q_33_groups_0 = const()[name = tensor("q_33_groups_0"), val = tensor(1)]; + tensor var_4479_weight_0_to_fp16 = const()[name = tensor("op_4479_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644316672)))]; + tensor var_4479_bias_0_to_fp16 = const()[name = tensor("op_4479_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647593536)))]; + tensor var_4479_cast_fp16 = conv(bias = var_4479_bias_0_to_fp16, dilations = q_33_dilations_0, groups = q_33_groups_0, pad = q_33_pad_0, pad_type = q_33_pad_type_0, strides = q_33_strides_0, weight = var_4479_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("op_4479_cast_fp16")]; + tensor k_33_pad_type_0 = const()[name = tensor("k_33_pad_type_0"), val = tensor("valid")]; + tensor k_33_strides_0 = const()[name = tensor("k_33_strides_0"), val = tensor([1, 1])]; + tensor k_33_pad_0 = const()[name = tensor("k_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_33_dilations_0 = const()[name = tensor("k_33_dilations_0"), val = tensor([1, 1])]; + tensor k_33_groups_0 = const()[name = tensor("k_33_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_key_weight_to_fp16 = const()[name = tensor("blocks_16_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647596160)))]; + tensor k_33_cast_fp16 = conv(dilations = k_33_dilations_0, groups = k_33_groups_0, pad = k_33_pad_0, pad_type = k_33_pad_type_0, strides = k_33_strides_0, weight = blocks_16_attn_key_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor var_4477_pad_type_0 = const()[name = tensor("op_4477_pad_type_0"), val = tensor("valid")]; + tensor var_4477_strides_0 = const()[name = tensor("op_4477_strides_0"), val = tensor([1, 1])]; + tensor var_4477_pad_0 = const()[name = tensor("op_4477_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4477_dilations_0 = const()[name = tensor("op_4477_dilations_0"), val = tensor([1, 1])]; + tensor var_4477_groups_0 = const()[name = tensor("op_4477_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_value_weight_to_fp16 = const()[name = tensor("blocks_16_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650873024)))]; + tensor blocks_16_attn_value_bias_to_fp16 = const()[name = tensor("blocks_16_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(654149888)))]; + tensor var_4477_cast_fp16 = conv(bias = blocks_16_attn_value_bias_to_fp16, dilations = var_4477_dilations_0, groups = var_4477_groups_0, pad = var_4477_pad_0, pad_type = var_4477_pad_type_0, strides = var_4477_strides_0, weight = blocks_16_attn_value_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("op_4477_cast_fp16")]; + tensor tile_48 = const()[name = tensor("tile_48"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4480_axis_0 = const()[name = tensor("op_4480_axis_0"), val = tensor(1)]; + tensor var_4480_cast_fp16_0, tensor var_4480_cast_fp16_1, tensor var_4480_cast_fp16_2, tensor var_4480_cast_fp16_3, tensor var_4480_cast_fp16_4, tensor var_4480_cast_fp16_5, tensor var_4480_cast_fp16_6, tensor var_4480_cast_fp16_7, tensor var_4480_cast_fp16_8, tensor var_4480_cast_fp16_9, tensor var_4480_cast_fp16_10, tensor var_4480_cast_fp16_11, tensor var_4480_cast_fp16_12, tensor var_4480_cast_fp16_13, tensor var_4480_cast_fp16_14, tensor var_4480_cast_fp16_15, tensor var_4480_cast_fp16_16, tensor var_4480_cast_fp16_17, tensor var_4480_cast_fp16_18, tensor var_4480_cast_fp16_19 = split(axis = var_4480_axis_0, split_sizes = tile_48, x = var_4479_cast_fp16)[name = tensor("op_4480_cast_fp16")]; + tensor var_4501_perm_0 = const()[name = tensor("op_4501_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_49 = const()[name = tensor("tile_49"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4502_axis_0 = const()[name = tensor("op_4502_axis_0"), val = tensor(3)]; + tensor var_4501_cast_fp16 = transpose(perm = var_4501_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_16")]; + tensor var_4502_cast_fp16_0, tensor var_4502_cast_fp16_1, tensor var_4502_cast_fp16_2, tensor var_4502_cast_fp16_3, tensor var_4502_cast_fp16_4, tensor var_4502_cast_fp16_5, tensor var_4502_cast_fp16_6, tensor var_4502_cast_fp16_7, tensor var_4502_cast_fp16_8, tensor var_4502_cast_fp16_9, tensor var_4502_cast_fp16_10, tensor var_4502_cast_fp16_11, tensor var_4502_cast_fp16_12, tensor var_4502_cast_fp16_13, tensor var_4502_cast_fp16_14, tensor var_4502_cast_fp16_15, tensor var_4502_cast_fp16_16, tensor var_4502_cast_fp16_17, tensor var_4502_cast_fp16_18, tensor var_4502_cast_fp16_19 = split(axis = var_4502_axis_0, split_sizes = tile_49, x = var_4501_cast_fp16)[name = tensor("op_4502_cast_fp16")]; + tensor tile_50 = const()[name = tensor("tile_50"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4523_axis_0 = const()[name = tensor("op_4523_axis_0"), val = tensor(1)]; + tensor var_4523_cast_fp16_0, tensor var_4523_cast_fp16_1, tensor var_4523_cast_fp16_2, tensor var_4523_cast_fp16_3, tensor var_4523_cast_fp16_4, tensor var_4523_cast_fp16_5, tensor var_4523_cast_fp16_6, tensor var_4523_cast_fp16_7, tensor var_4523_cast_fp16_8, tensor var_4523_cast_fp16_9, tensor var_4523_cast_fp16_10, tensor var_4523_cast_fp16_11, tensor var_4523_cast_fp16_12, tensor var_4523_cast_fp16_13, tensor var_4523_cast_fp16_14, tensor var_4523_cast_fp16_15, tensor var_4523_cast_fp16_16, tensor var_4523_cast_fp16_17, tensor var_4523_cast_fp16_18, tensor var_4523_cast_fp16_19 = split(axis = var_4523_axis_0, split_sizes = tile_50, x = var_4477_cast_fp16)[name = tensor("op_4523_cast_fp16")]; + tensor aw_641_equation_0 = const()[name = tensor("aw_641_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_641_cast_fp16 = einsum(equation = aw_641_equation_0, values = (var_4502_cast_fp16_0, var_4480_cast_fp16_0))[name = tensor("aw_641_cast_fp16")]; + tensor aw_643_equation_0 = const()[name = tensor("aw_643_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_643_cast_fp16 = einsum(equation = aw_643_equation_0, values = (var_4502_cast_fp16_1, var_4480_cast_fp16_1))[name = tensor("aw_643_cast_fp16")]; + tensor aw_645_equation_0 = const()[name = tensor("aw_645_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_645_cast_fp16 = einsum(equation = aw_645_equation_0, values = (var_4502_cast_fp16_2, var_4480_cast_fp16_2))[name = tensor("aw_645_cast_fp16")]; + tensor aw_647_equation_0 = const()[name = tensor("aw_647_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_647_cast_fp16 = einsum(equation = aw_647_equation_0, values = (var_4502_cast_fp16_3, var_4480_cast_fp16_3))[name = tensor("aw_647_cast_fp16")]; + tensor aw_649_equation_0 = const()[name = tensor("aw_649_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_649_cast_fp16 = einsum(equation = aw_649_equation_0, values = (var_4502_cast_fp16_4, var_4480_cast_fp16_4))[name = tensor("aw_649_cast_fp16")]; + tensor aw_651_equation_0 = const()[name = tensor("aw_651_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_651_cast_fp16 = einsum(equation = aw_651_equation_0, values = (var_4502_cast_fp16_5, var_4480_cast_fp16_5))[name = tensor("aw_651_cast_fp16")]; + tensor aw_653_equation_0 = const()[name = tensor("aw_653_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_653_cast_fp16 = einsum(equation = aw_653_equation_0, values = (var_4502_cast_fp16_6, var_4480_cast_fp16_6))[name = tensor("aw_653_cast_fp16")]; + tensor aw_655_equation_0 = const()[name = tensor("aw_655_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_655_cast_fp16 = einsum(equation = aw_655_equation_0, values = (var_4502_cast_fp16_7, var_4480_cast_fp16_7))[name = tensor("aw_655_cast_fp16")]; + tensor aw_657_equation_0 = const()[name = tensor("aw_657_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_657_cast_fp16 = einsum(equation = aw_657_equation_0, values = (var_4502_cast_fp16_8, var_4480_cast_fp16_8))[name = tensor("aw_657_cast_fp16")]; + tensor aw_659_equation_0 = const()[name = tensor("aw_659_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_659_cast_fp16 = einsum(equation = aw_659_equation_0, values = (var_4502_cast_fp16_9, var_4480_cast_fp16_9))[name = tensor("aw_659_cast_fp16")]; + tensor aw_661_equation_0 = const()[name = tensor("aw_661_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_661_cast_fp16 = einsum(equation = aw_661_equation_0, values = (var_4502_cast_fp16_10, var_4480_cast_fp16_10))[name = tensor("aw_661_cast_fp16")]; + tensor aw_663_equation_0 = const()[name = tensor("aw_663_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_663_cast_fp16 = einsum(equation = aw_663_equation_0, values = (var_4502_cast_fp16_11, var_4480_cast_fp16_11))[name = tensor("aw_663_cast_fp16")]; + tensor aw_665_equation_0 = const()[name = tensor("aw_665_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_665_cast_fp16 = einsum(equation = aw_665_equation_0, values = (var_4502_cast_fp16_12, var_4480_cast_fp16_12))[name = tensor("aw_665_cast_fp16")]; + tensor aw_667_equation_0 = const()[name = tensor("aw_667_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_667_cast_fp16 = einsum(equation = aw_667_equation_0, values = (var_4502_cast_fp16_13, var_4480_cast_fp16_13))[name = tensor("aw_667_cast_fp16")]; + tensor aw_669_equation_0 = const()[name = tensor("aw_669_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_669_cast_fp16 = einsum(equation = aw_669_equation_0, values = (var_4502_cast_fp16_14, var_4480_cast_fp16_14))[name = tensor("aw_669_cast_fp16")]; + tensor aw_671_equation_0 = const()[name = tensor("aw_671_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_671_cast_fp16 = einsum(equation = aw_671_equation_0, values = (var_4502_cast_fp16_15, var_4480_cast_fp16_15))[name = tensor("aw_671_cast_fp16")]; + tensor aw_673_equation_0 = const()[name = tensor("aw_673_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_673_cast_fp16 = einsum(equation = aw_673_equation_0, values = (var_4502_cast_fp16_16, var_4480_cast_fp16_16))[name = tensor("aw_673_cast_fp16")]; + tensor aw_675_equation_0 = const()[name = tensor("aw_675_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_675_cast_fp16 = einsum(equation = aw_675_equation_0, values = (var_4502_cast_fp16_17, var_4480_cast_fp16_17))[name = tensor("aw_675_cast_fp16")]; + tensor aw_677_equation_0 = const()[name = tensor("aw_677_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_677_cast_fp16 = einsum(equation = aw_677_equation_0, values = (var_4502_cast_fp16_18, var_4480_cast_fp16_18))[name = tensor("aw_677_cast_fp16")]; + tensor aw_679_equation_0 = const()[name = tensor("aw_679_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_679_cast_fp16 = einsum(equation = aw_679_equation_0, values = (var_4502_cast_fp16_19, var_4480_cast_fp16_19))[name = tensor("aw_679_cast_fp16")]; + tensor var_4584_cast_fp16 = softmax(axis = var_4428, x = aw_641_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = softmax(axis = var_4428, x = aw_643_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor var_4586_cast_fp16 = softmax(axis = var_4428, x = aw_645_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = softmax(axis = var_4428, x = aw_647_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor var_4588_cast_fp16 = softmax(axis = var_4428, x = aw_649_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = softmax(axis = var_4428, x = aw_651_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor var_4590_cast_fp16 = softmax(axis = var_4428, x = aw_653_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = softmax(axis = var_4428, x = aw_655_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor var_4592_cast_fp16 = softmax(axis = var_4428, x = aw_657_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = softmax(axis = var_4428, x = aw_659_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor var_4594_cast_fp16 = softmax(axis = var_4428, x = aw_661_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = softmax(axis = var_4428, x = aw_663_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor var_4596_cast_fp16 = softmax(axis = var_4428, x = aw_665_cast_fp16)[name = tensor("op_4596_cast_fp16")]; + tensor var_4597_cast_fp16 = softmax(axis = var_4428, x = aw_667_cast_fp16)[name = tensor("op_4597_cast_fp16")]; + tensor var_4598_cast_fp16 = softmax(axis = var_4428, x = aw_669_cast_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4599_cast_fp16 = softmax(axis = var_4428, x = aw_671_cast_fp16)[name = tensor("op_4599_cast_fp16")]; + tensor var_4600_cast_fp16 = softmax(axis = var_4428, x = aw_673_cast_fp16)[name = tensor("op_4600_cast_fp16")]; + tensor var_4601_cast_fp16 = softmax(axis = var_4428, x = aw_675_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor var_4602_cast_fp16 = softmax(axis = var_4428, x = aw_677_cast_fp16)[name = tensor("op_4602_cast_fp16")]; + tensor var_4603_cast_fp16 = softmax(axis = var_4428, x = aw_679_cast_fp16)[name = tensor("op_4603_cast_fp16")]; + tensor var_4605_equation_0 = const()[name = tensor("op_4605_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4605_cast_fp16 = einsum(equation = var_4605_equation_0, values = (var_4523_cast_fp16_0, var_4584_cast_fp16))[name = tensor("op_4605_cast_fp16")]; + tensor var_4607_equation_0 = const()[name = tensor("op_4607_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4607_cast_fp16 = einsum(equation = var_4607_equation_0, values = (var_4523_cast_fp16_1, var_4585_cast_fp16))[name = tensor("op_4607_cast_fp16")]; + tensor var_4609_equation_0 = const()[name = tensor("op_4609_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4609_cast_fp16 = einsum(equation = var_4609_equation_0, values = (var_4523_cast_fp16_2, var_4586_cast_fp16))[name = tensor("op_4609_cast_fp16")]; + tensor var_4611_equation_0 = const()[name = tensor("op_4611_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4611_cast_fp16 = einsum(equation = var_4611_equation_0, values = (var_4523_cast_fp16_3, var_4587_cast_fp16))[name = tensor("op_4611_cast_fp16")]; + tensor var_4613_equation_0 = const()[name = tensor("op_4613_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4613_cast_fp16 = einsum(equation = var_4613_equation_0, values = (var_4523_cast_fp16_4, var_4588_cast_fp16))[name = tensor("op_4613_cast_fp16")]; + tensor var_4615_equation_0 = const()[name = tensor("op_4615_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4615_cast_fp16 = einsum(equation = var_4615_equation_0, values = (var_4523_cast_fp16_5, var_4589_cast_fp16))[name = tensor("op_4615_cast_fp16")]; + tensor var_4617_equation_0 = const()[name = tensor("op_4617_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4617_cast_fp16 = einsum(equation = var_4617_equation_0, values = (var_4523_cast_fp16_6, var_4590_cast_fp16))[name = tensor("op_4617_cast_fp16")]; + tensor var_4619_equation_0 = const()[name = tensor("op_4619_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4619_cast_fp16 = einsum(equation = var_4619_equation_0, values = (var_4523_cast_fp16_7, var_4591_cast_fp16))[name = tensor("op_4619_cast_fp16")]; + tensor var_4621_equation_0 = const()[name = tensor("op_4621_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4621_cast_fp16 = einsum(equation = var_4621_equation_0, values = (var_4523_cast_fp16_8, var_4592_cast_fp16))[name = tensor("op_4621_cast_fp16")]; + tensor var_4623_equation_0 = const()[name = tensor("op_4623_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4623_cast_fp16 = einsum(equation = var_4623_equation_0, values = (var_4523_cast_fp16_9, var_4593_cast_fp16))[name = tensor("op_4623_cast_fp16")]; + tensor var_4625_equation_0 = const()[name = tensor("op_4625_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4625_cast_fp16 = einsum(equation = var_4625_equation_0, values = (var_4523_cast_fp16_10, var_4594_cast_fp16))[name = tensor("op_4625_cast_fp16")]; + tensor var_4627_equation_0 = const()[name = tensor("op_4627_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4627_cast_fp16 = einsum(equation = var_4627_equation_0, values = (var_4523_cast_fp16_11, var_4595_cast_fp16))[name = tensor("op_4627_cast_fp16")]; + tensor var_4629_equation_0 = const()[name = tensor("op_4629_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4629_cast_fp16 = einsum(equation = var_4629_equation_0, values = (var_4523_cast_fp16_12, var_4596_cast_fp16))[name = tensor("op_4629_cast_fp16")]; + tensor var_4631_equation_0 = const()[name = tensor("op_4631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4631_cast_fp16 = einsum(equation = var_4631_equation_0, values = (var_4523_cast_fp16_13, var_4597_cast_fp16))[name = tensor("op_4631_cast_fp16")]; + tensor var_4633_equation_0 = const()[name = tensor("op_4633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4633_cast_fp16 = einsum(equation = var_4633_equation_0, values = (var_4523_cast_fp16_14, var_4598_cast_fp16))[name = tensor("op_4633_cast_fp16")]; + tensor var_4635_equation_0 = const()[name = tensor("op_4635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4635_cast_fp16 = einsum(equation = var_4635_equation_0, values = (var_4523_cast_fp16_15, var_4599_cast_fp16))[name = tensor("op_4635_cast_fp16")]; + tensor var_4637_equation_0 = const()[name = tensor("op_4637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4637_cast_fp16 = einsum(equation = var_4637_equation_0, values = (var_4523_cast_fp16_16, var_4600_cast_fp16))[name = tensor("op_4637_cast_fp16")]; + tensor var_4639_equation_0 = const()[name = tensor("op_4639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4639_cast_fp16 = einsum(equation = var_4639_equation_0, values = (var_4523_cast_fp16_17, var_4601_cast_fp16))[name = tensor("op_4639_cast_fp16")]; + tensor var_4641_equation_0 = const()[name = tensor("op_4641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4641_cast_fp16 = einsum(equation = var_4641_equation_0, values = (var_4523_cast_fp16_18, var_4602_cast_fp16))[name = tensor("op_4641_cast_fp16")]; + tensor var_4643_equation_0 = const()[name = tensor("op_4643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4643_cast_fp16 = einsum(equation = var_4643_equation_0, values = (var_4523_cast_fp16_19, var_4603_cast_fp16))[name = tensor("op_4643_cast_fp16")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165_cast_fp16 = concat(axis = var_4428, interleave = input_165_interleave_0, values = (var_4605_cast_fp16, var_4607_cast_fp16, var_4609_cast_fp16, var_4611_cast_fp16, var_4613_cast_fp16, var_4615_cast_fp16, var_4617_cast_fp16, var_4619_cast_fp16, var_4621_cast_fp16, var_4623_cast_fp16, var_4625_cast_fp16, var_4627_cast_fp16, var_4629_cast_fp16, var_4631_cast_fp16, var_4633_cast_fp16, var_4635_cast_fp16, var_4637_cast_fp16, var_4639_cast_fp16, var_4641_cast_fp16, var_4643_cast_fp16))[name = tensor("input_165_cast_fp16")]; + tensor var_4652_pad_type_0 = const()[name = tensor("op_4652_pad_type_0"), val = tensor("valid")]; + tensor var_4652_strides_0 = const()[name = tensor("op_4652_strides_0"), val = tensor([1, 1])]; + tensor var_4652_pad_0 = const()[name = tensor("op_4652_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4652_dilations_0 = const()[name = tensor("op_4652_dilations_0"), val = tensor([1, 1])]; + tensor var_4652_groups_0 = const()[name = tensor("op_4652_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_out_weight_to_fp16 = const()[name = tensor("blocks_16_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(654152512)))]; + tensor blocks_16_attn_out_bias_to_fp16 = const()[name = tensor("blocks_16_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657429376)))]; + tensor var_4652_cast_fp16 = conv(bias = blocks_16_attn_out_bias_to_fp16, dilations = var_4652_dilations_0, groups = var_4652_groups_0, pad = var_4652_pad_0, pad_type = var_4652_pad_type_0, strides = var_4652_strides_0, weight = blocks_16_attn_out_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("op_4652_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = var_4652_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([1])]; + tensor input_167_gamma_0_to_fp16 = const()[name = tensor("input_167_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657432000)))]; + tensor input_167_beta_0_to_fp16 = const()[name = tensor("input_167_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657434624)))]; + tensor var_4662_to_fp16 = const()[name = tensor("op_4662_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = input_167_beta_0_to_fp16, epsilon = var_4662_to_fp16, gamma = input_167_gamma_0_to_fp16, x = inputs_67_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("valid")]; + tensor input_169_strides_0 = const()[name = tensor("input_169_strides_0"), val = tensor([1, 1])]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_169_dilations_0 = const()[name = tensor("input_169_dilations_0"), val = tensor([1, 1])]; + tensor input_169_groups_0 = const()[name = tensor("input_169_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657437248)))]; + tensor blocks_16_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670544512)))]; + tensor input_169_cast_fp16 = conv(bias = blocks_16_mlp_0_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = blocks_16_mlp_0_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_4688_pad_type_0 = const()[name = tensor("op_4688_pad_type_0"), val = tensor("valid")]; + tensor var_4688_strides_0 = const()[name = tensor("op_4688_strides_0"), val = tensor([1, 1])]; + tensor var_4688_pad_0 = const()[name = tensor("op_4688_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4688_dilations_0 = const()[name = tensor("op_4688_dilations_0"), val = tensor([1, 1])]; + tensor var_4688_groups_0 = const()[name = tensor("op_4688_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670554816)))]; + tensor blocks_16_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683662080)))]; + tensor var_4688_cast_fp16 = conv(bias = blocks_16_mlp_2_bias_to_fp16, dilations = var_4688_dilations_0, groups = var_4688_groups_0, pad = var_4688_pad_0, pad_type = var_4688_pad_type_0, strides = var_4688_strides_0, weight = blocks_16_mlp_2_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_4688_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4697 = const()[name = tensor("op_4697"), val = tensor(1)]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([1])]; + tensor input_173_gamma_0_to_fp16 = const()[name = tensor("input_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683664704)))]; + tensor input_173_beta_0_to_fp16 = const()[name = tensor("input_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683667328)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = input_173_beta_0_to_fp16, epsilon = var_4713_to_fp16, gamma = input_173_gamma_0_to_fp16, x = inputs_69_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor q_35_pad_type_0 = const()[name = tensor("q_35_pad_type_0"), val = tensor("valid")]; + tensor q_35_strides_0 = const()[name = tensor("q_35_strides_0"), val = tensor([1, 1])]; + tensor q_35_pad_0 = const()[name = tensor("q_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_35_dilations_0 = const()[name = tensor("q_35_dilations_0"), val = tensor([1, 1])]; + tensor q_35_groups_0 = const()[name = tensor("q_35_groups_0"), val = tensor(1)]; + tensor var_4748_weight_0_to_fp16 = const()[name = tensor("op_4748_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683669952)))]; + tensor var_4748_bias_0_to_fp16 = const()[name = tensor("op_4748_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686946816)))]; + tensor var_4748_cast_fp16 = conv(bias = var_4748_bias_0_to_fp16, dilations = q_35_dilations_0, groups = q_35_groups_0, pad = q_35_pad_0, pad_type = q_35_pad_type_0, strides = q_35_strides_0, weight = var_4748_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4748_cast_fp16")]; + tensor k_35_pad_type_0 = const()[name = tensor("k_35_pad_type_0"), val = tensor("valid")]; + tensor k_35_strides_0 = const()[name = tensor("k_35_strides_0"), val = tensor([1, 1])]; + tensor k_35_pad_0 = const()[name = tensor("k_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_35_dilations_0 = const()[name = tensor("k_35_dilations_0"), val = tensor([1, 1])]; + tensor k_35_groups_0 = const()[name = tensor("k_35_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_key_weight_to_fp16 = const()[name = tensor("blocks_17_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686949440)))]; + tensor k_35_cast_fp16 = conv(dilations = k_35_dilations_0, groups = k_35_groups_0, pad = k_35_pad_0, pad_type = k_35_pad_type_0, strides = k_35_strides_0, weight = blocks_17_attn_key_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_4746_pad_type_0 = const()[name = tensor("op_4746_pad_type_0"), val = tensor("valid")]; + tensor var_4746_strides_0 = const()[name = tensor("op_4746_strides_0"), val = tensor([1, 1])]; + tensor var_4746_pad_0 = const()[name = tensor("op_4746_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4746_dilations_0 = const()[name = tensor("op_4746_dilations_0"), val = tensor([1, 1])]; + tensor var_4746_groups_0 = const()[name = tensor("op_4746_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_value_weight_to_fp16 = const()[name = tensor("blocks_17_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(690226304)))]; + tensor blocks_17_attn_value_bias_to_fp16 = const()[name = tensor("blocks_17_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693503168)))]; + tensor var_4746_cast_fp16 = conv(bias = blocks_17_attn_value_bias_to_fp16, dilations = var_4746_dilations_0, groups = var_4746_groups_0, pad = var_4746_pad_0, pad_type = var_4746_pad_type_0, strides = var_4746_strides_0, weight = blocks_17_attn_value_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4746_cast_fp16")]; + tensor tile_51 = const()[name = tensor("tile_51"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4749_axis_0 = const()[name = tensor("op_4749_axis_0"), val = tensor(1)]; + tensor var_4749_cast_fp16_0, tensor var_4749_cast_fp16_1, tensor var_4749_cast_fp16_2, tensor var_4749_cast_fp16_3, tensor var_4749_cast_fp16_4, tensor var_4749_cast_fp16_5, tensor var_4749_cast_fp16_6, tensor var_4749_cast_fp16_7, tensor var_4749_cast_fp16_8, tensor var_4749_cast_fp16_9, tensor var_4749_cast_fp16_10, tensor var_4749_cast_fp16_11, tensor var_4749_cast_fp16_12, tensor var_4749_cast_fp16_13, tensor var_4749_cast_fp16_14, tensor var_4749_cast_fp16_15, tensor var_4749_cast_fp16_16, tensor var_4749_cast_fp16_17, tensor var_4749_cast_fp16_18, tensor var_4749_cast_fp16_19 = split(axis = var_4749_axis_0, split_sizes = tile_51, x = var_4748_cast_fp16)[name = tensor("op_4749_cast_fp16")]; + tensor var_4770_perm_0 = const()[name = tensor("op_4770_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_52 = const()[name = tensor("tile_52"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4771_axis_0 = const()[name = tensor("op_4771_axis_0"), val = tensor(3)]; + tensor var_4770_cast_fp16 = transpose(perm = var_4770_perm_0, x = k_35_cast_fp16)[name = tensor("transpose_15")]; + tensor var_4771_cast_fp16_0, tensor var_4771_cast_fp16_1, tensor var_4771_cast_fp16_2, tensor var_4771_cast_fp16_3, tensor var_4771_cast_fp16_4, tensor var_4771_cast_fp16_5, tensor var_4771_cast_fp16_6, tensor var_4771_cast_fp16_7, tensor var_4771_cast_fp16_8, tensor var_4771_cast_fp16_9, tensor var_4771_cast_fp16_10, tensor var_4771_cast_fp16_11, tensor var_4771_cast_fp16_12, tensor var_4771_cast_fp16_13, tensor var_4771_cast_fp16_14, tensor var_4771_cast_fp16_15, tensor var_4771_cast_fp16_16, tensor var_4771_cast_fp16_17, tensor var_4771_cast_fp16_18, tensor var_4771_cast_fp16_19 = split(axis = var_4771_axis_0, split_sizes = tile_52, x = var_4770_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor tile_53 = const()[name = tensor("tile_53"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4792_axis_0 = const()[name = tensor("op_4792_axis_0"), val = tensor(1)]; + tensor var_4792_cast_fp16_0, tensor var_4792_cast_fp16_1, tensor var_4792_cast_fp16_2, tensor var_4792_cast_fp16_3, tensor var_4792_cast_fp16_4, tensor var_4792_cast_fp16_5, tensor var_4792_cast_fp16_6, tensor var_4792_cast_fp16_7, tensor var_4792_cast_fp16_8, tensor var_4792_cast_fp16_9, tensor var_4792_cast_fp16_10, tensor var_4792_cast_fp16_11, tensor var_4792_cast_fp16_12, tensor var_4792_cast_fp16_13, tensor var_4792_cast_fp16_14, tensor var_4792_cast_fp16_15, tensor var_4792_cast_fp16_16, tensor var_4792_cast_fp16_17, tensor var_4792_cast_fp16_18, tensor var_4792_cast_fp16_19 = split(axis = var_4792_axis_0, split_sizes = tile_53, x = var_4746_cast_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor aw_681_equation_0 = const()[name = tensor("aw_681_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_681_cast_fp16 = einsum(equation = aw_681_equation_0, values = (var_4771_cast_fp16_0, var_4749_cast_fp16_0))[name = tensor("aw_681_cast_fp16")]; + tensor aw_683_equation_0 = const()[name = tensor("aw_683_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_683_cast_fp16 = einsum(equation = aw_683_equation_0, values = (var_4771_cast_fp16_1, var_4749_cast_fp16_1))[name = tensor("aw_683_cast_fp16")]; + tensor aw_685_equation_0 = const()[name = tensor("aw_685_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_685_cast_fp16 = einsum(equation = aw_685_equation_0, values = (var_4771_cast_fp16_2, var_4749_cast_fp16_2))[name = tensor("aw_685_cast_fp16")]; + tensor aw_687_equation_0 = const()[name = tensor("aw_687_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_687_cast_fp16 = einsum(equation = aw_687_equation_0, values = (var_4771_cast_fp16_3, var_4749_cast_fp16_3))[name = tensor("aw_687_cast_fp16")]; + tensor aw_689_equation_0 = const()[name = tensor("aw_689_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_689_cast_fp16 = einsum(equation = aw_689_equation_0, values = (var_4771_cast_fp16_4, var_4749_cast_fp16_4))[name = tensor("aw_689_cast_fp16")]; + tensor aw_691_equation_0 = const()[name = tensor("aw_691_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_691_cast_fp16 = einsum(equation = aw_691_equation_0, values = (var_4771_cast_fp16_5, var_4749_cast_fp16_5))[name = tensor("aw_691_cast_fp16")]; + tensor aw_693_equation_0 = const()[name = tensor("aw_693_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_693_cast_fp16 = einsum(equation = aw_693_equation_0, values = (var_4771_cast_fp16_6, var_4749_cast_fp16_6))[name = tensor("aw_693_cast_fp16")]; + tensor aw_695_equation_0 = const()[name = tensor("aw_695_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_695_cast_fp16 = einsum(equation = aw_695_equation_0, values = (var_4771_cast_fp16_7, var_4749_cast_fp16_7))[name = tensor("aw_695_cast_fp16")]; + tensor aw_697_equation_0 = const()[name = tensor("aw_697_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_697_cast_fp16 = einsum(equation = aw_697_equation_0, values = (var_4771_cast_fp16_8, var_4749_cast_fp16_8))[name = tensor("aw_697_cast_fp16")]; + tensor aw_699_equation_0 = const()[name = tensor("aw_699_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_699_cast_fp16 = einsum(equation = aw_699_equation_0, values = (var_4771_cast_fp16_9, var_4749_cast_fp16_9))[name = tensor("aw_699_cast_fp16")]; + tensor aw_701_equation_0 = const()[name = tensor("aw_701_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_701_cast_fp16 = einsum(equation = aw_701_equation_0, values = (var_4771_cast_fp16_10, var_4749_cast_fp16_10))[name = tensor("aw_701_cast_fp16")]; + tensor aw_703_equation_0 = const()[name = tensor("aw_703_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_703_cast_fp16 = einsum(equation = aw_703_equation_0, values = (var_4771_cast_fp16_11, var_4749_cast_fp16_11))[name = tensor("aw_703_cast_fp16")]; + tensor aw_705_equation_0 = const()[name = tensor("aw_705_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_705_cast_fp16 = einsum(equation = aw_705_equation_0, values = (var_4771_cast_fp16_12, var_4749_cast_fp16_12))[name = tensor("aw_705_cast_fp16")]; + tensor aw_707_equation_0 = const()[name = tensor("aw_707_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_707_cast_fp16 = einsum(equation = aw_707_equation_0, values = (var_4771_cast_fp16_13, var_4749_cast_fp16_13))[name = tensor("aw_707_cast_fp16")]; + tensor aw_709_equation_0 = const()[name = tensor("aw_709_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_709_cast_fp16 = einsum(equation = aw_709_equation_0, values = (var_4771_cast_fp16_14, var_4749_cast_fp16_14))[name = tensor("aw_709_cast_fp16")]; + tensor aw_711_equation_0 = const()[name = tensor("aw_711_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_711_cast_fp16 = einsum(equation = aw_711_equation_0, values = (var_4771_cast_fp16_15, var_4749_cast_fp16_15))[name = tensor("aw_711_cast_fp16")]; + tensor aw_713_equation_0 = const()[name = tensor("aw_713_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_713_cast_fp16 = einsum(equation = aw_713_equation_0, values = (var_4771_cast_fp16_16, var_4749_cast_fp16_16))[name = tensor("aw_713_cast_fp16")]; + tensor aw_715_equation_0 = const()[name = tensor("aw_715_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_715_cast_fp16 = einsum(equation = aw_715_equation_0, values = (var_4771_cast_fp16_17, var_4749_cast_fp16_17))[name = tensor("aw_715_cast_fp16")]; + tensor aw_717_equation_0 = const()[name = tensor("aw_717_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_717_cast_fp16 = einsum(equation = aw_717_equation_0, values = (var_4771_cast_fp16_18, var_4749_cast_fp16_18))[name = tensor("aw_717_cast_fp16")]; + tensor aw_719_equation_0 = const()[name = tensor("aw_719_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_719_cast_fp16 = einsum(equation = aw_719_equation_0, values = (var_4771_cast_fp16_19, var_4749_cast_fp16_19))[name = tensor("aw_719_cast_fp16")]; + tensor var_4853_cast_fp16 = softmax(axis = var_4697, x = aw_681_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = softmax(axis = var_4697, x = aw_683_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor var_4855_cast_fp16 = softmax(axis = var_4697, x = aw_685_cast_fp16)[name = tensor("op_4855_cast_fp16")]; + tensor var_4856_cast_fp16 = softmax(axis = var_4697, x = aw_687_cast_fp16)[name = tensor("op_4856_cast_fp16")]; + tensor var_4857_cast_fp16 = softmax(axis = var_4697, x = aw_689_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = softmax(axis = var_4697, x = aw_691_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor var_4859_cast_fp16 = softmax(axis = var_4697, x = aw_693_cast_fp16)[name = tensor("op_4859_cast_fp16")]; + tensor var_4860_cast_fp16 = softmax(axis = var_4697, x = aw_695_cast_fp16)[name = tensor("op_4860_cast_fp16")]; + tensor var_4861_cast_fp16 = softmax(axis = var_4697, x = aw_697_cast_fp16)[name = tensor("op_4861_cast_fp16")]; + tensor var_4862_cast_fp16 = softmax(axis = var_4697, x = aw_699_cast_fp16)[name = tensor("op_4862_cast_fp16")]; + tensor var_4863_cast_fp16 = softmax(axis = var_4697, x = aw_701_cast_fp16)[name = tensor("op_4863_cast_fp16")]; + tensor var_4864_cast_fp16 = softmax(axis = var_4697, x = aw_703_cast_fp16)[name = tensor("op_4864_cast_fp16")]; + tensor var_4865_cast_fp16 = softmax(axis = var_4697, x = aw_705_cast_fp16)[name = tensor("op_4865_cast_fp16")]; + tensor var_4866_cast_fp16 = softmax(axis = var_4697, x = aw_707_cast_fp16)[name = tensor("op_4866_cast_fp16")]; + tensor var_4867_cast_fp16 = softmax(axis = var_4697, x = aw_709_cast_fp16)[name = tensor("op_4867_cast_fp16")]; + tensor var_4868_cast_fp16 = softmax(axis = var_4697, x = aw_711_cast_fp16)[name = tensor("op_4868_cast_fp16")]; + tensor var_4869_cast_fp16 = softmax(axis = var_4697, x = aw_713_cast_fp16)[name = tensor("op_4869_cast_fp16")]; + tensor var_4870_cast_fp16 = softmax(axis = var_4697, x = aw_715_cast_fp16)[name = tensor("op_4870_cast_fp16")]; + tensor var_4871_cast_fp16 = softmax(axis = var_4697, x = aw_717_cast_fp16)[name = tensor("op_4871_cast_fp16")]; + tensor var_4872_cast_fp16 = softmax(axis = var_4697, x = aw_719_cast_fp16)[name = tensor("op_4872_cast_fp16")]; + tensor var_4874_equation_0 = const()[name = tensor("op_4874_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4874_cast_fp16 = einsum(equation = var_4874_equation_0, values = (var_4792_cast_fp16_0, var_4853_cast_fp16))[name = tensor("op_4874_cast_fp16")]; + tensor var_4876_equation_0 = const()[name = tensor("op_4876_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4876_cast_fp16 = einsum(equation = var_4876_equation_0, values = (var_4792_cast_fp16_1, var_4854_cast_fp16))[name = tensor("op_4876_cast_fp16")]; + tensor var_4878_equation_0 = const()[name = tensor("op_4878_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4878_cast_fp16 = einsum(equation = var_4878_equation_0, values = (var_4792_cast_fp16_2, var_4855_cast_fp16))[name = tensor("op_4878_cast_fp16")]; + tensor var_4880_equation_0 = const()[name = tensor("op_4880_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4880_cast_fp16 = einsum(equation = var_4880_equation_0, values = (var_4792_cast_fp16_3, var_4856_cast_fp16))[name = tensor("op_4880_cast_fp16")]; + tensor var_4882_equation_0 = const()[name = tensor("op_4882_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4882_cast_fp16 = einsum(equation = var_4882_equation_0, values = (var_4792_cast_fp16_4, var_4857_cast_fp16))[name = tensor("op_4882_cast_fp16")]; + tensor var_4884_equation_0 = const()[name = tensor("op_4884_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4884_cast_fp16 = einsum(equation = var_4884_equation_0, values = (var_4792_cast_fp16_5, var_4858_cast_fp16))[name = tensor("op_4884_cast_fp16")]; + tensor var_4886_equation_0 = const()[name = tensor("op_4886_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4886_cast_fp16 = einsum(equation = var_4886_equation_0, values = (var_4792_cast_fp16_6, var_4859_cast_fp16))[name = tensor("op_4886_cast_fp16")]; + tensor var_4888_equation_0 = const()[name = tensor("op_4888_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4888_cast_fp16 = einsum(equation = var_4888_equation_0, values = (var_4792_cast_fp16_7, var_4860_cast_fp16))[name = tensor("op_4888_cast_fp16")]; + tensor var_4890_equation_0 = const()[name = tensor("op_4890_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4890_cast_fp16 = einsum(equation = var_4890_equation_0, values = (var_4792_cast_fp16_8, var_4861_cast_fp16))[name = tensor("op_4890_cast_fp16")]; + tensor var_4892_equation_0 = const()[name = tensor("op_4892_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4892_cast_fp16 = einsum(equation = var_4892_equation_0, values = (var_4792_cast_fp16_9, var_4862_cast_fp16))[name = tensor("op_4892_cast_fp16")]; + tensor var_4894_equation_0 = const()[name = tensor("op_4894_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4894_cast_fp16 = einsum(equation = var_4894_equation_0, values = (var_4792_cast_fp16_10, var_4863_cast_fp16))[name = tensor("op_4894_cast_fp16")]; + tensor var_4896_equation_0 = const()[name = tensor("op_4896_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4896_cast_fp16 = einsum(equation = var_4896_equation_0, values = (var_4792_cast_fp16_11, var_4864_cast_fp16))[name = tensor("op_4896_cast_fp16")]; + tensor var_4898_equation_0 = const()[name = tensor("op_4898_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4898_cast_fp16 = einsum(equation = var_4898_equation_0, values = (var_4792_cast_fp16_12, var_4865_cast_fp16))[name = tensor("op_4898_cast_fp16")]; + tensor var_4900_equation_0 = const()[name = tensor("op_4900_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4900_cast_fp16 = einsum(equation = var_4900_equation_0, values = (var_4792_cast_fp16_13, var_4866_cast_fp16))[name = tensor("op_4900_cast_fp16")]; + tensor var_4902_equation_0 = const()[name = tensor("op_4902_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4902_cast_fp16 = einsum(equation = var_4902_equation_0, values = (var_4792_cast_fp16_14, var_4867_cast_fp16))[name = tensor("op_4902_cast_fp16")]; + tensor var_4904_equation_0 = const()[name = tensor("op_4904_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4904_cast_fp16 = einsum(equation = var_4904_equation_0, values = (var_4792_cast_fp16_15, var_4868_cast_fp16))[name = tensor("op_4904_cast_fp16")]; + tensor var_4906_equation_0 = const()[name = tensor("op_4906_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4906_cast_fp16 = einsum(equation = var_4906_equation_0, values = (var_4792_cast_fp16_16, var_4869_cast_fp16))[name = tensor("op_4906_cast_fp16")]; + tensor var_4908_equation_0 = const()[name = tensor("op_4908_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4908_cast_fp16 = einsum(equation = var_4908_equation_0, values = (var_4792_cast_fp16_17, var_4870_cast_fp16))[name = tensor("op_4908_cast_fp16")]; + tensor var_4910_equation_0 = const()[name = tensor("op_4910_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4910_cast_fp16 = einsum(equation = var_4910_equation_0, values = (var_4792_cast_fp16_18, var_4871_cast_fp16))[name = tensor("op_4910_cast_fp16")]; + tensor var_4912_equation_0 = const()[name = tensor("op_4912_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4912_cast_fp16 = einsum(equation = var_4912_equation_0, values = (var_4792_cast_fp16_19, var_4872_cast_fp16))[name = tensor("op_4912_cast_fp16")]; + tensor input_175_interleave_0 = const()[name = tensor("input_175_interleave_0"), val = tensor(false)]; + tensor input_175_cast_fp16 = concat(axis = var_4697, interleave = input_175_interleave_0, values = (var_4874_cast_fp16, var_4876_cast_fp16, var_4878_cast_fp16, var_4880_cast_fp16, var_4882_cast_fp16, var_4884_cast_fp16, var_4886_cast_fp16, var_4888_cast_fp16, var_4890_cast_fp16, var_4892_cast_fp16, var_4894_cast_fp16, var_4896_cast_fp16, var_4898_cast_fp16, var_4900_cast_fp16, var_4902_cast_fp16, var_4904_cast_fp16, var_4906_cast_fp16, var_4908_cast_fp16, var_4910_cast_fp16, var_4912_cast_fp16))[name = tensor("input_175_cast_fp16")]; + tensor var_4921_pad_type_0 = const()[name = tensor("op_4921_pad_type_0"), val = tensor("valid")]; + tensor var_4921_strides_0 = const()[name = tensor("op_4921_strides_0"), val = tensor([1, 1])]; + tensor var_4921_pad_0 = const()[name = tensor("op_4921_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4921_dilations_0 = const()[name = tensor("op_4921_dilations_0"), val = tensor([1, 1])]; + tensor var_4921_groups_0 = const()[name = tensor("op_4921_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_out_weight_to_fp16 = const()[name = tensor("blocks_17_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693505792)))]; + tensor blocks_17_attn_out_bias_to_fp16 = const()[name = tensor("blocks_17_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696782656)))]; + tensor var_4921_cast_fp16 = conv(bias = blocks_17_attn_out_bias_to_fp16, dilations = var_4921_dilations_0, groups = var_4921_groups_0, pad = var_4921_pad_0, pad_type = var_4921_pad_type_0, strides = var_4921_strides_0, weight = blocks_17_attn_out_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = var_4921_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([1])]; + tensor input_177_gamma_0_to_fp16 = const()[name = tensor("input_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696785280)))]; + tensor input_177_beta_0_to_fp16 = const()[name = tensor("input_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696787904)))]; + tensor var_4931_to_fp16 = const()[name = tensor("op_4931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = input_177_beta_0_to_fp16, epsilon = var_4931_to_fp16, gamma = input_177_gamma_0_to_fp16, x = inputs_71_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("valid")]; + tensor input_179_strides_0 = const()[name = tensor("input_179_strides_0"), val = tensor([1, 1])]; + tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_179_dilations_0 = const()[name = tensor("input_179_dilations_0"), val = tensor([1, 1])]; + tensor input_179_groups_0 = const()[name = tensor("input_179_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696790528)))]; + tensor blocks_17_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709897792)))]; + tensor input_179_cast_fp16 = conv(bias = blocks_17_mlp_0_bias_to_fp16, dilations = input_179_dilations_0, groups = input_179_groups_0, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = input_179_strides_0, weight = blocks_17_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4957_pad_type_0 = const()[name = tensor("op_4957_pad_type_0"), val = tensor("valid")]; + tensor var_4957_strides_0 = const()[name = tensor("op_4957_strides_0"), val = tensor([1, 1])]; + tensor var_4957_pad_0 = const()[name = tensor("op_4957_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4957_dilations_0 = const()[name = tensor("op_4957_dilations_0"), val = tensor([1, 1])]; + tensor var_4957_groups_0 = const()[name = tensor("op_4957_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709908096)))]; + tensor blocks_17_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723015360)))]; + tensor var_4957_cast_fp16 = conv(bias = blocks_17_mlp_2_bias_to_fp16, dilations = var_4957_dilations_0, groups = var_4957_groups_0, pad = var_4957_pad_0, pad_type = var_4957_pad_type_0, strides = var_4957_strides_0, weight = blocks_17_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("op_4957_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_4957_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4966 = const()[name = tensor("op_4966"), val = tensor(1)]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([1])]; + tensor input_183_gamma_0_to_fp16 = const()[name = tensor("input_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723017984)))]; + tensor input_183_beta_0_to_fp16 = const()[name = tensor("input_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723020608)))]; + tensor var_4982_to_fp16 = const()[name = tensor("op_4982_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = input_183_beta_0_to_fp16, epsilon = var_4982_to_fp16, gamma = input_183_gamma_0_to_fp16, x = inputs_73_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor q_37_pad_type_0 = const()[name = tensor("q_37_pad_type_0"), val = tensor("valid")]; + tensor q_37_strides_0 = const()[name = tensor("q_37_strides_0"), val = tensor([1, 1])]; + tensor q_37_pad_0 = const()[name = tensor("q_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_37_dilations_0 = const()[name = tensor("q_37_dilations_0"), val = tensor([1, 1])]; + tensor q_37_groups_0 = const()[name = tensor("q_37_groups_0"), val = tensor(1)]; + tensor var_5017_weight_0_to_fp16 = const()[name = tensor("op_5017_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723023232)))]; + tensor var_5017_bias_0_to_fp16 = const()[name = tensor("op_5017_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726300096)))]; + tensor var_5017_cast_fp16 = conv(bias = var_5017_bias_0_to_fp16, dilations = q_37_dilations_0, groups = q_37_groups_0, pad = q_37_pad_0, pad_type = q_37_pad_type_0, strides = q_37_strides_0, weight = var_5017_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("op_5017_cast_fp16")]; + tensor k_37_pad_type_0 = const()[name = tensor("k_37_pad_type_0"), val = tensor("valid")]; + tensor k_37_strides_0 = const()[name = tensor("k_37_strides_0"), val = tensor([1, 1])]; + tensor k_37_pad_0 = const()[name = tensor("k_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_37_dilations_0 = const()[name = tensor("k_37_dilations_0"), val = tensor([1, 1])]; + tensor k_37_groups_0 = const()[name = tensor("k_37_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_key_weight_to_fp16 = const()[name = tensor("blocks_18_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726302720)))]; + tensor k_37_cast_fp16 = conv(dilations = k_37_dilations_0, groups = k_37_groups_0, pad = k_37_pad_0, pad_type = k_37_pad_type_0, strides = k_37_strides_0, weight = blocks_18_attn_key_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor var_5015_pad_type_0 = const()[name = tensor("op_5015_pad_type_0"), val = tensor("valid")]; + tensor var_5015_strides_0 = const()[name = tensor("op_5015_strides_0"), val = tensor([1, 1])]; + tensor var_5015_pad_0 = const()[name = tensor("op_5015_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5015_dilations_0 = const()[name = tensor("op_5015_dilations_0"), val = tensor([1, 1])]; + tensor var_5015_groups_0 = const()[name = tensor("op_5015_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_value_weight_to_fp16 = const()[name = tensor("blocks_18_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729579584)))]; + tensor blocks_18_attn_value_bias_to_fp16 = const()[name = tensor("blocks_18_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732856448)))]; + tensor var_5015_cast_fp16 = conv(bias = blocks_18_attn_value_bias_to_fp16, dilations = var_5015_dilations_0, groups = var_5015_groups_0, pad = var_5015_pad_0, pad_type = var_5015_pad_type_0, strides = var_5015_strides_0, weight = blocks_18_attn_value_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor tile_54 = const()[name = tensor("tile_54"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5018_axis_0 = const()[name = tensor("op_5018_axis_0"), val = tensor(1)]; + tensor var_5018_cast_fp16_0, tensor var_5018_cast_fp16_1, tensor var_5018_cast_fp16_2, tensor var_5018_cast_fp16_3, tensor var_5018_cast_fp16_4, tensor var_5018_cast_fp16_5, tensor var_5018_cast_fp16_6, tensor var_5018_cast_fp16_7, tensor var_5018_cast_fp16_8, tensor var_5018_cast_fp16_9, tensor var_5018_cast_fp16_10, tensor var_5018_cast_fp16_11, tensor var_5018_cast_fp16_12, tensor var_5018_cast_fp16_13, tensor var_5018_cast_fp16_14, tensor var_5018_cast_fp16_15, tensor var_5018_cast_fp16_16, tensor var_5018_cast_fp16_17, tensor var_5018_cast_fp16_18, tensor var_5018_cast_fp16_19 = split(axis = var_5018_axis_0, split_sizes = tile_54, x = var_5017_cast_fp16)[name = tensor("op_5018_cast_fp16")]; + tensor var_5039_perm_0 = const()[name = tensor("op_5039_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_55 = const()[name = tensor("tile_55"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5040_axis_0 = const()[name = tensor("op_5040_axis_0"), val = tensor(3)]; + tensor var_5039_cast_fp16 = transpose(perm = var_5039_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_14")]; + tensor var_5040_cast_fp16_0, tensor var_5040_cast_fp16_1, tensor var_5040_cast_fp16_2, tensor var_5040_cast_fp16_3, tensor var_5040_cast_fp16_4, tensor var_5040_cast_fp16_5, tensor var_5040_cast_fp16_6, tensor var_5040_cast_fp16_7, tensor var_5040_cast_fp16_8, tensor var_5040_cast_fp16_9, tensor var_5040_cast_fp16_10, tensor var_5040_cast_fp16_11, tensor var_5040_cast_fp16_12, tensor var_5040_cast_fp16_13, tensor var_5040_cast_fp16_14, tensor var_5040_cast_fp16_15, tensor var_5040_cast_fp16_16, tensor var_5040_cast_fp16_17, tensor var_5040_cast_fp16_18, tensor var_5040_cast_fp16_19 = split(axis = var_5040_axis_0, split_sizes = tile_55, x = var_5039_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor tile_56 = const()[name = tensor("tile_56"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5061_axis_0 = const()[name = tensor("op_5061_axis_0"), val = tensor(1)]; + tensor var_5061_cast_fp16_0, tensor var_5061_cast_fp16_1, tensor var_5061_cast_fp16_2, tensor var_5061_cast_fp16_3, tensor var_5061_cast_fp16_4, tensor var_5061_cast_fp16_5, tensor var_5061_cast_fp16_6, tensor var_5061_cast_fp16_7, tensor var_5061_cast_fp16_8, tensor var_5061_cast_fp16_9, tensor var_5061_cast_fp16_10, tensor var_5061_cast_fp16_11, tensor var_5061_cast_fp16_12, tensor var_5061_cast_fp16_13, tensor var_5061_cast_fp16_14, tensor var_5061_cast_fp16_15, tensor var_5061_cast_fp16_16, tensor var_5061_cast_fp16_17, tensor var_5061_cast_fp16_18, tensor var_5061_cast_fp16_19 = split(axis = var_5061_axis_0, split_sizes = tile_56, x = var_5015_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor aw_721_equation_0 = const()[name = tensor("aw_721_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_721_cast_fp16 = einsum(equation = aw_721_equation_0, values = (var_5040_cast_fp16_0, var_5018_cast_fp16_0))[name = tensor("aw_721_cast_fp16")]; + tensor aw_723_equation_0 = const()[name = tensor("aw_723_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_723_cast_fp16 = einsum(equation = aw_723_equation_0, values = (var_5040_cast_fp16_1, var_5018_cast_fp16_1))[name = tensor("aw_723_cast_fp16")]; + tensor aw_725_equation_0 = const()[name = tensor("aw_725_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_725_cast_fp16 = einsum(equation = aw_725_equation_0, values = (var_5040_cast_fp16_2, var_5018_cast_fp16_2))[name = tensor("aw_725_cast_fp16")]; + tensor aw_727_equation_0 = const()[name = tensor("aw_727_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_727_cast_fp16 = einsum(equation = aw_727_equation_0, values = (var_5040_cast_fp16_3, var_5018_cast_fp16_3))[name = tensor("aw_727_cast_fp16")]; + tensor aw_729_equation_0 = const()[name = tensor("aw_729_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_729_cast_fp16 = einsum(equation = aw_729_equation_0, values = (var_5040_cast_fp16_4, var_5018_cast_fp16_4))[name = tensor("aw_729_cast_fp16")]; + tensor aw_731_equation_0 = const()[name = tensor("aw_731_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_731_cast_fp16 = einsum(equation = aw_731_equation_0, values = (var_5040_cast_fp16_5, var_5018_cast_fp16_5))[name = tensor("aw_731_cast_fp16")]; + tensor aw_733_equation_0 = const()[name = tensor("aw_733_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_733_cast_fp16 = einsum(equation = aw_733_equation_0, values = (var_5040_cast_fp16_6, var_5018_cast_fp16_6))[name = tensor("aw_733_cast_fp16")]; + tensor aw_735_equation_0 = const()[name = tensor("aw_735_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_735_cast_fp16 = einsum(equation = aw_735_equation_0, values = (var_5040_cast_fp16_7, var_5018_cast_fp16_7))[name = tensor("aw_735_cast_fp16")]; + tensor aw_737_equation_0 = const()[name = tensor("aw_737_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_737_cast_fp16 = einsum(equation = aw_737_equation_0, values = (var_5040_cast_fp16_8, var_5018_cast_fp16_8))[name = tensor("aw_737_cast_fp16")]; + tensor aw_739_equation_0 = const()[name = tensor("aw_739_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_739_cast_fp16 = einsum(equation = aw_739_equation_0, values = (var_5040_cast_fp16_9, var_5018_cast_fp16_9))[name = tensor("aw_739_cast_fp16")]; + tensor aw_741_equation_0 = const()[name = tensor("aw_741_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_741_cast_fp16 = einsum(equation = aw_741_equation_0, values = (var_5040_cast_fp16_10, var_5018_cast_fp16_10))[name = tensor("aw_741_cast_fp16")]; + tensor aw_743_equation_0 = const()[name = tensor("aw_743_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_743_cast_fp16 = einsum(equation = aw_743_equation_0, values = (var_5040_cast_fp16_11, var_5018_cast_fp16_11))[name = tensor("aw_743_cast_fp16")]; + tensor aw_745_equation_0 = const()[name = tensor("aw_745_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_745_cast_fp16 = einsum(equation = aw_745_equation_0, values = (var_5040_cast_fp16_12, var_5018_cast_fp16_12))[name = tensor("aw_745_cast_fp16")]; + tensor aw_747_equation_0 = const()[name = tensor("aw_747_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_747_cast_fp16 = einsum(equation = aw_747_equation_0, values = (var_5040_cast_fp16_13, var_5018_cast_fp16_13))[name = tensor("aw_747_cast_fp16")]; + tensor aw_749_equation_0 = const()[name = tensor("aw_749_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_749_cast_fp16 = einsum(equation = aw_749_equation_0, values = (var_5040_cast_fp16_14, var_5018_cast_fp16_14))[name = tensor("aw_749_cast_fp16")]; + tensor aw_751_equation_0 = const()[name = tensor("aw_751_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_751_cast_fp16 = einsum(equation = aw_751_equation_0, values = (var_5040_cast_fp16_15, var_5018_cast_fp16_15))[name = tensor("aw_751_cast_fp16")]; + tensor aw_753_equation_0 = const()[name = tensor("aw_753_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_753_cast_fp16 = einsum(equation = aw_753_equation_0, values = (var_5040_cast_fp16_16, var_5018_cast_fp16_16))[name = tensor("aw_753_cast_fp16")]; + tensor aw_755_equation_0 = const()[name = tensor("aw_755_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_755_cast_fp16 = einsum(equation = aw_755_equation_0, values = (var_5040_cast_fp16_17, var_5018_cast_fp16_17))[name = tensor("aw_755_cast_fp16")]; + tensor aw_757_equation_0 = const()[name = tensor("aw_757_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_757_cast_fp16 = einsum(equation = aw_757_equation_0, values = (var_5040_cast_fp16_18, var_5018_cast_fp16_18))[name = tensor("aw_757_cast_fp16")]; + tensor aw_759_equation_0 = const()[name = tensor("aw_759_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_759_cast_fp16 = einsum(equation = aw_759_equation_0, values = (var_5040_cast_fp16_19, var_5018_cast_fp16_19))[name = tensor("aw_759_cast_fp16")]; + tensor var_5122_cast_fp16 = softmax(axis = var_4966, x = aw_721_cast_fp16)[name = tensor("op_5122_cast_fp16")]; + tensor var_5123_cast_fp16 = softmax(axis = var_4966, x = aw_723_cast_fp16)[name = tensor("op_5123_cast_fp16")]; + tensor var_5124_cast_fp16 = softmax(axis = var_4966, x = aw_725_cast_fp16)[name = tensor("op_5124_cast_fp16")]; + tensor var_5125_cast_fp16 = softmax(axis = var_4966, x = aw_727_cast_fp16)[name = tensor("op_5125_cast_fp16")]; + tensor var_5126_cast_fp16 = softmax(axis = var_4966, x = aw_729_cast_fp16)[name = tensor("op_5126_cast_fp16")]; + tensor var_5127_cast_fp16 = softmax(axis = var_4966, x = aw_731_cast_fp16)[name = tensor("op_5127_cast_fp16")]; + tensor var_5128_cast_fp16 = softmax(axis = var_4966, x = aw_733_cast_fp16)[name = tensor("op_5128_cast_fp16")]; + tensor var_5129_cast_fp16 = softmax(axis = var_4966, x = aw_735_cast_fp16)[name = tensor("op_5129_cast_fp16")]; + tensor var_5130_cast_fp16 = softmax(axis = var_4966, x = aw_737_cast_fp16)[name = tensor("op_5130_cast_fp16")]; + tensor var_5131_cast_fp16 = softmax(axis = var_4966, x = aw_739_cast_fp16)[name = tensor("op_5131_cast_fp16")]; + tensor var_5132_cast_fp16 = softmax(axis = var_4966, x = aw_741_cast_fp16)[name = tensor("op_5132_cast_fp16")]; + tensor var_5133_cast_fp16 = softmax(axis = var_4966, x = aw_743_cast_fp16)[name = tensor("op_5133_cast_fp16")]; + tensor var_5134_cast_fp16 = softmax(axis = var_4966, x = aw_745_cast_fp16)[name = tensor("op_5134_cast_fp16")]; + tensor var_5135_cast_fp16 = softmax(axis = var_4966, x = aw_747_cast_fp16)[name = tensor("op_5135_cast_fp16")]; + tensor var_5136_cast_fp16 = softmax(axis = var_4966, x = aw_749_cast_fp16)[name = tensor("op_5136_cast_fp16")]; + tensor var_5137_cast_fp16 = softmax(axis = var_4966, x = aw_751_cast_fp16)[name = tensor("op_5137_cast_fp16")]; + tensor var_5138_cast_fp16 = softmax(axis = var_4966, x = aw_753_cast_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5139_cast_fp16 = softmax(axis = var_4966, x = aw_755_cast_fp16)[name = tensor("op_5139_cast_fp16")]; + tensor var_5140_cast_fp16 = softmax(axis = var_4966, x = aw_757_cast_fp16)[name = tensor("op_5140_cast_fp16")]; + tensor var_5141_cast_fp16 = softmax(axis = var_4966, x = aw_759_cast_fp16)[name = tensor("op_5141_cast_fp16")]; + tensor var_5143_equation_0 = const()[name = tensor("op_5143_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5143_cast_fp16 = einsum(equation = var_5143_equation_0, values = (var_5061_cast_fp16_0, var_5122_cast_fp16))[name = tensor("op_5143_cast_fp16")]; + tensor var_5145_equation_0 = const()[name = tensor("op_5145_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5145_cast_fp16 = einsum(equation = var_5145_equation_0, values = (var_5061_cast_fp16_1, var_5123_cast_fp16))[name = tensor("op_5145_cast_fp16")]; + tensor var_5147_equation_0 = const()[name = tensor("op_5147_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5147_cast_fp16 = einsum(equation = var_5147_equation_0, values = (var_5061_cast_fp16_2, var_5124_cast_fp16))[name = tensor("op_5147_cast_fp16")]; + tensor var_5149_equation_0 = const()[name = tensor("op_5149_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5149_cast_fp16 = einsum(equation = var_5149_equation_0, values = (var_5061_cast_fp16_3, var_5125_cast_fp16))[name = tensor("op_5149_cast_fp16")]; + tensor var_5151_equation_0 = const()[name = tensor("op_5151_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5151_cast_fp16 = einsum(equation = var_5151_equation_0, values = (var_5061_cast_fp16_4, var_5126_cast_fp16))[name = tensor("op_5151_cast_fp16")]; + tensor var_5153_equation_0 = const()[name = tensor("op_5153_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5153_cast_fp16 = einsum(equation = var_5153_equation_0, values = (var_5061_cast_fp16_5, var_5127_cast_fp16))[name = tensor("op_5153_cast_fp16")]; + tensor var_5155_equation_0 = const()[name = tensor("op_5155_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5155_cast_fp16 = einsum(equation = var_5155_equation_0, values = (var_5061_cast_fp16_6, var_5128_cast_fp16))[name = tensor("op_5155_cast_fp16")]; + tensor var_5157_equation_0 = const()[name = tensor("op_5157_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5157_cast_fp16 = einsum(equation = var_5157_equation_0, values = (var_5061_cast_fp16_7, var_5129_cast_fp16))[name = tensor("op_5157_cast_fp16")]; + tensor var_5159_equation_0 = const()[name = tensor("op_5159_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5159_cast_fp16 = einsum(equation = var_5159_equation_0, values = (var_5061_cast_fp16_8, var_5130_cast_fp16))[name = tensor("op_5159_cast_fp16")]; + tensor var_5161_equation_0 = const()[name = tensor("op_5161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5161_cast_fp16 = einsum(equation = var_5161_equation_0, values = (var_5061_cast_fp16_9, var_5131_cast_fp16))[name = tensor("op_5161_cast_fp16")]; + tensor var_5163_equation_0 = const()[name = tensor("op_5163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5163_cast_fp16 = einsum(equation = var_5163_equation_0, values = (var_5061_cast_fp16_10, var_5132_cast_fp16))[name = tensor("op_5163_cast_fp16")]; + tensor var_5165_equation_0 = const()[name = tensor("op_5165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5165_cast_fp16 = einsum(equation = var_5165_equation_0, values = (var_5061_cast_fp16_11, var_5133_cast_fp16))[name = tensor("op_5165_cast_fp16")]; + tensor var_5167_equation_0 = const()[name = tensor("op_5167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5167_cast_fp16 = einsum(equation = var_5167_equation_0, values = (var_5061_cast_fp16_12, var_5134_cast_fp16))[name = tensor("op_5167_cast_fp16")]; + tensor var_5169_equation_0 = const()[name = tensor("op_5169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5169_cast_fp16 = einsum(equation = var_5169_equation_0, values = (var_5061_cast_fp16_13, var_5135_cast_fp16))[name = tensor("op_5169_cast_fp16")]; + tensor var_5171_equation_0 = const()[name = tensor("op_5171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5171_cast_fp16 = einsum(equation = var_5171_equation_0, values = (var_5061_cast_fp16_14, var_5136_cast_fp16))[name = tensor("op_5171_cast_fp16")]; + tensor var_5173_equation_0 = const()[name = tensor("op_5173_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5173_cast_fp16 = einsum(equation = var_5173_equation_0, values = (var_5061_cast_fp16_15, var_5137_cast_fp16))[name = tensor("op_5173_cast_fp16")]; + tensor var_5175_equation_0 = const()[name = tensor("op_5175_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5175_cast_fp16 = einsum(equation = var_5175_equation_0, values = (var_5061_cast_fp16_16, var_5138_cast_fp16))[name = tensor("op_5175_cast_fp16")]; + tensor var_5177_equation_0 = const()[name = tensor("op_5177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5177_cast_fp16 = einsum(equation = var_5177_equation_0, values = (var_5061_cast_fp16_17, var_5139_cast_fp16))[name = tensor("op_5177_cast_fp16")]; + tensor var_5179_equation_0 = const()[name = tensor("op_5179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5179_cast_fp16 = einsum(equation = var_5179_equation_0, values = (var_5061_cast_fp16_18, var_5140_cast_fp16))[name = tensor("op_5179_cast_fp16")]; + tensor var_5181_equation_0 = const()[name = tensor("op_5181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5181_cast_fp16 = einsum(equation = var_5181_equation_0, values = (var_5061_cast_fp16_19, var_5141_cast_fp16))[name = tensor("op_5181_cast_fp16")]; + tensor input_185_interleave_0 = const()[name = tensor("input_185_interleave_0"), val = tensor(false)]; + tensor input_185_cast_fp16 = concat(axis = var_4966, interleave = input_185_interleave_0, values = (var_5143_cast_fp16, var_5145_cast_fp16, var_5147_cast_fp16, var_5149_cast_fp16, var_5151_cast_fp16, var_5153_cast_fp16, var_5155_cast_fp16, var_5157_cast_fp16, var_5159_cast_fp16, var_5161_cast_fp16, var_5163_cast_fp16, var_5165_cast_fp16, var_5167_cast_fp16, var_5169_cast_fp16, var_5171_cast_fp16, var_5173_cast_fp16, var_5175_cast_fp16, var_5177_cast_fp16, var_5179_cast_fp16, var_5181_cast_fp16))[name = tensor("input_185_cast_fp16")]; + tensor var_5190_pad_type_0 = const()[name = tensor("op_5190_pad_type_0"), val = tensor("valid")]; + tensor var_5190_strides_0 = const()[name = tensor("op_5190_strides_0"), val = tensor([1, 1])]; + tensor var_5190_pad_0 = const()[name = tensor("op_5190_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5190_dilations_0 = const()[name = tensor("op_5190_dilations_0"), val = tensor([1, 1])]; + tensor var_5190_groups_0 = const()[name = tensor("op_5190_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_out_weight_to_fp16 = const()[name = tensor("blocks_18_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732859072)))]; + tensor blocks_18_attn_out_bias_to_fp16 = const()[name = tensor("blocks_18_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736135936)))]; + tensor var_5190_cast_fp16 = conv(bias = blocks_18_attn_out_bias_to_fp16, dilations = var_5190_dilations_0, groups = var_5190_groups_0, pad = var_5190_pad_0, pad_type = var_5190_pad_type_0, strides = var_5190_strides_0, weight = blocks_18_attn_out_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("op_5190_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = var_5190_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor input_187_axes_0 = const()[name = tensor("input_187_axes_0"), val = tensor([1])]; + tensor input_187_gamma_0_to_fp16 = const()[name = tensor("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736138560)))]; + tensor input_187_beta_0_to_fp16 = const()[name = tensor("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736141184)))]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = input_187_beta_0_to_fp16, epsilon = var_5200_to_fp16, gamma = input_187_gamma_0_to_fp16, x = inputs_75_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("valid")]; + tensor input_189_strides_0 = const()[name = tensor("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = tensor("input_189_dilations_0"), val = tensor([1, 1])]; + tensor input_189_groups_0 = const()[name = tensor("input_189_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736143808)))]; + tensor blocks_18_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749251072)))]; + tensor input_189_cast_fp16 = conv(bias = blocks_18_mlp_0_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = blocks_18_mlp_0_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_5226_pad_type_0 = const()[name = tensor("op_5226_pad_type_0"), val = tensor("valid")]; + tensor var_5226_strides_0 = const()[name = tensor("op_5226_strides_0"), val = tensor([1, 1])]; + tensor var_5226_pad_0 = const()[name = tensor("op_5226_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5226_dilations_0 = const()[name = tensor("op_5226_dilations_0"), val = tensor([1, 1])]; + tensor var_5226_groups_0 = const()[name = tensor("op_5226_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749261376)))]; + tensor blocks_18_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762368640)))]; + tensor var_5226_cast_fp16 = conv(bias = blocks_18_mlp_2_bias_to_fp16, dilations = var_5226_dilations_0, groups = var_5226_groups_0, pad = var_5226_pad_0, pad_type = var_5226_pad_type_0, strides = var_5226_strides_0, weight = blocks_18_mlp_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("op_5226_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = var_5226_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor(1)]; + tensor input_193_axes_0 = const()[name = tensor("input_193_axes_0"), val = tensor([1])]; + tensor input_193_gamma_0_to_fp16 = const()[name = tensor("input_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762371264)))]; + tensor input_193_beta_0_to_fp16 = const()[name = tensor("input_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762373888)))]; + tensor var_5251_to_fp16 = const()[name = tensor("op_5251_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = input_193_beta_0_to_fp16, epsilon = var_5251_to_fp16, gamma = input_193_gamma_0_to_fp16, x = inputs_77_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor q_39_pad_type_0 = const()[name = tensor("q_39_pad_type_0"), val = tensor("valid")]; + tensor q_39_strides_0 = const()[name = tensor("q_39_strides_0"), val = tensor([1, 1])]; + tensor q_39_pad_0 = const()[name = tensor("q_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_39_dilations_0 = const()[name = tensor("q_39_dilations_0"), val = tensor([1, 1])]; + tensor q_39_groups_0 = const()[name = tensor("q_39_groups_0"), val = tensor(1)]; + tensor var_5286_weight_0_to_fp16 = const()[name = tensor("op_5286_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762376512)))]; + tensor var_5286_bias_0_to_fp16 = const()[name = tensor("op_5286_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765653376)))]; + tensor var_5286_cast_fp16 = conv(bias = var_5286_bias_0_to_fp16, dilations = q_39_dilations_0, groups = q_39_groups_0, pad = q_39_pad_0, pad_type = q_39_pad_type_0, strides = q_39_strides_0, weight = var_5286_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor k_39_pad_type_0 = const()[name = tensor("k_39_pad_type_0"), val = tensor("valid")]; + tensor k_39_strides_0 = const()[name = tensor("k_39_strides_0"), val = tensor([1, 1])]; + tensor k_39_pad_0 = const()[name = tensor("k_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_39_dilations_0 = const()[name = tensor("k_39_dilations_0"), val = tensor([1, 1])]; + tensor k_39_groups_0 = const()[name = tensor("k_39_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_key_weight_to_fp16 = const()[name = tensor("blocks_19_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765656000)))]; + tensor k_39_cast_fp16 = conv(dilations = k_39_dilations_0, groups = k_39_groups_0, pad = k_39_pad_0, pad_type = k_39_pad_type_0, strides = k_39_strides_0, weight = blocks_19_attn_key_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor var_5284_pad_type_0 = const()[name = tensor("op_5284_pad_type_0"), val = tensor("valid")]; + tensor var_5284_strides_0 = const()[name = tensor("op_5284_strides_0"), val = tensor([1, 1])]; + tensor var_5284_pad_0 = const()[name = tensor("op_5284_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5284_dilations_0 = const()[name = tensor("op_5284_dilations_0"), val = tensor([1, 1])]; + tensor var_5284_groups_0 = const()[name = tensor("op_5284_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_value_weight_to_fp16 = const()[name = tensor("blocks_19_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768932864)))]; + tensor blocks_19_attn_value_bias_to_fp16 = const()[name = tensor("blocks_19_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(772209728)))]; + tensor var_5284_cast_fp16 = conv(bias = blocks_19_attn_value_bias_to_fp16, dilations = var_5284_dilations_0, groups = var_5284_groups_0, pad = var_5284_pad_0, pad_type = var_5284_pad_type_0, strides = var_5284_strides_0, weight = blocks_19_attn_value_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor tile_57 = const()[name = tensor("tile_57"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5287_axis_0 = const()[name = tensor("op_5287_axis_0"), val = tensor(1)]; + tensor var_5287_cast_fp16_0, tensor var_5287_cast_fp16_1, tensor var_5287_cast_fp16_2, tensor var_5287_cast_fp16_3, tensor var_5287_cast_fp16_4, tensor var_5287_cast_fp16_5, tensor var_5287_cast_fp16_6, tensor var_5287_cast_fp16_7, tensor var_5287_cast_fp16_8, tensor var_5287_cast_fp16_9, tensor var_5287_cast_fp16_10, tensor var_5287_cast_fp16_11, tensor var_5287_cast_fp16_12, tensor var_5287_cast_fp16_13, tensor var_5287_cast_fp16_14, tensor var_5287_cast_fp16_15, tensor var_5287_cast_fp16_16, tensor var_5287_cast_fp16_17, tensor var_5287_cast_fp16_18, tensor var_5287_cast_fp16_19 = split(axis = var_5287_axis_0, split_sizes = tile_57, x = var_5286_cast_fp16)[name = tensor("op_5287_cast_fp16")]; + tensor var_5308_perm_0 = const()[name = tensor("op_5308_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_58 = const()[name = tensor("tile_58"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5309_axis_0 = const()[name = tensor("op_5309_axis_0"), val = tensor(3)]; + tensor var_5308_cast_fp16 = transpose(perm = var_5308_perm_0, x = k_39_cast_fp16)[name = tensor("transpose_13")]; + tensor var_5309_cast_fp16_0, tensor var_5309_cast_fp16_1, tensor var_5309_cast_fp16_2, tensor var_5309_cast_fp16_3, tensor var_5309_cast_fp16_4, tensor var_5309_cast_fp16_5, tensor var_5309_cast_fp16_6, tensor var_5309_cast_fp16_7, tensor var_5309_cast_fp16_8, tensor var_5309_cast_fp16_9, tensor var_5309_cast_fp16_10, tensor var_5309_cast_fp16_11, tensor var_5309_cast_fp16_12, tensor var_5309_cast_fp16_13, tensor var_5309_cast_fp16_14, tensor var_5309_cast_fp16_15, tensor var_5309_cast_fp16_16, tensor var_5309_cast_fp16_17, tensor var_5309_cast_fp16_18, tensor var_5309_cast_fp16_19 = split(axis = var_5309_axis_0, split_sizes = tile_58, x = var_5308_cast_fp16)[name = tensor("op_5309_cast_fp16")]; + tensor tile_59 = const()[name = tensor("tile_59"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5330_axis_0 = const()[name = tensor("op_5330_axis_0"), val = tensor(1)]; + tensor var_5330_cast_fp16_0, tensor var_5330_cast_fp16_1, tensor var_5330_cast_fp16_2, tensor var_5330_cast_fp16_3, tensor var_5330_cast_fp16_4, tensor var_5330_cast_fp16_5, tensor var_5330_cast_fp16_6, tensor var_5330_cast_fp16_7, tensor var_5330_cast_fp16_8, tensor var_5330_cast_fp16_9, tensor var_5330_cast_fp16_10, tensor var_5330_cast_fp16_11, tensor var_5330_cast_fp16_12, tensor var_5330_cast_fp16_13, tensor var_5330_cast_fp16_14, tensor var_5330_cast_fp16_15, tensor var_5330_cast_fp16_16, tensor var_5330_cast_fp16_17, tensor var_5330_cast_fp16_18, tensor var_5330_cast_fp16_19 = split(axis = var_5330_axis_0, split_sizes = tile_59, x = var_5284_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor aw_761_equation_0 = const()[name = tensor("aw_761_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_761_cast_fp16 = einsum(equation = aw_761_equation_0, values = (var_5309_cast_fp16_0, var_5287_cast_fp16_0))[name = tensor("aw_761_cast_fp16")]; + tensor aw_763_equation_0 = const()[name = tensor("aw_763_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_763_cast_fp16 = einsum(equation = aw_763_equation_0, values = (var_5309_cast_fp16_1, var_5287_cast_fp16_1))[name = tensor("aw_763_cast_fp16")]; + tensor aw_765_equation_0 = const()[name = tensor("aw_765_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_765_cast_fp16 = einsum(equation = aw_765_equation_0, values = (var_5309_cast_fp16_2, var_5287_cast_fp16_2))[name = tensor("aw_765_cast_fp16")]; + tensor aw_767_equation_0 = const()[name = tensor("aw_767_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_767_cast_fp16 = einsum(equation = aw_767_equation_0, values = (var_5309_cast_fp16_3, var_5287_cast_fp16_3))[name = tensor("aw_767_cast_fp16")]; + tensor aw_769_equation_0 = const()[name = tensor("aw_769_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_769_cast_fp16 = einsum(equation = aw_769_equation_0, values = (var_5309_cast_fp16_4, var_5287_cast_fp16_4))[name = tensor("aw_769_cast_fp16")]; + tensor aw_771_equation_0 = const()[name = tensor("aw_771_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_771_cast_fp16 = einsum(equation = aw_771_equation_0, values = (var_5309_cast_fp16_5, var_5287_cast_fp16_5))[name = tensor("aw_771_cast_fp16")]; + tensor aw_773_equation_0 = const()[name = tensor("aw_773_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_773_cast_fp16 = einsum(equation = aw_773_equation_0, values = (var_5309_cast_fp16_6, var_5287_cast_fp16_6))[name = tensor("aw_773_cast_fp16")]; + tensor aw_775_equation_0 = const()[name = tensor("aw_775_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_775_cast_fp16 = einsum(equation = aw_775_equation_0, values = (var_5309_cast_fp16_7, var_5287_cast_fp16_7))[name = tensor("aw_775_cast_fp16")]; + tensor aw_777_equation_0 = const()[name = tensor("aw_777_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_777_cast_fp16 = einsum(equation = aw_777_equation_0, values = (var_5309_cast_fp16_8, var_5287_cast_fp16_8))[name = tensor("aw_777_cast_fp16")]; + tensor aw_779_equation_0 = const()[name = tensor("aw_779_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_779_cast_fp16 = einsum(equation = aw_779_equation_0, values = (var_5309_cast_fp16_9, var_5287_cast_fp16_9))[name = tensor("aw_779_cast_fp16")]; + tensor aw_781_equation_0 = const()[name = tensor("aw_781_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_781_cast_fp16 = einsum(equation = aw_781_equation_0, values = (var_5309_cast_fp16_10, var_5287_cast_fp16_10))[name = tensor("aw_781_cast_fp16")]; + tensor aw_783_equation_0 = const()[name = tensor("aw_783_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_783_cast_fp16 = einsum(equation = aw_783_equation_0, values = (var_5309_cast_fp16_11, var_5287_cast_fp16_11))[name = tensor("aw_783_cast_fp16")]; + tensor aw_785_equation_0 = const()[name = tensor("aw_785_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_785_cast_fp16 = einsum(equation = aw_785_equation_0, values = (var_5309_cast_fp16_12, var_5287_cast_fp16_12))[name = tensor("aw_785_cast_fp16")]; + tensor aw_787_equation_0 = const()[name = tensor("aw_787_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_787_cast_fp16 = einsum(equation = aw_787_equation_0, values = (var_5309_cast_fp16_13, var_5287_cast_fp16_13))[name = tensor("aw_787_cast_fp16")]; + tensor aw_789_equation_0 = const()[name = tensor("aw_789_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_789_cast_fp16 = einsum(equation = aw_789_equation_0, values = (var_5309_cast_fp16_14, var_5287_cast_fp16_14))[name = tensor("aw_789_cast_fp16")]; + tensor aw_791_equation_0 = const()[name = tensor("aw_791_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_791_cast_fp16 = einsum(equation = aw_791_equation_0, values = (var_5309_cast_fp16_15, var_5287_cast_fp16_15))[name = tensor("aw_791_cast_fp16")]; + tensor aw_793_equation_0 = const()[name = tensor("aw_793_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_793_cast_fp16 = einsum(equation = aw_793_equation_0, values = (var_5309_cast_fp16_16, var_5287_cast_fp16_16))[name = tensor("aw_793_cast_fp16")]; + tensor aw_795_equation_0 = const()[name = tensor("aw_795_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_795_cast_fp16 = einsum(equation = aw_795_equation_0, values = (var_5309_cast_fp16_17, var_5287_cast_fp16_17))[name = tensor("aw_795_cast_fp16")]; + tensor aw_797_equation_0 = const()[name = tensor("aw_797_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_797_cast_fp16 = einsum(equation = aw_797_equation_0, values = (var_5309_cast_fp16_18, var_5287_cast_fp16_18))[name = tensor("aw_797_cast_fp16")]; + tensor aw_799_equation_0 = const()[name = tensor("aw_799_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_799_cast_fp16 = einsum(equation = aw_799_equation_0, values = (var_5309_cast_fp16_19, var_5287_cast_fp16_19))[name = tensor("aw_799_cast_fp16")]; + tensor var_5391_cast_fp16 = softmax(axis = var_5235, x = aw_761_cast_fp16)[name = tensor("op_5391_cast_fp16")]; + tensor var_5392_cast_fp16 = softmax(axis = var_5235, x = aw_763_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5393_cast_fp16 = softmax(axis = var_5235, x = aw_765_cast_fp16)[name = tensor("op_5393_cast_fp16")]; + tensor var_5394_cast_fp16 = softmax(axis = var_5235, x = aw_767_cast_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor var_5395_cast_fp16 = softmax(axis = var_5235, x = aw_769_cast_fp16)[name = tensor("op_5395_cast_fp16")]; + tensor var_5396_cast_fp16 = softmax(axis = var_5235, x = aw_771_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor var_5397_cast_fp16 = softmax(axis = var_5235, x = aw_773_cast_fp16)[name = tensor("op_5397_cast_fp16")]; + tensor var_5398_cast_fp16 = softmax(axis = var_5235, x = aw_775_cast_fp16)[name = tensor("op_5398_cast_fp16")]; + tensor var_5399_cast_fp16 = softmax(axis = var_5235, x = aw_777_cast_fp16)[name = tensor("op_5399_cast_fp16")]; + tensor var_5400_cast_fp16 = softmax(axis = var_5235, x = aw_779_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5401_cast_fp16 = softmax(axis = var_5235, x = aw_781_cast_fp16)[name = tensor("op_5401_cast_fp16")]; + tensor var_5402_cast_fp16 = softmax(axis = var_5235, x = aw_783_cast_fp16)[name = tensor("op_5402_cast_fp16")]; + tensor var_5403_cast_fp16 = softmax(axis = var_5235, x = aw_785_cast_fp16)[name = tensor("op_5403_cast_fp16")]; + tensor var_5404_cast_fp16 = softmax(axis = var_5235, x = aw_787_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor var_5405_cast_fp16 = softmax(axis = var_5235, x = aw_789_cast_fp16)[name = tensor("op_5405_cast_fp16")]; + tensor var_5406_cast_fp16 = softmax(axis = var_5235, x = aw_791_cast_fp16)[name = tensor("op_5406_cast_fp16")]; + tensor var_5407_cast_fp16 = softmax(axis = var_5235, x = aw_793_cast_fp16)[name = tensor("op_5407_cast_fp16")]; + tensor var_5408_cast_fp16 = softmax(axis = var_5235, x = aw_795_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor var_5409_cast_fp16 = softmax(axis = var_5235, x = aw_797_cast_fp16)[name = tensor("op_5409_cast_fp16")]; + tensor var_5410_cast_fp16 = softmax(axis = var_5235, x = aw_799_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5412_equation_0 = const()[name = tensor("op_5412_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5412_cast_fp16 = einsum(equation = var_5412_equation_0, values = (var_5330_cast_fp16_0, var_5391_cast_fp16))[name = tensor("op_5412_cast_fp16")]; + tensor var_5414_equation_0 = const()[name = tensor("op_5414_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5414_cast_fp16 = einsum(equation = var_5414_equation_0, values = (var_5330_cast_fp16_1, var_5392_cast_fp16))[name = tensor("op_5414_cast_fp16")]; + tensor var_5416_equation_0 = const()[name = tensor("op_5416_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5416_cast_fp16 = einsum(equation = var_5416_equation_0, values = (var_5330_cast_fp16_2, var_5393_cast_fp16))[name = tensor("op_5416_cast_fp16")]; + tensor var_5418_equation_0 = const()[name = tensor("op_5418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5418_cast_fp16 = einsum(equation = var_5418_equation_0, values = (var_5330_cast_fp16_3, var_5394_cast_fp16))[name = tensor("op_5418_cast_fp16")]; + tensor var_5420_equation_0 = const()[name = tensor("op_5420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5420_cast_fp16 = einsum(equation = var_5420_equation_0, values = (var_5330_cast_fp16_4, var_5395_cast_fp16))[name = tensor("op_5420_cast_fp16")]; + tensor var_5422_equation_0 = const()[name = tensor("op_5422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5422_cast_fp16 = einsum(equation = var_5422_equation_0, values = (var_5330_cast_fp16_5, var_5396_cast_fp16))[name = tensor("op_5422_cast_fp16")]; + tensor var_5424_equation_0 = const()[name = tensor("op_5424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5424_cast_fp16 = einsum(equation = var_5424_equation_0, values = (var_5330_cast_fp16_6, var_5397_cast_fp16))[name = tensor("op_5424_cast_fp16")]; + tensor var_5426_equation_0 = const()[name = tensor("op_5426_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5426_cast_fp16 = einsum(equation = var_5426_equation_0, values = (var_5330_cast_fp16_7, var_5398_cast_fp16))[name = tensor("op_5426_cast_fp16")]; + tensor var_5428_equation_0 = const()[name = tensor("op_5428_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5428_cast_fp16 = einsum(equation = var_5428_equation_0, values = (var_5330_cast_fp16_8, var_5399_cast_fp16))[name = tensor("op_5428_cast_fp16")]; + tensor var_5430_equation_0 = const()[name = tensor("op_5430_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5430_cast_fp16 = einsum(equation = var_5430_equation_0, values = (var_5330_cast_fp16_9, var_5400_cast_fp16))[name = tensor("op_5430_cast_fp16")]; + tensor var_5432_equation_0 = const()[name = tensor("op_5432_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5432_cast_fp16 = einsum(equation = var_5432_equation_0, values = (var_5330_cast_fp16_10, var_5401_cast_fp16))[name = tensor("op_5432_cast_fp16")]; + tensor var_5434_equation_0 = const()[name = tensor("op_5434_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5434_cast_fp16 = einsum(equation = var_5434_equation_0, values = (var_5330_cast_fp16_11, var_5402_cast_fp16))[name = tensor("op_5434_cast_fp16")]; + tensor var_5436_equation_0 = const()[name = tensor("op_5436_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5436_cast_fp16 = einsum(equation = var_5436_equation_0, values = (var_5330_cast_fp16_12, var_5403_cast_fp16))[name = tensor("op_5436_cast_fp16")]; + tensor var_5438_equation_0 = const()[name = tensor("op_5438_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5438_cast_fp16 = einsum(equation = var_5438_equation_0, values = (var_5330_cast_fp16_13, var_5404_cast_fp16))[name = tensor("op_5438_cast_fp16")]; + tensor var_5440_equation_0 = const()[name = tensor("op_5440_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5440_cast_fp16 = einsum(equation = var_5440_equation_0, values = (var_5330_cast_fp16_14, var_5405_cast_fp16))[name = tensor("op_5440_cast_fp16")]; + tensor var_5442_equation_0 = const()[name = tensor("op_5442_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5442_cast_fp16 = einsum(equation = var_5442_equation_0, values = (var_5330_cast_fp16_15, var_5406_cast_fp16))[name = tensor("op_5442_cast_fp16")]; + tensor var_5444_equation_0 = const()[name = tensor("op_5444_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5444_cast_fp16 = einsum(equation = var_5444_equation_0, values = (var_5330_cast_fp16_16, var_5407_cast_fp16))[name = tensor("op_5444_cast_fp16")]; + tensor var_5446_equation_0 = const()[name = tensor("op_5446_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5446_cast_fp16 = einsum(equation = var_5446_equation_0, values = (var_5330_cast_fp16_17, var_5408_cast_fp16))[name = tensor("op_5446_cast_fp16")]; + tensor var_5448_equation_0 = const()[name = tensor("op_5448_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5448_cast_fp16 = einsum(equation = var_5448_equation_0, values = (var_5330_cast_fp16_18, var_5409_cast_fp16))[name = tensor("op_5448_cast_fp16")]; + tensor var_5450_equation_0 = const()[name = tensor("op_5450_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5450_cast_fp16 = einsum(equation = var_5450_equation_0, values = (var_5330_cast_fp16_19, var_5410_cast_fp16))[name = tensor("op_5450_cast_fp16")]; + tensor input_195_interleave_0 = const()[name = tensor("input_195_interleave_0"), val = tensor(false)]; + tensor input_195_cast_fp16 = concat(axis = var_5235, interleave = input_195_interleave_0, values = (var_5412_cast_fp16, var_5414_cast_fp16, var_5416_cast_fp16, var_5418_cast_fp16, var_5420_cast_fp16, var_5422_cast_fp16, var_5424_cast_fp16, var_5426_cast_fp16, var_5428_cast_fp16, var_5430_cast_fp16, var_5432_cast_fp16, var_5434_cast_fp16, var_5436_cast_fp16, var_5438_cast_fp16, var_5440_cast_fp16, var_5442_cast_fp16, var_5444_cast_fp16, var_5446_cast_fp16, var_5448_cast_fp16, var_5450_cast_fp16))[name = tensor("input_195_cast_fp16")]; + tensor var_5459_pad_type_0 = const()[name = tensor("op_5459_pad_type_0"), val = tensor("valid")]; + tensor var_5459_strides_0 = const()[name = tensor("op_5459_strides_0"), val = tensor([1, 1])]; + tensor var_5459_pad_0 = const()[name = tensor("op_5459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5459_dilations_0 = const()[name = tensor("op_5459_dilations_0"), val = tensor([1, 1])]; + tensor var_5459_groups_0 = const()[name = tensor("op_5459_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_out_weight_to_fp16 = const()[name = tensor("blocks_19_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(772212352)))]; + tensor blocks_19_attn_out_bias_to_fp16 = const()[name = tensor("blocks_19_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775489216)))]; + tensor var_5459_cast_fp16 = conv(bias = blocks_19_attn_out_bias_to_fp16, dilations = var_5459_dilations_0, groups = var_5459_groups_0, pad = var_5459_pad_0, pad_type = var_5459_pad_type_0, strides = var_5459_strides_0, weight = blocks_19_attn_out_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("op_5459_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_5459_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([1])]; + tensor input_197_gamma_0_to_fp16 = const()[name = tensor("input_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775491840)))]; + tensor input_197_beta_0_to_fp16 = const()[name = tensor("input_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775494464)))]; + tensor var_5469_to_fp16 = const()[name = tensor("op_5469_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = input_197_beta_0_to_fp16, epsilon = var_5469_to_fp16, gamma = input_197_gamma_0_to_fp16, x = inputs_79_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("valid")]; + tensor input_199_strides_0 = const()[name = tensor("input_199_strides_0"), val = tensor([1, 1])]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_199_dilations_0 = const()[name = tensor("input_199_dilations_0"), val = tensor([1, 1])]; + tensor input_199_groups_0 = const()[name = tensor("input_199_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775497088)))]; + tensor blocks_19_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788604352)))]; + tensor input_199_cast_fp16 = conv(bias = blocks_19_mlp_0_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = blocks_19_mlp_0_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_5495_pad_type_0 = const()[name = tensor("op_5495_pad_type_0"), val = tensor("valid")]; + tensor var_5495_strides_0 = const()[name = tensor("op_5495_strides_0"), val = tensor([1, 1])]; + tensor var_5495_pad_0 = const()[name = tensor("op_5495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5495_dilations_0 = const()[name = tensor("op_5495_dilations_0"), val = tensor([1, 1])]; + tensor var_5495_groups_0 = const()[name = tensor("op_5495_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788614656)))]; + tensor blocks_19_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801721920)))]; + tensor var_5495_cast_fp16 = conv(bias = blocks_19_mlp_2_bias_to_fp16, dilations = var_5495_dilations_0, groups = var_5495_groups_0, pad = var_5495_pad_0, pad_type = var_5495_pad_type_0, strides = var_5495_strides_0, weight = blocks_19_mlp_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = var_5495_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_5504 = const()[name = tensor("op_5504"), val = tensor(1)]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([1])]; + tensor input_203_gamma_0_to_fp16 = const()[name = tensor("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801724544)))]; + tensor input_203_beta_0_to_fp16 = const()[name = tensor("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801727168)))]; + tensor var_5520_to_fp16 = const()[name = tensor("op_5520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = input_203_beta_0_to_fp16, epsilon = var_5520_to_fp16, gamma = input_203_gamma_0_to_fp16, x = inputs_81_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor q_41_pad_type_0 = const()[name = tensor("q_41_pad_type_0"), val = tensor("valid")]; + tensor q_41_strides_0 = const()[name = tensor("q_41_strides_0"), val = tensor([1, 1])]; + tensor q_41_pad_0 = const()[name = tensor("q_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_41_dilations_0 = const()[name = tensor("q_41_dilations_0"), val = tensor([1, 1])]; + tensor q_41_groups_0 = const()[name = tensor("q_41_groups_0"), val = tensor(1)]; + tensor var_5555_weight_0_to_fp16 = const()[name = tensor("op_5555_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801729792)))]; + tensor var_5555_bias_0_to_fp16 = const()[name = tensor("op_5555_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805006656)))]; + tensor var_5555_cast_fp16 = conv(bias = var_5555_bias_0_to_fp16, dilations = q_41_dilations_0, groups = q_41_groups_0, pad = q_41_pad_0, pad_type = q_41_pad_type_0, strides = q_41_strides_0, weight = var_5555_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("op_5555_cast_fp16")]; + tensor k_41_pad_type_0 = const()[name = tensor("k_41_pad_type_0"), val = tensor("valid")]; + tensor k_41_strides_0 = const()[name = tensor("k_41_strides_0"), val = tensor([1, 1])]; + tensor k_41_pad_0 = const()[name = tensor("k_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_41_dilations_0 = const()[name = tensor("k_41_dilations_0"), val = tensor([1, 1])]; + tensor k_41_groups_0 = const()[name = tensor("k_41_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_key_weight_to_fp16 = const()[name = tensor("blocks_20_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805009280)))]; + tensor k_41_cast_fp16 = conv(dilations = k_41_dilations_0, groups = k_41_groups_0, pad = k_41_pad_0, pad_type = k_41_pad_type_0, strides = k_41_strides_0, weight = blocks_20_attn_key_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor var_5553_pad_type_0 = const()[name = tensor("op_5553_pad_type_0"), val = tensor("valid")]; + tensor var_5553_strides_0 = const()[name = tensor("op_5553_strides_0"), val = tensor([1, 1])]; + tensor var_5553_pad_0 = const()[name = tensor("op_5553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5553_dilations_0 = const()[name = tensor("op_5553_dilations_0"), val = tensor([1, 1])]; + tensor var_5553_groups_0 = const()[name = tensor("op_5553_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_value_weight_to_fp16 = const()[name = tensor("blocks_20_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(808286144)))]; + tensor blocks_20_attn_value_bias_to_fp16 = const()[name = tensor("blocks_20_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811563008)))]; + tensor var_5553_cast_fp16 = conv(bias = blocks_20_attn_value_bias_to_fp16, dilations = var_5553_dilations_0, groups = var_5553_groups_0, pad = var_5553_pad_0, pad_type = var_5553_pad_type_0, strides = var_5553_strides_0, weight = blocks_20_attn_value_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("op_5553_cast_fp16")]; + tensor tile_60 = const()[name = tensor("tile_60"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5556_axis_0 = const()[name = tensor("op_5556_axis_0"), val = tensor(1)]; + tensor var_5556_cast_fp16_0, tensor var_5556_cast_fp16_1, tensor var_5556_cast_fp16_2, tensor var_5556_cast_fp16_3, tensor var_5556_cast_fp16_4, tensor var_5556_cast_fp16_5, tensor var_5556_cast_fp16_6, tensor var_5556_cast_fp16_7, tensor var_5556_cast_fp16_8, tensor var_5556_cast_fp16_9, tensor var_5556_cast_fp16_10, tensor var_5556_cast_fp16_11, tensor var_5556_cast_fp16_12, tensor var_5556_cast_fp16_13, tensor var_5556_cast_fp16_14, tensor var_5556_cast_fp16_15, tensor var_5556_cast_fp16_16, tensor var_5556_cast_fp16_17, tensor var_5556_cast_fp16_18, tensor var_5556_cast_fp16_19 = split(axis = var_5556_axis_0, split_sizes = tile_60, x = var_5555_cast_fp16)[name = tensor("op_5556_cast_fp16")]; + tensor var_5577_perm_0 = const()[name = tensor("op_5577_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_61 = const()[name = tensor("tile_61"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5578_axis_0 = const()[name = tensor("op_5578_axis_0"), val = tensor(3)]; + tensor var_5577_cast_fp16 = transpose(perm = var_5577_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_12")]; + tensor var_5578_cast_fp16_0, tensor var_5578_cast_fp16_1, tensor var_5578_cast_fp16_2, tensor var_5578_cast_fp16_3, tensor var_5578_cast_fp16_4, tensor var_5578_cast_fp16_5, tensor var_5578_cast_fp16_6, tensor var_5578_cast_fp16_7, tensor var_5578_cast_fp16_8, tensor var_5578_cast_fp16_9, tensor var_5578_cast_fp16_10, tensor var_5578_cast_fp16_11, tensor var_5578_cast_fp16_12, tensor var_5578_cast_fp16_13, tensor var_5578_cast_fp16_14, tensor var_5578_cast_fp16_15, tensor var_5578_cast_fp16_16, tensor var_5578_cast_fp16_17, tensor var_5578_cast_fp16_18, tensor var_5578_cast_fp16_19 = split(axis = var_5578_axis_0, split_sizes = tile_61, x = var_5577_cast_fp16)[name = tensor("op_5578_cast_fp16")]; + tensor tile_62 = const()[name = tensor("tile_62"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5599_axis_0 = const()[name = tensor("op_5599_axis_0"), val = tensor(1)]; + tensor var_5599_cast_fp16_0, tensor var_5599_cast_fp16_1, tensor var_5599_cast_fp16_2, tensor var_5599_cast_fp16_3, tensor var_5599_cast_fp16_4, tensor var_5599_cast_fp16_5, tensor var_5599_cast_fp16_6, tensor var_5599_cast_fp16_7, tensor var_5599_cast_fp16_8, tensor var_5599_cast_fp16_9, tensor var_5599_cast_fp16_10, tensor var_5599_cast_fp16_11, tensor var_5599_cast_fp16_12, tensor var_5599_cast_fp16_13, tensor var_5599_cast_fp16_14, tensor var_5599_cast_fp16_15, tensor var_5599_cast_fp16_16, tensor var_5599_cast_fp16_17, tensor var_5599_cast_fp16_18, tensor var_5599_cast_fp16_19 = split(axis = var_5599_axis_0, split_sizes = tile_62, x = var_5553_cast_fp16)[name = tensor("op_5599_cast_fp16")]; + tensor aw_801_equation_0 = const()[name = tensor("aw_801_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_801_cast_fp16 = einsum(equation = aw_801_equation_0, values = (var_5578_cast_fp16_0, var_5556_cast_fp16_0))[name = tensor("aw_801_cast_fp16")]; + tensor aw_803_equation_0 = const()[name = tensor("aw_803_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_803_cast_fp16 = einsum(equation = aw_803_equation_0, values = (var_5578_cast_fp16_1, var_5556_cast_fp16_1))[name = tensor("aw_803_cast_fp16")]; + tensor aw_805_equation_0 = const()[name = tensor("aw_805_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_805_cast_fp16 = einsum(equation = aw_805_equation_0, values = (var_5578_cast_fp16_2, var_5556_cast_fp16_2))[name = tensor("aw_805_cast_fp16")]; + tensor aw_807_equation_0 = const()[name = tensor("aw_807_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_807_cast_fp16 = einsum(equation = aw_807_equation_0, values = (var_5578_cast_fp16_3, var_5556_cast_fp16_3))[name = tensor("aw_807_cast_fp16")]; + tensor aw_809_equation_0 = const()[name = tensor("aw_809_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_809_cast_fp16 = einsum(equation = aw_809_equation_0, values = (var_5578_cast_fp16_4, var_5556_cast_fp16_4))[name = tensor("aw_809_cast_fp16")]; + tensor aw_811_equation_0 = const()[name = tensor("aw_811_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_811_cast_fp16 = einsum(equation = aw_811_equation_0, values = (var_5578_cast_fp16_5, var_5556_cast_fp16_5))[name = tensor("aw_811_cast_fp16")]; + tensor aw_813_equation_0 = const()[name = tensor("aw_813_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_813_cast_fp16 = einsum(equation = aw_813_equation_0, values = (var_5578_cast_fp16_6, var_5556_cast_fp16_6))[name = tensor("aw_813_cast_fp16")]; + tensor aw_815_equation_0 = const()[name = tensor("aw_815_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_815_cast_fp16 = einsum(equation = aw_815_equation_0, values = (var_5578_cast_fp16_7, var_5556_cast_fp16_7))[name = tensor("aw_815_cast_fp16")]; + tensor aw_817_equation_0 = const()[name = tensor("aw_817_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_817_cast_fp16 = einsum(equation = aw_817_equation_0, values = (var_5578_cast_fp16_8, var_5556_cast_fp16_8))[name = tensor("aw_817_cast_fp16")]; + tensor aw_819_equation_0 = const()[name = tensor("aw_819_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_819_cast_fp16 = einsum(equation = aw_819_equation_0, values = (var_5578_cast_fp16_9, var_5556_cast_fp16_9))[name = tensor("aw_819_cast_fp16")]; + tensor aw_821_equation_0 = const()[name = tensor("aw_821_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_821_cast_fp16 = einsum(equation = aw_821_equation_0, values = (var_5578_cast_fp16_10, var_5556_cast_fp16_10))[name = tensor("aw_821_cast_fp16")]; + tensor aw_823_equation_0 = const()[name = tensor("aw_823_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_823_cast_fp16 = einsum(equation = aw_823_equation_0, values = (var_5578_cast_fp16_11, var_5556_cast_fp16_11))[name = tensor("aw_823_cast_fp16")]; + tensor aw_825_equation_0 = const()[name = tensor("aw_825_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_825_cast_fp16 = einsum(equation = aw_825_equation_0, values = (var_5578_cast_fp16_12, var_5556_cast_fp16_12))[name = tensor("aw_825_cast_fp16")]; + tensor aw_827_equation_0 = const()[name = tensor("aw_827_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_827_cast_fp16 = einsum(equation = aw_827_equation_0, values = (var_5578_cast_fp16_13, var_5556_cast_fp16_13))[name = tensor("aw_827_cast_fp16")]; + tensor aw_829_equation_0 = const()[name = tensor("aw_829_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_829_cast_fp16 = einsum(equation = aw_829_equation_0, values = (var_5578_cast_fp16_14, var_5556_cast_fp16_14))[name = tensor("aw_829_cast_fp16")]; + tensor aw_831_equation_0 = const()[name = tensor("aw_831_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_831_cast_fp16 = einsum(equation = aw_831_equation_0, values = (var_5578_cast_fp16_15, var_5556_cast_fp16_15))[name = tensor("aw_831_cast_fp16")]; + tensor aw_833_equation_0 = const()[name = tensor("aw_833_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_833_cast_fp16 = einsum(equation = aw_833_equation_0, values = (var_5578_cast_fp16_16, var_5556_cast_fp16_16))[name = tensor("aw_833_cast_fp16")]; + tensor aw_835_equation_0 = const()[name = tensor("aw_835_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_835_cast_fp16 = einsum(equation = aw_835_equation_0, values = (var_5578_cast_fp16_17, var_5556_cast_fp16_17))[name = tensor("aw_835_cast_fp16")]; + tensor aw_837_equation_0 = const()[name = tensor("aw_837_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_837_cast_fp16 = einsum(equation = aw_837_equation_0, values = (var_5578_cast_fp16_18, var_5556_cast_fp16_18))[name = tensor("aw_837_cast_fp16")]; + tensor aw_839_equation_0 = const()[name = tensor("aw_839_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_839_cast_fp16 = einsum(equation = aw_839_equation_0, values = (var_5578_cast_fp16_19, var_5556_cast_fp16_19))[name = tensor("aw_839_cast_fp16")]; + tensor var_5660_cast_fp16 = softmax(axis = var_5504, x = aw_801_cast_fp16)[name = tensor("op_5660_cast_fp16")]; + tensor var_5661_cast_fp16 = softmax(axis = var_5504, x = aw_803_cast_fp16)[name = tensor("op_5661_cast_fp16")]; + tensor var_5662_cast_fp16 = softmax(axis = var_5504, x = aw_805_cast_fp16)[name = tensor("op_5662_cast_fp16")]; + tensor var_5663_cast_fp16 = softmax(axis = var_5504, x = aw_807_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor var_5664_cast_fp16 = softmax(axis = var_5504, x = aw_809_cast_fp16)[name = tensor("op_5664_cast_fp16")]; + tensor var_5665_cast_fp16 = softmax(axis = var_5504, x = aw_811_cast_fp16)[name = tensor("op_5665_cast_fp16")]; + tensor var_5666_cast_fp16 = softmax(axis = var_5504, x = aw_813_cast_fp16)[name = tensor("op_5666_cast_fp16")]; + tensor var_5667_cast_fp16 = softmax(axis = var_5504, x = aw_815_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = softmax(axis = var_5504, x = aw_817_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor var_5669_cast_fp16 = softmax(axis = var_5504, x = aw_819_cast_fp16)[name = tensor("op_5669_cast_fp16")]; + tensor var_5670_cast_fp16 = softmax(axis = var_5504, x = aw_821_cast_fp16)[name = tensor("op_5670_cast_fp16")]; + tensor var_5671_cast_fp16 = softmax(axis = var_5504, x = aw_823_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = softmax(axis = var_5504, x = aw_825_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor var_5673_cast_fp16 = softmax(axis = var_5504, x = aw_827_cast_fp16)[name = tensor("op_5673_cast_fp16")]; + tensor var_5674_cast_fp16 = softmax(axis = var_5504, x = aw_829_cast_fp16)[name = tensor("op_5674_cast_fp16")]; + tensor var_5675_cast_fp16 = softmax(axis = var_5504, x = aw_831_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = softmax(axis = var_5504, x = aw_833_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor var_5677_cast_fp16 = softmax(axis = var_5504, x = aw_835_cast_fp16)[name = tensor("op_5677_cast_fp16")]; + tensor var_5678_cast_fp16 = softmax(axis = var_5504, x = aw_837_cast_fp16)[name = tensor("op_5678_cast_fp16")]; + tensor var_5679_cast_fp16 = softmax(axis = var_5504, x = aw_839_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5681_equation_0 = const()[name = tensor("op_5681_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5681_cast_fp16 = einsum(equation = var_5681_equation_0, values = (var_5599_cast_fp16_0, var_5660_cast_fp16))[name = tensor("op_5681_cast_fp16")]; + tensor var_5683_equation_0 = const()[name = tensor("op_5683_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5683_cast_fp16 = einsum(equation = var_5683_equation_0, values = (var_5599_cast_fp16_1, var_5661_cast_fp16))[name = tensor("op_5683_cast_fp16")]; + tensor var_5685_equation_0 = const()[name = tensor("op_5685_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5685_cast_fp16 = einsum(equation = var_5685_equation_0, values = (var_5599_cast_fp16_2, var_5662_cast_fp16))[name = tensor("op_5685_cast_fp16")]; + tensor var_5687_equation_0 = const()[name = tensor("op_5687_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5687_cast_fp16 = einsum(equation = var_5687_equation_0, values = (var_5599_cast_fp16_3, var_5663_cast_fp16))[name = tensor("op_5687_cast_fp16")]; + tensor var_5689_equation_0 = const()[name = tensor("op_5689_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5689_cast_fp16 = einsum(equation = var_5689_equation_0, values = (var_5599_cast_fp16_4, var_5664_cast_fp16))[name = tensor("op_5689_cast_fp16")]; + tensor var_5691_equation_0 = const()[name = tensor("op_5691_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5691_cast_fp16 = einsum(equation = var_5691_equation_0, values = (var_5599_cast_fp16_5, var_5665_cast_fp16))[name = tensor("op_5691_cast_fp16")]; + tensor var_5693_equation_0 = const()[name = tensor("op_5693_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5693_cast_fp16 = einsum(equation = var_5693_equation_0, values = (var_5599_cast_fp16_6, var_5666_cast_fp16))[name = tensor("op_5693_cast_fp16")]; + tensor var_5695_equation_0 = const()[name = tensor("op_5695_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5695_cast_fp16 = einsum(equation = var_5695_equation_0, values = (var_5599_cast_fp16_7, var_5667_cast_fp16))[name = tensor("op_5695_cast_fp16")]; + tensor var_5697_equation_0 = const()[name = tensor("op_5697_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5697_cast_fp16 = einsum(equation = var_5697_equation_0, values = (var_5599_cast_fp16_8, var_5668_cast_fp16))[name = tensor("op_5697_cast_fp16")]; + tensor var_5699_equation_0 = const()[name = tensor("op_5699_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5699_cast_fp16 = einsum(equation = var_5699_equation_0, values = (var_5599_cast_fp16_9, var_5669_cast_fp16))[name = tensor("op_5699_cast_fp16")]; + tensor var_5701_equation_0 = const()[name = tensor("op_5701_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5701_cast_fp16 = einsum(equation = var_5701_equation_0, values = (var_5599_cast_fp16_10, var_5670_cast_fp16))[name = tensor("op_5701_cast_fp16")]; + tensor var_5703_equation_0 = const()[name = tensor("op_5703_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5703_cast_fp16 = einsum(equation = var_5703_equation_0, values = (var_5599_cast_fp16_11, var_5671_cast_fp16))[name = tensor("op_5703_cast_fp16")]; + tensor var_5705_equation_0 = const()[name = tensor("op_5705_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5705_cast_fp16 = einsum(equation = var_5705_equation_0, values = (var_5599_cast_fp16_12, var_5672_cast_fp16))[name = tensor("op_5705_cast_fp16")]; + tensor var_5707_equation_0 = const()[name = tensor("op_5707_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5707_cast_fp16 = einsum(equation = var_5707_equation_0, values = (var_5599_cast_fp16_13, var_5673_cast_fp16))[name = tensor("op_5707_cast_fp16")]; + tensor var_5709_equation_0 = const()[name = tensor("op_5709_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5709_cast_fp16 = einsum(equation = var_5709_equation_0, values = (var_5599_cast_fp16_14, var_5674_cast_fp16))[name = tensor("op_5709_cast_fp16")]; + tensor var_5711_equation_0 = const()[name = tensor("op_5711_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5711_cast_fp16 = einsum(equation = var_5711_equation_0, values = (var_5599_cast_fp16_15, var_5675_cast_fp16))[name = tensor("op_5711_cast_fp16")]; + tensor var_5713_equation_0 = const()[name = tensor("op_5713_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5713_cast_fp16 = einsum(equation = var_5713_equation_0, values = (var_5599_cast_fp16_16, var_5676_cast_fp16))[name = tensor("op_5713_cast_fp16")]; + tensor var_5715_equation_0 = const()[name = tensor("op_5715_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5715_cast_fp16 = einsum(equation = var_5715_equation_0, values = (var_5599_cast_fp16_17, var_5677_cast_fp16))[name = tensor("op_5715_cast_fp16")]; + tensor var_5717_equation_0 = const()[name = tensor("op_5717_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5717_cast_fp16 = einsum(equation = var_5717_equation_0, values = (var_5599_cast_fp16_18, var_5678_cast_fp16))[name = tensor("op_5717_cast_fp16")]; + tensor var_5719_equation_0 = const()[name = tensor("op_5719_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5719_cast_fp16 = einsum(equation = var_5719_equation_0, values = (var_5599_cast_fp16_19, var_5679_cast_fp16))[name = tensor("op_5719_cast_fp16")]; + tensor input_205_interleave_0 = const()[name = tensor("input_205_interleave_0"), val = tensor(false)]; + tensor input_205_cast_fp16 = concat(axis = var_5504, interleave = input_205_interleave_0, values = (var_5681_cast_fp16, var_5683_cast_fp16, var_5685_cast_fp16, var_5687_cast_fp16, var_5689_cast_fp16, var_5691_cast_fp16, var_5693_cast_fp16, var_5695_cast_fp16, var_5697_cast_fp16, var_5699_cast_fp16, var_5701_cast_fp16, var_5703_cast_fp16, var_5705_cast_fp16, var_5707_cast_fp16, var_5709_cast_fp16, var_5711_cast_fp16, var_5713_cast_fp16, var_5715_cast_fp16, var_5717_cast_fp16, var_5719_cast_fp16))[name = tensor("input_205_cast_fp16")]; + tensor var_5728_pad_type_0 = const()[name = tensor("op_5728_pad_type_0"), val = tensor("valid")]; + tensor var_5728_strides_0 = const()[name = tensor("op_5728_strides_0"), val = tensor([1, 1])]; + tensor var_5728_pad_0 = const()[name = tensor("op_5728_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5728_dilations_0 = const()[name = tensor("op_5728_dilations_0"), val = tensor([1, 1])]; + tensor var_5728_groups_0 = const()[name = tensor("op_5728_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_out_weight_to_fp16 = const()[name = tensor("blocks_20_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811565632)))]; + tensor blocks_20_attn_out_bias_to_fp16 = const()[name = tensor("blocks_20_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814842496)))]; + tensor var_5728_cast_fp16 = conv(bias = blocks_20_attn_out_bias_to_fp16, dilations = var_5728_dilations_0, groups = var_5728_groups_0, pad = var_5728_pad_0, pad_type = var_5728_pad_type_0, strides = var_5728_strides_0, weight = blocks_20_attn_out_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_5728_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([1])]; + tensor input_207_gamma_0_to_fp16 = const()[name = tensor("input_207_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814845120)))]; + tensor input_207_beta_0_to_fp16 = const()[name = tensor("input_207_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814847744)))]; + tensor var_5738_to_fp16 = const()[name = tensor("op_5738_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = input_207_beta_0_to_fp16, epsilon = var_5738_to_fp16, gamma = input_207_gamma_0_to_fp16, x = inputs_83_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("valid")]; + tensor input_209_strides_0 = const()[name = tensor("input_209_strides_0"), val = tensor([1, 1])]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_209_dilations_0 = const()[name = tensor("input_209_dilations_0"), val = tensor([1, 1])]; + tensor input_209_groups_0 = const()[name = tensor("input_209_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814850368)))]; + tensor blocks_20_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827957632)))]; + tensor input_209_cast_fp16 = conv(bias = blocks_20_mlp_0_bias_to_fp16, dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = blocks_20_mlp_0_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = input_209_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_5764_pad_type_0 = const()[name = tensor("op_5764_pad_type_0"), val = tensor("valid")]; + tensor var_5764_strides_0 = const()[name = tensor("op_5764_strides_0"), val = tensor([1, 1])]; + tensor var_5764_pad_0 = const()[name = tensor("op_5764_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5764_dilations_0 = const()[name = tensor("op_5764_dilations_0"), val = tensor([1, 1])]; + tensor var_5764_groups_0 = const()[name = tensor("op_5764_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827967936)))]; + tensor blocks_20_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841075200)))]; + tensor var_5764_cast_fp16 = conv(bias = blocks_20_mlp_2_bias_to_fp16, dilations = var_5764_dilations_0, groups = var_5764_groups_0, pad = var_5764_pad_0, pad_type = var_5764_pad_type_0, strides = var_5764_strides_0, weight = blocks_20_mlp_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("op_5764_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = var_5764_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_5773 = const()[name = tensor("op_5773"), val = tensor(1)]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([1])]; + tensor input_213_gamma_0_to_fp16 = const()[name = tensor("input_213_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841077824)))]; + tensor input_213_beta_0_to_fp16 = const()[name = tensor("input_213_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841080448)))]; + tensor var_5789_to_fp16 = const()[name = tensor("op_5789_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = input_213_beta_0_to_fp16, epsilon = var_5789_to_fp16, gamma = input_213_gamma_0_to_fp16, x = inputs_85_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor q_43_pad_type_0 = const()[name = tensor("q_43_pad_type_0"), val = tensor("valid")]; + tensor q_43_strides_0 = const()[name = tensor("q_43_strides_0"), val = tensor([1, 1])]; + tensor q_43_pad_0 = const()[name = tensor("q_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_43_dilations_0 = const()[name = tensor("q_43_dilations_0"), val = tensor([1, 1])]; + tensor q_43_groups_0 = const()[name = tensor("q_43_groups_0"), val = tensor(1)]; + tensor var_5824_weight_0_to_fp16 = const()[name = tensor("op_5824_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841083072)))]; + tensor var_5824_bias_0_to_fp16 = const()[name = tensor("op_5824_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844359936)))]; + tensor var_5824_cast_fp16 = conv(bias = var_5824_bias_0_to_fp16, dilations = q_43_dilations_0, groups = q_43_groups_0, pad = q_43_pad_0, pad_type = q_43_pad_type_0, strides = q_43_strides_0, weight = var_5824_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5824_cast_fp16")]; + tensor k_43_pad_type_0 = const()[name = tensor("k_43_pad_type_0"), val = tensor("valid")]; + tensor k_43_strides_0 = const()[name = tensor("k_43_strides_0"), val = tensor([1, 1])]; + tensor k_43_pad_0 = const()[name = tensor("k_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_43_dilations_0 = const()[name = tensor("k_43_dilations_0"), val = tensor([1, 1])]; + tensor k_43_groups_0 = const()[name = tensor("k_43_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_key_weight_to_fp16 = const()[name = tensor("blocks_21_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844362560)))]; + tensor k_43_cast_fp16 = conv(dilations = k_43_dilations_0, groups = k_43_groups_0, pad = k_43_pad_0, pad_type = k_43_pad_type_0, strides = k_43_strides_0, weight = blocks_21_attn_key_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor var_5822_pad_type_0 = const()[name = tensor("op_5822_pad_type_0"), val = tensor("valid")]; + tensor var_5822_strides_0 = const()[name = tensor("op_5822_strides_0"), val = tensor([1, 1])]; + tensor var_5822_pad_0 = const()[name = tensor("op_5822_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5822_dilations_0 = const()[name = tensor("op_5822_dilations_0"), val = tensor([1, 1])]; + tensor var_5822_groups_0 = const()[name = tensor("op_5822_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_value_weight_to_fp16 = const()[name = tensor("blocks_21_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847639424)))]; + tensor blocks_21_attn_value_bias_to_fp16 = const()[name = tensor("blocks_21_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850916288)))]; + tensor var_5822_cast_fp16 = conv(bias = blocks_21_attn_value_bias_to_fp16, dilations = var_5822_dilations_0, groups = var_5822_groups_0, pad = var_5822_pad_0, pad_type = var_5822_pad_type_0, strides = var_5822_strides_0, weight = blocks_21_attn_value_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5822_cast_fp16")]; + tensor tile_63 = const()[name = tensor("tile_63"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5825_axis_0 = const()[name = tensor("op_5825_axis_0"), val = tensor(1)]; + tensor var_5825_cast_fp16_0, tensor var_5825_cast_fp16_1, tensor var_5825_cast_fp16_2, tensor var_5825_cast_fp16_3, tensor var_5825_cast_fp16_4, tensor var_5825_cast_fp16_5, tensor var_5825_cast_fp16_6, tensor var_5825_cast_fp16_7, tensor var_5825_cast_fp16_8, tensor var_5825_cast_fp16_9, tensor var_5825_cast_fp16_10, tensor var_5825_cast_fp16_11, tensor var_5825_cast_fp16_12, tensor var_5825_cast_fp16_13, tensor var_5825_cast_fp16_14, tensor var_5825_cast_fp16_15, tensor var_5825_cast_fp16_16, tensor var_5825_cast_fp16_17, tensor var_5825_cast_fp16_18, tensor var_5825_cast_fp16_19 = split(axis = var_5825_axis_0, split_sizes = tile_63, x = var_5824_cast_fp16)[name = tensor("op_5825_cast_fp16")]; + tensor var_5846_perm_0 = const()[name = tensor("op_5846_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_64 = const()[name = tensor("tile_64"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5847_axis_0 = const()[name = tensor("op_5847_axis_0"), val = tensor(3)]; + tensor var_5846_cast_fp16 = transpose(perm = var_5846_perm_0, x = k_43_cast_fp16)[name = tensor("transpose_11")]; + tensor var_5847_cast_fp16_0, tensor var_5847_cast_fp16_1, tensor var_5847_cast_fp16_2, tensor var_5847_cast_fp16_3, tensor var_5847_cast_fp16_4, tensor var_5847_cast_fp16_5, tensor var_5847_cast_fp16_6, tensor var_5847_cast_fp16_7, tensor var_5847_cast_fp16_8, tensor var_5847_cast_fp16_9, tensor var_5847_cast_fp16_10, tensor var_5847_cast_fp16_11, tensor var_5847_cast_fp16_12, tensor var_5847_cast_fp16_13, tensor var_5847_cast_fp16_14, tensor var_5847_cast_fp16_15, tensor var_5847_cast_fp16_16, tensor var_5847_cast_fp16_17, tensor var_5847_cast_fp16_18, tensor var_5847_cast_fp16_19 = split(axis = var_5847_axis_0, split_sizes = tile_64, x = var_5846_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor tile_65 = const()[name = tensor("tile_65"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5868_axis_0 = const()[name = tensor("op_5868_axis_0"), val = tensor(1)]; + tensor var_5868_cast_fp16_0, tensor var_5868_cast_fp16_1, tensor var_5868_cast_fp16_2, tensor var_5868_cast_fp16_3, tensor var_5868_cast_fp16_4, tensor var_5868_cast_fp16_5, tensor var_5868_cast_fp16_6, tensor var_5868_cast_fp16_7, tensor var_5868_cast_fp16_8, tensor var_5868_cast_fp16_9, tensor var_5868_cast_fp16_10, tensor var_5868_cast_fp16_11, tensor var_5868_cast_fp16_12, tensor var_5868_cast_fp16_13, tensor var_5868_cast_fp16_14, tensor var_5868_cast_fp16_15, tensor var_5868_cast_fp16_16, tensor var_5868_cast_fp16_17, tensor var_5868_cast_fp16_18, tensor var_5868_cast_fp16_19 = split(axis = var_5868_axis_0, split_sizes = tile_65, x = var_5822_cast_fp16)[name = tensor("op_5868_cast_fp16")]; + tensor aw_841_equation_0 = const()[name = tensor("aw_841_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_841_cast_fp16 = einsum(equation = aw_841_equation_0, values = (var_5847_cast_fp16_0, var_5825_cast_fp16_0))[name = tensor("aw_841_cast_fp16")]; + tensor aw_843_equation_0 = const()[name = tensor("aw_843_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_843_cast_fp16 = einsum(equation = aw_843_equation_0, values = (var_5847_cast_fp16_1, var_5825_cast_fp16_1))[name = tensor("aw_843_cast_fp16")]; + tensor aw_845_equation_0 = const()[name = tensor("aw_845_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_845_cast_fp16 = einsum(equation = aw_845_equation_0, values = (var_5847_cast_fp16_2, var_5825_cast_fp16_2))[name = tensor("aw_845_cast_fp16")]; + tensor aw_847_equation_0 = const()[name = tensor("aw_847_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_847_cast_fp16 = einsum(equation = aw_847_equation_0, values = (var_5847_cast_fp16_3, var_5825_cast_fp16_3))[name = tensor("aw_847_cast_fp16")]; + tensor aw_849_equation_0 = const()[name = tensor("aw_849_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_849_cast_fp16 = einsum(equation = aw_849_equation_0, values = (var_5847_cast_fp16_4, var_5825_cast_fp16_4))[name = tensor("aw_849_cast_fp16")]; + tensor aw_851_equation_0 = const()[name = tensor("aw_851_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_851_cast_fp16 = einsum(equation = aw_851_equation_0, values = (var_5847_cast_fp16_5, var_5825_cast_fp16_5))[name = tensor("aw_851_cast_fp16")]; + tensor aw_853_equation_0 = const()[name = tensor("aw_853_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_853_cast_fp16 = einsum(equation = aw_853_equation_0, values = (var_5847_cast_fp16_6, var_5825_cast_fp16_6))[name = tensor("aw_853_cast_fp16")]; + tensor aw_855_equation_0 = const()[name = tensor("aw_855_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_855_cast_fp16 = einsum(equation = aw_855_equation_0, values = (var_5847_cast_fp16_7, var_5825_cast_fp16_7))[name = tensor("aw_855_cast_fp16")]; + tensor aw_857_equation_0 = const()[name = tensor("aw_857_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_857_cast_fp16 = einsum(equation = aw_857_equation_0, values = (var_5847_cast_fp16_8, var_5825_cast_fp16_8))[name = tensor("aw_857_cast_fp16")]; + tensor aw_859_equation_0 = const()[name = tensor("aw_859_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_859_cast_fp16 = einsum(equation = aw_859_equation_0, values = (var_5847_cast_fp16_9, var_5825_cast_fp16_9))[name = tensor("aw_859_cast_fp16")]; + tensor aw_861_equation_0 = const()[name = tensor("aw_861_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_861_cast_fp16 = einsum(equation = aw_861_equation_0, values = (var_5847_cast_fp16_10, var_5825_cast_fp16_10))[name = tensor("aw_861_cast_fp16")]; + tensor aw_863_equation_0 = const()[name = tensor("aw_863_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_863_cast_fp16 = einsum(equation = aw_863_equation_0, values = (var_5847_cast_fp16_11, var_5825_cast_fp16_11))[name = tensor("aw_863_cast_fp16")]; + tensor aw_865_equation_0 = const()[name = tensor("aw_865_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_865_cast_fp16 = einsum(equation = aw_865_equation_0, values = (var_5847_cast_fp16_12, var_5825_cast_fp16_12))[name = tensor("aw_865_cast_fp16")]; + tensor aw_867_equation_0 = const()[name = tensor("aw_867_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_867_cast_fp16 = einsum(equation = aw_867_equation_0, values = (var_5847_cast_fp16_13, var_5825_cast_fp16_13))[name = tensor("aw_867_cast_fp16")]; + tensor aw_869_equation_0 = const()[name = tensor("aw_869_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_869_cast_fp16 = einsum(equation = aw_869_equation_0, values = (var_5847_cast_fp16_14, var_5825_cast_fp16_14))[name = tensor("aw_869_cast_fp16")]; + tensor aw_871_equation_0 = const()[name = tensor("aw_871_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_871_cast_fp16 = einsum(equation = aw_871_equation_0, values = (var_5847_cast_fp16_15, var_5825_cast_fp16_15))[name = tensor("aw_871_cast_fp16")]; + tensor aw_873_equation_0 = const()[name = tensor("aw_873_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_873_cast_fp16 = einsum(equation = aw_873_equation_0, values = (var_5847_cast_fp16_16, var_5825_cast_fp16_16))[name = tensor("aw_873_cast_fp16")]; + tensor aw_875_equation_0 = const()[name = tensor("aw_875_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_875_cast_fp16 = einsum(equation = aw_875_equation_0, values = (var_5847_cast_fp16_17, var_5825_cast_fp16_17))[name = tensor("aw_875_cast_fp16")]; + tensor aw_877_equation_0 = const()[name = tensor("aw_877_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_877_cast_fp16 = einsum(equation = aw_877_equation_0, values = (var_5847_cast_fp16_18, var_5825_cast_fp16_18))[name = tensor("aw_877_cast_fp16")]; + tensor aw_879_equation_0 = const()[name = tensor("aw_879_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_879_cast_fp16 = einsum(equation = aw_879_equation_0, values = (var_5847_cast_fp16_19, var_5825_cast_fp16_19))[name = tensor("aw_879_cast_fp16")]; + tensor var_5929_cast_fp16 = softmax(axis = var_5773, x = aw_841_cast_fp16)[name = tensor("op_5929_cast_fp16")]; + tensor var_5930_cast_fp16 = softmax(axis = var_5773, x = aw_843_cast_fp16)[name = tensor("op_5930_cast_fp16")]; + tensor var_5931_cast_fp16 = softmax(axis = var_5773, x = aw_845_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5932_cast_fp16 = softmax(axis = var_5773, x = aw_847_cast_fp16)[name = tensor("op_5932_cast_fp16")]; + tensor var_5933_cast_fp16 = softmax(axis = var_5773, x = aw_849_cast_fp16)[name = tensor("op_5933_cast_fp16")]; + tensor var_5934_cast_fp16 = softmax(axis = var_5773, x = aw_851_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = softmax(axis = var_5773, x = aw_853_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor var_5936_cast_fp16 = softmax(axis = var_5773, x = aw_855_cast_fp16)[name = tensor("op_5936_cast_fp16")]; + tensor var_5937_cast_fp16 = softmax(axis = var_5773, x = aw_857_cast_fp16)[name = tensor("op_5937_cast_fp16")]; + tensor var_5938_cast_fp16 = softmax(axis = var_5773, x = aw_859_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = softmax(axis = var_5773, x = aw_861_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor var_5940_cast_fp16 = softmax(axis = var_5773, x = aw_863_cast_fp16)[name = tensor("op_5940_cast_fp16")]; + tensor var_5941_cast_fp16 = softmax(axis = var_5773, x = aw_865_cast_fp16)[name = tensor("op_5941_cast_fp16")]; + tensor var_5942_cast_fp16 = softmax(axis = var_5773, x = aw_867_cast_fp16)[name = tensor("op_5942_cast_fp16")]; + tensor var_5943_cast_fp16 = softmax(axis = var_5773, x = aw_869_cast_fp16)[name = tensor("op_5943_cast_fp16")]; + tensor var_5944_cast_fp16 = softmax(axis = var_5773, x = aw_871_cast_fp16)[name = tensor("op_5944_cast_fp16")]; + tensor var_5945_cast_fp16 = softmax(axis = var_5773, x = aw_873_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor var_5946_cast_fp16 = softmax(axis = var_5773, x = aw_875_cast_fp16)[name = tensor("op_5946_cast_fp16")]; + tensor var_5947_cast_fp16 = softmax(axis = var_5773, x = aw_877_cast_fp16)[name = tensor("op_5947_cast_fp16")]; + tensor var_5948_cast_fp16 = softmax(axis = var_5773, x = aw_879_cast_fp16)[name = tensor("op_5948_cast_fp16")]; + tensor var_5950_equation_0 = const()[name = tensor("op_5950_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5950_cast_fp16 = einsum(equation = var_5950_equation_0, values = (var_5868_cast_fp16_0, var_5929_cast_fp16))[name = tensor("op_5950_cast_fp16")]; + tensor var_5952_equation_0 = const()[name = tensor("op_5952_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5952_cast_fp16 = einsum(equation = var_5952_equation_0, values = (var_5868_cast_fp16_1, var_5930_cast_fp16))[name = tensor("op_5952_cast_fp16")]; + tensor var_5954_equation_0 = const()[name = tensor("op_5954_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5954_cast_fp16 = einsum(equation = var_5954_equation_0, values = (var_5868_cast_fp16_2, var_5931_cast_fp16))[name = tensor("op_5954_cast_fp16")]; + tensor var_5956_equation_0 = const()[name = tensor("op_5956_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5956_cast_fp16 = einsum(equation = var_5956_equation_0, values = (var_5868_cast_fp16_3, var_5932_cast_fp16))[name = tensor("op_5956_cast_fp16")]; + tensor var_5958_equation_0 = const()[name = tensor("op_5958_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5958_cast_fp16 = einsum(equation = var_5958_equation_0, values = (var_5868_cast_fp16_4, var_5933_cast_fp16))[name = tensor("op_5958_cast_fp16")]; + tensor var_5960_equation_0 = const()[name = tensor("op_5960_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5960_cast_fp16 = einsum(equation = var_5960_equation_0, values = (var_5868_cast_fp16_5, var_5934_cast_fp16))[name = tensor("op_5960_cast_fp16")]; + tensor var_5962_equation_0 = const()[name = tensor("op_5962_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5962_cast_fp16 = einsum(equation = var_5962_equation_0, values = (var_5868_cast_fp16_6, var_5935_cast_fp16))[name = tensor("op_5962_cast_fp16")]; + tensor var_5964_equation_0 = const()[name = tensor("op_5964_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5964_cast_fp16 = einsum(equation = var_5964_equation_0, values = (var_5868_cast_fp16_7, var_5936_cast_fp16))[name = tensor("op_5964_cast_fp16")]; + tensor var_5966_equation_0 = const()[name = tensor("op_5966_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5966_cast_fp16 = einsum(equation = var_5966_equation_0, values = (var_5868_cast_fp16_8, var_5937_cast_fp16))[name = tensor("op_5966_cast_fp16")]; + tensor var_5968_equation_0 = const()[name = tensor("op_5968_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5968_cast_fp16 = einsum(equation = var_5968_equation_0, values = (var_5868_cast_fp16_9, var_5938_cast_fp16))[name = tensor("op_5968_cast_fp16")]; + tensor var_5970_equation_0 = const()[name = tensor("op_5970_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5970_cast_fp16 = einsum(equation = var_5970_equation_0, values = (var_5868_cast_fp16_10, var_5939_cast_fp16))[name = tensor("op_5970_cast_fp16")]; + tensor var_5972_equation_0 = const()[name = tensor("op_5972_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5972_cast_fp16 = einsum(equation = var_5972_equation_0, values = (var_5868_cast_fp16_11, var_5940_cast_fp16))[name = tensor("op_5972_cast_fp16")]; + tensor var_5974_equation_0 = const()[name = tensor("op_5974_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5974_cast_fp16 = einsum(equation = var_5974_equation_0, values = (var_5868_cast_fp16_12, var_5941_cast_fp16))[name = tensor("op_5974_cast_fp16")]; + tensor var_5976_equation_0 = const()[name = tensor("op_5976_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5976_cast_fp16 = einsum(equation = var_5976_equation_0, values = (var_5868_cast_fp16_13, var_5942_cast_fp16))[name = tensor("op_5976_cast_fp16")]; + tensor var_5978_equation_0 = const()[name = tensor("op_5978_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5978_cast_fp16 = einsum(equation = var_5978_equation_0, values = (var_5868_cast_fp16_14, var_5943_cast_fp16))[name = tensor("op_5978_cast_fp16")]; + tensor var_5980_equation_0 = const()[name = tensor("op_5980_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5980_cast_fp16 = einsum(equation = var_5980_equation_0, values = (var_5868_cast_fp16_15, var_5944_cast_fp16))[name = tensor("op_5980_cast_fp16")]; + tensor var_5982_equation_0 = const()[name = tensor("op_5982_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5982_cast_fp16 = einsum(equation = var_5982_equation_0, values = (var_5868_cast_fp16_16, var_5945_cast_fp16))[name = tensor("op_5982_cast_fp16")]; + tensor var_5984_equation_0 = const()[name = tensor("op_5984_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5984_cast_fp16 = einsum(equation = var_5984_equation_0, values = (var_5868_cast_fp16_17, var_5946_cast_fp16))[name = tensor("op_5984_cast_fp16")]; + tensor var_5986_equation_0 = const()[name = tensor("op_5986_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5986_cast_fp16 = einsum(equation = var_5986_equation_0, values = (var_5868_cast_fp16_18, var_5947_cast_fp16))[name = tensor("op_5986_cast_fp16")]; + tensor var_5988_equation_0 = const()[name = tensor("op_5988_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5988_cast_fp16 = einsum(equation = var_5988_equation_0, values = (var_5868_cast_fp16_19, var_5948_cast_fp16))[name = tensor("op_5988_cast_fp16")]; + tensor input_215_interleave_0 = const()[name = tensor("input_215_interleave_0"), val = tensor(false)]; + tensor input_215_cast_fp16 = concat(axis = var_5773, interleave = input_215_interleave_0, values = (var_5950_cast_fp16, var_5952_cast_fp16, var_5954_cast_fp16, var_5956_cast_fp16, var_5958_cast_fp16, var_5960_cast_fp16, var_5962_cast_fp16, var_5964_cast_fp16, var_5966_cast_fp16, var_5968_cast_fp16, var_5970_cast_fp16, var_5972_cast_fp16, var_5974_cast_fp16, var_5976_cast_fp16, var_5978_cast_fp16, var_5980_cast_fp16, var_5982_cast_fp16, var_5984_cast_fp16, var_5986_cast_fp16, var_5988_cast_fp16))[name = tensor("input_215_cast_fp16")]; + tensor var_5997_pad_type_0 = const()[name = tensor("op_5997_pad_type_0"), val = tensor("valid")]; + tensor var_5997_strides_0 = const()[name = tensor("op_5997_strides_0"), val = tensor([1, 1])]; + tensor var_5997_pad_0 = const()[name = tensor("op_5997_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5997_dilations_0 = const()[name = tensor("op_5997_dilations_0"), val = tensor([1, 1])]; + tensor var_5997_groups_0 = const()[name = tensor("op_5997_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_out_weight_to_fp16 = const()[name = tensor("blocks_21_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850918912)))]; + tensor blocks_21_attn_out_bias_to_fp16 = const()[name = tensor("blocks_21_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854195776)))]; + tensor var_5997_cast_fp16 = conv(bias = blocks_21_attn_out_bias_to_fp16, dilations = var_5997_dilations_0, groups = var_5997_groups_0, pad = var_5997_pad_0, pad_type = var_5997_pad_type_0, strides = var_5997_strides_0, weight = blocks_21_attn_out_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("op_5997_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = var_5997_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor input_217_axes_0 = const()[name = tensor("input_217_axes_0"), val = tensor([1])]; + tensor input_217_gamma_0_to_fp16 = const()[name = tensor("input_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854198400)))]; + tensor input_217_beta_0_to_fp16 = const()[name = tensor("input_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854201024)))]; + tensor var_6007_to_fp16 = const()[name = tensor("op_6007_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_217_cast_fp16 = layer_norm(axes = input_217_axes_0, beta = input_217_beta_0_to_fp16, epsilon = var_6007_to_fp16, gamma = input_217_gamma_0_to_fp16, x = inputs_87_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("valid")]; + tensor input_219_strides_0 = const()[name = tensor("input_219_strides_0"), val = tensor([1, 1])]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_219_dilations_0 = const()[name = tensor("input_219_dilations_0"), val = tensor([1, 1])]; + tensor input_219_groups_0 = const()[name = tensor("input_219_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854203648)))]; + tensor blocks_21_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867310912)))]; + tensor input_219_cast_fp16 = conv(bias = blocks_21_mlp_0_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = blocks_21_mlp_0_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_6033_pad_type_0 = const()[name = tensor("op_6033_pad_type_0"), val = tensor("valid")]; + tensor var_6033_strides_0 = const()[name = tensor("op_6033_strides_0"), val = tensor([1, 1])]; + tensor var_6033_pad_0 = const()[name = tensor("op_6033_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6033_dilations_0 = const()[name = tensor("op_6033_dilations_0"), val = tensor([1, 1])]; + tensor var_6033_groups_0 = const()[name = tensor("op_6033_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867321216)))]; + tensor blocks_21_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880428480)))]; + tensor var_6033_cast_fp16 = conv(bias = blocks_21_mlp_2_bias_to_fp16, dilations = var_6033_dilations_0, groups = var_6033_groups_0, pad = var_6033_pad_0, pad_type = var_6033_pad_type_0, strides = var_6033_strides_0, weight = blocks_21_mlp_2_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("op_6033_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_6033_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(1)]; + tensor input_223_axes_0 = const()[name = tensor("input_223_axes_0"), val = tensor([1])]; + tensor input_223_gamma_0_to_fp16 = const()[name = tensor("input_223_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880431104)))]; + tensor input_223_beta_0_to_fp16 = const()[name = tensor("input_223_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880433728)))]; + tensor var_6058_to_fp16 = const()[name = tensor("op_6058_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_223_cast_fp16 = layer_norm(axes = input_223_axes_0, beta = input_223_beta_0_to_fp16, epsilon = var_6058_to_fp16, gamma = input_223_gamma_0_to_fp16, x = inputs_89_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor q_45_pad_type_0 = const()[name = tensor("q_45_pad_type_0"), val = tensor("valid")]; + tensor q_45_strides_0 = const()[name = tensor("q_45_strides_0"), val = tensor([1, 1])]; + tensor q_45_pad_0 = const()[name = tensor("q_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_45_dilations_0 = const()[name = tensor("q_45_dilations_0"), val = tensor([1, 1])]; + tensor q_45_groups_0 = const()[name = tensor("q_45_groups_0"), val = tensor(1)]; + tensor var_6093_weight_0_to_fp16 = const()[name = tensor("op_6093_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880436352)))]; + tensor var_6093_bias_0_to_fp16 = const()[name = tensor("op_6093_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883713216)))]; + tensor var_6093_cast_fp16 = conv(bias = var_6093_bias_0_to_fp16, dilations = q_45_dilations_0, groups = q_45_groups_0, pad = q_45_pad_0, pad_type = q_45_pad_type_0, strides = q_45_strides_0, weight = var_6093_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("op_6093_cast_fp16")]; + tensor k_45_pad_type_0 = const()[name = tensor("k_45_pad_type_0"), val = tensor("valid")]; + tensor k_45_strides_0 = const()[name = tensor("k_45_strides_0"), val = tensor([1, 1])]; + tensor k_45_pad_0 = const()[name = tensor("k_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_45_dilations_0 = const()[name = tensor("k_45_dilations_0"), val = tensor([1, 1])]; + tensor k_45_groups_0 = const()[name = tensor("k_45_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_key_weight_to_fp16 = const()[name = tensor("blocks_22_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883715840)))]; + tensor k_45_cast_fp16 = conv(dilations = k_45_dilations_0, groups = k_45_groups_0, pad = k_45_pad_0, pad_type = k_45_pad_type_0, strides = k_45_strides_0, weight = blocks_22_attn_key_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor var_6091_pad_type_0 = const()[name = tensor("op_6091_pad_type_0"), val = tensor("valid")]; + tensor var_6091_strides_0 = const()[name = tensor("op_6091_strides_0"), val = tensor([1, 1])]; + tensor var_6091_pad_0 = const()[name = tensor("op_6091_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6091_dilations_0 = const()[name = tensor("op_6091_dilations_0"), val = tensor([1, 1])]; + tensor var_6091_groups_0 = const()[name = tensor("op_6091_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_value_weight_to_fp16 = const()[name = tensor("blocks_22_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886992704)))]; + tensor blocks_22_attn_value_bias_to_fp16 = const()[name = tensor("blocks_22_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890269568)))]; + tensor var_6091_cast_fp16 = conv(bias = blocks_22_attn_value_bias_to_fp16, dilations = var_6091_dilations_0, groups = var_6091_groups_0, pad = var_6091_pad_0, pad_type = var_6091_pad_type_0, strides = var_6091_strides_0, weight = blocks_22_attn_value_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor tile_66 = const()[name = tensor("tile_66"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6094_axis_0 = const()[name = tensor("op_6094_axis_0"), val = tensor(1)]; + tensor var_6094_cast_fp16_0, tensor var_6094_cast_fp16_1, tensor var_6094_cast_fp16_2, tensor var_6094_cast_fp16_3, tensor var_6094_cast_fp16_4, tensor var_6094_cast_fp16_5, tensor var_6094_cast_fp16_6, tensor var_6094_cast_fp16_7, tensor var_6094_cast_fp16_8, tensor var_6094_cast_fp16_9, tensor var_6094_cast_fp16_10, tensor var_6094_cast_fp16_11, tensor var_6094_cast_fp16_12, tensor var_6094_cast_fp16_13, tensor var_6094_cast_fp16_14, tensor var_6094_cast_fp16_15, tensor var_6094_cast_fp16_16, tensor var_6094_cast_fp16_17, tensor var_6094_cast_fp16_18, tensor var_6094_cast_fp16_19 = split(axis = var_6094_axis_0, split_sizes = tile_66, x = var_6093_cast_fp16)[name = tensor("op_6094_cast_fp16")]; + tensor var_6115_perm_0 = const()[name = tensor("op_6115_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_67 = const()[name = tensor("tile_67"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6116_axis_0 = const()[name = tensor("op_6116_axis_0"), val = tensor(3)]; + tensor var_6115_cast_fp16 = transpose(perm = var_6115_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_10")]; + tensor var_6116_cast_fp16_0, tensor var_6116_cast_fp16_1, tensor var_6116_cast_fp16_2, tensor var_6116_cast_fp16_3, tensor var_6116_cast_fp16_4, tensor var_6116_cast_fp16_5, tensor var_6116_cast_fp16_6, tensor var_6116_cast_fp16_7, tensor var_6116_cast_fp16_8, tensor var_6116_cast_fp16_9, tensor var_6116_cast_fp16_10, tensor var_6116_cast_fp16_11, tensor var_6116_cast_fp16_12, tensor var_6116_cast_fp16_13, tensor var_6116_cast_fp16_14, tensor var_6116_cast_fp16_15, tensor var_6116_cast_fp16_16, tensor var_6116_cast_fp16_17, tensor var_6116_cast_fp16_18, tensor var_6116_cast_fp16_19 = split(axis = var_6116_axis_0, split_sizes = tile_67, x = var_6115_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor tile_68 = const()[name = tensor("tile_68"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6137_axis_0 = const()[name = tensor("op_6137_axis_0"), val = tensor(1)]; + tensor var_6137_cast_fp16_0, tensor var_6137_cast_fp16_1, tensor var_6137_cast_fp16_2, tensor var_6137_cast_fp16_3, tensor var_6137_cast_fp16_4, tensor var_6137_cast_fp16_5, tensor var_6137_cast_fp16_6, tensor var_6137_cast_fp16_7, tensor var_6137_cast_fp16_8, tensor var_6137_cast_fp16_9, tensor var_6137_cast_fp16_10, tensor var_6137_cast_fp16_11, tensor var_6137_cast_fp16_12, tensor var_6137_cast_fp16_13, tensor var_6137_cast_fp16_14, tensor var_6137_cast_fp16_15, tensor var_6137_cast_fp16_16, tensor var_6137_cast_fp16_17, tensor var_6137_cast_fp16_18, tensor var_6137_cast_fp16_19 = split(axis = var_6137_axis_0, split_sizes = tile_68, x = var_6091_cast_fp16)[name = tensor("op_6137_cast_fp16")]; + tensor aw_881_equation_0 = const()[name = tensor("aw_881_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_881_cast_fp16 = einsum(equation = aw_881_equation_0, values = (var_6116_cast_fp16_0, var_6094_cast_fp16_0))[name = tensor("aw_881_cast_fp16")]; + tensor aw_883_equation_0 = const()[name = tensor("aw_883_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_883_cast_fp16 = einsum(equation = aw_883_equation_0, values = (var_6116_cast_fp16_1, var_6094_cast_fp16_1))[name = tensor("aw_883_cast_fp16")]; + tensor aw_885_equation_0 = const()[name = tensor("aw_885_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_885_cast_fp16 = einsum(equation = aw_885_equation_0, values = (var_6116_cast_fp16_2, var_6094_cast_fp16_2))[name = tensor("aw_885_cast_fp16")]; + tensor aw_887_equation_0 = const()[name = tensor("aw_887_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_887_cast_fp16 = einsum(equation = aw_887_equation_0, values = (var_6116_cast_fp16_3, var_6094_cast_fp16_3))[name = tensor("aw_887_cast_fp16")]; + tensor aw_889_equation_0 = const()[name = tensor("aw_889_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_889_cast_fp16 = einsum(equation = aw_889_equation_0, values = (var_6116_cast_fp16_4, var_6094_cast_fp16_4))[name = tensor("aw_889_cast_fp16")]; + tensor aw_891_equation_0 = const()[name = tensor("aw_891_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_891_cast_fp16 = einsum(equation = aw_891_equation_0, values = (var_6116_cast_fp16_5, var_6094_cast_fp16_5))[name = tensor("aw_891_cast_fp16")]; + tensor aw_893_equation_0 = const()[name = tensor("aw_893_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_893_cast_fp16 = einsum(equation = aw_893_equation_0, values = (var_6116_cast_fp16_6, var_6094_cast_fp16_6))[name = tensor("aw_893_cast_fp16")]; + tensor aw_895_equation_0 = const()[name = tensor("aw_895_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_895_cast_fp16 = einsum(equation = aw_895_equation_0, values = (var_6116_cast_fp16_7, var_6094_cast_fp16_7))[name = tensor("aw_895_cast_fp16")]; + tensor aw_897_equation_0 = const()[name = tensor("aw_897_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_897_cast_fp16 = einsum(equation = aw_897_equation_0, values = (var_6116_cast_fp16_8, var_6094_cast_fp16_8))[name = tensor("aw_897_cast_fp16")]; + tensor aw_899_equation_0 = const()[name = tensor("aw_899_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_899_cast_fp16 = einsum(equation = aw_899_equation_0, values = (var_6116_cast_fp16_9, var_6094_cast_fp16_9))[name = tensor("aw_899_cast_fp16")]; + tensor aw_901_equation_0 = const()[name = tensor("aw_901_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_901_cast_fp16 = einsum(equation = aw_901_equation_0, values = (var_6116_cast_fp16_10, var_6094_cast_fp16_10))[name = tensor("aw_901_cast_fp16")]; + tensor aw_903_equation_0 = const()[name = tensor("aw_903_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_903_cast_fp16 = einsum(equation = aw_903_equation_0, values = (var_6116_cast_fp16_11, var_6094_cast_fp16_11))[name = tensor("aw_903_cast_fp16")]; + tensor aw_905_equation_0 = const()[name = tensor("aw_905_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_905_cast_fp16 = einsum(equation = aw_905_equation_0, values = (var_6116_cast_fp16_12, var_6094_cast_fp16_12))[name = tensor("aw_905_cast_fp16")]; + tensor aw_907_equation_0 = const()[name = tensor("aw_907_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_907_cast_fp16 = einsum(equation = aw_907_equation_0, values = (var_6116_cast_fp16_13, var_6094_cast_fp16_13))[name = tensor("aw_907_cast_fp16")]; + tensor aw_909_equation_0 = const()[name = tensor("aw_909_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_909_cast_fp16 = einsum(equation = aw_909_equation_0, values = (var_6116_cast_fp16_14, var_6094_cast_fp16_14))[name = tensor("aw_909_cast_fp16")]; + tensor aw_911_equation_0 = const()[name = tensor("aw_911_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_911_cast_fp16 = einsum(equation = aw_911_equation_0, values = (var_6116_cast_fp16_15, var_6094_cast_fp16_15))[name = tensor("aw_911_cast_fp16")]; + tensor aw_913_equation_0 = const()[name = tensor("aw_913_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_913_cast_fp16 = einsum(equation = aw_913_equation_0, values = (var_6116_cast_fp16_16, var_6094_cast_fp16_16))[name = tensor("aw_913_cast_fp16")]; + tensor aw_915_equation_0 = const()[name = tensor("aw_915_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_915_cast_fp16 = einsum(equation = aw_915_equation_0, values = (var_6116_cast_fp16_17, var_6094_cast_fp16_17))[name = tensor("aw_915_cast_fp16")]; + tensor aw_917_equation_0 = const()[name = tensor("aw_917_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_917_cast_fp16 = einsum(equation = aw_917_equation_0, values = (var_6116_cast_fp16_18, var_6094_cast_fp16_18))[name = tensor("aw_917_cast_fp16")]; + tensor aw_919_equation_0 = const()[name = tensor("aw_919_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_919_cast_fp16 = einsum(equation = aw_919_equation_0, values = (var_6116_cast_fp16_19, var_6094_cast_fp16_19))[name = tensor("aw_919_cast_fp16")]; + tensor var_6198_cast_fp16 = softmax(axis = var_6042, x = aw_881_cast_fp16)[name = tensor("op_6198_cast_fp16")]; + tensor var_6199_cast_fp16 = softmax(axis = var_6042, x = aw_883_cast_fp16)[name = tensor("op_6199_cast_fp16")]; + tensor var_6200_cast_fp16 = softmax(axis = var_6042, x = aw_885_cast_fp16)[name = tensor("op_6200_cast_fp16")]; + tensor var_6201_cast_fp16 = softmax(axis = var_6042, x = aw_887_cast_fp16)[name = tensor("op_6201_cast_fp16")]; + tensor var_6202_cast_fp16 = softmax(axis = var_6042, x = aw_889_cast_fp16)[name = tensor("op_6202_cast_fp16")]; + tensor var_6203_cast_fp16 = softmax(axis = var_6042, x = aw_891_cast_fp16)[name = tensor("op_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = softmax(axis = var_6042, x = aw_893_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_cast_fp16 = softmax(axis = var_6042, x = aw_895_cast_fp16)[name = tensor("op_6205_cast_fp16")]; + tensor var_6206_cast_fp16 = softmax(axis = var_6042, x = aw_897_cast_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor var_6207_cast_fp16 = softmax(axis = var_6042, x = aw_899_cast_fp16)[name = tensor("op_6207_cast_fp16")]; + tensor var_6208_cast_fp16 = softmax(axis = var_6042, x = aw_901_cast_fp16)[name = tensor("op_6208_cast_fp16")]; + tensor var_6209_cast_fp16 = softmax(axis = var_6042, x = aw_903_cast_fp16)[name = tensor("op_6209_cast_fp16")]; + tensor var_6210_cast_fp16 = softmax(axis = var_6042, x = aw_905_cast_fp16)[name = tensor("op_6210_cast_fp16")]; + tensor var_6211_cast_fp16 = softmax(axis = var_6042, x = aw_907_cast_fp16)[name = tensor("op_6211_cast_fp16")]; + tensor var_6212_cast_fp16 = softmax(axis = var_6042, x = aw_909_cast_fp16)[name = tensor("op_6212_cast_fp16")]; + tensor var_6213_cast_fp16 = softmax(axis = var_6042, x = aw_911_cast_fp16)[name = tensor("op_6213_cast_fp16")]; + tensor var_6214_cast_fp16 = softmax(axis = var_6042, x = aw_913_cast_fp16)[name = tensor("op_6214_cast_fp16")]; + tensor var_6215_cast_fp16 = softmax(axis = var_6042, x = aw_915_cast_fp16)[name = tensor("op_6215_cast_fp16")]; + tensor var_6216_cast_fp16 = softmax(axis = var_6042, x = aw_917_cast_fp16)[name = tensor("op_6216_cast_fp16")]; + tensor var_6217_cast_fp16 = softmax(axis = var_6042, x = aw_919_cast_fp16)[name = tensor("op_6217_cast_fp16")]; + tensor var_6219_equation_0 = const()[name = tensor("op_6219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6219_cast_fp16 = einsum(equation = var_6219_equation_0, values = (var_6137_cast_fp16_0, var_6198_cast_fp16))[name = tensor("op_6219_cast_fp16")]; + tensor var_6221_equation_0 = const()[name = tensor("op_6221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6221_cast_fp16 = einsum(equation = var_6221_equation_0, values = (var_6137_cast_fp16_1, var_6199_cast_fp16))[name = tensor("op_6221_cast_fp16")]; + tensor var_6223_equation_0 = const()[name = tensor("op_6223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6223_cast_fp16 = einsum(equation = var_6223_equation_0, values = (var_6137_cast_fp16_2, var_6200_cast_fp16))[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_equation_0 = const()[name = tensor("op_6225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6225_cast_fp16 = einsum(equation = var_6225_equation_0, values = (var_6137_cast_fp16_3, var_6201_cast_fp16))[name = tensor("op_6225_cast_fp16")]; + tensor var_6227_equation_0 = const()[name = tensor("op_6227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6227_cast_fp16 = einsum(equation = var_6227_equation_0, values = (var_6137_cast_fp16_4, var_6202_cast_fp16))[name = tensor("op_6227_cast_fp16")]; + tensor var_6229_equation_0 = const()[name = tensor("op_6229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6229_cast_fp16 = einsum(equation = var_6229_equation_0, values = (var_6137_cast_fp16_5, var_6203_cast_fp16))[name = tensor("op_6229_cast_fp16")]; + tensor var_6231_equation_0 = const()[name = tensor("op_6231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6231_cast_fp16 = einsum(equation = var_6231_equation_0, values = (var_6137_cast_fp16_6, var_6204_cast_fp16))[name = tensor("op_6231_cast_fp16")]; + tensor var_6233_equation_0 = const()[name = tensor("op_6233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6233_cast_fp16 = einsum(equation = var_6233_equation_0, values = (var_6137_cast_fp16_7, var_6205_cast_fp16))[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_equation_0 = const()[name = tensor("op_6235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6235_cast_fp16 = einsum(equation = var_6235_equation_0, values = (var_6137_cast_fp16_8, var_6206_cast_fp16))[name = tensor("op_6235_cast_fp16")]; + tensor var_6237_equation_0 = const()[name = tensor("op_6237_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6237_cast_fp16 = einsum(equation = var_6237_equation_0, values = (var_6137_cast_fp16_9, var_6207_cast_fp16))[name = tensor("op_6237_cast_fp16")]; + tensor var_6239_equation_0 = const()[name = tensor("op_6239_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6239_cast_fp16 = einsum(equation = var_6239_equation_0, values = (var_6137_cast_fp16_10, var_6208_cast_fp16))[name = tensor("op_6239_cast_fp16")]; + tensor var_6241_equation_0 = const()[name = tensor("op_6241_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6241_cast_fp16 = einsum(equation = var_6241_equation_0, values = (var_6137_cast_fp16_11, var_6209_cast_fp16))[name = tensor("op_6241_cast_fp16")]; + tensor var_6243_equation_0 = const()[name = tensor("op_6243_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6243_cast_fp16 = einsum(equation = var_6243_equation_0, values = (var_6137_cast_fp16_12, var_6210_cast_fp16))[name = tensor("op_6243_cast_fp16")]; + tensor var_6245_equation_0 = const()[name = tensor("op_6245_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6245_cast_fp16 = einsum(equation = var_6245_equation_0, values = (var_6137_cast_fp16_13, var_6211_cast_fp16))[name = tensor("op_6245_cast_fp16")]; + tensor var_6247_equation_0 = const()[name = tensor("op_6247_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6247_cast_fp16 = einsum(equation = var_6247_equation_0, values = (var_6137_cast_fp16_14, var_6212_cast_fp16))[name = tensor("op_6247_cast_fp16")]; + tensor var_6249_equation_0 = const()[name = tensor("op_6249_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6249_cast_fp16 = einsum(equation = var_6249_equation_0, values = (var_6137_cast_fp16_15, var_6213_cast_fp16))[name = tensor("op_6249_cast_fp16")]; + tensor var_6251_equation_0 = const()[name = tensor("op_6251_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6251_cast_fp16 = einsum(equation = var_6251_equation_0, values = (var_6137_cast_fp16_16, var_6214_cast_fp16))[name = tensor("op_6251_cast_fp16")]; + tensor var_6253_equation_0 = const()[name = tensor("op_6253_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6253_cast_fp16 = einsum(equation = var_6253_equation_0, values = (var_6137_cast_fp16_17, var_6215_cast_fp16))[name = tensor("op_6253_cast_fp16")]; + tensor var_6255_equation_0 = const()[name = tensor("op_6255_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6255_cast_fp16 = einsum(equation = var_6255_equation_0, values = (var_6137_cast_fp16_18, var_6216_cast_fp16))[name = tensor("op_6255_cast_fp16")]; + tensor var_6257_equation_0 = const()[name = tensor("op_6257_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6257_cast_fp16 = einsum(equation = var_6257_equation_0, values = (var_6137_cast_fp16_19, var_6217_cast_fp16))[name = tensor("op_6257_cast_fp16")]; + tensor input_225_interleave_0 = const()[name = tensor("input_225_interleave_0"), val = tensor(false)]; + tensor input_225_cast_fp16 = concat(axis = var_6042, interleave = input_225_interleave_0, values = (var_6219_cast_fp16, var_6221_cast_fp16, var_6223_cast_fp16, var_6225_cast_fp16, var_6227_cast_fp16, var_6229_cast_fp16, var_6231_cast_fp16, var_6233_cast_fp16, var_6235_cast_fp16, var_6237_cast_fp16, var_6239_cast_fp16, var_6241_cast_fp16, var_6243_cast_fp16, var_6245_cast_fp16, var_6247_cast_fp16, var_6249_cast_fp16, var_6251_cast_fp16, var_6253_cast_fp16, var_6255_cast_fp16, var_6257_cast_fp16))[name = tensor("input_225_cast_fp16")]; + tensor var_6266_pad_type_0 = const()[name = tensor("op_6266_pad_type_0"), val = tensor("valid")]; + tensor var_6266_strides_0 = const()[name = tensor("op_6266_strides_0"), val = tensor([1, 1])]; + tensor var_6266_pad_0 = const()[name = tensor("op_6266_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6266_dilations_0 = const()[name = tensor("op_6266_dilations_0"), val = tensor([1, 1])]; + tensor var_6266_groups_0 = const()[name = tensor("op_6266_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_out_weight_to_fp16 = const()[name = tensor("blocks_22_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890272192)))]; + tensor blocks_22_attn_out_bias_to_fp16 = const()[name = tensor("blocks_22_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893549056)))]; + tensor var_6266_cast_fp16 = conv(bias = blocks_22_attn_out_bias_to_fp16, dilations = var_6266_dilations_0, groups = var_6266_groups_0, pad = var_6266_pad_0, pad_type = var_6266_pad_type_0, strides = var_6266_strides_0, weight = blocks_22_attn_out_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("op_6266_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = var_6266_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([1])]; + tensor input_227_gamma_0_to_fp16 = const()[name = tensor("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893551680)))]; + tensor input_227_beta_0_to_fp16 = const()[name = tensor("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893554304)))]; + tensor var_6276_to_fp16 = const()[name = tensor("op_6276_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = input_227_beta_0_to_fp16, epsilon = var_6276_to_fp16, gamma = input_227_gamma_0_to_fp16, x = inputs_91_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("valid")]; + tensor input_229_strides_0 = const()[name = tensor("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = tensor("input_229_dilations_0"), val = tensor([1, 1])]; + tensor input_229_groups_0 = const()[name = tensor("input_229_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893556928)))]; + tensor blocks_22_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906664192)))]; + tensor input_229_cast_fp16 = conv(bias = blocks_22_mlp_0_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = blocks_22_mlp_0_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_6302_pad_type_0 = const()[name = tensor("op_6302_pad_type_0"), val = tensor("valid")]; + tensor var_6302_strides_0 = const()[name = tensor("op_6302_strides_0"), val = tensor([1, 1])]; + tensor var_6302_pad_0 = const()[name = tensor("op_6302_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6302_dilations_0 = const()[name = tensor("op_6302_dilations_0"), val = tensor([1, 1])]; + tensor var_6302_groups_0 = const()[name = tensor("op_6302_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906674496)))]; + tensor blocks_22_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919781760)))]; + tensor var_6302_cast_fp16 = conv(bias = blocks_22_mlp_2_bias_to_fp16, dilations = var_6302_dilations_0, groups = var_6302_groups_0, pad = var_6302_pad_0, pad_type = var_6302_pad_type_0, strides = var_6302_strides_0, weight = blocks_22_mlp_2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("op_6302_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_6302_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_6311 = const()[name = tensor("op_6311"), val = tensor(1)]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([1])]; + tensor input_233_gamma_0_to_fp16 = const()[name = tensor("input_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919784384)))]; + tensor input_233_beta_0_to_fp16 = const()[name = tensor("input_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919787008)))]; + tensor var_6327_to_fp16 = const()[name = tensor("op_6327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = input_233_beta_0_to_fp16, epsilon = var_6327_to_fp16, gamma = input_233_gamma_0_to_fp16, x = inputs_93_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor q_47_pad_type_0 = const()[name = tensor("q_47_pad_type_0"), val = tensor("valid")]; + tensor q_47_strides_0 = const()[name = tensor("q_47_strides_0"), val = tensor([1, 1])]; + tensor q_47_pad_0 = const()[name = tensor("q_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_47_dilations_0 = const()[name = tensor("q_47_dilations_0"), val = tensor([1, 1])]; + tensor q_47_groups_0 = const()[name = tensor("q_47_groups_0"), val = tensor(1)]; + tensor var_6362_weight_0_to_fp16 = const()[name = tensor("op_6362_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919789632)))]; + tensor var_6362_bias_0_to_fp16 = const()[name = tensor("op_6362_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(923066496)))]; + tensor var_6362_cast_fp16 = conv(bias = var_6362_bias_0_to_fp16, dilations = q_47_dilations_0, groups = q_47_groups_0, pad = q_47_pad_0, pad_type = q_47_pad_type_0, strides = q_47_strides_0, weight = var_6362_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("op_6362_cast_fp16")]; + tensor k_47_pad_type_0 = const()[name = tensor("k_47_pad_type_0"), val = tensor("valid")]; + tensor k_47_strides_0 = const()[name = tensor("k_47_strides_0"), val = tensor([1, 1])]; + tensor k_47_pad_0 = const()[name = tensor("k_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_47_dilations_0 = const()[name = tensor("k_47_dilations_0"), val = tensor([1, 1])]; + tensor k_47_groups_0 = const()[name = tensor("k_47_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_key_weight_to_fp16 = const()[name = tensor("blocks_23_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(923069120)))]; + tensor k_47_cast_fp16 = conv(dilations = k_47_dilations_0, groups = k_47_groups_0, pad = k_47_pad_0, pad_type = k_47_pad_type_0, strides = k_47_strides_0, weight = blocks_23_attn_key_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor var_6360_pad_type_0 = const()[name = tensor("op_6360_pad_type_0"), val = tensor("valid")]; + tensor var_6360_strides_0 = const()[name = tensor("op_6360_strides_0"), val = tensor([1, 1])]; + tensor var_6360_pad_0 = const()[name = tensor("op_6360_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6360_dilations_0 = const()[name = tensor("op_6360_dilations_0"), val = tensor([1, 1])]; + tensor var_6360_groups_0 = const()[name = tensor("op_6360_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_value_weight_to_fp16 = const()[name = tensor("blocks_23_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926345984)))]; + tensor blocks_23_attn_value_bias_to_fp16 = const()[name = tensor("blocks_23_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929622848)))]; + tensor var_6360_cast_fp16 = conv(bias = blocks_23_attn_value_bias_to_fp16, dilations = var_6360_dilations_0, groups = var_6360_groups_0, pad = var_6360_pad_0, pad_type = var_6360_pad_type_0, strides = var_6360_strides_0, weight = blocks_23_attn_value_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("op_6360_cast_fp16")]; + tensor tile_69 = const()[name = tensor("tile_69"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6363_axis_0 = const()[name = tensor("op_6363_axis_0"), val = tensor(1)]; + tensor var_6363_cast_fp16_0, tensor var_6363_cast_fp16_1, tensor var_6363_cast_fp16_2, tensor var_6363_cast_fp16_3, tensor var_6363_cast_fp16_4, tensor var_6363_cast_fp16_5, tensor var_6363_cast_fp16_6, tensor var_6363_cast_fp16_7, tensor var_6363_cast_fp16_8, tensor var_6363_cast_fp16_9, tensor var_6363_cast_fp16_10, tensor var_6363_cast_fp16_11, tensor var_6363_cast_fp16_12, tensor var_6363_cast_fp16_13, tensor var_6363_cast_fp16_14, tensor var_6363_cast_fp16_15, tensor var_6363_cast_fp16_16, tensor var_6363_cast_fp16_17, tensor var_6363_cast_fp16_18, tensor var_6363_cast_fp16_19 = split(axis = var_6363_axis_0, split_sizes = tile_69, x = var_6362_cast_fp16)[name = tensor("op_6363_cast_fp16")]; + tensor var_6384_perm_0 = const()[name = tensor("op_6384_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_70 = const()[name = tensor("tile_70"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6385_axis_0 = const()[name = tensor("op_6385_axis_0"), val = tensor(3)]; + tensor var_6384_cast_fp16 = transpose(perm = var_6384_perm_0, x = k_47_cast_fp16)[name = tensor("transpose_9")]; + tensor var_6385_cast_fp16_0, tensor var_6385_cast_fp16_1, tensor var_6385_cast_fp16_2, tensor var_6385_cast_fp16_3, tensor var_6385_cast_fp16_4, tensor var_6385_cast_fp16_5, tensor var_6385_cast_fp16_6, tensor var_6385_cast_fp16_7, tensor var_6385_cast_fp16_8, tensor var_6385_cast_fp16_9, tensor var_6385_cast_fp16_10, tensor var_6385_cast_fp16_11, tensor var_6385_cast_fp16_12, tensor var_6385_cast_fp16_13, tensor var_6385_cast_fp16_14, tensor var_6385_cast_fp16_15, tensor var_6385_cast_fp16_16, tensor var_6385_cast_fp16_17, tensor var_6385_cast_fp16_18, tensor var_6385_cast_fp16_19 = split(axis = var_6385_axis_0, split_sizes = tile_70, x = var_6384_cast_fp16)[name = tensor("op_6385_cast_fp16")]; + tensor tile_71 = const()[name = tensor("tile_71"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6406_axis_0 = const()[name = tensor("op_6406_axis_0"), val = tensor(1)]; + tensor var_6406_cast_fp16_0, tensor var_6406_cast_fp16_1, tensor var_6406_cast_fp16_2, tensor var_6406_cast_fp16_3, tensor var_6406_cast_fp16_4, tensor var_6406_cast_fp16_5, tensor var_6406_cast_fp16_6, tensor var_6406_cast_fp16_7, tensor var_6406_cast_fp16_8, tensor var_6406_cast_fp16_9, tensor var_6406_cast_fp16_10, tensor var_6406_cast_fp16_11, tensor var_6406_cast_fp16_12, tensor var_6406_cast_fp16_13, tensor var_6406_cast_fp16_14, tensor var_6406_cast_fp16_15, tensor var_6406_cast_fp16_16, tensor var_6406_cast_fp16_17, tensor var_6406_cast_fp16_18, tensor var_6406_cast_fp16_19 = split(axis = var_6406_axis_0, split_sizes = tile_71, x = var_6360_cast_fp16)[name = tensor("op_6406_cast_fp16")]; + tensor aw_921_equation_0 = const()[name = tensor("aw_921_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_921_cast_fp16 = einsum(equation = aw_921_equation_0, values = (var_6385_cast_fp16_0, var_6363_cast_fp16_0))[name = tensor("aw_921_cast_fp16")]; + tensor aw_923_equation_0 = const()[name = tensor("aw_923_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_923_cast_fp16 = einsum(equation = aw_923_equation_0, values = (var_6385_cast_fp16_1, var_6363_cast_fp16_1))[name = tensor("aw_923_cast_fp16")]; + tensor aw_925_equation_0 = const()[name = tensor("aw_925_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_925_cast_fp16 = einsum(equation = aw_925_equation_0, values = (var_6385_cast_fp16_2, var_6363_cast_fp16_2))[name = tensor("aw_925_cast_fp16")]; + tensor aw_927_equation_0 = const()[name = tensor("aw_927_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_927_cast_fp16 = einsum(equation = aw_927_equation_0, values = (var_6385_cast_fp16_3, var_6363_cast_fp16_3))[name = tensor("aw_927_cast_fp16")]; + tensor aw_929_equation_0 = const()[name = tensor("aw_929_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_929_cast_fp16 = einsum(equation = aw_929_equation_0, values = (var_6385_cast_fp16_4, var_6363_cast_fp16_4))[name = tensor("aw_929_cast_fp16")]; + tensor aw_931_equation_0 = const()[name = tensor("aw_931_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_931_cast_fp16 = einsum(equation = aw_931_equation_0, values = (var_6385_cast_fp16_5, var_6363_cast_fp16_5))[name = tensor("aw_931_cast_fp16")]; + tensor aw_933_equation_0 = const()[name = tensor("aw_933_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_933_cast_fp16 = einsum(equation = aw_933_equation_0, values = (var_6385_cast_fp16_6, var_6363_cast_fp16_6))[name = tensor("aw_933_cast_fp16")]; + tensor aw_935_equation_0 = const()[name = tensor("aw_935_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_935_cast_fp16 = einsum(equation = aw_935_equation_0, values = (var_6385_cast_fp16_7, var_6363_cast_fp16_7))[name = tensor("aw_935_cast_fp16")]; + tensor aw_937_equation_0 = const()[name = tensor("aw_937_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_937_cast_fp16 = einsum(equation = aw_937_equation_0, values = (var_6385_cast_fp16_8, var_6363_cast_fp16_8))[name = tensor("aw_937_cast_fp16")]; + tensor aw_939_equation_0 = const()[name = tensor("aw_939_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_939_cast_fp16 = einsum(equation = aw_939_equation_0, values = (var_6385_cast_fp16_9, var_6363_cast_fp16_9))[name = tensor("aw_939_cast_fp16")]; + tensor aw_941_equation_0 = const()[name = tensor("aw_941_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_941_cast_fp16 = einsum(equation = aw_941_equation_0, values = (var_6385_cast_fp16_10, var_6363_cast_fp16_10))[name = tensor("aw_941_cast_fp16")]; + tensor aw_943_equation_0 = const()[name = tensor("aw_943_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_943_cast_fp16 = einsum(equation = aw_943_equation_0, values = (var_6385_cast_fp16_11, var_6363_cast_fp16_11))[name = tensor("aw_943_cast_fp16")]; + tensor aw_945_equation_0 = const()[name = tensor("aw_945_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_945_cast_fp16 = einsum(equation = aw_945_equation_0, values = (var_6385_cast_fp16_12, var_6363_cast_fp16_12))[name = tensor("aw_945_cast_fp16")]; + tensor aw_947_equation_0 = const()[name = tensor("aw_947_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_947_cast_fp16 = einsum(equation = aw_947_equation_0, values = (var_6385_cast_fp16_13, var_6363_cast_fp16_13))[name = tensor("aw_947_cast_fp16")]; + tensor aw_949_equation_0 = const()[name = tensor("aw_949_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_949_cast_fp16 = einsum(equation = aw_949_equation_0, values = (var_6385_cast_fp16_14, var_6363_cast_fp16_14))[name = tensor("aw_949_cast_fp16")]; + tensor aw_951_equation_0 = const()[name = tensor("aw_951_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_951_cast_fp16 = einsum(equation = aw_951_equation_0, values = (var_6385_cast_fp16_15, var_6363_cast_fp16_15))[name = tensor("aw_951_cast_fp16")]; + tensor aw_953_equation_0 = const()[name = tensor("aw_953_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_953_cast_fp16 = einsum(equation = aw_953_equation_0, values = (var_6385_cast_fp16_16, var_6363_cast_fp16_16))[name = tensor("aw_953_cast_fp16")]; + tensor aw_955_equation_0 = const()[name = tensor("aw_955_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_955_cast_fp16 = einsum(equation = aw_955_equation_0, values = (var_6385_cast_fp16_17, var_6363_cast_fp16_17))[name = tensor("aw_955_cast_fp16")]; + tensor aw_957_equation_0 = const()[name = tensor("aw_957_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_957_cast_fp16 = einsum(equation = aw_957_equation_0, values = (var_6385_cast_fp16_18, var_6363_cast_fp16_18))[name = tensor("aw_957_cast_fp16")]; + tensor aw_959_equation_0 = const()[name = tensor("aw_959_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_959_cast_fp16 = einsum(equation = aw_959_equation_0, values = (var_6385_cast_fp16_19, var_6363_cast_fp16_19))[name = tensor("aw_959_cast_fp16")]; + tensor var_6467_cast_fp16 = softmax(axis = var_6311, x = aw_921_cast_fp16)[name = tensor("op_6467_cast_fp16")]; + tensor var_6468_cast_fp16 = softmax(axis = var_6311, x = aw_923_cast_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6469_cast_fp16 = softmax(axis = var_6311, x = aw_925_cast_fp16)[name = tensor("op_6469_cast_fp16")]; + tensor var_6470_cast_fp16 = softmax(axis = var_6311, x = aw_927_cast_fp16)[name = tensor("op_6470_cast_fp16")]; + tensor var_6471_cast_fp16 = softmax(axis = var_6311, x = aw_929_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6472_cast_fp16 = softmax(axis = var_6311, x = aw_931_cast_fp16)[name = tensor("op_6472_cast_fp16")]; + tensor var_6473_cast_fp16 = softmax(axis = var_6311, x = aw_933_cast_fp16)[name = tensor("op_6473_cast_fp16")]; + tensor var_6474_cast_fp16 = softmax(axis = var_6311, x = aw_935_cast_fp16)[name = tensor("op_6474_cast_fp16")]; + tensor var_6475_cast_fp16 = softmax(axis = var_6311, x = aw_937_cast_fp16)[name = tensor("op_6475_cast_fp16")]; + tensor var_6476_cast_fp16 = softmax(axis = var_6311, x = aw_939_cast_fp16)[name = tensor("op_6476_cast_fp16")]; + tensor var_6477_cast_fp16 = softmax(axis = var_6311, x = aw_941_cast_fp16)[name = tensor("op_6477_cast_fp16")]; + tensor var_6478_cast_fp16 = softmax(axis = var_6311, x = aw_943_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6479_cast_fp16 = softmax(axis = var_6311, x = aw_945_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor var_6480_cast_fp16 = softmax(axis = var_6311, x = aw_947_cast_fp16)[name = tensor("op_6480_cast_fp16")]; + tensor var_6481_cast_fp16 = softmax(axis = var_6311, x = aw_949_cast_fp16)[name = tensor("op_6481_cast_fp16")]; + tensor var_6482_cast_fp16 = softmax(axis = var_6311, x = aw_951_cast_fp16)[name = tensor("op_6482_cast_fp16")]; + tensor var_6483_cast_fp16 = softmax(axis = var_6311, x = aw_953_cast_fp16)[name = tensor("op_6483_cast_fp16")]; + tensor var_6484_cast_fp16 = softmax(axis = var_6311, x = aw_955_cast_fp16)[name = tensor("op_6484_cast_fp16")]; + tensor var_6485_cast_fp16 = softmax(axis = var_6311, x = aw_957_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor var_6486_cast_fp16 = softmax(axis = var_6311, x = aw_959_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6488_equation_0 = const()[name = tensor("op_6488_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6488_cast_fp16 = einsum(equation = var_6488_equation_0, values = (var_6406_cast_fp16_0, var_6467_cast_fp16))[name = tensor("op_6488_cast_fp16")]; + tensor var_6490_equation_0 = const()[name = tensor("op_6490_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6490_cast_fp16 = einsum(equation = var_6490_equation_0, values = (var_6406_cast_fp16_1, var_6468_cast_fp16))[name = tensor("op_6490_cast_fp16")]; + tensor var_6492_equation_0 = const()[name = tensor("op_6492_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6492_cast_fp16 = einsum(equation = var_6492_equation_0, values = (var_6406_cast_fp16_2, var_6469_cast_fp16))[name = tensor("op_6492_cast_fp16")]; + tensor var_6494_equation_0 = const()[name = tensor("op_6494_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6494_cast_fp16 = einsum(equation = var_6494_equation_0, values = (var_6406_cast_fp16_3, var_6470_cast_fp16))[name = tensor("op_6494_cast_fp16")]; + tensor var_6496_equation_0 = const()[name = tensor("op_6496_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6496_cast_fp16 = einsum(equation = var_6496_equation_0, values = (var_6406_cast_fp16_4, var_6471_cast_fp16))[name = tensor("op_6496_cast_fp16")]; + tensor var_6498_equation_0 = const()[name = tensor("op_6498_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6498_cast_fp16 = einsum(equation = var_6498_equation_0, values = (var_6406_cast_fp16_5, var_6472_cast_fp16))[name = tensor("op_6498_cast_fp16")]; + tensor var_6500_equation_0 = const()[name = tensor("op_6500_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6500_cast_fp16 = einsum(equation = var_6500_equation_0, values = (var_6406_cast_fp16_6, var_6473_cast_fp16))[name = tensor("op_6500_cast_fp16")]; + tensor var_6502_equation_0 = const()[name = tensor("op_6502_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6502_cast_fp16 = einsum(equation = var_6502_equation_0, values = (var_6406_cast_fp16_7, var_6474_cast_fp16))[name = tensor("op_6502_cast_fp16")]; + tensor var_6504_equation_0 = const()[name = tensor("op_6504_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6504_cast_fp16 = einsum(equation = var_6504_equation_0, values = (var_6406_cast_fp16_8, var_6475_cast_fp16))[name = tensor("op_6504_cast_fp16")]; + tensor var_6506_equation_0 = const()[name = tensor("op_6506_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6506_cast_fp16 = einsum(equation = var_6506_equation_0, values = (var_6406_cast_fp16_9, var_6476_cast_fp16))[name = tensor("op_6506_cast_fp16")]; + tensor var_6508_equation_0 = const()[name = tensor("op_6508_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6508_cast_fp16 = einsum(equation = var_6508_equation_0, values = (var_6406_cast_fp16_10, var_6477_cast_fp16))[name = tensor("op_6508_cast_fp16")]; + tensor var_6510_equation_0 = const()[name = tensor("op_6510_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6510_cast_fp16 = einsum(equation = var_6510_equation_0, values = (var_6406_cast_fp16_11, var_6478_cast_fp16))[name = tensor("op_6510_cast_fp16")]; + tensor var_6512_equation_0 = const()[name = tensor("op_6512_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6512_cast_fp16 = einsum(equation = var_6512_equation_0, values = (var_6406_cast_fp16_12, var_6479_cast_fp16))[name = tensor("op_6512_cast_fp16")]; + tensor var_6514_equation_0 = const()[name = tensor("op_6514_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6514_cast_fp16 = einsum(equation = var_6514_equation_0, values = (var_6406_cast_fp16_13, var_6480_cast_fp16))[name = tensor("op_6514_cast_fp16")]; + tensor var_6516_equation_0 = const()[name = tensor("op_6516_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6516_cast_fp16 = einsum(equation = var_6516_equation_0, values = (var_6406_cast_fp16_14, var_6481_cast_fp16))[name = tensor("op_6516_cast_fp16")]; + tensor var_6518_equation_0 = const()[name = tensor("op_6518_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6518_cast_fp16 = einsum(equation = var_6518_equation_0, values = (var_6406_cast_fp16_15, var_6482_cast_fp16))[name = tensor("op_6518_cast_fp16")]; + tensor var_6520_equation_0 = const()[name = tensor("op_6520_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6520_cast_fp16 = einsum(equation = var_6520_equation_0, values = (var_6406_cast_fp16_16, var_6483_cast_fp16))[name = tensor("op_6520_cast_fp16")]; + tensor var_6522_equation_0 = const()[name = tensor("op_6522_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6522_cast_fp16 = einsum(equation = var_6522_equation_0, values = (var_6406_cast_fp16_17, var_6484_cast_fp16))[name = tensor("op_6522_cast_fp16")]; + tensor var_6524_equation_0 = const()[name = tensor("op_6524_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6524_cast_fp16 = einsum(equation = var_6524_equation_0, values = (var_6406_cast_fp16_18, var_6485_cast_fp16))[name = tensor("op_6524_cast_fp16")]; + tensor var_6526_equation_0 = const()[name = tensor("op_6526_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6526_cast_fp16 = einsum(equation = var_6526_equation_0, values = (var_6406_cast_fp16_19, var_6486_cast_fp16))[name = tensor("op_6526_cast_fp16")]; + tensor input_235_interleave_0 = const()[name = tensor("input_235_interleave_0"), val = tensor(false)]; + tensor input_235_cast_fp16 = concat(axis = var_6311, interleave = input_235_interleave_0, values = (var_6488_cast_fp16, var_6490_cast_fp16, var_6492_cast_fp16, var_6494_cast_fp16, var_6496_cast_fp16, var_6498_cast_fp16, var_6500_cast_fp16, var_6502_cast_fp16, var_6504_cast_fp16, var_6506_cast_fp16, var_6508_cast_fp16, var_6510_cast_fp16, var_6512_cast_fp16, var_6514_cast_fp16, var_6516_cast_fp16, var_6518_cast_fp16, var_6520_cast_fp16, var_6522_cast_fp16, var_6524_cast_fp16, var_6526_cast_fp16))[name = tensor("input_235_cast_fp16")]; + tensor var_6535_pad_type_0 = const()[name = tensor("op_6535_pad_type_0"), val = tensor("valid")]; + tensor var_6535_strides_0 = const()[name = tensor("op_6535_strides_0"), val = tensor([1, 1])]; + tensor var_6535_pad_0 = const()[name = tensor("op_6535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6535_dilations_0 = const()[name = tensor("op_6535_dilations_0"), val = tensor([1, 1])]; + tensor var_6535_groups_0 = const()[name = tensor("op_6535_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_out_weight_to_fp16 = const()[name = tensor("blocks_23_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929625472)))]; + tensor blocks_23_attn_out_bias_to_fp16 = const()[name = tensor("blocks_23_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932902336)))]; + tensor var_6535_cast_fp16 = conv(bias = blocks_23_attn_out_bias_to_fp16, dilations = var_6535_dilations_0, groups = var_6535_groups_0, pad = var_6535_pad_0, pad_type = var_6535_pad_type_0, strides = var_6535_strides_0, weight = blocks_23_attn_out_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("op_6535_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = var_6535_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([1])]; + tensor input_237_gamma_0_to_fp16 = const()[name = tensor("input_237_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932904960)))]; + tensor input_237_beta_0_to_fp16 = const()[name = tensor("input_237_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932907584)))]; + tensor var_6545_to_fp16 = const()[name = tensor("op_6545_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = input_237_beta_0_to_fp16, epsilon = var_6545_to_fp16, gamma = input_237_gamma_0_to_fp16, x = inputs_95_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("valid")]; + tensor input_239_strides_0 = const()[name = tensor("input_239_strides_0"), val = tensor([1, 1])]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_239_dilations_0 = const()[name = tensor("input_239_dilations_0"), val = tensor([1, 1])]; + tensor input_239_groups_0 = const()[name = tensor("input_239_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932910208)))]; + tensor blocks_23_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(946017472)))]; + tensor input_239_cast_fp16 = conv(bias = blocks_23_mlp_0_bias_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = blocks_23_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = input_239_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_6571_pad_type_0 = const()[name = tensor("op_6571_pad_type_0"), val = tensor("valid")]; + tensor var_6571_strides_0 = const()[name = tensor("op_6571_strides_0"), val = tensor([1, 1])]; + tensor var_6571_pad_0 = const()[name = tensor("op_6571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6571_dilations_0 = const()[name = tensor("op_6571_dilations_0"), val = tensor([1, 1])]; + tensor var_6571_groups_0 = const()[name = tensor("op_6571_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(946027776)))]; + tensor blocks_23_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959135040)))]; + tensor var_6571_cast_fp16 = conv(bias = blocks_23_mlp_2_bias_to_fp16, dilations = var_6571_dilations_0, groups = var_6571_groups_0, pad = var_6571_pad_0, pad_type = var_6571_pad_type_0, strides = var_6571_strides_0, weight = blocks_23_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("op_6571_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = var_6571_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_6580 = const()[name = tensor("op_6580"), val = tensor(1)]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([1])]; + tensor input_243_gamma_0_to_fp16 = const()[name = tensor("input_243_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959137664)))]; + tensor input_243_beta_0_to_fp16 = const()[name = tensor("input_243_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959140288)))]; + tensor var_6596_to_fp16 = const()[name = tensor("op_6596_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = input_243_beta_0_to_fp16, epsilon = var_6596_to_fp16, gamma = input_243_gamma_0_to_fp16, x = inputs_97_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor q_49_pad_type_0 = const()[name = tensor("q_49_pad_type_0"), val = tensor("valid")]; + tensor q_49_strides_0 = const()[name = tensor("q_49_strides_0"), val = tensor([1, 1])]; + tensor q_49_pad_0 = const()[name = tensor("q_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_49_dilations_0 = const()[name = tensor("q_49_dilations_0"), val = tensor([1, 1])]; + tensor q_49_groups_0 = const()[name = tensor("q_49_groups_0"), val = tensor(1)]; + tensor var_6631_weight_0_to_fp16 = const()[name = tensor("op_6631_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959142912)))]; + tensor var_6631_bias_0_to_fp16 = const()[name = tensor("op_6631_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962419776)))]; + tensor var_6631_cast_fp16 = conv(bias = var_6631_bias_0_to_fp16, dilations = q_49_dilations_0, groups = q_49_groups_0, pad = q_49_pad_0, pad_type = q_49_pad_type_0, strides = q_49_strides_0, weight = var_6631_weight_0_to_fp16, x = input_243_cast_fp16)[name = tensor("op_6631_cast_fp16")]; + tensor k_49_pad_type_0 = const()[name = tensor("k_49_pad_type_0"), val = tensor("valid")]; + tensor k_49_strides_0 = const()[name = tensor("k_49_strides_0"), val = tensor([1, 1])]; + tensor k_49_pad_0 = const()[name = tensor("k_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_49_dilations_0 = const()[name = tensor("k_49_dilations_0"), val = tensor([1, 1])]; + tensor k_49_groups_0 = const()[name = tensor("k_49_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_key_weight_to_fp16 = const()[name = tensor("blocks_24_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962422400)))]; + tensor k_49_cast_fp16 = conv(dilations = k_49_dilations_0, groups = k_49_groups_0, pad = k_49_pad_0, pad_type = k_49_pad_type_0, strides = k_49_strides_0, weight = blocks_24_attn_key_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor var_6629_pad_type_0 = const()[name = tensor("op_6629_pad_type_0"), val = tensor("valid")]; + tensor var_6629_strides_0 = const()[name = tensor("op_6629_strides_0"), val = tensor([1, 1])]; + tensor var_6629_pad_0 = const()[name = tensor("op_6629_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6629_dilations_0 = const()[name = tensor("op_6629_dilations_0"), val = tensor([1, 1])]; + tensor var_6629_groups_0 = const()[name = tensor("op_6629_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_value_weight_to_fp16 = const()[name = tensor("blocks_24_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965699264)))]; + tensor blocks_24_attn_value_bias_to_fp16 = const()[name = tensor("blocks_24_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968976128)))]; + tensor var_6629_cast_fp16 = conv(bias = blocks_24_attn_value_bias_to_fp16, dilations = var_6629_dilations_0, groups = var_6629_groups_0, pad = var_6629_pad_0, pad_type = var_6629_pad_type_0, strides = var_6629_strides_0, weight = blocks_24_attn_value_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor tile_72 = const()[name = tensor("tile_72"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6632_axis_0 = const()[name = tensor("op_6632_axis_0"), val = tensor(1)]; + tensor var_6632_cast_fp16_0, tensor var_6632_cast_fp16_1, tensor var_6632_cast_fp16_2, tensor var_6632_cast_fp16_3, tensor var_6632_cast_fp16_4, tensor var_6632_cast_fp16_5, tensor var_6632_cast_fp16_6, tensor var_6632_cast_fp16_7, tensor var_6632_cast_fp16_8, tensor var_6632_cast_fp16_9, tensor var_6632_cast_fp16_10, tensor var_6632_cast_fp16_11, tensor var_6632_cast_fp16_12, tensor var_6632_cast_fp16_13, tensor var_6632_cast_fp16_14, tensor var_6632_cast_fp16_15, tensor var_6632_cast_fp16_16, tensor var_6632_cast_fp16_17, tensor var_6632_cast_fp16_18, tensor var_6632_cast_fp16_19 = split(axis = var_6632_axis_0, split_sizes = tile_72, x = var_6631_cast_fp16)[name = tensor("op_6632_cast_fp16")]; + tensor var_6653_perm_0 = const()[name = tensor("op_6653_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_73 = const()[name = tensor("tile_73"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6654_axis_0 = const()[name = tensor("op_6654_axis_0"), val = tensor(3)]; + tensor var_6653_cast_fp16 = transpose(perm = var_6653_perm_0, x = k_49_cast_fp16)[name = tensor("transpose_8")]; + tensor var_6654_cast_fp16_0, tensor var_6654_cast_fp16_1, tensor var_6654_cast_fp16_2, tensor var_6654_cast_fp16_3, tensor var_6654_cast_fp16_4, tensor var_6654_cast_fp16_5, tensor var_6654_cast_fp16_6, tensor var_6654_cast_fp16_7, tensor var_6654_cast_fp16_8, tensor var_6654_cast_fp16_9, tensor var_6654_cast_fp16_10, tensor var_6654_cast_fp16_11, tensor var_6654_cast_fp16_12, tensor var_6654_cast_fp16_13, tensor var_6654_cast_fp16_14, tensor var_6654_cast_fp16_15, tensor var_6654_cast_fp16_16, tensor var_6654_cast_fp16_17, tensor var_6654_cast_fp16_18, tensor var_6654_cast_fp16_19 = split(axis = var_6654_axis_0, split_sizes = tile_73, x = var_6653_cast_fp16)[name = tensor("op_6654_cast_fp16")]; + tensor tile_74 = const()[name = tensor("tile_74"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6675_axis_0 = const()[name = tensor("op_6675_axis_0"), val = tensor(1)]; + tensor var_6675_cast_fp16_0, tensor var_6675_cast_fp16_1, tensor var_6675_cast_fp16_2, tensor var_6675_cast_fp16_3, tensor var_6675_cast_fp16_4, tensor var_6675_cast_fp16_5, tensor var_6675_cast_fp16_6, tensor var_6675_cast_fp16_7, tensor var_6675_cast_fp16_8, tensor var_6675_cast_fp16_9, tensor var_6675_cast_fp16_10, tensor var_6675_cast_fp16_11, tensor var_6675_cast_fp16_12, tensor var_6675_cast_fp16_13, tensor var_6675_cast_fp16_14, tensor var_6675_cast_fp16_15, tensor var_6675_cast_fp16_16, tensor var_6675_cast_fp16_17, tensor var_6675_cast_fp16_18, tensor var_6675_cast_fp16_19 = split(axis = var_6675_axis_0, split_sizes = tile_74, x = var_6629_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor aw_961_equation_0 = const()[name = tensor("aw_961_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_961_cast_fp16 = einsum(equation = aw_961_equation_0, values = (var_6654_cast_fp16_0, var_6632_cast_fp16_0))[name = tensor("aw_961_cast_fp16")]; + tensor aw_963_equation_0 = const()[name = tensor("aw_963_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_963_cast_fp16 = einsum(equation = aw_963_equation_0, values = (var_6654_cast_fp16_1, var_6632_cast_fp16_1))[name = tensor("aw_963_cast_fp16")]; + tensor aw_965_equation_0 = const()[name = tensor("aw_965_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_965_cast_fp16 = einsum(equation = aw_965_equation_0, values = (var_6654_cast_fp16_2, var_6632_cast_fp16_2))[name = tensor("aw_965_cast_fp16")]; + tensor aw_967_equation_0 = const()[name = tensor("aw_967_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_967_cast_fp16 = einsum(equation = aw_967_equation_0, values = (var_6654_cast_fp16_3, var_6632_cast_fp16_3))[name = tensor("aw_967_cast_fp16")]; + tensor aw_969_equation_0 = const()[name = tensor("aw_969_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_969_cast_fp16 = einsum(equation = aw_969_equation_0, values = (var_6654_cast_fp16_4, var_6632_cast_fp16_4))[name = tensor("aw_969_cast_fp16")]; + tensor aw_971_equation_0 = const()[name = tensor("aw_971_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_971_cast_fp16 = einsum(equation = aw_971_equation_0, values = (var_6654_cast_fp16_5, var_6632_cast_fp16_5))[name = tensor("aw_971_cast_fp16")]; + tensor aw_973_equation_0 = const()[name = tensor("aw_973_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_973_cast_fp16 = einsum(equation = aw_973_equation_0, values = (var_6654_cast_fp16_6, var_6632_cast_fp16_6))[name = tensor("aw_973_cast_fp16")]; + tensor aw_975_equation_0 = const()[name = tensor("aw_975_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_975_cast_fp16 = einsum(equation = aw_975_equation_0, values = (var_6654_cast_fp16_7, var_6632_cast_fp16_7))[name = tensor("aw_975_cast_fp16")]; + tensor aw_977_equation_0 = const()[name = tensor("aw_977_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_977_cast_fp16 = einsum(equation = aw_977_equation_0, values = (var_6654_cast_fp16_8, var_6632_cast_fp16_8))[name = tensor("aw_977_cast_fp16")]; + tensor aw_979_equation_0 = const()[name = tensor("aw_979_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_979_cast_fp16 = einsum(equation = aw_979_equation_0, values = (var_6654_cast_fp16_9, var_6632_cast_fp16_9))[name = tensor("aw_979_cast_fp16")]; + tensor aw_981_equation_0 = const()[name = tensor("aw_981_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_981_cast_fp16 = einsum(equation = aw_981_equation_0, values = (var_6654_cast_fp16_10, var_6632_cast_fp16_10))[name = tensor("aw_981_cast_fp16")]; + tensor aw_983_equation_0 = const()[name = tensor("aw_983_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_983_cast_fp16 = einsum(equation = aw_983_equation_0, values = (var_6654_cast_fp16_11, var_6632_cast_fp16_11))[name = tensor("aw_983_cast_fp16")]; + tensor aw_985_equation_0 = const()[name = tensor("aw_985_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_985_cast_fp16 = einsum(equation = aw_985_equation_0, values = (var_6654_cast_fp16_12, var_6632_cast_fp16_12))[name = tensor("aw_985_cast_fp16")]; + tensor aw_987_equation_0 = const()[name = tensor("aw_987_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_987_cast_fp16 = einsum(equation = aw_987_equation_0, values = (var_6654_cast_fp16_13, var_6632_cast_fp16_13))[name = tensor("aw_987_cast_fp16")]; + tensor aw_989_equation_0 = const()[name = tensor("aw_989_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_989_cast_fp16 = einsum(equation = aw_989_equation_0, values = (var_6654_cast_fp16_14, var_6632_cast_fp16_14))[name = tensor("aw_989_cast_fp16")]; + tensor aw_991_equation_0 = const()[name = tensor("aw_991_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_991_cast_fp16 = einsum(equation = aw_991_equation_0, values = (var_6654_cast_fp16_15, var_6632_cast_fp16_15))[name = tensor("aw_991_cast_fp16")]; + tensor aw_993_equation_0 = const()[name = tensor("aw_993_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_993_cast_fp16 = einsum(equation = aw_993_equation_0, values = (var_6654_cast_fp16_16, var_6632_cast_fp16_16))[name = tensor("aw_993_cast_fp16")]; + tensor aw_995_equation_0 = const()[name = tensor("aw_995_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_995_cast_fp16 = einsum(equation = aw_995_equation_0, values = (var_6654_cast_fp16_17, var_6632_cast_fp16_17))[name = tensor("aw_995_cast_fp16")]; + tensor aw_997_equation_0 = const()[name = tensor("aw_997_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_997_cast_fp16 = einsum(equation = aw_997_equation_0, values = (var_6654_cast_fp16_18, var_6632_cast_fp16_18))[name = tensor("aw_997_cast_fp16")]; + tensor aw_999_equation_0 = const()[name = tensor("aw_999_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_999_cast_fp16 = einsum(equation = aw_999_equation_0, values = (var_6654_cast_fp16_19, var_6632_cast_fp16_19))[name = tensor("aw_999_cast_fp16")]; + tensor var_6736_cast_fp16 = softmax(axis = var_6580, x = aw_961_cast_fp16)[name = tensor("op_6736_cast_fp16")]; + tensor var_6737_cast_fp16 = softmax(axis = var_6580, x = aw_963_cast_fp16)[name = tensor("op_6737_cast_fp16")]; + tensor var_6738_cast_fp16 = softmax(axis = var_6580, x = aw_965_cast_fp16)[name = tensor("op_6738_cast_fp16")]; + tensor var_6739_cast_fp16 = softmax(axis = var_6580, x = aw_967_cast_fp16)[name = tensor("op_6739_cast_fp16")]; + tensor var_6740_cast_fp16 = softmax(axis = var_6580, x = aw_969_cast_fp16)[name = tensor("op_6740_cast_fp16")]; + tensor var_6741_cast_fp16 = softmax(axis = var_6580, x = aw_971_cast_fp16)[name = tensor("op_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = softmax(axis = var_6580, x = aw_973_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_cast_fp16 = softmax(axis = var_6580, x = aw_975_cast_fp16)[name = tensor("op_6743_cast_fp16")]; + tensor var_6744_cast_fp16 = softmax(axis = var_6580, x = aw_977_cast_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor var_6745_cast_fp16 = softmax(axis = var_6580, x = aw_979_cast_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6746_cast_fp16 = softmax(axis = var_6580, x = aw_981_cast_fp16)[name = tensor("op_6746_cast_fp16")]; + tensor var_6747_cast_fp16 = softmax(axis = var_6580, x = aw_983_cast_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6748_cast_fp16 = softmax(axis = var_6580, x = aw_985_cast_fp16)[name = tensor("op_6748_cast_fp16")]; + tensor var_6749_cast_fp16 = softmax(axis = var_6580, x = aw_987_cast_fp16)[name = tensor("op_6749_cast_fp16")]; + tensor var_6750_cast_fp16 = softmax(axis = var_6580, x = aw_989_cast_fp16)[name = tensor("op_6750_cast_fp16")]; + tensor var_6751_cast_fp16 = softmax(axis = var_6580, x = aw_991_cast_fp16)[name = tensor("op_6751_cast_fp16")]; + tensor var_6752_cast_fp16 = softmax(axis = var_6580, x = aw_993_cast_fp16)[name = tensor("op_6752_cast_fp16")]; + tensor var_6753_cast_fp16 = softmax(axis = var_6580, x = aw_995_cast_fp16)[name = tensor("op_6753_cast_fp16")]; + tensor var_6754_cast_fp16 = softmax(axis = var_6580, x = aw_997_cast_fp16)[name = tensor("op_6754_cast_fp16")]; + tensor var_6755_cast_fp16 = softmax(axis = var_6580, x = aw_999_cast_fp16)[name = tensor("op_6755_cast_fp16")]; + tensor var_6757_equation_0 = const()[name = tensor("op_6757_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6757_cast_fp16 = einsum(equation = var_6757_equation_0, values = (var_6675_cast_fp16_0, var_6736_cast_fp16))[name = tensor("op_6757_cast_fp16")]; + tensor var_6759_equation_0 = const()[name = tensor("op_6759_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6759_cast_fp16 = einsum(equation = var_6759_equation_0, values = (var_6675_cast_fp16_1, var_6737_cast_fp16))[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_equation_0 = const()[name = tensor("op_6761_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6761_cast_fp16 = einsum(equation = var_6761_equation_0, values = (var_6675_cast_fp16_2, var_6738_cast_fp16))[name = tensor("op_6761_cast_fp16")]; + tensor var_6763_equation_0 = const()[name = tensor("op_6763_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6763_cast_fp16 = einsum(equation = var_6763_equation_0, values = (var_6675_cast_fp16_3, var_6739_cast_fp16))[name = tensor("op_6763_cast_fp16")]; + tensor var_6765_equation_0 = const()[name = tensor("op_6765_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6765_cast_fp16 = einsum(equation = var_6765_equation_0, values = (var_6675_cast_fp16_4, var_6740_cast_fp16))[name = tensor("op_6765_cast_fp16")]; + tensor var_6767_equation_0 = const()[name = tensor("op_6767_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6767_cast_fp16 = einsum(equation = var_6767_equation_0, values = (var_6675_cast_fp16_5, var_6741_cast_fp16))[name = tensor("op_6767_cast_fp16")]; + tensor var_6769_equation_0 = const()[name = tensor("op_6769_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6769_cast_fp16 = einsum(equation = var_6769_equation_0, values = (var_6675_cast_fp16_6, var_6742_cast_fp16))[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_equation_0 = const()[name = tensor("op_6771_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6771_cast_fp16 = einsum(equation = var_6771_equation_0, values = (var_6675_cast_fp16_7, var_6743_cast_fp16))[name = tensor("op_6771_cast_fp16")]; + tensor var_6773_equation_0 = const()[name = tensor("op_6773_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6773_cast_fp16 = einsum(equation = var_6773_equation_0, values = (var_6675_cast_fp16_8, var_6744_cast_fp16))[name = tensor("op_6773_cast_fp16")]; + tensor var_6775_equation_0 = const()[name = tensor("op_6775_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6775_cast_fp16 = einsum(equation = var_6775_equation_0, values = (var_6675_cast_fp16_9, var_6745_cast_fp16))[name = tensor("op_6775_cast_fp16")]; + tensor var_6777_equation_0 = const()[name = tensor("op_6777_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6777_cast_fp16 = einsum(equation = var_6777_equation_0, values = (var_6675_cast_fp16_10, var_6746_cast_fp16))[name = tensor("op_6777_cast_fp16")]; + tensor var_6779_equation_0 = const()[name = tensor("op_6779_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6779_cast_fp16 = einsum(equation = var_6779_equation_0, values = (var_6675_cast_fp16_11, var_6747_cast_fp16))[name = tensor("op_6779_cast_fp16")]; + tensor var_6781_equation_0 = const()[name = tensor("op_6781_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6781_cast_fp16 = einsum(equation = var_6781_equation_0, values = (var_6675_cast_fp16_12, var_6748_cast_fp16))[name = tensor("op_6781_cast_fp16")]; + tensor var_6783_equation_0 = const()[name = tensor("op_6783_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6783_cast_fp16 = einsum(equation = var_6783_equation_0, values = (var_6675_cast_fp16_13, var_6749_cast_fp16))[name = tensor("op_6783_cast_fp16")]; + tensor var_6785_equation_0 = const()[name = tensor("op_6785_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6785_cast_fp16 = einsum(equation = var_6785_equation_0, values = (var_6675_cast_fp16_14, var_6750_cast_fp16))[name = tensor("op_6785_cast_fp16")]; + tensor var_6787_equation_0 = const()[name = tensor("op_6787_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6787_cast_fp16 = einsum(equation = var_6787_equation_0, values = (var_6675_cast_fp16_15, var_6751_cast_fp16))[name = tensor("op_6787_cast_fp16")]; + tensor var_6789_equation_0 = const()[name = tensor("op_6789_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6789_cast_fp16 = einsum(equation = var_6789_equation_0, values = (var_6675_cast_fp16_16, var_6752_cast_fp16))[name = tensor("op_6789_cast_fp16")]; + tensor var_6791_equation_0 = const()[name = tensor("op_6791_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6791_cast_fp16 = einsum(equation = var_6791_equation_0, values = (var_6675_cast_fp16_17, var_6753_cast_fp16))[name = tensor("op_6791_cast_fp16")]; + tensor var_6793_equation_0 = const()[name = tensor("op_6793_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6793_cast_fp16 = einsum(equation = var_6793_equation_0, values = (var_6675_cast_fp16_18, var_6754_cast_fp16))[name = tensor("op_6793_cast_fp16")]; + tensor var_6795_equation_0 = const()[name = tensor("op_6795_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_6795_cast_fp16 = einsum(equation = var_6795_equation_0, values = (var_6675_cast_fp16_19, var_6755_cast_fp16))[name = tensor("op_6795_cast_fp16")]; + tensor input_245_interleave_0 = const()[name = tensor("input_245_interleave_0"), val = tensor(false)]; + tensor input_245_cast_fp16 = concat(axis = var_6580, interleave = input_245_interleave_0, values = (var_6757_cast_fp16, var_6759_cast_fp16, var_6761_cast_fp16, var_6763_cast_fp16, var_6765_cast_fp16, var_6767_cast_fp16, var_6769_cast_fp16, var_6771_cast_fp16, var_6773_cast_fp16, var_6775_cast_fp16, var_6777_cast_fp16, var_6779_cast_fp16, var_6781_cast_fp16, var_6783_cast_fp16, var_6785_cast_fp16, var_6787_cast_fp16, var_6789_cast_fp16, var_6791_cast_fp16, var_6793_cast_fp16, var_6795_cast_fp16))[name = tensor("input_245_cast_fp16")]; + tensor var_6804_pad_type_0 = const()[name = tensor("op_6804_pad_type_0"), val = tensor("valid")]; + tensor var_6804_strides_0 = const()[name = tensor("op_6804_strides_0"), val = tensor([1, 1])]; + tensor var_6804_pad_0 = const()[name = tensor("op_6804_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6804_dilations_0 = const()[name = tensor("op_6804_dilations_0"), val = tensor([1, 1])]; + tensor var_6804_groups_0 = const()[name = tensor("op_6804_groups_0"), val = tensor(1)]; + tensor blocks_24_attn_out_weight_to_fp16 = const()[name = tensor("blocks_24_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968978752)))]; + tensor blocks_24_attn_out_bias_to_fp16 = const()[name = tensor("blocks_24_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972255616)))]; + tensor var_6804_cast_fp16 = conv(bias = blocks_24_attn_out_bias_to_fp16, dilations = var_6804_dilations_0, groups = var_6804_groups_0, pad = var_6804_pad_0, pad_type = var_6804_pad_type_0, strides = var_6804_strides_0, weight = blocks_24_attn_out_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("op_6804_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_6804_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor input_247_axes_0 = const()[name = tensor("input_247_axes_0"), val = tensor([1])]; + tensor input_247_gamma_0_to_fp16 = const()[name = tensor("input_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972258240)))]; + tensor input_247_beta_0_to_fp16 = const()[name = tensor("input_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972260864)))]; + tensor var_6814_to_fp16 = const()[name = tensor("op_6814_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, beta = input_247_beta_0_to_fp16, epsilon = var_6814_to_fp16, gamma = input_247_gamma_0_to_fp16, x = inputs_99_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_pad_type_0 = const()[name = tensor("input_249_pad_type_0"), val = tensor("valid")]; + tensor input_249_strides_0 = const()[name = tensor("input_249_strides_0"), val = tensor([1, 1])]; + tensor input_249_pad_0 = const()[name = tensor("input_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_249_dilations_0 = const()[name = tensor("input_249_dilations_0"), val = tensor([1, 1])]; + tensor input_249_groups_0 = const()[name = tensor("input_249_groups_0"), val = tensor(1)]; + tensor blocks_24_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972263488)))]; + tensor blocks_24_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985370752)))]; + tensor input_249_cast_fp16 = conv(bias = blocks_24_mlp_0_bias_to_fp16, dilations = input_249_dilations_0, groups = input_249_groups_0, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = input_249_strides_0, weight = blocks_24_mlp_0_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_6840_pad_type_0 = const()[name = tensor("op_6840_pad_type_0"), val = tensor("valid")]; + tensor var_6840_strides_0 = const()[name = tensor("op_6840_strides_0"), val = tensor([1, 1])]; + tensor var_6840_pad_0 = const()[name = tensor("op_6840_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6840_dilations_0 = const()[name = tensor("op_6840_dilations_0"), val = tensor([1, 1])]; + tensor var_6840_groups_0 = const()[name = tensor("op_6840_groups_0"), val = tensor(1)]; + tensor blocks_24_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985381056)))]; + tensor blocks_24_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998488320)))]; + tensor var_6840_cast_fp16 = conv(bias = blocks_24_mlp_2_bias_to_fp16, dilations = var_6840_dilations_0, groups = var_6840_groups_0, pad = var_6840_pad_0, pad_type = var_6840_pad_type_0, strides = var_6840_strides_0, weight = blocks_24_mlp_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("op_6840_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = var_6840_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_6849 = const()[name = tensor("op_6849"), val = tensor(1)]; + tensor input_253_axes_0 = const()[name = tensor("input_253_axes_0"), val = tensor([1])]; + tensor input_253_gamma_0_to_fp16 = const()[name = tensor("input_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998490944)))]; + tensor input_253_beta_0_to_fp16 = const()[name = tensor("input_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998493568)))]; + tensor var_6865_to_fp16 = const()[name = tensor("op_6865_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_253_cast_fp16 = layer_norm(axes = input_253_axes_0, beta = input_253_beta_0_to_fp16, epsilon = var_6865_to_fp16, gamma = input_253_gamma_0_to_fp16, x = inputs_101_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor q_51_pad_type_0 = const()[name = tensor("q_51_pad_type_0"), val = tensor("valid")]; + tensor q_51_strides_0 = const()[name = tensor("q_51_strides_0"), val = tensor([1, 1])]; + tensor q_51_pad_0 = const()[name = tensor("q_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_51_dilations_0 = const()[name = tensor("q_51_dilations_0"), val = tensor([1, 1])]; + tensor q_51_groups_0 = const()[name = tensor("q_51_groups_0"), val = tensor(1)]; + tensor var_6900_weight_0_to_fp16 = const()[name = tensor("op_6900_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998496192)))]; + tensor var_6900_bias_0_to_fp16 = const()[name = tensor("op_6900_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001773056)))]; + tensor var_6900_cast_fp16 = conv(bias = var_6900_bias_0_to_fp16, dilations = q_51_dilations_0, groups = q_51_groups_0, pad = q_51_pad_0, pad_type = q_51_pad_type_0, strides = q_51_strides_0, weight = var_6900_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor k_51_pad_type_0 = const()[name = tensor("k_51_pad_type_0"), val = tensor("valid")]; + tensor k_51_strides_0 = const()[name = tensor("k_51_strides_0"), val = tensor([1, 1])]; + tensor k_51_pad_0 = const()[name = tensor("k_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_51_dilations_0 = const()[name = tensor("k_51_dilations_0"), val = tensor([1, 1])]; + tensor k_51_groups_0 = const()[name = tensor("k_51_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_key_weight_to_fp16 = const()[name = tensor("blocks_25_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001775680)))]; + tensor k_51_cast_fp16 = conv(dilations = k_51_dilations_0, groups = k_51_groups_0, pad = k_51_pad_0, pad_type = k_51_pad_type_0, strides = k_51_strides_0, weight = blocks_25_attn_key_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor var_6898_pad_type_0 = const()[name = tensor("op_6898_pad_type_0"), val = tensor("valid")]; + tensor var_6898_strides_0 = const()[name = tensor("op_6898_strides_0"), val = tensor([1, 1])]; + tensor var_6898_pad_0 = const()[name = tensor("op_6898_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6898_dilations_0 = const()[name = tensor("op_6898_dilations_0"), val = tensor([1, 1])]; + tensor var_6898_groups_0 = const()[name = tensor("op_6898_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_value_weight_to_fp16 = const()[name = tensor("blocks_25_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1005052544)))]; + tensor blocks_25_attn_value_bias_to_fp16 = const()[name = tensor("blocks_25_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008329408)))]; + tensor var_6898_cast_fp16 = conv(bias = blocks_25_attn_value_bias_to_fp16, dilations = var_6898_dilations_0, groups = var_6898_groups_0, pad = var_6898_pad_0, pad_type = var_6898_pad_type_0, strides = var_6898_strides_0, weight = blocks_25_attn_value_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor tile_75 = const()[name = tensor("tile_75"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6901_axis_0 = const()[name = tensor("op_6901_axis_0"), val = tensor(1)]; + tensor var_6901_cast_fp16_0, tensor var_6901_cast_fp16_1, tensor var_6901_cast_fp16_2, tensor var_6901_cast_fp16_3, tensor var_6901_cast_fp16_4, tensor var_6901_cast_fp16_5, tensor var_6901_cast_fp16_6, tensor var_6901_cast_fp16_7, tensor var_6901_cast_fp16_8, tensor var_6901_cast_fp16_9, tensor var_6901_cast_fp16_10, tensor var_6901_cast_fp16_11, tensor var_6901_cast_fp16_12, tensor var_6901_cast_fp16_13, tensor var_6901_cast_fp16_14, tensor var_6901_cast_fp16_15, tensor var_6901_cast_fp16_16, tensor var_6901_cast_fp16_17, tensor var_6901_cast_fp16_18, tensor var_6901_cast_fp16_19 = split(axis = var_6901_axis_0, split_sizes = tile_75, x = var_6900_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor var_6922_perm_0 = const()[name = tensor("op_6922_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_76 = const()[name = tensor("tile_76"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6923_axis_0 = const()[name = tensor("op_6923_axis_0"), val = tensor(3)]; + tensor var_6922_cast_fp16 = transpose(perm = var_6922_perm_0, x = k_51_cast_fp16)[name = tensor("transpose_7")]; + tensor var_6923_cast_fp16_0, tensor var_6923_cast_fp16_1, tensor var_6923_cast_fp16_2, tensor var_6923_cast_fp16_3, tensor var_6923_cast_fp16_4, tensor var_6923_cast_fp16_5, tensor var_6923_cast_fp16_6, tensor var_6923_cast_fp16_7, tensor var_6923_cast_fp16_8, tensor var_6923_cast_fp16_9, tensor var_6923_cast_fp16_10, tensor var_6923_cast_fp16_11, tensor var_6923_cast_fp16_12, tensor var_6923_cast_fp16_13, tensor var_6923_cast_fp16_14, tensor var_6923_cast_fp16_15, tensor var_6923_cast_fp16_16, tensor var_6923_cast_fp16_17, tensor var_6923_cast_fp16_18, tensor var_6923_cast_fp16_19 = split(axis = var_6923_axis_0, split_sizes = tile_76, x = var_6922_cast_fp16)[name = tensor("op_6923_cast_fp16")]; + tensor tile_77 = const()[name = tensor("tile_77"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_6944_axis_0 = const()[name = tensor("op_6944_axis_0"), val = tensor(1)]; + tensor var_6944_cast_fp16_0, tensor var_6944_cast_fp16_1, tensor var_6944_cast_fp16_2, tensor var_6944_cast_fp16_3, tensor var_6944_cast_fp16_4, tensor var_6944_cast_fp16_5, tensor var_6944_cast_fp16_6, tensor var_6944_cast_fp16_7, tensor var_6944_cast_fp16_8, tensor var_6944_cast_fp16_9, tensor var_6944_cast_fp16_10, tensor var_6944_cast_fp16_11, tensor var_6944_cast_fp16_12, tensor var_6944_cast_fp16_13, tensor var_6944_cast_fp16_14, tensor var_6944_cast_fp16_15, tensor var_6944_cast_fp16_16, tensor var_6944_cast_fp16_17, tensor var_6944_cast_fp16_18, tensor var_6944_cast_fp16_19 = split(axis = var_6944_axis_0, split_sizes = tile_77, x = var_6898_cast_fp16)[name = tensor("op_6944_cast_fp16")]; + tensor aw_1001_equation_0 = const()[name = tensor("aw_1001_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1001_cast_fp16 = einsum(equation = aw_1001_equation_0, values = (var_6923_cast_fp16_0, var_6901_cast_fp16_0))[name = tensor("aw_1001_cast_fp16")]; + tensor aw_1003_equation_0 = const()[name = tensor("aw_1003_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1003_cast_fp16 = einsum(equation = aw_1003_equation_0, values = (var_6923_cast_fp16_1, var_6901_cast_fp16_1))[name = tensor("aw_1003_cast_fp16")]; + tensor aw_1005_equation_0 = const()[name = tensor("aw_1005_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1005_cast_fp16 = einsum(equation = aw_1005_equation_0, values = (var_6923_cast_fp16_2, var_6901_cast_fp16_2))[name = tensor("aw_1005_cast_fp16")]; + tensor aw_1007_equation_0 = const()[name = tensor("aw_1007_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1007_cast_fp16 = einsum(equation = aw_1007_equation_0, values = (var_6923_cast_fp16_3, var_6901_cast_fp16_3))[name = tensor("aw_1007_cast_fp16")]; + tensor aw_1009_equation_0 = const()[name = tensor("aw_1009_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1009_cast_fp16 = einsum(equation = aw_1009_equation_0, values = (var_6923_cast_fp16_4, var_6901_cast_fp16_4))[name = tensor("aw_1009_cast_fp16")]; + tensor aw_1011_equation_0 = const()[name = tensor("aw_1011_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1011_cast_fp16 = einsum(equation = aw_1011_equation_0, values = (var_6923_cast_fp16_5, var_6901_cast_fp16_5))[name = tensor("aw_1011_cast_fp16")]; + tensor aw_1013_equation_0 = const()[name = tensor("aw_1013_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1013_cast_fp16 = einsum(equation = aw_1013_equation_0, values = (var_6923_cast_fp16_6, var_6901_cast_fp16_6))[name = tensor("aw_1013_cast_fp16")]; + tensor aw_1015_equation_0 = const()[name = tensor("aw_1015_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1015_cast_fp16 = einsum(equation = aw_1015_equation_0, values = (var_6923_cast_fp16_7, var_6901_cast_fp16_7))[name = tensor("aw_1015_cast_fp16")]; + tensor aw_1017_equation_0 = const()[name = tensor("aw_1017_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1017_cast_fp16 = einsum(equation = aw_1017_equation_0, values = (var_6923_cast_fp16_8, var_6901_cast_fp16_8))[name = tensor("aw_1017_cast_fp16")]; + tensor aw_1019_equation_0 = const()[name = tensor("aw_1019_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1019_cast_fp16 = einsum(equation = aw_1019_equation_0, values = (var_6923_cast_fp16_9, var_6901_cast_fp16_9))[name = tensor("aw_1019_cast_fp16")]; + tensor aw_1021_equation_0 = const()[name = tensor("aw_1021_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1021_cast_fp16 = einsum(equation = aw_1021_equation_0, values = (var_6923_cast_fp16_10, var_6901_cast_fp16_10))[name = tensor("aw_1021_cast_fp16")]; + tensor aw_1023_equation_0 = const()[name = tensor("aw_1023_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1023_cast_fp16 = einsum(equation = aw_1023_equation_0, values = (var_6923_cast_fp16_11, var_6901_cast_fp16_11))[name = tensor("aw_1023_cast_fp16")]; + tensor aw_1025_equation_0 = const()[name = tensor("aw_1025_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1025_cast_fp16 = einsum(equation = aw_1025_equation_0, values = (var_6923_cast_fp16_12, var_6901_cast_fp16_12))[name = tensor("aw_1025_cast_fp16")]; + tensor aw_1027_equation_0 = const()[name = tensor("aw_1027_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1027_cast_fp16 = einsum(equation = aw_1027_equation_0, values = (var_6923_cast_fp16_13, var_6901_cast_fp16_13))[name = tensor("aw_1027_cast_fp16")]; + tensor aw_1029_equation_0 = const()[name = tensor("aw_1029_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1029_cast_fp16 = einsum(equation = aw_1029_equation_0, values = (var_6923_cast_fp16_14, var_6901_cast_fp16_14))[name = tensor("aw_1029_cast_fp16")]; + tensor aw_1031_equation_0 = const()[name = tensor("aw_1031_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1031_cast_fp16 = einsum(equation = aw_1031_equation_0, values = (var_6923_cast_fp16_15, var_6901_cast_fp16_15))[name = tensor("aw_1031_cast_fp16")]; + tensor aw_1033_equation_0 = const()[name = tensor("aw_1033_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1033_cast_fp16 = einsum(equation = aw_1033_equation_0, values = (var_6923_cast_fp16_16, var_6901_cast_fp16_16))[name = tensor("aw_1033_cast_fp16")]; + tensor aw_1035_equation_0 = const()[name = tensor("aw_1035_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1035_cast_fp16 = einsum(equation = aw_1035_equation_0, values = (var_6923_cast_fp16_17, var_6901_cast_fp16_17))[name = tensor("aw_1035_cast_fp16")]; + tensor aw_1037_equation_0 = const()[name = tensor("aw_1037_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1037_cast_fp16 = einsum(equation = aw_1037_equation_0, values = (var_6923_cast_fp16_18, var_6901_cast_fp16_18))[name = tensor("aw_1037_cast_fp16")]; + tensor aw_1039_equation_0 = const()[name = tensor("aw_1039_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1039_cast_fp16 = einsum(equation = aw_1039_equation_0, values = (var_6923_cast_fp16_19, var_6901_cast_fp16_19))[name = tensor("aw_1039_cast_fp16")]; + tensor var_7005_cast_fp16 = softmax(axis = var_6849, x = aw_1001_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7006_cast_fp16 = softmax(axis = var_6849, x = aw_1003_cast_fp16)[name = tensor("op_7006_cast_fp16")]; + tensor var_7007_cast_fp16 = softmax(axis = var_6849, x = aw_1005_cast_fp16)[name = tensor("op_7007_cast_fp16")]; + tensor var_7008_cast_fp16 = softmax(axis = var_6849, x = aw_1007_cast_fp16)[name = tensor("op_7008_cast_fp16")]; + tensor var_7009_cast_fp16 = softmax(axis = var_6849, x = aw_1009_cast_fp16)[name = tensor("op_7009_cast_fp16")]; + tensor var_7010_cast_fp16 = softmax(axis = var_6849, x = aw_1011_cast_fp16)[name = tensor("op_7010_cast_fp16")]; + tensor var_7011_cast_fp16 = softmax(axis = var_6849, x = aw_1013_cast_fp16)[name = tensor("op_7011_cast_fp16")]; + tensor var_7012_cast_fp16 = softmax(axis = var_6849, x = aw_1015_cast_fp16)[name = tensor("op_7012_cast_fp16")]; + tensor var_7013_cast_fp16 = softmax(axis = var_6849, x = aw_1017_cast_fp16)[name = tensor("op_7013_cast_fp16")]; + tensor var_7014_cast_fp16 = softmax(axis = var_6849, x = aw_1019_cast_fp16)[name = tensor("op_7014_cast_fp16")]; + tensor var_7015_cast_fp16 = softmax(axis = var_6849, x = aw_1021_cast_fp16)[name = tensor("op_7015_cast_fp16")]; + tensor var_7016_cast_fp16 = softmax(axis = var_6849, x = aw_1023_cast_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7017_cast_fp16 = softmax(axis = var_6849, x = aw_1025_cast_fp16)[name = tensor("op_7017_cast_fp16")]; + tensor var_7018_cast_fp16 = softmax(axis = var_6849, x = aw_1027_cast_fp16)[name = tensor("op_7018_cast_fp16")]; + tensor var_7019_cast_fp16 = softmax(axis = var_6849, x = aw_1029_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor var_7020_cast_fp16 = softmax(axis = var_6849, x = aw_1031_cast_fp16)[name = tensor("op_7020_cast_fp16")]; + tensor var_7021_cast_fp16 = softmax(axis = var_6849, x = aw_1033_cast_fp16)[name = tensor("op_7021_cast_fp16")]; + tensor var_7022_cast_fp16 = softmax(axis = var_6849, x = aw_1035_cast_fp16)[name = tensor("op_7022_cast_fp16")]; + tensor var_7023_cast_fp16 = softmax(axis = var_6849, x = aw_1037_cast_fp16)[name = tensor("op_7023_cast_fp16")]; + tensor var_7024_cast_fp16 = softmax(axis = var_6849, x = aw_1039_cast_fp16)[name = tensor("op_7024_cast_fp16")]; + tensor var_7026_equation_0 = const()[name = tensor("op_7026_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7026_cast_fp16 = einsum(equation = var_7026_equation_0, values = (var_6944_cast_fp16_0, var_7005_cast_fp16))[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_equation_0 = const()[name = tensor("op_7028_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7028_cast_fp16 = einsum(equation = var_7028_equation_0, values = (var_6944_cast_fp16_1, var_7006_cast_fp16))[name = tensor("op_7028_cast_fp16")]; + tensor var_7030_equation_0 = const()[name = tensor("op_7030_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7030_cast_fp16 = einsum(equation = var_7030_equation_0, values = (var_6944_cast_fp16_2, var_7007_cast_fp16))[name = tensor("op_7030_cast_fp16")]; + tensor var_7032_equation_0 = const()[name = tensor("op_7032_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7032_cast_fp16 = einsum(equation = var_7032_equation_0, values = (var_6944_cast_fp16_3, var_7008_cast_fp16))[name = tensor("op_7032_cast_fp16")]; + tensor var_7034_equation_0 = const()[name = tensor("op_7034_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7034_cast_fp16 = einsum(equation = var_7034_equation_0, values = (var_6944_cast_fp16_4, var_7009_cast_fp16))[name = tensor("op_7034_cast_fp16")]; + tensor var_7036_equation_0 = const()[name = tensor("op_7036_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7036_cast_fp16 = einsum(equation = var_7036_equation_0, values = (var_6944_cast_fp16_5, var_7010_cast_fp16))[name = tensor("op_7036_cast_fp16")]; + tensor var_7038_equation_0 = const()[name = tensor("op_7038_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7038_cast_fp16 = einsum(equation = var_7038_equation_0, values = (var_6944_cast_fp16_6, var_7011_cast_fp16))[name = tensor("op_7038_cast_fp16")]; + tensor var_7040_equation_0 = const()[name = tensor("op_7040_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7040_cast_fp16 = einsum(equation = var_7040_equation_0, values = (var_6944_cast_fp16_7, var_7012_cast_fp16))[name = tensor("op_7040_cast_fp16")]; + tensor var_7042_equation_0 = const()[name = tensor("op_7042_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7042_cast_fp16 = einsum(equation = var_7042_equation_0, values = (var_6944_cast_fp16_8, var_7013_cast_fp16))[name = tensor("op_7042_cast_fp16")]; + tensor var_7044_equation_0 = const()[name = tensor("op_7044_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7044_cast_fp16 = einsum(equation = var_7044_equation_0, values = (var_6944_cast_fp16_9, var_7014_cast_fp16))[name = tensor("op_7044_cast_fp16")]; + tensor var_7046_equation_0 = const()[name = tensor("op_7046_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7046_cast_fp16 = einsum(equation = var_7046_equation_0, values = (var_6944_cast_fp16_10, var_7015_cast_fp16))[name = tensor("op_7046_cast_fp16")]; + tensor var_7048_equation_0 = const()[name = tensor("op_7048_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7048_cast_fp16 = einsum(equation = var_7048_equation_0, values = (var_6944_cast_fp16_11, var_7016_cast_fp16))[name = tensor("op_7048_cast_fp16")]; + tensor var_7050_equation_0 = const()[name = tensor("op_7050_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7050_cast_fp16 = einsum(equation = var_7050_equation_0, values = (var_6944_cast_fp16_12, var_7017_cast_fp16))[name = tensor("op_7050_cast_fp16")]; + tensor var_7052_equation_0 = const()[name = tensor("op_7052_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7052_cast_fp16 = einsum(equation = var_7052_equation_0, values = (var_6944_cast_fp16_13, var_7018_cast_fp16))[name = tensor("op_7052_cast_fp16")]; + tensor var_7054_equation_0 = const()[name = tensor("op_7054_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7054_cast_fp16 = einsum(equation = var_7054_equation_0, values = (var_6944_cast_fp16_14, var_7019_cast_fp16))[name = tensor("op_7054_cast_fp16")]; + tensor var_7056_equation_0 = const()[name = tensor("op_7056_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7056_cast_fp16 = einsum(equation = var_7056_equation_0, values = (var_6944_cast_fp16_15, var_7020_cast_fp16))[name = tensor("op_7056_cast_fp16")]; + tensor var_7058_equation_0 = const()[name = tensor("op_7058_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7058_cast_fp16 = einsum(equation = var_7058_equation_0, values = (var_6944_cast_fp16_16, var_7021_cast_fp16))[name = tensor("op_7058_cast_fp16")]; + tensor var_7060_equation_0 = const()[name = tensor("op_7060_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7060_cast_fp16 = einsum(equation = var_7060_equation_0, values = (var_6944_cast_fp16_17, var_7022_cast_fp16))[name = tensor("op_7060_cast_fp16")]; + tensor var_7062_equation_0 = const()[name = tensor("op_7062_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7062_cast_fp16 = einsum(equation = var_7062_equation_0, values = (var_6944_cast_fp16_18, var_7023_cast_fp16))[name = tensor("op_7062_cast_fp16")]; + tensor var_7064_equation_0 = const()[name = tensor("op_7064_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7064_cast_fp16 = einsum(equation = var_7064_equation_0, values = (var_6944_cast_fp16_19, var_7024_cast_fp16))[name = tensor("op_7064_cast_fp16")]; + tensor input_255_interleave_0 = const()[name = tensor("input_255_interleave_0"), val = tensor(false)]; + tensor input_255_cast_fp16 = concat(axis = var_6849, interleave = input_255_interleave_0, values = (var_7026_cast_fp16, var_7028_cast_fp16, var_7030_cast_fp16, var_7032_cast_fp16, var_7034_cast_fp16, var_7036_cast_fp16, var_7038_cast_fp16, var_7040_cast_fp16, var_7042_cast_fp16, var_7044_cast_fp16, var_7046_cast_fp16, var_7048_cast_fp16, var_7050_cast_fp16, var_7052_cast_fp16, var_7054_cast_fp16, var_7056_cast_fp16, var_7058_cast_fp16, var_7060_cast_fp16, var_7062_cast_fp16, var_7064_cast_fp16))[name = tensor("input_255_cast_fp16")]; + tensor var_7073_pad_type_0 = const()[name = tensor("op_7073_pad_type_0"), val = tensor("valid")]; + tensor var_7073_strides_0 = const()[name = tensor("op_7073_strides_0"), val = tensor([1, 1])]; + tensor var_7073_pad_0 = const()[name = tensor("op_7073_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7073_dilations_0 = const()[name = tensor("op_7073_dilations_0"), val = tensor([1, 1])]; + tensor var_7073_groups_0 = const()[name = tensor("op_7073_groups_0"), val = tensor(1)]; + tensor blocks_25_attn_out_weight_to_fp16 = const()[name = tensor("blocks_25_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008332032)))]; + tensor blocks_25_attn_out_bias_to_fp16 = const()[name = tensor("blocks_25_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011608896)))]; + tensor var_7073_cast_fp16 = conv(bias = blocks_25_attn_out_bias_to_fp16, dilations = var_7073_dilations_0, groups = var_7073_groups_0, pad = var_7073_pad_0, pad_type = var_7073_pad_type_0, strides = var_7073_strides_0, weight = blocks_25_attn_out_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("op_7073_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_7073_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([1])]; + tensor input_257_gamma_0_to_fp16 = const()[name = tensor("input_257_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011611520)))]; + tensor input_257_beta_0_to_fp16 = const()[name = tensor("input_257_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011614144)))]; + tensor var_7083_to_fp16 = const()[name = tensor("op_7083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = input_257_beta_0_to_fp16, epsilon = var_7083_to_fp16, gamma = input_257_gamma_0_to_fp16, x = inputs_103_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("valid")]; + tensor input_259_strides_0 = const()[name = tensor("input_259_strides_0"), val = tensor([1, 1])]; + tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_259_dilations_0 = const()[name = tensor("input_259_dilations_0"), val = tensor([1, 1])]; + tensor input_259_groups_0 = const()[name = tensor("input_259_groups_0"), val = tensor(1)]; + tensor blocks_25_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011616768)))]; + tensor blocks_25_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024724032)))]; + tensor input_259_cast_fp16 = conv(bias = blocks_25_mlp_0_bias_to_fp16, dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = blocks_25_mlp_0_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_mode_0 = const()[name = tensor("input_261_mode_0"), val = tensor("EXACT")]; + tensor input_261_cast_fp16 = gelu(mode = input_261_mode_0, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_7109_pad_type_0 = const()[name = tensor("op_7109_pad_type_0"), val = tensor("valid")]; + tensor var_7109_strides_0 = const()[name = tensor("op_7109_strides_0"), val = tensor([1, 1])]; + tensor var_7109_pad_0 = const()[name = tensor("op_7109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7109_dilations_0 = const()[name = tensor("op_7109_dilations_0"), val = tensor([1, 1])]; + tensor var_7109_groups_0 = const()[name = tensor("op_7109_groups_0"), val = tensor(1)]; + tensor blocks_25_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024734336)))]; + tensor blocks_25_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037841600)))]; + tensor var_7109_cast_fp16 = conv(bias = blocks_25_mlp_2_bias_to_fp16, dilations = var_7109_dilations_0, groups = var_7109_groups_0, pad = var_7109_pad_0, pad_type = var_7109_pad_type_0, strides = var_7109_strides_0, weight = blocks_25_mlp_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("op_7109_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = var_7109_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_7118 = const()[name = tensor("op_7118"), val = tensor(1)]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([1])]; + tensor input_263_gamma_0_to_fp16 = const()[name = tensor("input_263_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037844224)))]; + tensor input_263_beta_0_to_fp16 = const()[name = tensor("input_263_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037846848)))]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = input_263_beta_0_to_fp16, epsilon = var_7134_to_fp16, gamma = input_263_gamma_0_to_fp16, x = inputs_105_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor q_53_pad_type_0 = const()[name = tensor("q_53_pad_type_0"), val = tensor("valid")]; + tensor q_53_strides_0 = const()[name = tensor("q_53_strides_0"), val = tensor([1, 1])]; + tensor q_53_pad_0 = const()[name = tensor("q_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_53_dilations_0 = const()[name = tensor("q_53_dilations_0"), val = tensor([1, 1])]; + tensor q_53_groups_0 = const()[name = tensor("q_53_groups_0"), val = tensor(1)]; + tensor var_7169_weight_0_to_fp16 = const()[name = tensor("op_7169_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037849472)))]; + tensor var_7169_bias_0_to_fp16 = const()[name = tensor("op_7169_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1041126336)))]; + tensor var_7169_cast_fp16 = conv(bias = var_7169_bias_0_to_fp16, dilations = q_53_dilations_0, groups = q_53_groups_0, pad = q_53_pad_0, pad_type = q_53_pad_type_0, strides = q_53_strides_0, weight = var_7169_weight_0_to_fp16, x = input_263_cast_fp16)[name = tensor("op_7169_cast_fp16")]; + tensor k_53_pad_type_0 = const()[name = tensor("k_53_pad_type_0"), val = tensor("valid")]; + tensor k_53_strides_0 = const()[name = tensor("k_53_strides_0"), val = tensor([1, 1])]; + tensor k_53_pad_0 = const()[name = tensor("k_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_53_dilations_0 = const()[name = tensor("k_53_dilations_0"), val = tensor([1, 1])]; + tensor k_53_groups_0 = const()[name = tensor("k_53_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_key_weight_to_fp16 = const()[name = tensor("blocks_26_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1041128960)))]; + tensor k_53_cast_fp16 = conv(dilations = k_53_dilations_0, groups = k_53_groups_0, pad = k_53_pad_0, pad_type = k_53_pad_type_0, strides = k_53_strides_0, weight = blocks_26_attn_key_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor var_7167_pad_type_0 = const()[name = tensor("op_7167_pad_type_0"), val = tensor("valid")]; + tensor var_7167_strides_0 = const()[name = tensor("op_7167_strides_0"), val = tensor([1, 1])]; + tensor var_7167_pad_0 = const()[name = tensor("op_7167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7167_dilations_0 = const()[name = tensor("op_7167_dilations_0"), val = tensor([1, 1])]; + tensor var_7167_groups_0 = const()[name = tensor("op_7167_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_value_weight_to_fp16 = const()[name = tensor("blocks_26_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044405824)))]; + tensor blocks_26_attn_value_bias_to_fp16 = const()[name = tensor("blocks_26_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047682688)))]; + tensor var_7167_cast_fp16 = conv(bias = blocks_26_attn_value_bias_to_fp16, dilations = var_7167_dilations_0, groups = var_7167_groups_0, pad = var_7167_pad_0, pad_type = var_7167_pad_type_0, strides = var_7167_strides_0, weight = blocks_26_attn_value_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("op_7167_cast_fp16")]; + tensor tile_78 = const()[name = tensor("tile_78"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7170_axis_0 = const()[name = tensor("op_7170_axis_0"), val = tensor(1)]; + tensor var_7170_cast_fp16_0, tensor var_7170_cast_fp16_1, tensor var_7170_cast_fp16_2, tensor var_7170_cast_fp16_3, tensor var_7170_cast_fp16_4, tensor var_7170_cast_fp16_5, tensor var_7170_cast_fp16_6, tensor var_7170_cast_fp16_7, tensor var_7170_cast_fp16_8, tensor var_7170_cast_fp16_9, tensor var_7170_cast_fp16_10, tensor var_7170_cast_fp16_11, tensor var_7170_cast_fp16_12, tensor var_7170_cast_fp16_13, tensor var_7170_cast_fp16_14, tensor var_7170_cast_fp16_15, tensor var_7170_cast_fp16_16, tensor var_7170_cast_fp16_17, tensor var_7170_cast_fp16_18, tensor var_7170_cast_fp16_19 = split(axis = var_7170_axis_0, split_sizes = tile_78, x = var_7169_cast_fp16)[name = tensor("op_7170_cast_fp16")]; + tensor var_7191_perm_0 = const()[name = tensor("op_7191_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_79 = const()[name = tensor("tile_79"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7192_axis_0 = const()[name = tensor("op_7192_axis_0"), val = tensor(3)]; + tensor var_7191_cast_fp16 = transpose(perm = var_7191_perm_0, x = k_53_cast_fp16)[name = tensor("transpose_6")]; + tensor var_7192_cast_fp16_0, tensor var_7192_cast_fp16_1, tensor var_7192_cast_fp16_2, tensor var_7192_cast_fp16_3, tensor var_7192_cast_fp16_4, tensor var_7192_cast_fp16_5, tensor var_7192_cast_fp16_6, tensor var_7192_cast_fp16_7, tensor var_7192_cast_fp16_8, tensor var_7192_cast_fp16_9, tensor var_7192_cast_fp16_10, tensor var_7192_cast_fp16_11, tensor var_7192_cast_fp16_12, tensor var_7192_cast_fp16_13, tensor var_7192_cast_fp16_14, tensor var_7192_cast_fp16_15, tensor var_7192_cast_fp16_16, tensor var_7192_cast_fp16_17, tensor var_7192_cast_fp16_18, tensor var_7192_cast_fp16_19 = split(axis = var_7192_axis_0, split_sizes = tile_79, x = var_7191_cast_fp16)[name = tensor("op_7192_cast_fp16")]; + tensor tile_80 = const()[name = tensor("tile_80"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7213_axis_0 = const()[name = tensor("op_7213_axis_0"), val = tensor(1)]; + tensor var_7213_cast_fp16_0, tensor var_7213_cast_fp16_1, tensor var_7213_cast_fp16_2, tensor var_7213_cast_fp16_3, tensor var_7213_cast_fp16_4, tensor var_7213_cast_fp16_5, tensor var_7213_cast_fp16_6, tensor var_7213_cast_fp16_7, tensor var_7213_cast_fp16_8, tensor var_7213_cast_fp16_9, tensor var_7213_cast_fp16_10, tensor var_7213_cast_fp16_11, tensor var_7213_cast_fp16_12, tensor var_7213_cast_fp16_13, tensor var_7213_cast_fp16_14, tensor var_7213_cast_fp16_15, tensor var_7213_cast_fp16_16, tensor var_7213_cast_fp16_17, tensor var_7213_cast_fp16_18, tensor var_7213_cast_fp16_19 = split(axis = var_7213_axis_0, split_sizes = tile_80, x = var_7167_cast_fp16)[name = tensor("op_7213_cast_fp16")]; + tensor aw_1041_equation_0 = const()[name = tensor("aw_1041_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1041_cast_fp16 = einsum(equation = aw_1041_equation_0, values = (var_7192_cast_fp16_0, var_7170_cast_fp16_0))[name = tensor("aw_1041_cast_fp16")]; + tensor aw_1043_equation_0 = const()[name = tensor("aw_1043_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1043_cast_fp16 = einsum(equation = aw_1043_equation_0, values = (var_7192_cast_fp16_1, var_7170_cast_fp16_1))[name = tensor("aw_1043_cast_fp16")]; + tensor aw_1045_equation_0 = const()[name = tensor("aw_1045_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1045_cast_fp16 = einsum(equation = aw_1045_equation_0, values = (var_7192_cast_fp16_2, var_7170_cast_fp16_2))[name = tensor("aw_1045_cast_fp16")]; + tensor aw_1047_equation_0 = const()[name = tensor("aw_1047_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1047_cast_fp16 = einsum(equation = aw_1047_equation_0, values = (var_7192_cast_fp16_3, var_7170_cast_fp16_3))[name = tensor("aw_1047_cast_fp16")]; + tensor aw_1049_equation_0 = const()[name = tensor("aw_1049_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1049_cast_fp16 = einsum(equation = aw_1049_equation_0, values = (var_7192_cast_fp16_4, var_7170_cast_fp16_4))[name = tensor("aw_1049_cast_fp16")]; + tensor aw_1051_equation_0 = const()[name = tensor("aw_1051_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1051_cast_fp16 = einsum(equation = aw_1051_equation_0, values = (var_7192_cast_fp16_5, var_7170_cast_fp16_5))[name = tensor("aw_1051_cast_fp16")]; + tensor aw_1053_equation_0 = const()[name = tensor("aw_1053_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1053_cast_fp16 = einsum(equation = aw_1053_equation_0, values = (var_7192_cast_fp16_6, var_7170_cast_fp16_6))[name = tensor("aw_1053_cast_fp16")]; + tensor aw_1055_equation_0 = const()[name = tensor("aw_1055_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1055_cast_fp16 = einsum(equation = aw_1055_equation_0, values = (var_7192_cast_fp16_7, var_7170_cast_fp16_7))[name = tensor("aw_1055_cast_fp16")]; + tensor aw_1057_equation_0 = const()[name = tensor("aw_1057_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1057_cast_fp16 = einsum(equation = aw_1057_equation_0, values = (var_7192_cast_fp16_8, var_7170_cast_fp16_8))[name = tensor("aw_1057_cast_fp16")]; + tensor aw_1059_equation_0 = const()[name = tensor("aw_1059_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1059_cast_fp16 = einsum(equation = aw_1059_equation_0, values = (var_7192_cast_fp16_9, var_7170_cast_fp16_9))[name = tensor("aw_1059_cast_fp16")]; + tensor aw_1061_equation_0 = const()[name = tensor("aw_1061_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1061_cast_fp16 = einsum(equation = aw_1061_equation_0, values = (var_7192_cast_fp16_10, var_7170_cast_fp16_10))[name = tensor("aw_1061_cast_fp16")]; + tensor aw_1063_equation_0 = const()[name = tensor("aw_1063_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1063_cast_fp16 = einsum(equation = aw_1063_equation_0, values = (var_7192_cast_fp16_11, var_7170_cast_fp16_11))[name = tensor("aw_1063_cast_fp16")]; + tensor aw_1065_equation_0 = const()[name = tensor("aw_1065_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1065_cast_fp16 = einsum(equation = aw_1065_equation_0, values = (var_7192_cast_fp16_12, var_7170_cast_fp16_12))[name = tensor("aw_1065_cast_fp16")]; + tensor aw_1067_equation_0 = const()[name = tensor("aw_1067_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1067_cast_fp16 = einsum(equation = aw_1067_equation_0, values = (var_7192_cast_fp16_13, var_7170_cast_fp16_13))[name = tensor("aw_1067_cast_fp16")]; + tensor aw_1069_equation_0 = const()[name = tensor("aw_1069_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1069_cast_fp16 = einsum(equation = aw_1069_equation_0, values = (var_7192_cast_fp16_14, var_7170_cast_fp16_14))[name = tensor("aw_1069_cast_fp16")]; + tensor aw_1071_equation_0 = const()[name = tensor("aw_1071_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1071_cast_fp16 = einsum(equation = aw_1071_equation_0, values = (var_7192_cast_fp16_15, var_7170_cast_fp16_15))[name = tensor("aw_1071_cast_fp16")]; + tensor aw_1073_equation_0 = const()[name = tensor("aw_1073_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1073_cast_fp16 = einsum(equation = aw_1073_equation_0, values = (var_7192_cast_fp16_16, var_7170_cast_fp16_16))[name = tensor("aw_1073_cast_fp16")]; + tensor aw_1075_equation_0 = const()[name = tensor("aw_1075_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1075_cast_fp16 = einsum(equation = aw_1075_equation_0, values = (var_7192_cast_fp16_17, var_7170_cast_fp16_17))[name = tensor("aw_1075_cast_fp16")]; + tensor aw_1077_equation_0 = const()[name = tensor("aw_1077_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1077_cast_fp16 = einsum(equation = aw_1077_equation_0, values = (var_7192_cast_fp16_18, var_7170_cast_fp16_18))[name = tensor("aw_1077_cast_fp16")]; + tensor aw_1079_equation_0 = const()[name = tensor("aw_1079_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1079_cast_fp16 = einsum(equation = aw_1079_equation_0, values = (var_7192_cast_fp16_19, var_7170_cast_fp16_19))[name = tensor("aw_1079_cast_fp16")]; + tensor var_7274_cast_fp16 = softmax(axis = var_7118, x = aw_1041_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor var_7275_cast_fp16 = softmax(axis = var_7118, x = aw_1043_cast_fp16)[name = tensor("op_7275_cast_fp16")]; + tensor var_7276_cast_fp16 = softmax(axis = var_7118, x = aw_1045_cast_fp16)[name = tensor("op_7276_cast_fp16")]; + tensor var_7277_cast_fp16 = softmax(axis = var_7118, x = aw_1047_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7278_cast_fp16 = softmax(axis = var_7118, x = aw_1049_cast_fp16)[name = tensor("op_7278_cast_fp16")]; + tensor var_7279_cast_fp16 = softmax(axis = var_7118, x = aw_1051_cast_fp16)[name = tensor("op_7279_cast_fp16")]; + tensor var_7280_cast_fp16 = softmax(axis = var_7118, x = aw_1053_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor var_7281_cast_fp16 = softmax(axis = var_7118, x = aw_1055_cast_fp16)[name = tensor("op_7281_cast_fp16")]; + tensor var_7282_cast_fp16 = softmax(axis = var_7118, x = aw_1057_cast_fp16)[name = tensor("op_7282_cast_fp16")]; + tensor var_7283_cast_fp16 = softmax(axis = var_7118, x = aw_1059_cast_fp16)[name = tensor("op_7283_cast_fp16")]; + tensor var_7284_cast_fp16 = softmax(axis = var_7118, x = aw_1061_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = softmax(axis = var_7118, x = aw_1063_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7286_cast_fp16 = softmax(axis = var_7118, x = aw_1065_cast_fp16)[name = tensor("op_7286_cast_fp16")]; + tensor var_7287_cast_fp16 = softmax(axis = var_7118, x = aw_1067_cast_fp16)[name = tensor("op_7287_cast_fp16")]; + tensor var_7288_cast_fp16 = softmax(axis = var_7118, x = aw_1069_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = softmax(axis = var_7118, x = aw_1071_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor var_7290_cast_fp16 = softmax(axis = var_7118, x = aw_1073_cast_fp16)[name = tensor("op_7290_cast_fp16")]; + tensor var_7291_cast_fp16 = softmax(axis = var_7118, x = aw_1075_cast_fp16)[name = tensor("op_7291_cast_fp16")]; + tensor var_7292_cast_fp16 = softmax(axis = var_7118, x = aw_1077_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = softmax(axis = var_7118, x = aw_1079_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7295_equation_0 = const()[name = tensor("op_7295_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7295_cast_fp16 = einsum(equation = var_7295_equation_0, values = (var_7213_cast_fp16_0, var_7274_cast_fp16))[name = tensor("op_7295_cast_fp16")]; + tensor var_7297_equation_0 = const()[name = tensor("op_7297_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7297_cast_fp16 = einsum(equation = var_7297_equation_0, values = (var_7213_cast_fp16_1, var_7275_cast_fp16))[name = tensor("op_7297_cast_fp16")]; + tensor var_7299_equation_0 = const()[name = tensor("op_7299_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7299_cast_fp16 = einsum(equation = var_7299_equation_0, values = (var_7213_cast_fp16_2, var_7276_cast_fp16))[name = tensor("op_7299_cast_fp16")]; + tensor var_7301_equation_0 = const()[name = tensor("op_7301_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7301_cast_fp16 = einsum(equation = var_7301_equation_0, values = (var_7213_cast_fp16_3, var_7277_cast_fp16))[name = tensor("op_7301_cast_fp16")]; + tensor var_7303_equation_0 = const()[name = tensor("op_7303_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7303_cast_fp16 = einsum(equation = var_7303_equation_0, values = (var_7213_cast_fp16_4, var_7278_cast_fp16))[name = tensor("op_7303_cast_fp16")]; + tensor var_7305_equation_0 = const()[name = tensor("op_7305_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7305_cast_fp16 = einsum(equation = var_7305_equation_0, values = (var_7213_cast_fp16_5, var_7279_cast_fp16))[name = tensor("op_7305_cast_fp16")]; + tensor var_7307_equation_0 = const()[name = tensor("op_7307_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7307_cast_fp16 = einsum(equation = var_7307_equation_0, values = (var_7213_cast_fp16_6, var_7280_cast_fp16))[name = tensor("op_7307_cast_fp16")]; + tensor var_7309_equation_0 = const()[name = tensor("op_7309_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7309_cast_fp16 = einsum(equation = var_7309_equation_0, values = (var_7213_cast_fp16_7, var_7281_cast_fp16))[name = tensor("op_7309_cast_fp16")]; + tensor var_7311_equation_0 = const()[name = tensor("op_7311_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7311_cast_fp16 = einsum(equation = var_7311_equation_0, values = (var_7213_cast_fp16_8, var_7282_cast_fp16))[name = tensor("op_7311_cast_fp16")]; + tensor var_7313_equation_0 = const()[name = tensor("op_7313_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7313_cast_fp16 = einsum(equation = var_7313_equation_0, values = (var_7213_cast_fp16_9, var_7283_cast_fp16))[name = tensor("op_7313_cast_fp16")]; + tensor var_7315_equation_0 = const()[name = tensor("op_7315_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7315_cast_fp16 = einsum(equation = var_7315_equation_0, values = (var_7213_cast_fp16_10, var_7284_cast_fp16))[name = tensor("op_7315_cast_fp16")]; + tensor var_7317_equation_0 = const()[name = tensor("op_7317_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7317_cast_fp16 = einsum(equation = var_7317_equation_0, values = (var_7213_cast_fp16_11, var_7285_cast_fp16))[name = tensor("op_7317_cast_fp16")]; + tensor var_7319_equation_0 = const()[name = tensor("op_7319_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7319_cast_fp16 = einsum(equation = var_7319_equation_0, values = (var_7213_cast_fp16_12, var_7286_cast_fp16))[name = tensor("op_7319_cast_fp16")]; + tensor var_7321_equation_0 = const()[name = tensor("op_7321_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7321_cast_fp16 = einsum(equation = var_7321_equation_0, values = (var_7213_cast_fp16_13, var_7287_cast_fp16))[name = tensor("op_7321_cast_fp16")]; + tensor var_7323_equation_0 = const()[name = tensor("op_7323_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7323_cast_fp16 = einsum(equation = var_7323_equation_0, values = (var_7213_cast_fp16_14, var_7288_cast_fp16))[name = tensor("op_7323_cast_fp16")]; + tensor var_7325_equation_0 = const()[name = tensor("op_7325_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7325_cast_fp16 = einsum(equation = var_7325_equation_0, values = (var_7213_cast_fp16_15, var_7289_cast_fp16))[name = tensor("op_7325_cast_fp16")]; + tensor var_7327_equation_0 = const()[name = tensor("op_7327_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7327_cast_fp16 = einsum(equation = var_7327_equation_0, values = (var_7213_cast_fp16_16, var_7290_cast_fp16))[name = tensor("op_7327_cast_fp16")]; + tensor var_7329_equation_0 = const()[name = tensor("op_7329_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7329_cast_fp16 = einsum(equation = var_7329_equation_0, values = (var_7213_cast_fp16_17, var_7291_cast_fp16))[name = tensor("op_7329_cast_fp16")]; + tensor var_7331_equation_0 = const()[name = tensor("op_7331_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7331_cast_fp16 = einsum(equation = var_7331_equation_0, values = (var_7213_cast_fp16_18, var_7292_cast_fp16))[name = tensor("op_7331_cast_fp16")]; + tensor var_7333_equation_0 = const()[name = tensor("op_7333_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7333_cast_fp16 = einsum(equation = var_7333_equation_0, values = (var_7213_cast_fp16_19, var_7293_cast_fp16))[name = tensor("op_7333_cast_fp16")]; + tensor input_265_interleave_0 = const()[name = tensor("input_265_interleave_0"), val = tensor(false)]; + tensor input_265_cast_fp16 = concat(axis = var_7118, interleave = input_265_interleave_0, values = (var_7295_cast_fp16, var_7297_cast_fp16, var_7299_cast_fp16, var_7301_cast_fp16, var_7303_cast_fp16, var_7305_cast_fp16, var_7307_cast_fp16, var_7309_cast_fp16, var_7311_cast_fp16, var_7313_cast_fp16, var_7315_cast_fp16, var_7317_cast_fp16, var_7319_cast_fp16, var_7321_cast_fp16, var_7323_cast_fp16, var_7325_cast_fp16, var_7327_cast_fp16, var_7329_cast_fp16, var_7331_cast_fp16, var_7333_cast_fp16))[name = tensor("input_265_cast_fp16")]; + tensor var_7342_pad_type_0 = const()[name = tensor("op_7342_pad_type_0"), val = tensor("valid")]; + tensor var_7342_strides_0 = const()[name = tensor("op_7342_strides_0"), val = tensor([1, 1])]; + tensor var_7342_pad_0 = const()[name = tensor("op_7342_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7342_dilations_0 = const()[name = tensor("op_7342_dilations_0"), val = tensor([1, 1])]; + tensor var_7342_groups_0 = const()[name = tensor("op_7342_groups_0"), val = tensor(1)]; + tensor blocks_26_attn_out_weight_to_fp16 = const()[name = tensor("blocks_26_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047685312)))]; + tensor blocks_26_attn_out_bias_to_fp16 = const()[name = tensor("blocks_26_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050962176)))]; + tensor var_7342_cast_fp16 = conv(bias = blocks_26_attn_out_bias_to_fp16, dilations = var_7342_dilations_0, groups = var_7342_groups_0, pad = var_7342_pad_0, pad_type = var_7342_pad_type_0, strides = var_7342_strides_0, weight = blocks_26_attn_out_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("op_7342_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = var_7342_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([1])]; + tensor input_267_gamma_0_to_fp16 = const()[name = tensor("input_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050964800)))]; + tensor input_267_beta_0_to_fp16 = const()[name = tensor("input_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050967424)))]; + tensor var_7352_to_fp16 = const()[name = tensor("op_7352_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = input_267_beta_0_to_fp16, epsilon = var_7352_to_fp16, gamma = input_267_gamma_0_to_fp16, x = inputs_107_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_pad_type_0 = const()[name = tensor("input_269_pad_type_0"), val = tensor("valid")]; + tensor input_269_strides_0 = const()[name = tensor("input_269_strides_0"), val = tensor([1, 1])]; + tensor input_269_pad_0 = const()[name = tensor("input_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_269_dilations_0 = const()[name = tensor("input_269_dilations_0"), val = tensor([1, 1])]; + tensor input_269_groups_0 = const()[name = tensor("input_269_groups_0"), val = tensor(1)]; + tensor blocks_26_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050970048)))]; + tensor blocks_26_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1064077312)))]; + tensor input_269_cast_fp16 = conv(bias = blocks_26_mlp_0_bias_to_fp16, dilations = input_269_dilations_0, groups = input_269_groups_0, pad = input_269_pad_0, pad_type = input_269_pad_type_0, strides = input_269_strides_0, weight = blocks_26_mlp_0_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor input_271_mode_0 = const()[name = tensor("input_271_mode_0"), val = tensor("EXACT")]; + tensor input_271_cast_fp16 = gelu(mode = input_271_mode_0, x = input_269_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor var_7378_pad_type_0 = const()[name = tensor("op_7378_pad_type_0"), val = tensor("valid")]; + tensor var_7378_strides_0 = const()[name = tensor("op_7378_strides_0"), val = tensor([1, 1])]; + tensor var_7378_pad_0 = const()[name = tensor("op_7378_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7378_dilations_0 = const()[name = tensor("op_7378_dilations_0"), val = tensor([1, 1])]; + tensor var_7378_groups_0 = const()[name = tensor("op_7378_groups_0"), val = tensor(1)]; + tensor blocks_26_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1064087616)))]; + tensor blocks_26_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077194880)))]; + tensor var_7378_cast_fp16 = conv(bias = blocks_26_mlp_2_bias_to_fp16, dilations = var_7378_dilations_0, groups = var_7378_groups_0, pad = var_7378_pad_0, pad_type = var_7378_pad_type_0, strides = var_7378_strides_0, weight = blocks_26_mlp_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("op_7378_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_7378_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_7387 = const()[name = tensor("op_7387"), val = tensor(1)]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([1])]; + tensor input_273_gamma_0_to_fp16 = const()[name = tensor("input_273_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077197504)))]; + tensor input_273_beta_0_to_fp16 = const()[name = tensor("input_273_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077200128)))]; + tensor var_7403_to_fp16 = const()[name = tensor("op_7403_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = input_273_beta_0_to_fp16, epsilon = var_7403_to_fp16, gamma = input_273_gamma_0_to_fp16, x = inputs_109_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor q_55_pad_type_0 = const()[name = tensor("q_55_pad_type_0"), val = tensor("valid")]; + tensor q_55_strides_0 = const()[name = tensor("q_55_strides_0"), val = tensor([1, 1])]; + tensor q_55_pad_0 = const()[name = tensor("q_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_55_dilations_0 = const()[name = tensor("q_55_dilations_0"), val = tensor([1, 1])]; + tensor q_55_groups_0 = const()[name = tensor("q_55_groups_0"), val = tensor(1)]; + tensor var_7438_weight_0_to_fp16 = const()[name = tensor("op_7438_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077202752)))]; + tensor var_7438_bias_0_to_fp16 = const()[name = tensor("op_7438_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080479616)))]; + tensor var_7438_cast_fp16 = conv(bias = var_7438_bias_0_to_fp16, dilations = q_55_dilations_0, groups = q_55_groups_0, pad = q_55_pad_0, pad_type = q_55_pad_type_0, strides = q_55_strides_0, weight = var_7438_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("op_7438_cast_fp16")]; + tensor k_55_pad_type_0 = const()[name = tensor("k_55_pad_type_0"), val = tensor("valid")]; + tensor k_55_strides_0 = const()[name = tensor("k_55_strides_0"), val = tensor([1, 1])]; + tensor k_55_pad_0 = const()[name = tensor("k_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_55_dilations_0 = const()[name = tensor("k_55_dilations_0"), val = tensor([1, 1])]; + tensor k_55_groups_0 = const()[name = tensor("k_55_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_key_weight_to_fp16 = const()[name = tensor("blocks_27_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080482240)))]; + tensor k_55_cast_fp16 = conv(dilations = k_55_dilations_0, groups = k_55_groups_0, pad = k_55_pad_0, pad_type = k_55_pad_type_0, strides = k_55_strides_0, weight = blocks_27_attn_key_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor var_7436_pad_type_0 = const()[name = tensor("op_7436_pad_type_0"), val = tensor("valid")]; + tensor var_7436_strides_0 = const()[name = tensor("op_7436_strides_0"), val = tensor([1, 1])]; + tensor var_7436_pad_0 = const()[name = tensor("op_7436_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7436_dilations_0 = const()[name = tensor("op_7436_dilations_0"), val = tensor([1, 1])]; + tensor var_7436_groups_0 = const()[name = tensor("op_7436_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_value_weight_to_fp16 = const()[name = tensor("blocks_27_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083759104)))]; + tensor blocks_27_attn_value_bias_to_fp16 = const()[name = tensor("blocks_27_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087035968)))]; + tensor var_7436_cast_fp16 = conv(bias = blocks_27_attn_value_bias_to_fp16, dilations = var_7436_dilations_0, groups = var_7436_groups_0, pad = var_7436_pad_0, pad_type = var_7436_pad_type_0, strides = var_7436_strides_0, weight = blocks_27_attn_value_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("op_7436_cast_fp16")]; + tensor tile_81 = const()[name = tensor("tile_81"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_cast_fp16_0, tensor var_7439_cast_fp16_1, tensor var_7439_cast_fp16_2, tensor var_7439_cast_fp16_3, tensor var_7439_cast_fp16_4, tensor var_7439_cast_fp16_5, tensor var_7439_cast_fp16_6, tensor var_7439_cast_fp16_7, tensor var_7439_cast_fp16_8, tensor var_7439_cast_fp16_9, tensor var_7439_cast_fp16_10, tensor var_7439_cast_fp16_11, tensor var_7439_cast_fp16_12, tensor var_7439_cast_fp16_13, tensor var_7439_cast_fp16_14, tensor var_7439_cast_fp16_15, tensor var_7439_cast_fp16_16, tensor var_7439_cast_fp16_17, tensor var_7439_cast_fp16_18, tensor var_7439_cast_fp16_19 = split(axis = var_7439_axis_0, split_sizes = tile_81, x = var_7438_cast_fp16)[name = tensor("op_7439_cast_fp16")]; + tensor var_7460_perm_0 = const()[name = tensor("op_7460_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_82 = const()[name = tensor("tile_82"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7461_axis_0 = const()[name = tensor("op_7461_axis_0"), val = tensor(3)]; + tensor var_7460_cast_fp16 = transpose(perm = var_7460_perm_0, x = k_55_cast_fp16)[name = tensor("transpose_5")]; + tensor var_7461_cast_fp16_0, tensor var_7461_cast_fp16_1, tensor var_7461_cast_fp16_2, tensor var_7461_cast_fp16_3, tensor var_7461_cast_fp16_4, tensor var_7461_cast_fp16_5, tensor var_7461_cast_fp16_6, tensor var_7461_cast_fp16_7, tensor var_7461_cast_fp16_8, tensor var_7461_cast_fp16_9, tensor var_7461_cast_fp16_10, tensor var_7461_cast_fp16_11, tensor var_7461_cast_fp16_12, tensor var_7461_cast_fp16_13, tensor var_7461_cast_fp16_14, tensor var_7461_cast_fp16_15, tensor var_7461_cast_fp16_16, tensor var_7461_cast_fp16_17, tensor var_7461_cast_fp16_18, tensor var_7461_cast_fp16_19 = split(axis = var_7461_axis_0, split_sizes = tile_82, x = var_7460_cast_fp16)[name = tensor("op_7461_cast_fp16")]; + tensor tile_83 = const()[name = tensor("tile_83"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7482_axis_0 = const()[name = tensor("op_7482_axis_0"), val = tensor(1)]; + tensor var_7482_cast_fp16_0, tensor var_7482_cast_fp16_1, tensor var_7482_cast_fp16_2, tensor var_7482_cast_fp16_3, tensor var_7482_cast_fp16_4, tensor var_7482_cast_fp16_5, tensor var_7482_cast_fp16_6, tensor var_7482_cast_fp16_7, tensor var_7482_cast_fp16_8, tensor var_7482_cast_fp16_9, tensor var_7482_cast_fp16_10, tensor var_7482_cast_fp16_11, tensor var_7482_cast_fp16_12, tensor var_7482_cast_fp16_13, tensor var_7482_cast_fp16_14, tensor var_7482_cast_fp16_15, tensor var_7482_cast_fp16_16, tensor var_7482_cast_fp16_17, tensor var_7482_cast_fp16_18, tensor var_7482_cast_fp16_19 = split(axis = var_7482_axis_0, split_sizes = tile_83, x = var_7436_cast_fp16)[name = tensor("op_7482_cast_fp16")]; + tensor aw_1081_equation_0 = const()[name = tensor("aw_1081_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1081_cast_fp16 = einsum(equation = aw_1081_equation_0, values = (var_7461_cast_fp16_0, var_7439_cast_fp16_0))[name = tensor("aw_1081_cast_fp16")]; + tensor aw_1083_equation_0 = const()[name = tensor("aw_1083_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1083_cast_fp16 = einsum(equation = aw_1083_equation_0, values = (var_7461_cast_fp16_1, var_7439_cast_fp16_1))[name = tensor("aw_1083_cast_fp16")]; + tensor aw_1085_equation_0 = const()[name = tensor("aw_1085_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1085_cast_fp16 = einsum(equation = aw_1085_equation_0, values = (var_7461_cast_fp16_2, var_7439_cast_fp16_2))[name = tensor("aw_1085_cast_fp16")]; + tensor aw_1087_equation_0 = const()[name = tensor("aw_1087_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1087_cast_fp16 = einsum(equation = aw_1087_equation_0, values = (var_7461_cast_fp16_3, var_7439_cast_fp16_3))[name = tensor("aw_1087_cast_fp16")]; + tensor aw_1089_equation_0 = const()[name = tensor("aw_1089_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1089_cast_fp16 = einsum(equation = aw_1089_equation_0, values = (var_7461_cast_fp16_4, var_7439_cast_fp16_4))[name = tensor("aw_1089_cast_fp16")]; + tensor aw_1091_equation_0 = const()[name = tensor("aw_1091_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1091_cast_fp16 = einsum(equation = aw_1091_equation_0, values = (var_7461_cast_fp16_5, var_7439_cast_fp16_5))[name = tensor("aw_1091_cast_fp16")]; + tensor aw_1093_equation_0 = const()[name = tensor("aw_1093_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1093_cast_fp16 = einsum(equation = aw_1093_equation_0, values = (var_7461_cast_fp16_6, var_7439_cast_fp16_6))[name = tensor("aw_1093_cast_fp16")]; + tensor aw_1095_equation_0 = const()[name = tensor("aw_1095_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1095_cast_fp16 = einsum(equation = aw_1095_equation_0, values = (var_7461_cast_fp16_7, var_7439_cast_fp16_7))[name = tensor("aw_1095_cast_fp16")]; + tensor aw_1097_equation_0 = const()[name = tensor("aw_1097_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1097_cast_fp16 = einsum(equation = aw_1097_equation_0, values = (var_7461_cast_fp16_8, var_7439_cast_fp16_8))[name = tensor("aw_1097_cast_fp16")]; + tensor aw_1099_equation_0 = const()[name = tensor("aw_1099_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1099_cast_fp16 = einsum(equation = aw_1099_equation_0, values = (var_7461_cast_fp16_9, var_7439_cast_fp16_9))[name = tensor("aw_1099_cast_fp16")]; + tensor aw_1101_equation_0 = const()[name = tensor("aw_1101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1101_cast_fp16 = einsum(equation = aw_1101_equation_0, values = (var_7461_cast_fp16_10, var_7439_cast_fp16_10))[name = tensor("aw_1101_cast_fp16")]; + tensor aw_1103_equation_0 = const()[name = tensor("aw_1103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1103_cast_fp16 = einsum(equation = aw_1103_equation_0, values = (var_7461_cast_fp16_11, var_7439_cast_fp16_11))[name = tensor("aw_1103_cast_fp16")]; + tensor aw_1105_equation_0 = const()[name = tensor("aw_1105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1105_cast_fp16 = einsum(equation = aw_1105_equation_0, values = (var_7461_cast_fp16_12, var_7439_cast_fp16_12))[name = tensor("aw_1105_cast_fp16")]; + tensor aw_1107_equation_0 = const()[name = tensor("aw_1107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1107_cast_fp16 = einsum(equation = aw_1107_equation_0, values = (var_7461_cast_fp16_13, var_7439_cast_fp16_13))[name = tensor("aw_1107_cast_fp16")]; + tensor aw_1109_equation_0 = const()[name = tensor("aw_1109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1109_cast_fp16 = einsum(equation = aw_1109_equation_0, values = (var_7461_cast_fp16_14, var_7439_cast_fp16_14))[name = tensor("aw_1109_cast_fp16")]; + tensor aw_1111_equation_0 = const()[name = tensor("aw_1111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1111_cast_fp16 = einsum(equation = aw_1111_equation_0, values = (var_7461_cast_fp16_15, var_7439_cast_fp16_15))[name = tensor("aw_1111_cast_fp16")]; + tensor aw_1113_equation_0 = const()[name = tensor("aw_1113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1113_cast_fp16 = einsum(equation = aw_1113_equation_0, values = (var_7461_cast_fp16_16, var_7439_cast_fp16_16))[name = tensor("aw_1113_cast_fp16")]; + tensor aw_1115_equation_0 = const()[name = tensor("aw_1115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1115_cast_fp16 = einsum(equation = aw_1115_equation_0, values = (var_7461_cast_fp16_17, var_7439_cast_fp16_17))[name = tensor("aw_1115_cast_fp16")]; + tensor aw_1117_equation_0 = const()[name = tensor("aw_1117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1117_cast_fp16 = einsum(equation = aw_1117_equation_0, values = (var_7461_cast_fp16_18, var_7439_cast_fp16_18))[name = tensor("aw_1117_cast_fp16")]; + tensor aw_1119_equation_0 = const()[name = tensor("aw_1119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1119_cast_fp16 = einsum(equation = aw_1119_equation_0, values = (var_7461_cast_fp16_19, var_7439_cast_fp16_19))[name = tensor("aw_1119_cast_fp16")]; + tensor var_7543_cast_fp16 = softmax(axis = var_7387, x = aw_1081_cast_fp16)[name = tensor("op_7543_cast_fp16")]; + tensor var_7544_cast_fp16 = softmax(axis = var_7387, x = aw_1083_cast_fp16)[name = tensor("op_7544_cast_fp16")]; + tensor var_7545_cast_fp16 = softmax(axis = var_7387, x = aw_1085_cast_fp16)[name = tensor("op_7545_cast_fp16")]; + tensor var_7546_cast_fp16 = softmax(axis = var_7387, x = aw_1087_cast_fp16)[name = tensor("op_7546_cast_fp16")]; + tensor var_7547_cast_fp16 = softmax(axis = var_7387, x = aw_1089_cast_fp16)[name = tensor("op_7547_cast_fp16")]; + tensor var_7548_cast_fp16 = softmax(axis = var_7387, x = aw_1091_cast_fp16)[name = tensor("op_7548_cast_fp16")]; + tensor var_7549_cast_fp16 = softmax(axis = var_7387, x = aw_1093_cast_fp16)[name = tensor("op_7549_cast_fp16")]; + tensor var_7550_cast_fp16 = softmax(axis = var_7387, x = aw_1095_cast_fp16)[name = tensor("op_7550_cast_fp16")]; + tensor var_7551_cast_fp16 = softmax(axis = var_7387, x = aw_1097_cast_fp16)[name = tensor("op_7551_cast_fp16")]; + tensor var_7552_cast_fp16 = softmax(axis = var_7387, x = aw_1099_cast_fp16)[name = tensor("op_7552_cast_fp16")]; + tensor var_7553_cast_fp16 = softmax(axis = var_7387, x = aw_1101_cast_fp16)[name = tensor("op_7553_cast_fp16")]; + tensor var_7554_cast_fp16 = softmax(axis = var_7387, x = aw_1103_cast_fp16)[name = tensor("op_7554_cast_fp16")]; + tensor var_7555_cast_fp16 = softmax(axis = var_7387, x = aw_1105_cast_fp16)[name = tensor("op_7555_cast_fp16")]; + tensor var_7556_cast_fp16 = softmax(axis = var_7387, x = aw_1107_cast_fp16)[name = tensor("op_7556_cast_fp16")]; + tensor var_7557_cast_fp16 = softmax(axis = var_7387, x = aw_1109_cast_fp16)[name = tensor("op_7557_cast_fp16")]; + tensor var_7558_cast_fp16 = softmax(axis = var_7387, x = aw_1111_cast_fp16)[name = tensor("op_7558_cast_fp16")]; + tensor var_7559_cast_fp16 = softmax(axis = var_7387, x = aw_1113_cast_fp16)[name = tensor("op_7559_cast_fp16")]; + tensor var_7560_cast_fp16 = softmax(axis = var_7387, x = aw_1115_cast_fp16)[name = tensor("op_7560_cast_fp16")]; + tensor var_7561_cast_fp16 = softmax(axis = var_7387, x = aw_1117_cast_fp16)[name = tensor("op_7561_cast_fp16")]; + tensor var_7562_cast_fp16 = softmax(axis = var_7387, x = aw_1119_cast_fp16)[name = tensor("op_7562_cast_fp16")]; + tensor var_7564_equation_0 = const()[name = tensor("op_7564_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7564_cast_fp16 = einsum(equation = var_7564_equation_0, values = (var_7482_cast_fp16_0, var_7543_cast_fp16))[name = tensor("op_7564_cast_fp16")]; + tensor var_7566_equation_0 = const()[name = tensor("op_7566_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7566_cast_fp16 = einsum(equation = var_7566_equation_0, values = (var_7482_cast_fp16_1, var_7544_cast_fp16))[name = tensor("op_7566_cast_fp16")]; + tensor var_7568_equation_0 = const()[name = tensor("op_7568_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7568_cast_fp16 = einsum(equation = var_7568_equation_0, values = (var_7482_cast_fp16_2, var_7545_cast_fp16))[name = tensor("op_7568_cast_fp16")]; + tensor var_7570_equation_0 = const()[name = tensor("op_7570_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7570_cast_fp16 = einsum(equation = var_7570_equation_0, values = (var_7482_cast_fp16_3, var_7546_cast_fp16))[name = tensor("op_7570_cast_fp16")]; + tensor var_7572_equation_0 = const()[name = tensor("op_7572_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7572_cast_fp16 = einsum(equation = var_7572_equation_0, values = (var_7482_cast_fp16_4, var_7547_cast_fp16))[name = tensor("op_7572_cast_fp16")]; + tensor var_7574_equation_0 = const()[name = tensor("op_7574_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7574_cast_fp16 = einsum(equation = var_7574_equation_0, values = (var_7482_cast_fp16_5, var_7548_cast_fp16))[name = tensor("op_7574_cast_fp16")]; + tensor var_7576_equation_0 = const()[name = tensor("op_7576_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7576_cast_fp16 = einsum(equation = var_7576_equation_0, values = (var_7482_cast_fp16_6, var_7549_cast_fp16))[name = tensor("op_7576_cast_fp16")]; + tensor var_7578_equation_0 = const()[name = tensor("op_7578_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7578_cast_fp16 = einsum(equation = var_7578_equation_0, values = (var_7482_cast_fp16_7, var_7550_cast_fp16))[name = tensor("op_7578_cast_fp16")]; + tensor var_7580_equation_0 = const()[name = tensor("op_7580_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7580_cast_fp16 = einsum(equation = var_7580_equation_0, values = (var_7482_cast_fp16_8, var_7551_cast_fp16))[name = tensor("op_7580_cast_fp16")]; + tensor var_7582_equation_0 = const()[name = tensor("op_7582_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7582_cast_fp16 = einsum(equation = var_7582_equation_0, values = (var_7482_cast_fp16_9, var_7552_cast_fp16))[name = tensor("op_7582_cast_fp16")]; + tensor var_7584_equation_0 = const()[name = tensor("op_7584_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7584_cast_fp16 = einsum(equation = var_7584_equation_0, values = (var_7482_cast_fp16_10, var_7553_cast_fp16))[name = tensor("op_7584_cast_fp16")]; + tensor var_7586_equation_0 = const()[name = tensor("op_7586_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7586_cast_fp16 = einsum(equation = var_7586_equation_0, values = (var_7482_cast_fp16_11, var_7554_cast_fp16))[name = tensor("op_7586_cast_fp16")]; + tensor var_7588_equation_0 = const()[name = tensor("op_7588_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7588_cast_fp16 = einsum(equation = var_7588_equation_0, values = (var_7482_cast_fp16_12, var_7555_cast_fp16))[name = tensor("op_7588_cast_fp16")]; + tensor var_7590_equation_0 = const()[name = tensor("op_7590_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7590_cast_fp16 = einsum(equation = var_7590_equation_0, values = (var_7482_cast_fp16_13, var_7556_cast_fp16))[name = tensor("op_7590_cast_fp16")]; + tensor var_7592_equation_0 = const()[name = tensor("op_7592_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7592_cast_fp16 = einsum(equation = var_7592_equation_0, values = (var_7482_cast_fp16_14, var_7557_cast_fp16))[name = tensor("op_7592_cast_fp16")]; + tensor var_7594_equation_0 = const()[name = tensor("op_7594_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7594_cast_fp16 = einsum(equation = var_7594_equation_0, values = (var_7482_cast_fp16_15, var_7558_cast_fp16))[name = tensor("op_7594_cast_fp16")]; + tensor var_7596_equation_0 = const()[name = tensor("op_7596_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7596_cast_fp16 = einsum(equation = var_7596_equation_0, values = (var_7482_cast_fp16_16, var_7559_cast_fp16))[name = tensor("op_7596_cast_fp16")]; + tensor var_7598_equation_0 = const()[name = tensor("op_7598_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7598_cast_fp16 = einsum(equation = var_7598_equation_0, values = (var_7482_cast_fp16_17, var_7560_cast_fp16))[name = tensor("op_7598_cast_fp16")]; + tensor var_7600_equation_0 = const()[name = tensor("op_7600_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7600_cast_fp16 = einsum(equation = var_7600_equation_0, values = (var_7482_cast_fp16_18, var_7561_cast_fp16))[name = tensor("op_7600_cast_fp16")]; + tensor var_7602_equation_0 = const()[name = tensor("op_7602_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7602_cast_fp16 = einsum(equation = var_7602_equation_0, values = (var_7482_cast_fp16_19, var_7562_cast_fp16))[name = tensor("op_7602_cast_fp16")]; + tensor input_275_interleave_0 = const()[name = tensor("input_275_interleave_0"), val = tensor(false)]; + tensor input_275_cast_fp16 = concat(axis = var_7387, interleave = input_275_interleave_0, values = (var_7564_cast_fp16, var_7566_cast_fp16, var_7568_cast_fp16, var_7570_cast_fp16, var_7572_cast_fp16, var_7574_cast_fp16, var_7576_cast_fp16, var_7578_cast_fp16, var_7580_cast_fp16, var_7582_cast_fp16, var_7584_cast_fp16, var_7586_cast_fp16, var_7588_cast_fp16, var_7590_cast_fp16, var_7592_cast_fp16, var_7594_cast_fp16, var_7596_cast_fp16, var_7598_cast_fp16, var_7600_cast_fp16, var_7602_cast_fp16))[name = tensor("input_275_cast_fp16")]; + tensor var_7611_pad_type_0 = const()[name = tensor("op_7611_pad_type_0"), val = tensor("valid")]; + tensor var_7611_strides_0 = const()[name = tensor("op_7611_strides_0"), val = tensor([1, 1])]; + tensor var_7611_pad_0 = const()[name = tensor("op_7611_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7611_dilations_0 = const()[name = tensor("op_7611_dilations_0"), val = tensor([1, 1])]; + tensor var_7611_groups_0 = const()[name = tensor("op_7611_groups_0"), val = tensor(1)]; + tensor blocks_27_attn_out_weight_to_fp16 = const()[name = tensor("blocks_27_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087038592)))]; + tensor blocks_27_attn_out_bias_to_fp16 = const()[name = tensor("blocks_27_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090315456)))]; + tensor var_7611_cast_fp16 = conv(bias = blocks_27_attn_out_bias_to_fp16, dilations = var_7611_dilations_0, groups = var_7611_groups_0, pad = var_7611_pad_0, pad_type = var_7611_pad_type_0, strides = var_7611_strides_0, weight = blocks_27_attn_out_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("op_7611_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = var_7611_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor input_277_axes_0 = const()[name = tensor("input_277_axes_0"), val = tensor([1])]; + tensor input_277_gamma_0_to_fp16 = const()[name = tensor("input_277_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090318080)))]; + tensor input_277_beta_0_to_fp16 = const()[name = tensor("input_277_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090320704)))]; + tensor var_7621_to_fp16 = const()[name = tensor("op_7621_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_277_cast_fp16 = layer_norm(axes = input_277_axes_0, beta = input_277_beta_0_to_fp16, epsilon = var_7621_to_fp16, gamma = input_277_gamma_0_to_fp16, x = inputs_111_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_pad_type_0 = const()[name = tensor("input_279_pad_type_0"), val = tensor("valid")]; + tensor input_279_strides_0 = const()[name = tensor("input_279_strides_0"), val = tensor([1, 1])]; + tensor input_279_pad_0 = const()[name = tensor("input_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_279_dilations_0 = const()[name = tensor("input_279_dilations_0"), val = tensor([1, 1])]; + tensor input_279_groups_0 = const()[name = tensor("input_279_groups_0"), val = tensor(1)]; + tensor blocks_27_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090323328)))]; + tensor blocks_27_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103430592)))]; + tensor input_279_cast_fp16 = conv(bias = blocks_27_mlp_0_bias_to_fp16, dilations = input_279_dilations_0, groups = input_279_groups_0, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = input_279_strides_0, weight = blocks_27_mlp_0_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_mode_0 = const()[name = tensor("input_281_mode_0"), val = tensor("EXACT")]; + tensor input_281_cast_fp16 = gelu(mode = input_281_mode_0, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_7647_pad_type_0 = const()[name = tensor("op_7647_pad_type_0"), val = tensor("valid")]; + tensor var_7647_strides_0 = const()[name = tensor("op_7647_strides_0"), val = tensor([1, 1])]; + tensor var_7647_pad_0 = const()[name = tensor("op_7647_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7647_dilations_0 = const()[name = tensor("op_7647_dilations_0"), val = tensor([1, 1])]; + tensor var_7647_groups_0 = const()[name = tensor("op_7647_groups_0"), val = tensor(1)]; + tensor blocks_27_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103440896)))]; + tensor blocks_27_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116548160)))]; + tensor var_7647_cast_fp16 = conv(bias = blocks_27_mlp_2_bias_to_fp16, dilations = var_7647_dilations_0, groups = var_7647_groups_0, pad = var_7647_pad_0, pad_type = var_7647_pad_type_0, strides = var_7647_strides_0, weight = blocks_27_mlp_2_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("op_7647_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_7647_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_7656 = const()[name = tensor("op_7656"), val = tensor(1)]; + tensor input_283_axes_0 = const()[name = tensor("input_283_axes_0"), val = tensor([1])]; + tensor input_283_gamma_0_to_fp16 = const()[name = tensor("input_283_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116550784)))]; + tensor input_283_beta_0_to_fp16 = const()[name = tensor("input_283_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116553408)))]; + tensor var_7672_to_fp16 = const()[name = tensor("op_7672_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = input_283_beta_0_to_fp16, epsilon = var_7672_to_fp16, gamma = input_283_gamma_0_to_fp16, x = inputs_113_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor q_57_pad_type_0 = const()[name = tensor("q_57_pad_type_0"), val = tensor("valid")]; + tensor q_57_strides_0 = const()[name = tensor("q_57_strides_0"), val = tensor([1, 1])]; + tensor q_57_pad_0 = const()[name = tensor("q_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_57_dilations_0 = const()[name = tensor("q_57_dilations_0"), val = tensor([1, 1])]; + tensor q_57_groups_0 = const()[name = tensor("q_57_groups_0"), val = tensor(1)]; + tensor var_7707_weight_0_to_fp16 = const()[name = tensor("op_7707_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116556032)))]; + tensor var_7707_bias_0_to_fp16 = const()[name = tensor("op_7707_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119832896)))]; + tensor var_7707_cast_fp16 = conv(bias = var_7707_bias_0_to_fp16, dilations = q_57_dilations_0, groups = q_57_groups_0, pad = q_57_pad_0, pad_type = q_57_pad_type_0, strides = q_57_strides_0, weight = var_7707_weight_0_to_fp16, x = input_283_cast_fp16)[name = tensor("op_7707_cast_fp16")]; + tensor k_57_pad_type_0 = const()[name = tensor("k_57_pad_type_0"), val = tensor("valid")]; + tensor k_57_strides_0 = const()[name = tensor("k_57_strides_0"), val = tensor([1, 1])]; + tensor k_57_pad_0 = const()[name = tensor("k_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_57_dilations_0 = const()[name = tensor("k_57_dilations_0"), val = tensor([1, 1])]; + tensor k_57_groups_0 = const()[name = tensor("k_57_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_key_weight_to_fp16 = const()[name = tensor("blocks_28_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119835520)))]; + tensor k_57_cast_fp16 = conv(dilations = k_57_dilations_0, groups = k_57_groups_0, pad = k_57_pad_0, pad_type = k_57_pad_type_0, strides = k_57_strides_0, weight = blocks_28_attn_key_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor var_7705_pad_type_0 = const()[name = tensor("op_7705_pad_type_0"), val = tensor("valid")]; + tensor var_7705_strides_0 = const()[name = tensor("op_7705_strides_0"), val = tensor([1, 1])]; + tensor var_7705_pad_0 = const()[name = tensor("op_7705_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7705_dilations_0 = const()[name = tensor("op_7705_dilations_0"), val = tensor([1, 1])]; + tensor var_7705_groups_0 = const()[name = tensor("op_7705_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_value_weight_to_fp16 = const()[name = tensor("blocks_28_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1123112384)))]; + tensor blocks_28_attn_value_bias_to_fp16 = const()[name = tensor("blocks_28_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126389248)))]; + tensor var_7705_cast_fp16 = conv(bias = blocks_28_attn_value_bias_to_fp16, dilations = var_7705_dilations_0, groups = var_7705_groups_0, pad = var_7705_pad_0, pad_type = var_7705_pad_type_0, strides = var_7705_strides_0, weight = blocks_28_attn_value_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("op_7705_cast_fp16")]; + tensor tile_84 = const()[name = tensor("tile_84"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7708_axis_0 = const()[name = tensor("op_7708_axis_0"), val = tensor(1)]; + tensor var_7708_cast_fp16_0, tensor var_7708_cast_fp16_1, tensor var_7708_cast_fp16_2, tensor var_7708_cast_fp16_3, tensor var_7708_cast_fp16_4, tensor var_7708_cast_fp16_5, tensor var_7708_cast_fp16_6, tensor var_7708_cast_fp16_7, tensor var_7708_cast_fp16_8, tensor var_7708_cast_fp16_9, tensor var_7708_cast_fp16_10, tensor var_7708_cast_fp16_11, tensor var_7708_cast_fp16_12, tensor var_7708_cast_fp16_13, tensor var_7708_cast_fp16_14, tensor var_7708_cast_fp16_15, tensor var_7708_cast_fp16_16, tensor var_7708_cast_fp16_17, tensor var_7708_cast_fp16_18, tensor var_7708_cast_fp16_19 = split(axis = var_7708_axis_0, split_sizes = tile_84, x = var_7707_cast_fp16)[name = tensor("op_7708_cast_fp16")]; + tensor var_7729_perm_0 = const()[name = tensor("op_7729_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_85 = const()[name = tensor("tile_85"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7730_axis_0 = const()[name = tensor("op_7730_axis_0"), val = tensor(3)]; + tensor var_7729_cast_fp16 = transpose(perm = var_7729_perm_0, x = k_57_cast_fp16)[name = tensor("transpose_4")]; + tensor var_7730_cast_fp16_0, tensor var_7730_cast_fp16_1, tensor var_7730_cast_fp16_2, tensor var_7730_cast_fp16_3, tensor var_7730_cast_fp16_4, tensor var_7730_cast_fp16_5, tensor var_7730_cast_fp16_6, tensor var_7730_cast_fp16_7, tensor var_7730_cast_fp16_8, tensor var_7730_cast_fp16_9, tensor var_7730_cast_fp16_10, tensor var_7730_cast_fp16_11, tensor var_7730_cast_fp16_12, tensor var_7730_cast_fp16_13, tensor var_7730_cast_fp16_14, tensor var_7730_cast_fp16_15, tensor var_7730_cast_fp16_16, tensor var_7730_cast_fp16_17, tensor var_7730_cast_fp16_18, tensor var_7730_cast_fp16_19 = split(axis = var_7730_axis_0, split_sizes = tile_85, x = var_7729_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor tile_86 = const()[name = tensor("tile_86"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7751_axis_0 = const()[name = tensor("op_7751_axis_0"), val = tensor(1)]; + tensor var_7751_cast_fp16_0, tensor var_7751_cast_fp16_1, tensor var_7751_cast_fp16_2, tensor var_7751_cast_fp16_3, tensor var_7751_cast_fp16_4, tensor var_7751_cast_fp16_5, tensor var_7751_cast_fp16_6, tensor var_7751_cast_fp16_7, tensor var_7751_cast_fp16_8, tensor var_7751_cast_fp16_9, tensor var_7751_cast_fp16_10, tensor var_7751_cast_fp16_11, tensor var_7751_cast_fp16_12, tensor var_7751_cast_fp16_13, tensor var_7751_cast_fp16_14, tensor var_7751_cast_fp16_15, tensor var_7751_cast_fp16_16, tensor var_7751_cast_fp16_17, tensor var_7751_cast_fp16_18, tensor var_7751_cast_fp16_19 = split(axis = var_7751_axis_0, split_sizes = tile_86, x = var_7705_cast_fp16)[name = tensor("op_7751_cast_fp16")]; + tensor aw_1121_equation_0 = const()[name = tensor("aw_1121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1121_cast_fp16 = einsum(equation = aw_1121_equation_0, values = (var_7730_cast_fp16_0, var_7708_cast_fp16_0))[name = tensor("aw_1121_cast_fp16")]; + tensor aw_1123_equation_0 = const()[name = tensor("aw_1123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1123_cast_fp16 = einsum(equation = aw_1123_equation_0, values = (var_7730_cast_fp16_1, var_7708_cast_fp16_1))[name = tensor("aw_1123_cast_fp16")]; + tensor aw_1125_equation_0 = const()[name = tensor("aw_1125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1125_cast_fp16 = einsum(equation = aw_1125_equation_0, values = (var_7730_cast_fp16_2, var_7708_cast_fp16_2))[name = tensor("aw_1125_cast_fp16")]; + tensor aw_1127_equation_0 = const()[name = tensor("aw_1127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1127_cast_fp16 = einsum(equation = aw_1127_equation_0, values = (var_7730_cast_fp16_3, var_7708_cast_fp16_3))[name = tensor("aw_1127_cast_fp16")]; + tensor aw_1129_equation_0 = const()[name = tensor("aw_1129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1129_cast_fp16 = einsum(equation = aw_1129_equation_0, values = (var_7730_cast_fp16_4, var_7708_cast_fp16_4))[name = tensor("aw_1129_cast_fp16")]; + tensor aw_1131_equation_0 = const()[name = tensor("aw_1131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1131_cast_fp16 = einsum(equation = aw_1131_equation_0, values = (var_7730_cast_fp16_5, var_7708_cast_fp16_5))[name = tensor("aw_1131_cast_fp16")]; + tensor aw_1133_equation_0 = const()[name = tensor("aw_1133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1133_cast_fp16 = einsum(equation = aw_1133_equation_0, values = (var_7730_cast_fp16_6, var_7708_cast_fp16_6))[name = tensor("aw_1133_cast_fp16")]; + tensor aw_1135_equation_0 = const()[name = tensor("aw_1135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1135_cast_fp16 = einsum(equation = aw_1135_equation_0, values = (var_7730_cast_fp16_7, var_7708_cast_fp16_7))[name = tensor("aw_1135_cast_fp16")]; + tensor aw_1137_equation_0 = const()[name = tensor("aw_1137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1137_cast_fp16 = einsum(equation = aw_1137_equation_0, values = (var_7730_cast_fp16_8, var_7708_cast_fp16_8))[name = tensor("aw_1137_cast_fp16")]; + tensor aw_1139_equation_0 = const()[name = tensor("aw_1139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1139_cast_fp16 = einsum(equation = aw_1139_equation_0, values = (var_7730_cast_fp16_9, var_7708_cast_fp16_9))[name = tensor("aw_1139_cast_fp16")]; + tensor aw_1141_equation_0 = const()[name = tensor("aw_1141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1141_cast_fp16 = einsum(equation = aw_1141_equation_0, values = (var_7730_cast_fp16_10, var_7708_cast_fp16_10))[name = tensor("aw_1141_cast_fp16")]; + tensor aw_1143_equation_0 = const()[name = tensor("aw_1143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1143_cast_fp16 = einsum(equation = aw_1143_equation_0, values = (var_7730_cast_fp16_11, var_7708_cast_fp16_11))[name = tensor("aw_1143_cast_fp16")]; + tensor aw_1145_equation_0 = const()[name = tensor("aw_1145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1145_cast_fp16 = einsum(equation = aw_1145_equation_0, values = (var_7730_cast_fp16_12, var_7708_cast_fp16_12))[name = tensor("aw_1145_cast_fp16")]; + tensor aw_1147_equation_0 = const()[name = tensor("aw_1147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1147_cast_fp16 = einsum(equation = aw_1147_equation_0, values = (var_7730_cast_fp16_13, var_7708_cast_fp16_13))[name = tensor("aw_1147_cast_fp16")]; + tensor aw_1149_equation_0 = const()[name = tensor("aw_1149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1149_cast_fp16 = einsum(equation = aw_1149_equation_0, values = (var_7730_cast_fp16_14, var_7708_cast_fp16_14))[name = tensor("aw_1149_cast_fp16")]; + tensor aw_1151_equation_0 = const()[name = tensor("aw_1151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1151_cast_fp16 = einsum(equation = aw_1151_equation_0, values = (var_7730_cast_fp16_15, var_7708_cast_fp16_15))[name = tensor("aw_1151_cast_fp16")]; + tensor aw_1153_equation_0 = const()[name = tensor("aw_1153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1153_cast_fp16 = einsum(equation = aw_1153_equation_0, values = (var_7730_cast_fp16_16, var_7708_cast_fp16_16))[name = tensor("aw_1153_cast_fp16")]; + tensor aw_1155_equation_0 = const()[name = tensor("aw_1155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1155_cast_fp16 = einsum(equation = aw_1155_equation_0, values = (var_7730_cast_fp16_17, var_7708_cast_fp16_17))[name = tensor("aw_1155_cast_fp16")]; + tensor aw_1157_equation_0 = const()[name = tensor("aw_1157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1157_cast_fp16 = einsum(equation = aw_1157_equation_0, values = (var_7730_cast_fp16_18, var_7708_cast_fp16_18))[name = tensor("aw_1157_cast_fp16")]; + tensor aw_1159_equation_0 = const()[name = tensor("aw_1159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1159_cast_fp16 = einsum(equation = aw_1159_equation_0, values = (var_7730_cast_fp16_19, var_7708_cast_fp16_19))[name = tensor("aw_1159_cast_fp16")]; + tensor var_7812_cast_fp16 = softmax(axis = var_7656, x = aw_1121_cast_fp16)[name = tensor("op_7812_cast_fp16")]; + tensor var_7813_cast_fp16 = softmax(axis = var_7656, x = aw_1123_cast_fp16)[name = tensor("op_7813_cast_fp16")]; + tensor var_7814_cast_fp16 = softmax(axis = var_7656, x = aw_1125_cast_fp16)[name = tensor("op_7814_cast_fp16")]; + tensor var_7815_cast_fp16 = softmax(axis = var_7656, x = aw_1127_cast_fp16)[name = tensor("op_7815_cast_fp16")]; + tensor var_7816_cast_fp16 = softmax(axis = var_7656, x = aw_1129_cast_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor var_7817_cast_fp16 = softmax(axis = var_7656, x = aw_1131_cast_fp16)[name = tensor("op_7817_cast_fp16")]; + tensor var_7818_cast_fp16 = softmax(axis = var_7656, x = aw_1133_cast_fp16)[name = tensor("op_7818_cast_fp16")]; + tensor var_7819_cast_fp16 = softmax(axis = var_7656, x = aw_1135_cast_fp16)[name = tensor("op_7819_cast_fp16")]; + tensor var_7820_cast_fp16 = softmax(axis = var_7656, x = aw_1137_cast_fp16)[name = tensor("op_7820_cast_fp16")]; + tensor var_7821_cast_fp16 = softmax(axis = var_7656, x = aw_1139_cast_fp16)[name = tensor("op_7821_cast_fp16")]; + tensor var_7822_cast_fp16 = softmax(axis = var_7656, x = aw_1141_cast_fp16)[name = tensor("op_7822_cast_fp16")]; + tensor var_7823_cast_fp16 = softmax(axis = var_7656, x = aw_1143_cast_fp16)[name = tensor("op_7823_cast_fp16")]; + tensor var_7824_cast_fp16 = softmax(axis = var_7656, x = aw_1145_cast_fp16)[name = tensor("op_7824_cast_fp16")]; + tensor var_7825_cast_fp16 = softmax(axis = var_7656, x = aw_1147_cast_fp16)[name = tensor("op_7825_cast_fp16")]; + tensor var_7826_cast_fp16 = softmax(axis = var_7656, x = aw_1149_cast_fp16)[name = tensor("op_7826_cast_fp16")]; + tensor var_7827_cast_fp16 = softmax(axis = var_7656, x = aw_1151_cast_fp16)[name = tensor("op_7827_cast_fp16")]; + tensor var_7828_cast_fp16 = softmax(axis = var_7656, x = aw_1153_cast_fp16)[name = tensor("op_7828_cast_fp16")]; + tensor var_7829_cast_fp16 = softmax(axis = var_7656, x = aw_1155_cast_fp16)[name = tensor("op_7829_cast_fp16")]; + tensor var_7830_cast_fp16 = softmax(axis = var_7656, x = aw_1157_cast_fp16)[name = tensor("op_7830_cast_fp16")]; + tensor var_7831_cast_fp16 = softmax(axis = var_7656, x = aw_1159_cast_fp16)[name = tensor("op_7831_cast_fp16")]; + tensor var_7833_equation_0 = const()[name = tensor("op_7833_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7833_cast_fp16 = einsum(equation = var_7833_equation_0, values = (var_7751_cast_fp16_0, var_7812_cast_fp16))[name = tensor("op_7833_cast_fp16")]; + tensor var_7835_equation_0 = const()[name = tensor("op_7835_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7835_cast_fp16 = einsum(equation = var_7835_equation_0, values = (var_7751_cast_fp16_1, var_7813_cast_fp16))[name = tensor("op_7835_cast_fp16")]; + tensor var_7837_equation_0 = const()[name = tensor("op_7837_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7837_cast_fp16 = einsum(equation = var_7837_equation_0, values = (var_7751_cast_fp16_2, var_7814_cast_fp16))[name = tensor("op_7837_cast_fp16")]; + tensor var_7839_equation_0 = const()[name = tensor("op_7839_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7839_cast_fp16 = einsum(equation = var_7839_equation_0, values = (var_7751_cast_fp16_3, var_7815_cast_fp16))[name = tensor("op_7839_cast_fp16")]; + tensor var_7841_equation_0 = const()[name = tensor("op_7841_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7841_cast_fp16 = einsum(equation = var_7841_equation_0, values = (var_7751_cast_fp16_4, var_7816_cast_fp16))[name = tensor("op_7841_cast_fp16")]; + tensor var_7843_equation_0 = const()[name = tensor("op_7843_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7843_cast_fp16 = einsum(equation = var_7843_equation_0, values = (var_7751_cast_fp16_5, var_7817_cast_fp16))[name = tensor("op_7843_cast_fp16")]; + tensor var_7845_equation_0 = const()[name = tensor("op_7845_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7845_cast_fp16 = einsum(equation = var_7845_equation_0, values = (var_7751_cast_fp16_6, var_7818_cast_fp16))[name = tensor("op_7845_cast_fp16")]; + tensor var_7847_equation_0 = const()[name = tensor("op_7847_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7847_cast_fp16 = einsum(equation = var_7847_equation_0, values = (var_7751_cast_fp16_7, var_7819_cast_fp16))[name = tensor("op_7847_cast_fp16")]; + tensor var_7849_equation_0 = const()[name = tensor("op_7849_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7849_cast_fp16 = einsum(equation = var_7849_equation_0, values = (var_7751_cast_fp16_8, var_7820_cast_fp16))[name = tensor("op_7849_cast_fp16")]; + tensor var_7851_equation_0 = const()[name = tensor("op_7851_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7851_cast_fp16 = einsum(equation = var_7851_equation_0, values = (var_7751_cast_fp16_9, var_7821_cast_fp16))[name = tensor("op_7851_cast_fp16")]; + tensor var_7853_equation_0 = const()[name = tensor("op_7853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7853_cast_fp16 = einsum(equation = var_7853_equation_0, values = (var_7751_cast_fp16_10, var_7822_cast_fp16))[name = tensor("op_7853_cast_fp16")]; + tensor var_7855_equation_0 = const()[name = tensor("op_7855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7855_cast_fp16 = einsum(equation = var_7855_equation_0, values = (var_7751_cast_fp16_11, var_7823_cast_fp16))[name = tensor("op_7855_cast_fp16")]; + tensor var_7857_equation_0 = const()[name = tensor("op_7857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7857_cast_fp16 = einsum(equation = var_7857_equation_0, values = (var_7751_cast_fp16_12, var_7824_cast_fp16))[name = tensor("op_7857_cast_fp16")]; + tensor var_7859_equation_0 = const()[name = tensor("op_7859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7859_cast_fp16 = einsum(equation = var_7859_equation_0, values = (var_7751_cast_fp16_13, var_7825_cast_fp16))[name = tensor("op_7859_cast_fp16")]; + tensor var_7861_equation_0 = const()[name = tensor("op_7861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7861_cast_fp16 = einsum(equation = var_7861_equation_0, values = (var_7751_cast_fp16_14, var_7826_cast_fp16))[name = tensor("op_7861_cast_fp16")]; + tensor var_7863_equation_0 = const()[name = tensor("op_7863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7863_cast_fp16 = einsum(equation = var_7863_equation_0, values = (var_7751_cast_fp16_15, var_7827_cast_fp16))[name = tensor("op_7863_cast_fp16")]; + tensor var_7865_equation_0 = const()[name = tensor("op_7865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7865_cast_fp16 = einsum(equation = var_7865_equation_0, values = (var_7751_cast_fp16_16, var_7828_cast_fp16))[name = tensor("op_7865_cast_fp16")]; + tensor var_7867_equation_0 = const()[name = tensor("op_7867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7867_cast_fp16 = einsum(equation = var_7867_equation_0, values = (var_7751_cast_fp16_17, var_7829_cast_fp16))[name = tensor("op_7867_cast_fp16")]; + tensor var_7869_equation_0 = const()[name = tensor("op_7869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7869_cast_fp16 = einsum(equation = var_7869_equation_0, values = (var_7751_cast_fp16_18, var_7830_cast_fp16))[name = tensor("op_7869_cast_fp16")]; + tensor var_7871_equation_0 = const()[name = tensor("op_7871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_7871_cast_fp16 = einsum(equation = var_7871_equation_0, values = (var_7751_cast_fp16_19, var_7831_cast_fp16))[name = tensor("op_7871_cast_fp16")]; + tensor input_285_interleave_0 = const()[name = tensor("input_285_interleave_0"), val = tensor(false)]; + tensor input_285_cast_fp16 = concat(axis = var_7656, interleave = input_285_interleave_0, values = (var_7833_cast_fp16, var_7835_cast_fp16, var_7837_cast_fp16, var_7839_cast_fp16, var_7841_cast_fp16, var_7843_cast_fp16, var_7845_cast_fp16, var_7847_cast_fp16, var_7849_cast_fp16, var_7851_cast_fp16, var_7853_cast_fp16, var_7855_cast_fp16, var_7857_cast_fp16, var_7859_cast_fp16, var_7861_cast_fp16, var_7863_cast_fp16, var_7865_cast_fp16, var_7867_cast_fp16, var_7869_cast_fp16, var_7871_cast_fp16))[name = tensor("input_285_cast_fp16")]; + tensor var_7880_pad_type_0 = const()[name = tensor("op_7880_pad_type_0"), val = tensor("valid")]; + tensor var_7880_strides_0 = const()[name = tensor("op_7880_strides_0"), val = tensor([1, 1])]; + tensor var_7880_pad_0 = const()[name = tensor("op_7880_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7880_dilations_0 = const()[name = tensor("op_7880_dilations_0"), val = tensor([1, 1])]; + tensor var_7880_groups_0 = const()[name = tensor("op_7880_groups_0"), val = tensor(1)]; + tensor blocks_28_attn_out_weight_to_fp16 = const()[name = tensor("blocks_28_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126391872)))]; + tensor blocks_28_attn_out_bias_to_fp16 = const()[name = tensor("blocks_28_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129668736)))]; + tensor var_7880_cast_fp16 = conv(bias = blocks_28_attn_out_bias_to_fp16, dilations = var_7880_dilations_0, groups = var_7880_groups_0, pad = var_7880_pad_0, pad_type = var_7880_pad_type_0, strides = var_7880_strides_0, weight = blocks_28_attn_out_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("op_7880_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = var_7880_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([1])]; + tensor input_287_gamma_0_to_fp16 = const()[name = tensor("input_287_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129671360)))]; + tensor input_287_beta_0_to_fp16 = const()[name = tensor("input_287_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129673984)))]; + tensor var_7890_to_fp16 = const()[name = tensor("op_7890_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = input_287_beta_0_to_fp16, epsilon = var_7890_to_fp16, gamma = input_287_gamma_0_to_fp16, x = inputs_115_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_pad_type_0 = const()[name = tensor("input_289_pad_type_0"), val = tensor("valid")]; + tensor input_289_strides_0 = const()[name = tensor("input_289_strides_0"), val = tensor([1, 1])]; + tensor input_289_pad_0 = const()[name = tensor("input_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_289_dilations_0 = const()[name = tensor("input_289_dilations_0"), val = tensor([1, 1])]; + tensor input_289_groups_0 = const()[name = tensor("input_289_groups_0"), val = tensor(1)]; + tensor blocks_28_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129676608)))]; + tensor blocks_28_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142783872)))]; + tensor input_289_cast_fp16 = conv(bias = blocks_28_mlp_0_bias_to_fp16, dilations = input_289_dilations_0, groups = input_289_groups_0, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = input_289_strides_0, weight = blocks_28_mlp_0_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor input_291_mode_0 = const()[name = tensor("input_291_mode_0"), val = tensor("EXACT")]; + tensor input_291_cast_fp16 = gelu(mode = input_291_mode_0, x = input_289_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor var_7916_pad_type_0 = const()[name = tensor("op_7916_pad_type_0"), val = tensor("valid")]; + tensor var_7916_strides_0 = const()[name = tensor("op_7916_strides_0"), val = tensor([1, 1])]; + tensor var_7916_pad_0 = const()[name = tensor("op_7916_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7916_dilations_0 = const()[name = tensor("op_7916_dilations_0"), val = tensor([1, 1])]; + tensor var_7916_groups_0 = const()[name = tensor("op_7916_groups_0"), val = tensor(1)]; + tensor blocks_28_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142794176)))]; + tensor blocks_28_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155901440)))]; + tensor var_7916_cast_fp16 = conv(bias = blocks_28_mlp_2_bias_to_fp16, dilations = var_7916_dilations_0, groups = var_7916_groups_0, pad = var_7916_pad_0, pad_type = var_7916_pad_type_0, strides = var_7916_strides_0, weight = blocks_28_mlp_2_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("op_7916_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = var_7916_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_7925 = const()[name = tensor("op_7925"), val = tensor(1)]; + tensor input_293_axes_0 = const()[name = tensor("input_293_axes_0"), val = tensor([1])]; + tensor input_293_gamma_0_to_fp16 = const()[name = tensor("input_293_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155904064)))]; + tensor input_293_beta_0_to_fp16 = const()[name = tensor("input_293_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155906688)))]; + tensor var_7941_to_fp16 = const()[name = tensor("op_7941_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = input_293_beta_0_to_fp16, epsilon = var_7941_to_fp16, gamma = input_293_gamma_0_to_fp16, x = inputs_117_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor q_59_pad_type_0 = const()[name = tensor("q_59_pad_type_0"), val = tensor("valid")]; + tensor q_59_strides_0 = const()[name = tensor("q_59_strides_0"), val = tensor([1, 1])]; + tensor q_59_pad_0 = const()[name = tensor("q_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_59_dilations_0 = const()[name = tensor("q_59_dilations_0"), val = tensor([1, 1])]; + tensor q_59_groups_0 = const()[name = tensor("q_59_groups_0"), val = tensor(1)]; + tensor var_7976_weight_0_to_fp16 = const()[name = tensor("op_7976_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155909312)))]; + tensor var_7976_bias_0_to_fp16 = const()[name = tensor("op_7976_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159186176)))]; + tensor var_7976_cast_fp16 = conv(bias = var_7976_bias_0_to_fp16, dilations = q_59_dilations_0, groups = q_59_groups_0, pad = q_59_pad_0, pad_type = q_59_pad_type_0, strides = q_59_strides_0, weight = var_7976_weight_0_to_fp16, x = input_293_cast_fp16)[name = tensor("op_7976_cast_fp16")]; + tensor k_59_pad_type_0 = const()[name = tensor("k_59_pad_type_0"), val = tensor("valid")]; + tensor k_59_strides_0 = const()[name = tensor("k_59_strides_0"), val = tensor([1, 1])]; + tensor k_59_pad_0 = const()[name = tensor("k_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_59_dilations_0 = const()[name = tensor("k_59_dilations_0"), val = tensor([1, 1])]; + tensor k_59_groups_0 = const()[name = tensor("k_59_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_key_weight_to_fp16 = const()[name = tensor("blocks_29_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159188800)))]; + tensor k_59_cast_fp16 = conv(dilations = k_59_dilations_0, groups = k_59_groups_0, pad = k_59_pad_0, pad_type = k_59_pad_type_0, strides = k_59_strides_0, weight = blocks_29_attn_key_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor var_7974_pad_type_0 = const()[name = tensor("op_7974_pad_type_0"), val = tensor("valid")]; + tensor var_7974_strides_0 = const()[name = tensor("op_7974_strides_0"), val = tensor([1, 1])]; + tensor var_7974_pad_0 = const()[name = tensor("op_7974_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_7974_dilations_0 = const()[name = tensor("op_7974_dilations_0"), val = tensor([1, 1])]; + tensor var_7974_groups_0 = const()[name = tensor("op_7974_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_value_weight_to_fp16 = const()[name = tensor("blocks_29_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162465664)))]; + tensor blocks_29_attn_value_bias_to_fp16 = const()[name = tensor("blocks_29_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165742528)))]; + tensor var_7974_cast_fp16 = conv(bias = blocks_29_attn_value_bias_to_fp16, dilations = var_7974_dilations_0, groups = var_7974_groups_0, pad = var_7974_pad_0, pad_type = var_7974_pad_type_0, strides = var_7974_strides_0, weight = blocks_29_attn_value_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("op_7974_cast_fp16")]; + tensor tile_87 = const()[name = tensor("tile_87"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7977_axis_0 = const()[name = tensor("op_7977_axis_0"), val = tensor(1)]; + tensor var_7977_cast_fp16_0, tensor var_7977_cast_fp16_1, tensor var_7977_cast_fp16_2, tensor var_7977_cast_fp16_3, tensor var_7977_cast_fp16_4, tensor var_7977_cast_fp16_5, tensor var_7977_cast_fp16_6, tensor var_7977_cast_fp16_7, tensor var_7977_cast_fp16_8, tensor var_7977_cast_fp16_9, tensor var_7977_cast_fp16_10, tensor var_7977_cast_fp16_11, tensor var_7977_cast_fp16_12, tensor var_7977_cast_fp16_13, tensor var_7977_cast_fp16_14, tensor var_7977_cast_fp16_15, tensor var_7977_cast_fp16_16, tensor var_7977_cast_fp16_17, tensor var_7977_cast_fp16_18, tensor var_7977_cast_fp16_19 = split(axis = var_7977_axis_0, split_sizes = tile_87, x = var_7976_cast_fp16)[name = tensor("op_7977_cast_fp16")]; + tensor var_7998_perm_0 = const()[name = tensor("op_7998_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_88 = const()[name = tensor("tile_88"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_7999_axis_0 = const()[name = tensor("op_7999_axis_0"), val = tensor(3)]; + tensor var_7998_cast_fp16 = transpose(perm = var_7998_perm_0, x = k_59_cast_fp16)[name = tensor("transpose_3")]; + tensor var_7999_cast_fp16_0, tensor var_7999_cast_fp16_1, tensor var_7999_cast_fp16_2, tensor var_7999_cast_fp16_3, tensor var_7999_cast_fp16_4, tensor var_7999_cast_fp16_5, tensor var_7999_cast_fp16_6, tensor var_7999_cast_fp16_7, tensor var_7999_cast_fp16_8, tensor var_7999_cast_fp16_9, tensor var_7999_cast_fp16_10, tensor var_7999_cast_fp16_11, tensor var_7999_cast_fp16_12, tensor var_7999_cast_fp16_13, tensor var_7999_cast_fp16_14, tensor var_7999_cast_fp16_15, tensor var_7999_cast_fp16_16, tensor var_7999_cast_fp16_17, tensor var_7999_cast_fp16_18, tensor var_7999_cast_fp16_19 = split(axis = var_7999_axis_0, split_sizes = tile_88, x = var_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor tile_89 = const()[name = tensor("tile_89"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8020_axis_0 = const()[name = tensor("op_8020_axis_0"), val = tensor(1)]; + tensor var_8020_cast_fp16_0, tensor var_8020_cast_fp16_1, tensor var_8020_cast_fp16_2, tensor var_8020_cast_fp16_3, tensor var_8020_cast_fp16_4, tensor var_8020_cast_fp16_5, tensor var_8020_cast_fp16_6, tensor var_8020_cast_fp16_7, tensor var_8020_cast_fp16_8, tensor var_8020_cast_fp16_9, tensor var_8020_cast_fp16_10, tensor var_8020_cast_fp16_11, tensor var_8020_cast_fp16_12, tensor var_8020_cast_fp16_13, tensor var_8020_cast_fp16_14, tensor var_8020_cast_fp16_15, tensor var_8020_cast_fp16_16, tensor var_8020_cast_fp16_17, tensor var_8020_cast_fp16_18, tensor var_8020_cast_fp16_19 = split(axis = var_8020_axis_0, split_sizes = tile_89, x = var_7974_cast_fp16)[name = tensor("op_8020_cast_fp16")]; + tensor aw_1161_equation_0 = const()[name = tensor("aw_1161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1161_cast_fp16 = einsum(equation = aw_1161_equation_0, values = (var_7999_cast_fp16_0, var_7977_cast_fp16_0))[name = tensor("aw_1161_cast_fp16")]; + tensor aw_1163_equation_0 = const()[name = tensor("aw_1163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1163_cast_fp16 = einsum(equation = aw_1163_equation_0, values = (var_7999_cast_fp16_1, var_7977_cast_fp16_1))[name = tensor("aw_1163_cast_fp16")]; + tensor aw_1165_equation_0 = const()[name = tensor("aw_1165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1165_cast_fp16 = einsum(equation = aw_1165_equation_0, values = (var_7999_cast_fp16_2, var_7977_cast_fp16_2))[name = tensor("aw_1165_cast_fp16")]; + tensor aw_1167_equation_0 = const()[name = tensor("aw_1167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1167_cast_fp16 = einsum(equation = aw_1167_equation_0, values = (var_7999_cast_fp16_3, var_7977_cast_fp16_3))[name = tensor("aw_1167_cast_fp16")]; + tensor aw_1169_equation_0 = const()[name = tensor("aw_1169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1169_cast_fp16 = einsum(equation = aw_1169_equation_0, values = (var_7999_cast_fp16_4, var_7977_cast_fp16_4))[name = tensor("aw_1169_cast_fp16")]; + tensor aw_1171_equation_0 = const()[name = tensor("aw_1171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1171_cast_fp16 = einsum(equation = aw_1171_equation_0, values = (var_7999_cast_fp16_5, var_7977_cast_fp16_5))[name = tensor("aw_1171_cast_fp16")]; + tensor aw_1173_equation_0 = const()[name = tensor("aw_1173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1173_cast_fp16 = einsum(equation = aw_1173_equation_0, values = (var_7999_cast_fp16_6, var_7977_cast_fp16_6))[name = tensor("aw_1173_cast_fp16")]; + tensor aw_1175_equation_0 = const()[name = tensor("aw_1175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1175_cast_fp16 = einsum(equation = aw_1175_equation_0, values = (var_7999_cast_fp16_7, var_7977_cast_fp16_7))[name = tensor("aw_1175_cast_fp16")]; + tensor aw_1177_equation_0 = const()[name = tensor("aw_1177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1177_cast_fp16 = einsum(equation = aw_1177_equation_0, values = (var_7999_cast_fp16_8, var_7977_cast_fp16_8))[name = tensor("aw_1177_cast_fp16")]; + tensor aw_1179_equation_0 = const()[name = tensor("aw_1179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1179_cast_fp16 = einsum(equation = aw_1179_equation_0, values = (var_7999_cast_fp16_9, var_7977_cast_fp16_9))[name = tensor("aw_1179_cast_fp16")]; + tensor aw_1181_equation_0 = const()[name = tensor("aw_1181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1181_cast_fp16 = einsum(equation = aw_1181_equation_0, values = (var_7999_cast_fp16_10, var_7977_cast_fp16_10))[name = tensor("aw_1181_cast_fp16")]; + tensor aw_1183_equation_0 = const()[name = tensor("aw_1183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1183_cast_fp16 = einsum(equation = aw_1183_equation_0, values = (var_7999_cast_fp16_11, var_7977_cast_fp16_11))[name = tensor("aw_1183_cast_fp16")]; + tensor aw_1185_equation_0 = const()[name = tensor("aw_1185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1185_cast_fp16 = einsum(equation = aw_1185_equation_0, values = (var_7999_cast_fp16_12, var_7977_cast_fp16_12))[name = tensor("aw_1185_cast_fp16")]; + tensor aw_1187_equation_0 = const()[name = tensor("aw_1187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1187_cast_fp16 = einsum(equation = aw_1187_equation_0, values = (var_7999_cast_fp16_13, var_7977_cast_fp16_13))[name = tensor("aw_1187_cast_fp16")]; + tensor aw_1189_equation_0 = const()[name = tensor("aw_1189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1189_cast_fp16 = einsum(equation = aw_1189_equation_0, values = (var_7999_cast_fp16_14, var_7977_cast_fp16_14))[name = tensor("aw_1189_cast_fp16")]; + tensor aw_1191_equation_0 = const()[name = tensor("aw_1191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1191_cast_fp16 = einsum(equation = aw_1191_equation_0, values = (var_7999_cast_fp16_15, var_7977_cast_fp16_15))[name = tensor("aw_1191_cast_fp16")]; + tensor aw_1193_equation_0 = const()[name = tensor("aw_1193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1193_cast_fp16 = einsum(equation = aw_1193_equation_0, values = (var_7999_cast_fp16_16, var_7977_cast_fp16_16))[name = tensor("aw_1193_cast_fp16")]; + tensor aw_1195_equation_0 = const()[name = tensor("aw_1195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1195_cast_fp16 = einsum(equation = aw_1195_equation_0, values = (var_7999_cast_fp16_17, var_7977_cast_fp16_17))[name = tensor("aw_1195_cast_fp16")]; + tensor aw_1197_equation_0 = const()[name = tensor("aw_1197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1197_cast_fp16 = einsum(equation = aw_1197_equation_0, values = (var_7999_cast_fp16_18, var_7977_cast_fp16_18))[name = tensor("aw_1197_cast_fp16")]; + tensor aw_1199_equation_0 = const()[name = tensor("aw_1199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1199_cast_fp16 = einsum(equation = aw_1199_equation_0, values = (var_7999_cast_fp16_19, var_7977_cast_fp16_19))[name = tensor("aw_1199_cast_fp16")]; + tensor var_8081_cast_fp16 = softmax(axis = var_7925, x = aw_1161_cast_fp16)[name = tensor("op_8081_cast_fp16")]; + tensor var_8082_cast_fp16 = softmax(axis = var_7925, x = aw_1163_cast_fp16)[name = tensor("op_8082_cast_fp16")]; + tensor var_8083_cast_fp16 = softmax(axis = var_7925, x = aw_1165_cast_fp16)[name = tensor("op_8083_cast_fp16")]; + tensor var_8084_cast_fp16 = softmax(axis = var_7925, x = aw_1167_cast_fp16)[name = tensor("op_8084_cast_fp16")]; + tensor var_8085_cast_fp16 = softmax(axis = var_7925, x = aw_1169_cast_fp16)[name = tensor("op_8085_cast_fp16")]; + tensor var_8086_cast_fp16 = softmax(axis = var_7925, x = aw_1171_cast_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8087_cast_fp16 = softmax(axis = var_7925, x = aw_1173_cast_fp16)[name = tensor("op_8087_cast_fp16")]; + tensor var_8088_cast_fp16 = softmax(axis = var_7925, x = aw_1175_cast_fp16)[name = tensor("op_8088_cast_fp16")]; + tensor var_8089_cast_fp16 = softmax(axis = var_7925, x = aw_1177_cast_fp16)[name = tensor("op_8089_cast_fp16")]; + tensor var_8090_cast_fp16 = softmax(axis = var_7925, x = aw_1179_cast_fp16)[name = tensor("op_8090_cast_fp16")]; + tensor var_8091_cast_fp16 = softmax(axis = var_7925, x = aw_1181_cast_fp16)[name = tensor("op_8091_cast_fp16")]; + tensor var_8092_cast_fp16 = softmax(axis = var_7925, x = aw_1183_cast_fp16)[name = tensor("op_8092_cast_fp16")]; + tensor var_8093_cast_fp16 = softmax(axis = var_7925, x = aw_1185_cast_fp16)[name = tensor("op_8093_cast_fp16")]; + tensor var_8094_cast_fp16 = softmax(axis = var_7925, x = aw_1187_cast_fp16)[name = tensor("op_8094_cast_fp16")]; + tensor var_8095_cast_fp16 = softmax(axis = var_7925, x = aw_1189_cast_fp16)[name = tensor("op_8095_cast_fp16")]; + tensor var_8096_cast_fp16 = softmax(axis = var_7925, x = aw_1191_cast_fp16)[name = tensor("op_8096_cast_fp16")]; + tensor var_8097_cast_fp16 = softmax(axis = var_7925, x = aw_1193_cast_fp16)[name = tensor("op_8097_cast_fp16")]; + tensor var_8098_cast_fp16 = softmax(axis = var_7925, x = aw_1195_cast_fp16)[name = tensor("op_8098_cast_fp16")]; + tensor var_8099_cast_fp16 = softmax(axis = var_7925, x = aw_1197_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8100_cast_fp16 = softmax(axis = var_7925, x = aw_1199_cast_fp16)[name = tensor("op_8100_cast_fp16")]; + tensor var_8102_equation_0 = const()[name = tensor("op_8102_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8102_cast_fp16 = einsum(equation = var_8102_equation_0, values = (var_8020_cast_fp16_0, var_8081_cast_fp16))[name = tensor("op_8102_cast_fp16")]; + tensor var_8104_equation_0 = const()[name = tensor("op_8104_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8104_cast_fp16 = einsum(equation = var_8104_equation_0, values = (var_8020_cast_fp16_1, var_8082_cast_fp16))[name = tensor("op_8104_cast_fp16")]; + tensor var_8106_equation_0 = const()[name = tensor("op_8106_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8106_cast_fp16 = einsum(equation = var_8106_equation_0, values = (var_8020_cast_fp16_2, var_8083_cast_fp16))[name = tensor("op_8106_cast_fp16")]; + tensor var_8108_equation_0 = const()[name = tensor("op_8108_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8108_cast_fp16 = einsum(equation = var_8108_equation_0, values = (var_8020_cast_fp16_3, var_8084_cast_fp16))[name = tensor("op_8108_cast_fp16")]; + tensor var_8110_equation_0 = const()[name = tensor("op_8110_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8110_cast_fp16 = einsum(equation = var_8110_equation_0, values = (var_8020_cast_fp16_4, var_8085_cast_fp16))[name = tensor("op_8110_cast_fp16")]; + tensor var_8112_equation_0 = const()[name = tensor("op_8112_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8112_cast_fp16 = einsum(equation = var_8112_equation_0, values = (var_8020_cast_fp16_5, var_8086_cast_fp16))[name = tensor("op_8112_cast_fp16")]; + tensor var_8114_equation_0 = const()[name = tensor("op_8114_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8114_cast_fp16 = einsum(equation = var_8114_equation_0, values = (var_8020_cast_fp16_6, var_8087_cast_fp16))[name = tensor("op_8114_cast_fp16")]; + tensor var_8116_equation_0 = const()[name = tensor("op_8116_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8116_cast_fp16 = einsum(equation = var_8116_equation_0, values = (var_8020_cast_fp16_7, var_8088_cast_fp16))[name = tensor("op_8116_cast_fp16")]; + tensor var_8118_equation_0 = const()[name = tensor("op_8118_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8118_cast_fp16 = einsum(equation = var_8118_equation_0, values = (var_8020_cast_fp16_8, var_8089_cast_fp16))[name = tensor("op_8118_cast_fp16")]; + tensor var_8120_equation_0 = const()[name = tensor("op_8120_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8120_cast_fp16 = einsum(equation = var_8120_equation_0, values = (var_8020_cast_fp16_9, var_8090_cast_fp16))[name = tensor("op_8120_cast_fp16")]; + tensor var_8122_equation_0 = const()[name = tensor("op_8122_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8122_cast_fp16 = einsum(equation = var_8122_equation_0, values = (var_8020_cast_fp16_10, var_8091_cast_fp16))[name = tensor("op_8122_cast_fp16")]; + tensor var_8124_equation_0 = const()[name = tensor("op_8124_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8124_cast_fp16 = einsum(equation = var_8124_equation_0, values = (var_8020_cast_fp16_11, var_8092_cast_fp16))[name = tensor("op_8124_cast_fp16")]; + tensor var_8126_equation_0 = const()[name = tensor("op_8126_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8126_cast_fp16 = einsum(equation = var_8126_equation_0, values = (var_8020_cast_fp16_12, var_8093_cast_fp16))[name = tensor("op_8126_cast_fp16")]; + tensor var_8128_equation_0 = const()[name = tensor("op_8128_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8128_cast_fp16 = einsum(equation = var_8128_equation_0, values = (var_8020_cast_fp16_13, var_8094_cast_fp16))[name = tensor("op_8128_cast_fp16")]; + tensor var_8130_equation_0 = const()[name = tensor("op_8130_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8130_cast_fp16 = einsum(equation = var_8130_equation_0, values = (var_8020_cast_fp16_14, var_8095_cast_fp16))[name = tensor("op_8130_cast_fp16")]; + tensor var_8132_equation_0 = const()[name = tensor("op_8132_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8132_cast_fp16 = einsum(equation = var_8132_equation_0, values = (var_8020_cast_fp16_15, var_8096_cast_fp16))[name = tensor("op_8132_cast_fp16")]; + tensor var_8134_equation_0 = const()[name = tensor("op_8134_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8134_cast_fp16 = einsum(equation = var_8134_equation_0, values = (var_8020_cast_fp16_16, var_8097_cast_fp16))[name = tensor("op_8134_cast_fp16")]; + tensor var_8136_equation_0 = const()[name = tensor("op_8136_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8136_cast_fp16 = einsum(equation = var_8136_equation_0, values = (var_8020_cast_fp16_17, var_8098_cast_fp16))[name = tensor("op_8136_cast_fp16")]; + tensor var_8138_equation_0 = const()[name = tensor("op_8138_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8138_cast_fp16 = einsum(equation = var_8138_equation_0, values = (var_8020_cast_fp16_18, var_8099_cast_fp16))[name = tensor("op_8138_cast_fp16")]; + tensor var_8140_equation_0 = const()[name = tensor("op_8140_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8140_cast_fp16 = einsum(equation = var_8140_equation_0, values = (var_8020_cast_fp16_19, var_8100_cast_fp16))[name = tensor("op_8140_cast_fp16")]; + tensor input_295_interleave_0 = const()[name = tensor("input_295_interleave_0"), val = tensor(false)]; + tensor input_295_cast_fp16 = concat(axis = var_7925, interleave = input_295_interleave_0, values = (var_8102_cast_fp16, var_8104_cast_fp16, var_8106_cast_fp16, var_8108_cast_fp16, var_8110_cast_fp16, var_8112_cast_fp16, var_8114_cast_fp16, var_8116_cast_fp16, var_8118_cast_fp16, var_8120_cast_fp16, var_8122_cast_fp16, var_8124_cast_fp16, var_8126_cast_fp16, var_8128_cast_fp16, var_8130_cast_fp16, var_8132_cast_fp16, var_8134_cast_fp16, var_8136_cast_fp16, var_8138_cast_fp16, var_8140_cast_fp16))[name = tensor("input_295_cast_fp16")]; + tensor var_8149_pad_type_0 = const()[name = tensor("op_8149_pad_type_0"), val = tensor("valid")]; + tensor var_8149_strides_0 = const()[name = tensor("op_8149_strides_0"), val = tensor([1, 1])]; + tensor var_8149_pad_0 = const()[name = tensor("op_8149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8149_dilations_0 = const()[name = tensor("op_8149_dilations_0"), val = tensor([1, 1])]; + tensor var_8149_groups_0 = const()[name = tensor("op_8149_groups_0"), val = tensor(1)]; + tensor blocks_29_attn_out_weight_to_fp16 = const()[name = tensor("blocks_29_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165745152)))]; + tensor blocks_29_attn_out_bias_to_fp16 = const()[name = tensor("blocks_29_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169022016)))]; + tensor var_8149_cast_fp16 = conv(bias = blocks_29_attn_out_bias_to_fp16, dilations = var_8149_dilations_0, groups = var_8149_groups_0, pad = var_8149_pad_0, pad_type = var_8149_pad_type_0, strides = var_8149_strides_0, weight = blocks_29_attn_out_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("op_8149_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_8149_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor input_297_axes_0 = const()[name = tensor("input_297_axes_0"), val = tensor([1])]; + tensor input_297_gamma_0_to_fp16 = const()[name = tensor("input_297_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169024640)))]; + tensor input_297_beta_0_to_fp16 = const()[name = tensor("input_297_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169027264)))]; + tensor var_8159_to_fp16 = const()[name = tensor("op_8159_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_297_cast_fp16 = layer_norm(axes = input_297_axes_0, beta = input_297_beta_0_to_fp16, epsilon = var_8159_to_fp16, gamma = input_297_gamma_0_to_fp16, x = inputs_119_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_pad_type_0 = const()[name = tensor("input_299_pad_type_0"), val = tensor("valid")]; + tensor input_299_strides_0 = const()[name = tensor("input_299_strides_0"), val = tensor([1, 1])]; + tensor input_299_pad_0 = const()[name = tensor("input_299_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_299_dilations_0 = const()[name = tensor("input_299_dilations_0"), val = tensor([1, 1])]; + tensor input_299_groups_0 = const()[name = tensor("input_299_groups_0"), val = tensor(1)]; + tensor blocks_29_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169029888)))]; + tensor blocks_29_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1182137152)))]; + tensor input_299_cast_fp16 = conv(bias = blocks_29_mlp_0_bias_to_fp16, dilations = input_299_dilations_0, groups = input_299_groups_0, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = input_299_strides_0, weight = blocks_29_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor input_301_mode_0 = const()[name = tensor("input_301_mode_0"), val = tensor("EXACT")]; + tensor input_301_cast_fp16 = gelu(mode = input_301_mode_0, x = input_299_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_8185_pad_type_0 = const()[name = tensor("op_8185_pad_type_0"), val = tensor("valid")]; + tensor var_8185_strides_0 = const()[name = tensor("op_8185_strides_0"), val = tensor([1, 1])]; + tensor var_8185_pad_0 = const()[name = tensor("op_8185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8185_dilations_0 = const()[name = tensor("op_8185_dilations_0"), val = tensor([1, 1])]; + tensor var_8185_groups_0 = const()[name = tensor("op_8185_groups_0"), val = tensor(1)]; + tensor blocks_29_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1182147456)))]; + tensor blocks_29_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195254720)))]; + tensor var_8185_cast_fp16 = conv(bias = blocks_29_mlp_2_bias_to_fp16, dilations = var_8185_dilations_0, groups = var_8185_groups_0, pad = var_8185_pad_0, pad_type = var_8185_pad_type_0, strides = var_8185_strides_0, weight = blocks_29_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = var_8185_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_8194 = const()[name = tensor("op_8194"), val = tensor(1)]; + tensor input_303_axes_0 = const()[name = tensor("input_303_axes_0"), val = tensor([1])]; + tensor input_303_gamma_0_to_fp16 = const()[name = tensor("input_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195257344)))]; + tensor input_303_beta_0_to_fp16 = const()[name = tensor("input_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195259968)))]; + tensor var_8210_to_fp16 = const()[name = tensor("op_8210_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_303_cast_fp16 = layer_norm(axes = input_303_axes_0, beta = input_303_beta_0_to_fp16, epsilon = var_8210_to_fp16, gamma = input_303_gamma_0_to_fp16, x = inputs_121_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor q_61_pad_type_0 = const()[name = tensor("q_61_pad_type_0"), val = tensor("valid")]; + tensor q_61_strides_0 = const()[name = tensor("q_61_strides_0"), val = tensor([1, 1])]; + tensor q_61_pad_0 = const()[name = tensor("q_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_61_dilations_0 = const()[name = tensor("q_61_dilations_0"), val = tensor([1, 1])]; + tensor q_61_groups_0 = const()[name = tensor("q_61_groups_0"), val = tensor(1)]; + tensor var_8245_weight_0_to_fp16 = const()[name = tensor("op_8245_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195262592)))]; + tensor var_8245_bias_0_to_fp16 = const()[name = tensor("op_8245_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198539456)))]; + tensor var_8245_cast_fp16 = conv(bias = var_8245_bias_0_to_fp16, dilations = q_61_dilations_0, groups = q_61_groups_0, pad = q_61_pad_0, pad_type = q_61_pad_type_0, strides = q_61_strides_0, weight = var_8245_weight_0_to_fp16, x = input_303_cast_fp16)[name = tensor("op_8245_cast_fp16")]; + tensor k_61_pad_type_0 = const()[name = tensor("k_61_pad_type_0"), val = tensor("valid")]; + tensor k_61_strides_0 = const()[name = tensor("k_61_strides_0"), val = tensor([1, 1])]; + tensor k_61_pad_0 = const()[name = tensor("k_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_61_dilations_0 = const()[name = tensor("k_61_dilations_0"), val = tensor([1, 1])]; + tensor k_61_groups_0 = const()[name = tensor("k_61_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_key_weight_to_fp16 = const()[name = tensor("blocks_30_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198542080)))]; + tensor k_61_cast_fp16 = conv(dilations = k_61_dilations_0, groups = k_61_groups_0, pad = k_61_pad_0, pad_type = k_61_pad_type_0, strides = k_61_strides_0, weight = blocks_30_attn_key_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor var_8243_pad_type_0 = const()[name = tensor("op_8243_pad_type_0"), val = tensor("valid")]; + tensor var_8243_strides_0 = const()[name = tensor("op_8243_strides_0"), val = tensor([1, 1])]; + tensor var_8243_pad_0 = const()[name = tensor("op_8243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8243_dilations_0 = const()[name = tensor("op_8243_dilations_0"), val = tensor([1, 1])]; + tensor var_8243_groups_0 = const()[name = tensor("op_8243_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_value_weight_to_fp16 = const()[name = tensor("blocks_30_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201818944)))]; + tensor blocks_30_attn_value_bias_to_fp16 = const()[name = tensor("blocks_30_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1205095808)))]; + tensor var_8243_cast_fp16 = conv(bias = blocks_30_attn_value_bias_to_fp16, dilations = var_8243_dilations_0, groups = var_8243_groups_0, pad = var_8243_pad_0, pad_type = var_8243_pad_type_0, strides = var_8243_strides_0, weight = blocks_30_attn_value_weight_to_fp16, x = input_303_cast_fp16)[name = tensor("op_8243_cast_fp16")]; + tensor tile_90 = const()[name = tensor("tile_90"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8246_axis_0 = const()[name = tensor("op_8246_axis_0"), val = tensor(1)]; + tensor var_8246_cast_fp16_0, tensor var_8246_cast_fp16_1, tensor var_8246_cast_fp16_2, tensor var_8246_cast_fp16_3, tensor var_8246_cast_fp16_4, tensor var_8246_cast_fp16_5, tensor var_8246_cast_fp16_6, tensor var_8246_cast_fp16_7, tensor var_8246_cast_fp16_8, tensor var_8246_cast_fp16_9, tensor var_8246_cast_fp16_10, tensor var_8246_cast_fp16_11, tensor var_8246_cast_fp16_12, tensor var_8246_cast_fp16_13, tensor var_8246_cast_fp16_14, tensor var_8246_cast_fp16_15, tensor var_8246_cast_fp16_16, tensor var_8246_cast_fp16_17, tensor var_8246_cast_fp16_18, tensor var_8246_cast_fp16_19 = split(axis = var_8246_axis_0, split_sizes = tile_90, x = var_8245_cast_fp16)[name = tensor("op_8246_cast_fp16")]; + tensor var_8267_perm_0 = const()[name = tensor("op_8267_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_91 = const()[name = tensor("tile_91"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8268_axis_0 = const()[name = tensor("op_8268_axis_0"), val = tensor(3)]; + tensor var_8267_cast_fp16 = transpose(perm = var_8267_perm_0, x = k_61_cast_fp16)[name = tensor("transpose_2")]; + tensor var_8268_cast_fp16_0, tensor var_8268_cast_fp16_1, tensor var_8268_cast_fp16_2, tensor var_8268_cast_fp16_3, tensor var_8268_cast_fp16_4, tensor var_8268_cast_fp16_5, tensor var_8268_cast_fp16_6, tensor var_8268_cast_fp16_7, tensor var_8268_cast_fp16_8, tensor var_8268_cast_fp16_9, tensor var_8268_cast_fp16_10, tensor var_8268_cast_fp16_11, tensor var_8268_cast_fp16_12, tensor var_8268_cast_fp16_13, tensor var_8268_cast_fp16_14, tensor var_8268_cast_fp16_15, tensor var_8268_cast_fp16_16, tensor var_8268_cast_fp16_17, tensor var_8268_cast_fp16_18, tensor var_8268_cast_fp16_19 = split(axis = var_8268_axis_0, split_sizes = tile_91, x = var_8267_cast_fp16)[name = tensor("op_8268_cast_fp16")]; + tensor tile_92 = const()[name = tensor("tile_92"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8289_axis_0 = const()[name = tensor("op_8289_axis_0"), val = tensor(1)]; + tensor var_8289_cast_fp16_0, tensor var_8289_cast_fp16_1, tensor var_8289_cast_fp16_2, tensor var_8289_cast_fp16_3, tensor var_8289_cast_fp16_4, tensor var_8289_cast_fp16_5, tensor var_8289_cast_fp16_6, tensor var_8289_cast_fp16_7, tensor var_8289_cast_fp16_8, tensor var_8289_cast_fp16_9, tensor var_8289_cast_fp16_10, tensor var_8289_cast_fp16_11, tensor var_8289_cast_fp16_12, tensor var_8289_cast_fp16_13, tensor var_8289_cast_fp16_14, tensor var_8289_cast_fp16_15, tensor var_8289_cast_fp16_16, tensor var_8289_cast_fp16_17, tensor var_8289_cast_fp16_18, tensor var_8289_cast_fp16_19 = split(axis = var_8289_axis_0, split_sizes = tile_92, x = var_8243_cast_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor aw_1201_equation_0 = const()[name = tensor("aw_1201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1201_cast_fp16 = einsum(equation = aw_1201_equation_0, values = (var_8268_cast_fp16_0, var_8246_cast_fp16_0))[name = tensor("aw_1201_cast_fp16")]; + tensor aw_1203_equation_0 = const()[name = tensor("aw_1203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1203_cast_fp16 = einsum(equation = aw_1203_equation_0, values = (var_8268_cast_fp16_1, var_8246_cast_fp16_1))[name = tensor("aw_1203_cast_fp16")]; + tensor aw_1205_equation_0 = const()[name = tensor("aw_1205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1205_cast_fp16 = einsum(equation = aw_1205_equation_0, values = (var_8268_cast_fp16_2, var_8246_cast_fp16_2))[name = tensor("aw_1205_cast_fp16")]; + tensor aw_1207_equation_0 = const()[name = tensor("aw_1207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1207_cast_fp16 = einsum(equation = aw_1207_equation_0, values = (var_8268_cast_fp16_3, var_8246_cast_fp16_3))[name = tensor("aw_1207_cast_fp16")]; + tensor aw_1209_equation_0 = const()[name = tensor("aw_1209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1209_cast_fp16 = einsum(equation = aw_1209_equation_0, values = (var_8268_cast_fp16_4, var_8246_cast_fp16_4))[name = tensor("aw_1209_cast_fp16")]; + tensor aw_1211_equation_0 = const()[name = tensor("aw_1211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1211_cast_fp16 = einsum(equation = aw_1211_equation_0, values = (var_8268_cast_fp16_5, var_8246_cast_fp16_5))[name = tensor("aw_1211_cast_fp16")]; + tensor aw_1213_equation_0 = const()[name = tensor("aw_1213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1213_cast_fp16 = einsum(equation = aw_1213_equation_0, values = (var_8268_cast_fp16_6, var_8246_cast_fp16_6))[name = tensor("aw_1213_cast_fp16")]; + tensor aw_1215_equation_0 = const()[name = tensor("aw_1215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1215_cast_fp16 = einsum(equation = aw_1215_equation_0, values = (var_8268_cast_fp16_7, var_8246_cast_fp16_7))[name = tensor("aw_1215_cast_fp16")]; + tensor aw_1217_equation_0 = const()[name = tensor("aw_1217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1217_cast_fp16 = einsum(equation = aw_1217_equation_0, values = (var_8268_cast_fp16_8, var_8246_cast_fp16_8))[name = tensor("aw_1217_cast_fp16")]; + tensor aw_1219_equation_0 = const()[name = tensor("aw_1219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1219_cast_fp16 = einsum(equation = aw_1219_equation_0, values = (var_8268_cast_fp16_9, var_8246_cast_fp16_9))[name = tensor("aw_1219_cast_fp16")]; + tensor aw_1221_equation_0 = const()[name = tensor("aw_1221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1221_cast_fp16 = einsum(equation = aw_1221_equation_0, values = (var_8268_cast_fp16_10, var_8246_cast_fp16_10))[name = tensor("aw_1221_cast_fp16")]; + tensor aw_1223_equation_0 = const()[name = tensor("aw_1223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1223_cast_fp16 = einsum(equation = aw_1223_equation_0, values = (var_8268_cast_fp16_11, var_8246_cast_fp16_11))[name = tensor("aw_1223_cast_fp16")]; + tensor aw_1225_equation_0 = const()[name = tensor("aw_1225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1225_cast_fp16 = einsum(equation = aw_1225_equation_0, values = (var_8268_cast_fp16_12, var_8246_cast_fp16_12))[name = tensor("aw_1225_cast_fp16")]; + tensor aw_1227_equation_0 = const()[name = tensor("aw_1227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1227_cast_fp16 = einsum(equation = aw_1227_equation_0, values = (var_8268_cast_fp16_13, var_8246_cast_fp16_13))[name = tensor("aw_1227_cast_fp16")]; + tensor aw_1229_equation_0 = const()[name = tensor("aw_1229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1229_cast_fp16 = einsum(equation = aw_1229_equation_0, values = (var_8268_cast_fp16_14, var_8246_cast_fp16_14))[name = tensor("aw_1229_cast_fp16")]; + tensor aw_1231_equation_0 = const()[name = tensor("aw_1231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1231_cast_fp16 = einsum(equation = aw_1231_equation_0, values = (var_8268_cast_fp16_15, var_8246_cast_fp16_15))[name = tensor("aw_1231_cast_fp16")]; + tensor aw_1233_equation_0 = const()[name = tensor("aw_1233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1233_cast_fp16 = einsum(equation = aw_1233_equation_0, values = (var_8268_cast_fp16_16, var_8246_cast_fp16_16))[name = tensor("aw_1233_cast_fp16")]; + tensor aw_1235_equation_0 = const()[name = tensor("aw_1235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1235_cast_fp16 = einsum(equation = aw_1235_equation_0, values = (var_8268_cast_fp16_17, var_8246_cast_fp16_17))[name = tensor("aw_1235_cast_fp16")]; + tensor aw_1237_equation_0 = const()[name = tensor("aw_1237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1237_cast_fp16 = einsum(equation = aw_1237_equation_0, values = (var_8268_cast_fp16_18, var_8246_cast_fp16_18))[name = tensor("aw_1237_cast_fp16")]; + tensor aw_1239_equation_0 = const()[name = tensor("aw_1239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1239_cast_fp16 = einsum(equation = aw_1239_equation_0, values = (var_8268_cast_fp16_19, var_8246_cast_fp16_19))[name = tensor("aw_1239_cast_fp16")]; + tensor var_8350_cast_fp16 = softmax(axis = var_8194, x = aw_1201_cast_fp16)[name = tensor("op_8350_cast_fp16")]; + tensor var_8351_cast_fp16 = softmax(axis = var_8194, x = aw_1203_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = softmax(axis = var_8194, x = aw_1205_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor var_8353_cast_fp16 = softmax(axis = var_8194, x = aw_1207_cast_fp16)[name = tensor("op_8353_cast_fp16")]; + tensor var_8354_cast_fp16 = softmax(axis = var_8194, x = aw_1209_cast_fp16)[name = tensor("op_8354_cast_fp16")]; + tensor var_8355_cast_fp16 = softmax(axis = var_8194, x = aw_1211_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = softmax(axis = var_8194, x = aw_1213_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor var_8357_cast_fp16 = softmax(axis = var_8194, x = aw_1215_cast_fp16)[name = tensor("op_8357_cast_fp16")]; + tensor var_8358_cast_fp16 = softmax(axis = var_8194, x = aw_1217_cast_fp16)[name = tensor("op_8358_cast_fp16")]; + tensor var_8359_cast_fp16 = softmax(axis = var_8194, x = aw_1219_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = softmax(axis = var_8194, x = aw_1221_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor var_8361_cast_fp16 = softmax(axis = var_8194, x = aw_1223_cast_fp16)[name = tensor("op_8361_cast_fp16")]; + tensor var_8362_cast_fp16 = softmax(axis = var_8194, x = aw_1225_cast_fp16)[name = tensor("op_8362_cast_fp16")]; + tensor var_8363_cast_fp16 = softmax(axis = var_8194, x = aw_1227_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = softmax(axis = var_8194, x = aw_1229_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor var_8365_cast_fp16 = softmax(axis = var_8194, x = aw_1231_cast_fp16)[name = tensor("op_8365_cast_fp16")]; + tensor var_8366_cast_fp16 = softmax(axis = var_8194, x = aw_1233_cast_fp16)[name = tensor("op_8366_cast_fp16")]; + tensor var_8367_cast_fp16 = softmax(axis = var_8194, x = aw_1235_cast_fp16)[name = tensor("op_8367_cast_fp16")]; + tensor var_8368_cast_fp16 = softmax(axis = var_8194, x = aw_1237_cast_fp16)[name = tensor("op_8368_cast_fp16")]; + tensor var_8369_cast_fp16 = softmax(axis = var_8194, x = aw_1239_cast_fp16)[name = tensor("op_8369_cast_fp16")]; + tensor var_8371_equation_0 = const()[name = tensor("op_8371_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8371_cast_fp16 = einsum(equation = var_8371_equation_0, values = (var_8289_cast_fp16_0, var_8350_cast_fp16))[name = tensor("op_8371_cast_fp16")]; + tensor var_8373_equation_0 = const()[name = tensor("op_8373_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8373_cast_fp16 = einsum(equation = var_8373_equation_0, values = (var_8289_cast_fp16_1, var_8351_cast_fp16))[name = tensor("op_8373_cast_fp16")]; + tensor var_8375_equation_0 = const()[name = tensor("op_8375_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8375_cast_fp16 = einsum(equation = var_8375_equation_0, values = (var_8289_cast_fp16_2, var_8352_cast_fp16))[name = tensor("op_8375_cast_fp16")]; + tensor var_8377_equation_0 = const()[name = tensor("op_8377_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8377_cast_fp16 = einsum(equation = var_8377_equation_0, values = (var_8289_cast_fp16_3, var_8353_cast_fp16))[name = tensor("op_8377_cast_fp16")]; + tensor var_8379_equation_0 = const()[name = tensor("op_8379_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8379_cast_fp16 = einsum(equation = var_8379_equation_0, values = (var_8289_cast_fp16_4, var_8354_cast_fp16))[name = tensor("op_8379_cast_fp16")]; + tensor var_8381_equation_0 = const()[name = tensor("op_8381_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8381_cast_fp16 = einsum(equation = var_8381_equation_0, values = (var_8289_cast_fp16_5, var_8355_cast_fp16))[name = tensor("op_8381_cast_fp16")]; + tensor var_8383_equation_0 = const()[name = tensor("op_8383_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8383_cast_fp16 = einsum(equation = var_8383_equation_0, values = (var_8289_cast_fp16_6, var_8356_cast_fp16))[name = tensor("op_8383_cast_fp16")]; + tensor var_8385_equation_0 = const()[name = tensor("op_8385_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8385_cast_fp16 = einsum(equation = var_8385_equation_0, values = (var_8289_cast_fp16_7, var_8357_cast_fp16))[name = tensor("op_8385_cast_fp16")]; + tensor var_8387_equation_0 = const()[name = tensor("op_8387_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8387_cast_fp16 = einsum(equation = var_8387_equation_0, values = (var_8289_cast_fp16_8, var_8358_cast_fp16))[name = tensor("op_8387_cast_fp16")]; + tensor var_8389_equation_0 = const()[name = tensor("op_8389_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8389_cast_fp16 = einsum(equation = var_8389_equation_0, values = (var_8289_cast_fp16_9, var_8359_cast_fp16))[name = tensor("op_8389_cast_fp16")]; + tensor var_8391_equation_0 = const()[name = tensor("op_8391_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8391_cast_fp16 = einsum(equation = var_8391_equation_0, values = (var_8289_cast_fp16_10, var_8360_cast_fp16))[name = tensor("op_8391_cast_fp16")]; + tensor var_8393_equation_0 = const()[name = tensor("op_8393_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8393_cast_fp16 = einsum(equation = var_8393_equation_0, values = (var_8289_cast_fp16_11, var_8361_cast_fp16))[name = tensor("op_8393_cast_fp16")]; + tensor var_8395_equation_0 = const()[name = tensor("op_8395_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8395_cast_fp16 = einsum(equation = var_8395_equation_0, values = (var_8289_cast_fp16_12, var_8362_cast_fp16))[name = tensor("op_8395_cast_fp16")]; + tensor var_8397_equation_0 = const()[name = tensor("op_8397_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8397_cast_fp16 = einsum(equation = var_8397_equation_0, values = (var_8289_cast_fp16_13, var_8363_cast_fp16))[name = tensor("op_8397_cast_fp16")]; + tensor var_8399_equation_0 = const()[name = tensor("op_8399_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8399_cast_fp16 = einsum(equation = var_8399_equation_0, values = (var_8289_cast_fp16_14, var_8364_cast_fp16))[name = tensor("op_8399_cast_fp16")]; + tensor var_8401_equation_0 = const()[name = tensor("op_8401_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8401_cast_fp16 = einsum(equation = var_8401_equation_0, values = (var_8289_cast_fp16_15, var_8365_cast_fp16))[name = tensor("op_8401_cast_fp16")]; + tensor var_8403_equation_0 = const()[name = tensor("op_8403_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8403_cast_fp16 = einsum(equation = var_8403_equation_0, values = (var_8289_cast_fp16_16, var_8366_cast_fp16))[name = tensor("op_8403_cast_fp16")]; + tensor var_8405_equation_0 = const()[name = tensor("op_8405_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8405_cast_fp16 = einsum(equation = var_8405_equation_0, values = (var_8289_cast_fp16_17, var_8367_cast_fp16))[name = tensor("op_8405_cast_fp16")]; + tensor var_8407_equation_0 = const()[name = tensor("op_8407_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8407_cast_fp16 = einsum(equation = var_8407_equation_0, values = (var_8289_cast_fp16_18, var_8368_cast_fp16))[name = tensor("op_8407_cast_fp16")]; + tensor var_8409_equation_0 = const()[name = tensor("op_8409_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8409_cast_fp16 = einsum(equation = var_8409_equation_0, values = (var_8289_cast_fp16_19, var_8369_cast_fp16))[name = tensor("op_8409_cast_fp16")]; + tensor input_305_interleave_0 = const()[name = tensor("input_305_interleave_0"), val = tensor(false)]; + tensor input_305_cast_fp16 = concat(axis = var_8194, interleave = input_305_interleave_0, values = (var_8371_cast_fp16, var_8373_cast_fp16, var_8375_cast_fp16, var_8377_cast_fp16, var_8379_cast_fp16, var_8381_cast_fp16, var_8383_cast_fp16, var_8385_cast_fp16, var_8387_cast_fp16, var_8389_cast_fp16, var_8391_cast_fp16, var_8393_cast_fp16, var_8395_cast_fp16, var_8397_cast_fp16, var_8399_cast_fp16, var_8401_cast_fp16, var_8403_cast_fp16, var_8405_cast_fp16, var_8407_cast_fp16, var_8409_cast_fp16))[name = tensor("input_305_cast_fp16")]; + tensor var_8418_pad_type_0 = const()[name = tensor("op_8418_pad_type_0"), val = tensor("valid")]; + tensor var_8418_strides_0 = const()[name = tensor("op_8418_strides_0"), val = tensor([1, 1])]; + tensor var_8418_pad_0 = const()[name = tensor("op_8418_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8418_dilations_0 = const()[name = tensor("op_8418_dilations_0"), val = tensor([1, 1])]; + tensor var_8418_groups_0 = const()[name = tensor("op_8418_groups_0"), val = tensor(1)]; + tensor blocks_30_attn_out_weight_to_fp16 = const()[name = tensor("blocks_30_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1205098432)))]; + tensor blocks_30_attn_out_bias_to_fp16 = const()[name = tensor("blocks_30_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208375296)))]; + tensor var_8418_cast_fp16 = conv(bias = blocks_30_attn_out_bias_to_fp16, dilations = var_8418_dilations_0, groups = var_8418_groups_0, pad = var_8418_pad_0, pad_type = var_8418_pad_type_0, strides = var_8418_strides_0, weight = blocks_30_attn_out_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_8418_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor input_307_axes_0 = const()[name = tensor("input_307_axes_0"), val = tensor([1])]; + tensor input_307_gamma_0_to_fp16 = const()[name = tensor("input_307_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208377920)))]; + tensor input_307_beta_0_to_fp16 = const()[name = tensor("input_307_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208380544)))]; + tensor var_8428_to_fp16 = const()[name = tensor("op_8428_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = input_307_beta_0_to_fp16, epsilon = var_8428_to_fp16, gamma = input_307_gamma_0_to_fp16, x = inputs_123_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor input_309_pad_type_0 = const()[name = tensor("input_309_pad_type_0"), val = tensor("valid")]; + tensor input_309_strides_0 = const()[name = tensor("input_309_strides_0"), val = tensor([1, 1])]; + tensor input_309_pad_0 = const()[name = tensor("input_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_309_dilations_0 = const()[name = tensor("input_309_dilations_0"), val = tensor([1, 1])]; + tensor input_309_groups_0 = const()[name = tensor("input_309_groups_0"), val = tensor(1)]; + tensor blocks_30_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208383168)))]; + tensor blocks_30_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221490432)))]; + tensor input_309_cast_fp16 = conv(bias = blocks_30_mlp_0_bias_to_fp16, dilations = input_309_dilations_0, groups = input_309_groups_0, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = input_309_strides_0, weight = blocks_30_mlp_0_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor input_311_mode_0 = const()[name = tensor("input_311_mode_0"), val = tensor("EXACT")]; + tensor input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = input_309_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor var_8454_pad_type_0 = const()[name = tensor("op_8454_pad_type_0"), val = tensor("valid")]; + tensor var_8454_strides_0 = const()[name = tensor("op_8454_strides_0"), val = tensor([1, 1])]; + tensor var_8454_pad_0 = const()[name = tensor("op_8454_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8454_dilations_0 = const()[name = tensor("op_8454_dilations_0"), val = tensor([1, 1])]; + tensor var_8454_groups_0 = const()[name = tensor("op_8454_groups_0"), val = tensor(1)]; + tensor blocks_30_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221500736)))]; + tensor blocks_30_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234608000)))]; + tensor var_8454_cast_fp16 = conv(bias = blocks_30_mlp_2_bias_to_fp16, dilations = var_8454_dilations_0, groups = var_8454_groups_0, pad = var_8454_pad_0, pad_type = var_8454_pad_type_0, strides = var_8454_strides_0, weight = blocks_30_mlp_2_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = var_8454_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_8463 = const()[name = tensor("op_8463"), val = tensor(1)]; + tensor input_313_axes_0 = const()[name = tensor("input_313_axes_0"), val = tensor([1])]; + tensor input_313_gamma_0_to_fp16 = const()[name = tensor("input_313_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234610624)))]; + tensor input_313_beta_0_to_fp16 = const()[name = tensor("input_313_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234613248)))]; + tensor var_8479_to_fp16 = const()[name = tensor("op_8479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_313_cast_fp16 = layer_norm(axes = input_313_axes_0, beta = input_313_beta_0_to_fp16, epsilon = var_8479_to_fp16, gamma = input_313_gamma_0_to_fp16, x = inputs_125_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_8514_weight_0_to_fp16 = const()[name = tensor("op_8514_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234615872)))]; + tensor var_8514_bias_0_to_fp16 = const()[name = tensor("op_8514_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237892736)))]; + tensor var_8514_cast_fp16 = conv(bias = var_8514_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_8514_weight_0_to_fp16, x = input_313_cast_fp16)[name = tensor("op_8514_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_key_weight_to_fp16 = const()[name = tensor("blocks_31_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237895360)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_31_attn_key_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_8512_pad_type_0 = const()[name = tensor("op_8512_pad_type_0"), val = tensor("valid")]; + tensor var_8512_strides_0 = const()[name = tensor("op_8512_strides_0"), val = tensor([1, 1])]; + tensor var_8512_pad_0 = const()[name = tensor("op_8512_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8512_dilations_0 = const()[name = tensor("op_8512_dilations_0"), val = tensor([1, 1])]; + tensor var_8512_groups_0 = const()[name = tensor("op_8512_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_value_weight_to_fp16 = const()[name = tensor("blocks_31_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1241172224)))]; + tensor blocks_31_attn_value_bias_to_fp16 = const()[name = tensor("blocks_31_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244449088)))]; + tensor var_8512_cast_fp16 = conv(bias = blocks_31_attn_value_bias_to_fp16, dilations = var_8512_dilations_0, groups = var_8512_groups_0, pad = var_8512_pad_0, pad_type = var_8512_pad_type_0, strides = var_8512_strides_0, weight = blocks_31_attn_value_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("op_8512_cast_fp16")]; + tensor tile_93 = const()[name = tensor("tile_93"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8515_axis_0 = const()[name = tensor("op_8515_axis_0"), val = tensor(1)]; + tensor var_8515_cast_fp16_0, tensor var_8515_cast_fp16_1, tensor var_8515_cast_fp16_2, tensor var_8515_cast_fp16_3, tensor var_8515_cast_fp16_4, tensor var_8515_cast_fp16_5, tensor var_8515_cast_fp16_6, tensor var_8515_cast_fp16_7, tensor var_8515_cast_fp16_8, tensor var_8515_cast_fp16_9, tensor var_8515_cast_fp16_10, tensor var_8515_cast_fp16_11, tensor var_8515_cast_fp16_12, tensor var_8515_cast_fp16_13, tensor var_8515_cast_fp16_14, tensor var_8515_cast_fp16_15, tensor var_8515_cast_fp16_16, tensor var_8515_cast_fp16_17, tensor var_8515_cast_fp16_18, tensor var_8515_cast_fp16_19 = split(axis = var_8515_axis_0, split_sizes = tile_93, x = var_8514_cast_fp16)[name = tensor("op_8515_cast_fp16")]; + tensor var_8536_perm_0 = const()[name = tensor("op_8536_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_94 = const()[name = tensor("tile_94"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8537_axis_0 = const()[name = tensor("op_8537_axis_0"), val = tensor(3)]; + tensor var_8536_cast_fp16 = transpose(perm = var_8536_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_8537_cast_fp16_0, tensor var_8537_cast_fp16_1, tensor var_8537_cast_fp16_2, tensor var_8537_cast_fp16_3, tensor var_8537_cast_fp16_4, tensor var_8537_cast_fp16_5, tensor var_8537_cast_fp16_6, tensor var_8537_cast_fp16_7, tensor var_8537_cast_fp16_8, tensor var_8537_cast_fp16_9, tensor var_8537_cast_fp16_10, tensor var_8537_cast_fp16_11, tensor var_8537_cast_fp16_12, tensor var_8537_cast_fp16_13, tensor var_8537_cast_fp16_14, tensor var_8537_cast_fp16_15, tensor var_8537_cast_fp16_16, tensor var_8537_cast_fp16_17, tensor var_8537_cast_fp16_18, tensor var_8537_cast_fp16_19 = split(axis = var_8537_axis_0, split_sizes = tile_94, x = var_8536_cast_fp16)[name = tensor("op_8537_cast_fp16")]; + tensor tile_95 = const()[name = tensor("tile_95"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_8558_axis_0 = const()[name = tensor("op_8558_axis_0"), val = tensor(1)]; + tensor var_8558_cast_fp16_0, tensor var_8558_cast_fp16_1, tensor var_8558_cast_fp16_2, tensor var_8558_cast_fp16_3, tensor var_8558_cast_fp16_4, tensor var_8558_cast_fp16_5, tensor var_8558_cast_fp16_6, tensor var_8558_cast_fp16_7, tensor var_8558_cast_fp16_8, tensor var_8558_cast_fp16_9, tensor var_8558_cast_fp16_10, tensor var_8558_cast_fp16_11, tensor var_8558_cast_fp16_12, tensor var_8558_cast_fp16_13, tensor var_8558_cast_fp16_14, tensor var_8558_cast_fp16_15, tensor var_8558_cast_fp16_16, tensor var_8558_cast_fp16_17, tensor var_8558_cast_fp16_18, tensor var_8558_cast_fp16_19 = split(axis = var_8558_axis_0, split_sizes = tile_95, x = var_8512_cast_fp16)[name = tensor("op_8558_cast_fp16")]; + tensor aw_1241_equation_0 = const()[name = tensor("aw_1241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1241_cast_fp16 = einsum(equation = aw_1241_equation_0, values = (var_8537_cast_fp16_0, var_8515_cast_fp16_0))[name = tensor("aw_1241_cast_fp16")]; + tensor aw_1243_equation_0 = const()[name = tensor("aw_1243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1243_cast_fp16 = einsum(equation = aw_1243_equation_0, values = (var_8537_cast_fp16_1, var_8515_cast_fp16_1))[name = tensor("aw_1243_cast_fp16")]; + tensor aw_1245_equation_0 = const()[name = tensor("aw_1245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1245_cast_fp16 = einsum(equation = aw_1245_equation_0, values = (var_8537_cast_fp16_2, var_8515_cast_fp16_2))[name = tensor("aw_1245_cast_fp16")]; + tensor aw_1247_equation_0 = const()[name = tensor("aw_1247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1247_cast_fp16 = einsum(equation = aw_1247_equation_0, values = (var_8537_cast_fp16_3, var_8515_cast_fp16_3))[name = tensor("aw_1247_cast_fp16")]; + tensor aw_1249_equation_0 = const()[name = tensor("aw_1249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1249_cast_fp16 = einsum(equation = aw_1249_equation_0, values = (var_8537_cast_fp16_4, var_8515_cast_fp16_4))[name = tensor("aw_1249_cast_fp16")]; + tensor aw_1251_equation_0 = const()[name = tensor("aw_1251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1251_cast_fp16 = einsum(equation = aw_1251_equation_0, values = (var_8537_cast_fp16_5, var_8515_cast_fp16_5))[name = tensor("aw_1251_cast_fp16")]; + tensor aw_1253_equation_0 = const()[name = tensor("aw_1253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1253_cast_fp16 = einsum(equation = aw_1253_equation_0, values = (var_8537_cast_fp16_6, var_8515_cast_fp16_6))[name = tensor("aw_1253_cast_fp16")]; + tensor aw_1255_equation_0 = const()[name = tensor("aw_1255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1255_cast_fp16 = einsum(equation = aw_1255_equation_0, values = (var_8537_cast_fp16_7, var_8515_cast_fp16_7))[name = tensor("aw_1255_cast_fp16")]; + tensor aw_1257_equation_0 = const()[name = tensor("aw_1257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1257_cast_fp16 = einsum(equation = aw_1257_equation_0, values = (var_8537_cast_fp16_8, var_8515_cast_fp16_8))[name = tensor("aw_1257_cast_fp16")]; + tensor aw_1259_equation_0 = const()[name = tensor("aw_1259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1259_cast_fp16 = einsum(equation = aw_1259_equation_0, values = (var_8537_cast_fp16_9, var_8515_cast_fp16_9))[name = tensor("aw_1259_cast_fp16")]; + tensor aw_1261_equation_0 = const()[name = tensor("aw_1261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1261_cast_fp16 = einsum(equation = aw_1261_equation_0, values = (var_8537_cast_fp16_10, var_8515_cast_fp16_10))[name = tensor("aw_1261_cast_fp16")]; + tensor aw_1263_equation_0 = const()[name = tensor("aw_1263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1263_cast_fp16 = einsum(equation = aw_1263_equation_0, values = (var_8537_cast_fp16_11, var_8515_cast_fp16_11))[name = tensor("aw_1263_cast_fp16")]; + tensor aw_1265_equation_0 = const()[name = tensor("aw_1265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1265_cast_fp16 = einsum(equation = aw_1265_equation_0, values = (var_8537_cast_fp16_12, var_8515_cast_fp16_12))[name = tensor("aw_1265_cast_fp16")]; + tensor aw_1267_equation_0 = const()[name = tensor("aw_1267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1267_cast_fp16 = einsum(equation = aw_1267_equation_0, values = (var_8537_cast_fp16_13, var_8515_cast_fp16_13))[name = tensor("aw_1267_cast_fp16")]; + tensor aw_1269_equation_0 = const()[name = tensor("aw_1269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1269_cast_fp16 = einsum(equation = aw_1269_equation_0, values = (var_8537_cast_fp16_14, var_8515_cast_fp16_14))[name = tensor("aw_1269_cast_fp16")]; + tensor aw_1271_equation_0 = const()[name = tensor("aw_1271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1271_cast_fp16 = einsum(equation = aw_1271_equation_0, values = (var_8537_cast_fp16_15, var_8515_cast_fp16_15))[name = tensor("aw_1271_cast_fp16")]; + tensor aw_1273_equation_0 = const()[name = tensor("aw_1273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1273_cast_fp16 = einsum(equation = aw_1273_equation_0, values = (var_8537_cast_fp16_16, var_8515_cast_fp16_16))[name = tensor("aw_1273_cast_fp16")]; + tensor aw_1275_equation_0 = const()[name = tensor("aw_1275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1275_cast_fp16 = einsum(equation = aw_1275_equation_0, values = (var_8537_cast_fp16_17, var_8515_cast_fp16_17))[name = tensor("aw_1275_cast_fp16")]; + tensor aw_1277_equation_0 = const()[name = tensor("aw_1277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1277_cast_fp16 = einsum(equation = aw_1277_equation_0, values = (var_8537_cast_fp16_18, var_8515_cast_fp16_18))[name = tensor("aw_1277_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_8537_cast_fp16_19, var_8515_cast_fp16_19))[name = tensor("aw_cast_fp16")]; + tensor var_8619_cast_fp16 = softmax(axis = var_8463, x = aw_1241_cast_fp16)[name = tensor("op_8619_cast_fp16")]; + tensor var_8620_cast_fp16 = softmax(axis = var_8463, x = aw_1243_cast_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor var_8621_cast_fp16 = softmax(axis = var_8463, x = aw_1245_cast_fp16)[name = tensor("op_8621_cast_fp16")]; + tensor var_8622_cast_fp16 = softmax(axis = var_8463, x = aw_1247_cast_fp16)[name = tensor("op_8622_cast_fp16")]; + tensor var_8623_cast_fp16 = softmax(axis = var_8463, x = aw_1249_cast_fp16)[name = tensor("op_8623_cast_fp16")]; + tensor var_8624_cast_fp16 = softmax(axis = var_8463, x = aw_1251_cast_fp16)[name = tensor("op_8624_cast_fp16")]; + tensor var_8625_cast_fp16 = softmax(axis = var_8463, x = aw_1253_cast_fp16)[name = tensor("op_8625_cast_fp16")]; + tensor var_8626_cast_fp16 = softmax(axis = var_8463, x = aw_1255_cast_fp16)[name = tensor("op_8626_cast_fp16")]; + tensor var_8627_cast_fp16 = softmax(axis = var_8463, x = aw_1257_cast_fp16)[name = tensor("op_8627_cast_fp16")]; + tensor var_8628_cast_fp16 = softmax(axis = var_8463, x = aw_1259_cast_fp16)[name = tensor("op_8628_cast_fp16")]; + tensor var_8629_cast_fp16 = softmax(axis = var_8463, x = aw_1261_cast_fp16)[name = tensor("op_8629_cast_fp16")]; + tensor var_8630_cast_fp16 = softmax(axis = var_8463, x = aw_1263_cast_fp16)[name = tensor("op_8630_cast_fp16")]; + tensor var_8631_cast_fp16 = softmax(axis = var_8463, x = aw_1265_cast_fp16)[name = tensor("op_8631_cast_fp16")]; + tensor var_8632_cast_fp16 = softmax(axis = var_8463, x = aw_1267_cast_fp16)[name = tensor("op_8632_cast_fp16")]; + tensor var_8633_cast_fp16 = softmax(axis = var_8463, x = aw_1269_cast_fp16)[name = tensor("op_8633_cast_fp16")]; + tensor var_8634_cast_fp16 = softmax(axis = var_8463, x = aw_1271_cast_fp16)[name = tensor("op_8634_cast_fp16")]; + tensor var_8635_cast_fp16 = softmax(axis = var_8463, x = aw_1273_cast_fp16)[name = tensor("op_8635_cast_fp16")]; + tensor var_8636_cast_fp16 = softmax(axis = var_8463, x = aw_1275_cast_fp16)[name = tensor("op_8636_cast_fp16")]; + tensor var_8637_cast_fp16 = softmax(axis = var_8463, x = aw_1277_cast_fp16)[name = tensor("op_8637_cast_fp16")]; + tensor var_8638_cast_fp16 = softmax(axis = var_8463, x = aw_cast_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor var_8640_equation_0 = const()[name = tensor("op_8640_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8640_cast_fp16 = einsum(equation = var_8640_equation_0, values = (var_8558_cast_fp16_0, var_8619_cast_fp16))[name = tensor("op_8640_cast_fp16")]; + tensor var_8642_equation_0 = const()[name = tensor("op_8642_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8642_cast_fp16 = einsum(equation = var_8642_equation_0, values = (var_8558_cast_fp16_1, var_8620_cast_fp16))[name = tensor("op_8642_cast_fp16")]; + tensor var_8644_equation_0 = const()[name = tensor("op_8644_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8644_cast_fp16 = einsum(equation = var_8644_equation_0, values = (var_8558_cast_fp16_2, var_8621_cast_fp16))[name = tensor("op_8644_cast_fp16")]; + tensor var_8646_equation_0 = const()[name = tensor("op_8646_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8646_cast_fp16 = einsum(equation = var_8646_equation_0, values = (var_8558_cast_fp16_3, var_8622_cast_fp16))[name = tensor("op_8646_cast_fp16")]; + tensor var_8648_equation_0 = const()[name = tensor("op_8648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8648_cast_fp16 = einsum(equation = var_8648_equation_0, values = (var_8558_cast_fp16_4, var_8623_cast_fp16))[name = tensor("op_8648_cast_fp16")]; + tensor var_8650_equation_0 = const()[name = tensor("op_8650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8650_cast_fp16 = einsum(equation = var_8650_equation_0, values = (var_8558_cast_fp16_5, var_8624_cast_fp16))[name = tensor("op_8650_cast_fp16")]; + tensor var_8652_equation_0 = const()[name = tensor("op_8652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8652_cast_fp16 = einsum(equation = var_8652_equation_0, values = (var_8558_cast_fp16_6, var_8625_cast_fp16))[name = tensor("op_8652_cast_fp16")]; + tensor var_8654_equation_0 = const()[name = tensor("op_8654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8654_cast_fp16 = einsum(equation = var_8654_equation_0, values = (var_8558_cast_fp16_7, var_8626_cast_fp16))[name = tensor("op_8654_cast_fp16")]; + tensor var_8656_equation_0 = const()[name = tensor("op_8656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8656_cast_fp16 = einsum(equation = var_8656_equation_0, values = (var_8558_cast_fp16_8, var_8627_cast_fp16))[name = tensor("op_8656_cast_fp16")]; + tensor var_8658_equation_0 = const()[name = tensor("op_8658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8658_cast_fp16 = einsum(equation = var_8658_equation_0, values = (var_8558_cast_fp16_9, var_8628_cast_fp16))[name = tensor("op_8658_cast_fp16")]; + tensor var_8660_equation_0 = const()[name = tensor("op_8660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8660_cast_fp16 = einsum(equation = var_8660_equation_0, values = (var_8558_cast_fp16_10, var_8629_cast_fp16))[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_equation_0 = const()[name = tensor("op_8662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8662_cast_fp16 = einsum(equation = var_8662_equation_0, values = (var_8558_cast_fp16_11, var_8630_cast_fp16))[name = tensor("op_8662_cast_fp16")]; + tensor var_8664_equation_0 = const()[name = tensor("op_8664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8664_cast_fp16 = einsum(equation = var_8664_equation_0, values = (var_8558_cast_fp16_12, var_8631_cast_fp16))[name = tensor("op_8664_cast_fp16")]; + tensor var_8666_equation_0 = const()[name = tensor("op_8666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8666_cast_fp16 = einsum(equation = var_8666_equation_0, values = (var_8558_cast_fp16_13, var_8632_cast_fp16))[name = tensor("op_8666_cast_fp16")]; + tensor var_8668_equation_0 = const()[name = tensor("op_8668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8668_cast_fp16 = einsum(equation = var_8668_equation_0, values = (var_8558_cast_fp16_14, var_8633_cast_fp16))[name = tensor("op_8668_cast_fp16")]; + tensor var_8670_equation_0 = const()[name = tensor("op_8670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8670_cast_fp16 = einsum(equation = var_8670_equation_0, values = (var_8558_cast_fp16_15, var_8634_cast_fp16))[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_equation_0 = const()[name = tensor("op_8672_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8672_cast_fp16 = einsum(equation = var_8672_equation_0, values = (var_8558_cast_fp16_16, var_8635_cast_fp16))[name = tensor("op_8672_cast_fp16")]; + tensor var_8674_equation_0 = const()[name = tensor("op_8674_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8674_cast_fp16 = einsum(equation = var_8674_equation_0, values = (var_8558_cast_fp16_17, var_8636_cast_fp16))[name = tensor("op_8674_cast_fp16")]; + tensor var_8676_equation_0 = const()[name = tensor("op_8676_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8676_cast_fp16 = einsum(equation = var_8676_equation_0, values = (var_8558_cast_fp16_18, var_8637_cast_fp16))[name = tensor("op_8676_cast_fp16")]; + tensor var_8678_equation_0 = const()[name = tensor("op_8678_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_8678_cast_fp16 = einsum(equation = var_8678_equation_0, values = (var_8558_cast_fp16_19, var_8638_cast_fp16))[name = tensor("op_8678_cast_fp16")]; + tensor input_315_interleave_0 = const()[name = tensor("input_315_interleave_0"), val = tensor(false)]; + tensor input_315_cast_fp16 = concat(axis = var_8463, interleave = input_315_interleave_0, values = (var_8640_cast_fp16, var_8642_cast_fp16, var_8644_cast_fp16, var_8646_cast_fp16, var_8648_cast_fp16, var_8650_cast_fp16, var_8652_cast_fp16, var_8654_cast_fp16, var_8656_cast_fp16, var_8658_cast_fp16, var_8660_cast_fp16, var_8662_cast_fp16, var_8664_cast_fp16, var_8666_cast_fp16, var_8668_cast_fp16, var_8670_cast_fp16, var_8672_cast_fp16, var_8674_cast_fp16, var_8676_cast_fp16, var_8678_cast_fp16))[name = tensor("input_315_cast_fp16")]; + tensor var_8687_pad_type_0 = const()[name = tensor("op_8687_pad_type_0"), val = tensor("valid")]; + tensor var_8687_strides_0 = const()[name = tensor("op_8687_strides_0"), val = tensor([1, 1])]; + tensor var_8687_pad_0 = const()[name = tensor("op_8687_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8687_dilations_0 = const()[name = tensor("op_8687_dilations_0"), val = tensor([1, 1])]; + tensor var_8687_groups_0 = const()[name = tensor("op_8687_groups_0"), val = tensor(1)]; + tensor blocks_31_attn_out_weight_to_fp16 = const()[name = tensor("blocks_31_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244451712)))]; + tensor blocks_31_attn_out_bias_to_fp16 = const()[name = tensor("blocks_31_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247728576)))]; + tensor var_8687_cast_fp16 = conv(bias = blocks_31_attn_out_bias_to_fp16, dilations = var_8687_dilations_0, groups = var_8687_groups_0, pad = var_8687_pad_0, pad_type = var_8687_pad_type_0, strides = var_8687_strides_0, weight = blocks_31_attn_out_weight_to_fp16, x = input_315_cast_fp16)[name = tensor("op_8687_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = var_8687_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor input_317_axes_0 = const()[name = tensor("input_317_axes_0"), val = tensor([1])]; + tensor input_317_gamma_0_to_fp16 = const()[name = tensor("input_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247731200)))]; + tensor input_317_beta_0_to_fp16 = const()[name = tensor("input_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247733824)))]; + tensor var_8697_to_fp16 = const()[name = tensor("op_8697_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = input_317_beta_0_to_fp16, epsilon = var_8697_to_fp16, gamma = input_317_gamma_0_to_fp16, x = inputs_127_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_319_pad_type_0 = const()[name = tensor("input_319_pad_type_0"), val = tensor("valid")]; + tensor input_319_strides_0 = const()[name = tensor("input_319_strides_0"), val = tensor([1, 1])]; + tensor input_319_pad_0 = const()[name = tensor("input_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_319_dilations_0 = const()[name = tensor("input_319_dilations_0"), val = tensor([1, 1])]; + tensor input_319_groups_0 = const()[name = tensor("input_319_groups_0"), val = tensor(1)]; + tensor blocks_31_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247736448)))]; + tensor blocks_31_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260843712)))]; + tensor input_319_cast_fp16 = conv(bias = blocks_31_mlp_0_bias_to_fp16, dilations = input_319_dilations_0, groups = input_319_groups_0, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = input_319_strides_0, weight = blocks_31_mlp_0_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_319_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_8723_pad_type_0 = const()[name = tensor("op_8723_pad_type_0"), val = tensor("valid")]; + tensor var_8723_strides_0 = const()[name = tensor("op_8723_strides_0"), val = tensor([1, 1])]; + tensor var_8723_pad_0 = const()[name = tensor("op_8723_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_8723_dilations_0 = const()[name = tensor("op_8723_dilations_0"), val = tensor([1, 1])]; + tensor var_8723_groups_0 = const()[name = tensor("op_8723_groups_0"), val = tensor(1)]; + tensor blocks_31_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260854016)))]; + tensor blocks_31_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273961280)))]; + tensor var_8723_cast_fp16 = conv(bias = blocks_31_mlp_2_bias_to_fp16, dilations = var_8723_dilations_0, groups = var_8723_groups_0, pad = var_8723_pad_0, pad_type = var_8723_pad_type_0, strides = var_8723_strides_0, weight = blocks_31_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_8723_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_8723_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273963904)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273966528)))]; + tensor var_8737_to_fp16 = const()[name = tensor("op_8737_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_8737_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_8748_axes_0 = const()[name = tensor("op_8748_axes_0"), val = tensor([2])]; + tensor var_8748_cast_fp16 = squeeze(axes = var_8748_axes_0, x = x_cast_fp16)[name = tensor("op_8748_cast_fp16")]; + tensor var_8751_perm_0 = const()[name = tensor("op_8751_perm_0"), val = tensor([0, 2, 1])]; + tensor var_8751_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8751_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_8751_cast_fp16 = transpose(perm = var_8751_perm_0, x = var_8748_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_8751_cast_fp16_to_fp32_dtype_0, x = var_8751_cast_fp16)[name = tensor("cast_131")]; + } -> (output); +} \ No newline at end of file diff --git a/large-v3/ggml-large-v3-encoder.mlmodelc/weights/weight.bin b/large-v3/ggml-large-v3-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d3c319e2907a1d272802edc0414226f7dd4697fd --- /dev/null +++ b/large-v3/ggml-large-v3-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:806bd7aef5df068fea795e1af8a671cf8817f42f5179e3624632bfbbcbad869f +size 1273969152 diff --git a/large-v3/ggml-large-v3.bin b/large-v3/ggml-large-v3.bin new file mode 100644 index 0000000000000000000000000000000000000000..30488f6b9eeae93e026c978ac7a3190274732ea2 --- /dev/null +++ b/large-v3/ggml-large-v3.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64d182b440b98d5203c4f9bd541544d84c605196c4f7b845dfa11fb23594d1e2 +size 3095033483 diff --git a/medium.en/.DS_Store b/medium.en/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..b8941f18e809afaa65bda9ee38ee3e1f58bc17e0 Binary files /dev/null and b/medium.en/.DS_Store differ diff --git a/medium.en/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin b/medium.en/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..328e5c6707519542579df61365ef4c4391c16353 --- /dev/null +++ b/medium.en/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:423209035d398f1e6945478cc49cf8e83022a3dc4e00d735c54289b3f62cdf89 +size 243 diff --git a/medium.en/ggml-medium.en-encoder.mlmodelc/coremldata.bin b/medium.en/ggml-medium.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a7e2b54475c4dbc86435a318e0608d48945bae16 --- /dev/null +++ b/medium.en/ggml-medium.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d20a4fe17a031efc213c5c295df6967c6e87eba9cca3f07fa63c2beb835ca420 +size 320 diff --git a/medium.en/ggml-medium.en-encoder.mlmodelc/metadata.json b/medium.en/ggml-medium.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..d641992607e37ca4cd71dc6a55c6bba8dd73f78b --- /dev/null +++ b/medium.en/ggml-medium.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1500, 1024]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 24, + "Gelu" : 26, + "LayerNorm" : 49, + "Transpose" : 25, + "Softmax" : 384, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 49, + "Einsum" : 768, + "ExpandDims" : 1, + "Split" : 72, + "Conv" : 146 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.source" : "torch==2.2.2", + "com.github.apple.coremltools.version" : "8.3.0" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_medium_en", + "method" : "predict" + } +] \ No newline at end of file diff --git a/medium.en/ggml-medium.en-encoder.mlmodelc/model.mil b/medium.en/ggml-medium.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..e6dfb1231cc22b9b4af2039b332f526ac8800584 --- /dev/null +++ b/medium.en/ggml-medium.en-encoder.mlmodelc/model.mil @@ -0,0 +1,3763 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_68_pad_type_0 = const()[name = tensor("op_68_pad_type_0"), val = tensor("custom")]; + tensor var_68_pad_0 = const()[name = tensor("op_68_pad_0"), val = tensor([1, 1])]; + tensor var_68_strides_0 = const()[name = tensor("op_68_strides_0"), val = tensor([1])]; + tensor var_68_dilations_0 = const()[name = tensor("op_68_dilations_0"), val = tensor([1])]; + tensor var_68_groups_0 = const()[name = tensor("op_68_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491648)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_100")]; + tensor var_68_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_68_dilations_0, groups = var_68_groups_0, pad = var_68_pad_0, pad_type = var_68_pad_type_0, strides = var_68_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_68_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_68_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_86_pad_type_0 = const()[name = tensor("op_86_pad_type_0"), val = tensor("custom")]; + tensor var_86_pad_0 = const()[name = tensor("op_86_pad_0"), val = tensor([1, 1])]; + tensor var_86_strides_0 = const()[name = tensor("op_86_strides_0"), val = tensor([2])]; + tensor var_86_dilations_0 = const()[name = tensor("op_86_dilations_0"), val = tensor([1])]; + tensor var_86_groups_0 = const()[name = tensor("op_86_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493760)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6785280)))]; + tensor var_86_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_86_dilations_0, groups = var_86_groups_0, pad = var_86_pad_0, pad_type = var_86_pad_type_0, strides = var_86_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_86_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_86_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_91_to_fp16 = const()[name = tensor("op_91_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6787392)))]; + tensor var_93_cast_fp16 = add(x = x_3_cast_fp16, y = var_91_to_fp16)[name = tensor("op_93_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_93_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_108 = const()[name = tensor("op_108"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9859456)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9861568)))]; + tensor var_124_to_fp16 = const()[name = tensor("op_124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_124_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_159_weight_0_to_fp16 = const()[name = tensor("op_159_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9863680)))]; + tensor var_159_bias_0_to_fp16 = const()[name = tensor("op_159_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11960896)))]; + tensor var_159_cast_fp16 = conv(bias = var_159_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_159_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_159_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11963008)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_157_pad_type_0 = const()[name = tensor("op_157_pad_type_0"), val = tensor("valid")]; + tensor var_157_strides_0 = const()[name = tensor("op_157_strides_0"), val = tensor([1, 1])]; + tensor var_157_pad_0 = const()[name = tensor("op_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_157_dilations_0 = const()[name = tensor("op_157_dilations_0"), val = tensor([1, 1])]; + tensor var_157_groups_0 = const()[name = tensor("op_157_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14060224)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16157440)))]; + tensor var_157_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_157_dilations_0, groups = var_157_groups_0, pad = var_157_pad_0, pad_type = var_157_pad_type_0, strides = var_157_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_157_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_160_axis_0 = const()[name = tensor("op_160_axis_0"), val = tensor(1)]; + tensor var_160_cast_fp16_0, tensor var_160_cast_fp16_1, tensor var_160_cast_fp16_2, tensor var_160_cast_fp16_3, tensor var_160_cast_fp16_4, tensor var_160_cast_fp16_5, tensor var_160_cast_fp16_6, tensor var_160_cast_fp16_7, tensor var_160_cast_fp16_8, tensor var_160_cast_fp16_9, tensor var_160_cast_fp16_10, tensor var_160_cast_fp16_11, tensor var_160_cast_fp16_12, tensor var_160_cast_fp16_13, tensor var_160_cast_fp16_14, tensor var_160_cast_fp16_15 = split(axis = var_160_axis_0, split_sizes = tile_0, x = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_177_perm_0 = const()[name = tensor("op_177_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_178_axis_0 = const()[name = tensor("op_178_axis_0"), val = tensor(3)]; + tensor var_177_cast_fp16 = transpose(perm = var_177_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_24")]; + tensor var_178_cast_fp16_0, tensor var_178_cast_fp16_1, tensor var_178_cast_fp16_2, tensor var_178_cast_fp16_3, tensor var_178_cast_fp16_4, tensor var_178_cast_fp16_5, tensor var_178_cast_fp16_6, tensor var_178_cast_fp16_7, tensor var_178_cast_fp16_8, tensor var_178_cast_fp16_9, tensor var_178_cast_fp16_10, tensor var_178_cast_fp16_11, tensor var_178_cast_fp16_12, tensor var_178_cast_fp16_13, tensor var_178_cast_fp16_14, tensor var_178_cast_fp16_15 = split(axis = var_178_axis_0, split_sizes = tile_1, x = var_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_195_axis_0 = const()[name = tensor("op_195_axis_0"), val = tensor(1)]; + tensor var_195_cast_fp16_0, tensor var_195_cast_fp16_1, tensor var_195_cast_fp16_2, tensor var_195_cast_fp16_3, tensor var_195_cast_fp16_4, tensor var_195_cast_fp16_5, tensor var_195_cast_fp16_6, tensor var_195_cast_fp16_7, tensor var_195_cast_fp16_8, tensor var_195_cast_fp16_9, tensor var_195_cast_fp16_10, tensor var_195_cast_fp16_11, tensor var_195_cast_fp16_12, tensor var_195_cast_fp16_13, tensor var_195_cast_fp16_14, tensor var_195_cast_fp16_15 = split(axis = var_195_axis_0, split_sizes = tile_2, x = var_157_cast_fp16)[name = tensor("op_195_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_178_cast_fp16_0, var_160_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_178_cast_fp16_1, var_160_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_178_cast_fp16_2, var_160_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_178_cast_fp16_3, var_160_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_178_cast_fp16_4, var_160_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_178_cast_fp16_5, var_160_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_178_cast_fp16_6, var_160_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_178_cast_fp16_7, var_160_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_178_cast_fp16_8, var_160_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_178_cast_fp16_9, var_160_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_178_cast_fp16_10, var_160_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_178_cast_fp16_11, var_160_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_178_cast_fp16_12, var_160_cast_fp16_12))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_178_cast_fp16_13, var_160_cast_fp16_13))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_178_cast_fp16_14, var_160_cast_fp16_14))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_178_cast_fp16_15, var_160_cast_fp16_15))[name = tensor("aw_31_cast_fp16")]; + tensor var_244_cast_fp16 = softmax(axis = var_108, x = aw_1_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor var_245_cast_fp16 = softmax(axis = var_108, x = aw_3_cast_fp16)[name = tensor("op_245_cast_fp16")]; + tensor var_246_cast_fp16 = softmax(axis = var_108, x = aw_5_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = softmax(axis = var_108, x = aw_7_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor var_248_cast_fp16 = softmax(axis = var_108, x = aw_9_cast_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_249_cast_fp16 = softmax(axis = var_108, x = aw_11_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor var_250_cast_fp16 = softmax(axis = var_108, x = aw_13_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = softmax(axis = var_108, x = aw_15_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor var_252_cast_fp16 = softmax(axis = var_108, x = aw_17_cast_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_cast_fp16 = softmax(axis = var_108, x = aw_19_cast_fp16)[name = tensor("op_253_cast_fp16")]; + tensor var_254_cast_fp16 = softmax(axis = var_108, x = aw_21_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = softmax(axis = var_108, x = aw_23_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_cast_fp16 = softmax(axis = var_108, x = aw_25_cast_fp16)[name = tensor("op_256_cast_fp16")]; + tensor var_257_cast_fp16 = softmax(axis = var_108, x = aw_27_cast_fp16)[name = tensor("op_257_cast_fp16")]; + tensor var_258_cast_fp16 = softmax(axis = var_108, x = aw_29_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = softmax(axis = var_108, x = aw_31_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor var_261_equation_0 = const()[name = tensor("op_261_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_261_cast_fp16 = einsum(equation = var_261_equation_0, values = (var_195_cast_fp16_0, var_244_cast_fp16))[name = tensor("op_261_cast_fp16")]; + tensor var_263_equation_0 = const()[name = tensor("op_263_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_263_cast_fp16 = einsum(equation = var_263_equation_0, values = (var_195_cast_fp16_1, var_245_cast_fp16))[name = tensor("op_263_cast_fp16")]; + tensor var_265_equation_0 = const()[name = tensor("op_265_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_265_cast_fp16 = einsum(equation = var_265_equation_0, values = (var_195_cast_fp16_2, var_246_cast_fp16))[name = tensor("op_265_cast_fp16")]; + tensor var_267_equation_0 = const()[name = tensor("op_267_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_267_cast_fp16 = einsum(equation = var_267_equation_0, values = (var_195_cast_fp16_3, var_247_cast_fp16))[name = tensor("op_267_cast_fp16")]; + tensor var_269_equation_0 = const()[name = tensor("op_269_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_269_cast_fp16 = einsum(equation = var_269_equation_0, values = (var_195_cast_fp16_4, var_248_cast_fp16))[name = tensor("op_269_cast_fp16")]; + tensor var_271_equation_0 = const()[name = tensor("op_271_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_271_cast_fp16 = einsum(equation = var_271_equation_0, values = (var_195_cast_fp16_5, var_249_cast_fp16))[name = tensor("op_271_cast_fp16")]; + tensor var_273_equation_0 = const()[name = tensor("op_273_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_273_cast_fp16 = einsum(equation = var_273_equation_0, values = (var_195_cast_fp16_6, var_250_cast_fp16))[name = tensor("op_273_cast_fp16")]; + tensor var_275_equation_0 = const()[name = tensor("op_275_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_275_cast_fp16 = einsum(equation = var_275_equation_0, values = (var_195_cast_fp16_7, var_251_cast_fp16))[name = tensor("op_275_cast_fp16")]; + tensor var_277_equation_0 = const()[name = tensor("op_277_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_277_cast_fp16 = einsum(equation = var_277_equation_0, values = (var_195_cast_fp16_8, var_252_cast_fp16))[name = tensor("op_277_cast_fp16")]; + tensor var_279_equation_0 = const()[name = tensor("op_279_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_279_cast_fp16 = einsum(equation = var_279_equation_0, values = (var_195_cast_fp16_9, var_253_cast_fp16))[name = tensor("op_279_cast_fp16")]; + tensor var_281_equation_0 = const()[name = tensor("op_281_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_281_cast_fp16 = einsum(equation = var_281_equation_0, values = (var_195_cast_fp16_10, var_254_cast_fp16))[name = tensor("op_281_cast_fp16")]; + tensor var_283_equation_0 = const()[name = tensor("op_283_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_283_cast_fp16 = einsum(equation = var_283_equation_0, values = (var_195_cast_fp16_11, var_255_cast_fp16))[name = tensor("op_283_cast_fp16")]; + tensor var_285_equation_0 = const()[name = tensor("op_285_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_285_cast_fp16 = einsum(equation = var_285_equation_0, values = (var_195_cast_fp16_12, var_256_cast_fp16))[name = tensor("op_285_cast_fp16")]; + tensor var_287_equation_0 = const()[name = tensor("op_287_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_287_cast_fp16 = einsum(equation = var_287_equation_0, values = (var_195_cast_fp16_13, var_257_cast_fp16))[name = tensor("op_287_cast_fp16")]; + tensor var_289_equation_0 = const()[name = tensor("op_289_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_289_cast_fp16 = einsum(equation = var_289_equation_0, values = (var_195_cast_fp16_14, var_258_cast_fp16))[name = tensor("op_289_cast_fp16")]; + tensor var_291_equation_0 = const()[name = tensor("op_291_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_291_cast_fp16 = einsum(equation = var_291_equation_0, values = (var_195_cast_fp16_15, var_259_cast_fp16))[name = tensor("op_291_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_108, interleave = input_5_interleave_0, values = (var_261_cast_fp16, var_263_cast_fp16, var_265_cast_fp16, var_267_cast_fp16, var_269_cast_fp16, var_271_cast_fp16, var_273_cast_fp16, var_275_cast_fp16, var_277_cast_fp16, var_279_cast_fp16, var_281_cast_fp16, var_283_cast_fp16, var_285_cast_fp16, var_287_cast_fp16, var_289_cast_fp16, var_291_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_300_pad_type_0 = const()[name = tensor("op_300_pad_type_0"), val = tensor("valid")]; + tensor var_300_strides_0 = const()[name = tensor("op_300_strides_0"), val = tensor([1, 1])]; + tensor var_300_pad_0 = const()[name = tensor("op_300_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_300_dilations_0 = const()[name = tensor("op_300_dilations_0"), val = tensor([1, 1])]; + tensor var_300_groups_0 = const()[name = tensor("op_300_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16159552)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18256768)))]; + tensor var_300_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_300_dilations_0, groups = var_300_groups_0, pad = var_300_pad_0, pad_type = var_300_pad_type_0, strides = var_300_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_300_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_300_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18258880)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18260992)))]; + tensor var_310_to_fp16 = const()[name = tensor("op_310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_310_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18263104)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26651776)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_336_pad_type_0 = const()[name = tensor("op_336_pad_type_0"), val = tensor("valid")]; + tensor var_336_strides_0 = const()[name = tensor("op_336_strides_0"), val = tensor([1, 1])]; + tensor var_336_pad_0 = const()[name = tensor("op_336_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_336_dilations_0 = const()[name = tensor("op_336_dilations_0"), val = tensor([1, 1])]; + tensor var_336_groups_0 = const()[name = tensor("op_336_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26660032)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35048704)))]; + tensor var_336_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_336_dilations_0, groups = var_336_groups_0, pad = var_336_pad_0, pad_type = var_336_pad_type_0, strides = var_336_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_336_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35050816)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35052928)))]; + tensor var_361_to_fp16 = const()[name = tensor("op_361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_361_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_396_weight_0_to_fp16 = const()[name = tensor("op_396_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35055040)))]; + tensor var_396_bias_0_to_fp16 = const()[name = tensor("op_396_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37152256)))]; + tensor var_396_cast_fp16 = conv(bias = var_396_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_396_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_396_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37154368)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_394_pad_type_0 = const()[name = tensor("op_394_pad_type_0"), val = tensor("valid")]; + tensor var_394_strides_0 = const()[name = tensor("op_394_strides_0"), val = tensor([1, 1])]; + tensor var_394_pad_0 = const()[name = tensor("op_394_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_394_dilations_0 = const()[name = tensor("op_394_dilations_0"), val = tensor([1, 1])]; + tensor var_394_groups_0 = const()[name = tensor("op_394_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39251584)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41348800)))]; + tensor var_394_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_394_dilations_0, groups = var_394_groups_0, pad = var_394_pad_0, pad_type = var_394_pad_type_0, strides = var_394_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_397_axis_0 = const()[name = tensor("op_397_axis_0"), val = tensor(1)]; + tensor var_397_cast_fp16_0, tensor var_397_cast_fp16_1, tensor var_397_cast_fp16_2, tensor var_397_cast_fp16_3, tensor var_397_cast_fp16_4, tensor var_397_cast_fp16_5, tensor var_397_cast_fp16_6, tensor var_397_cast_fp16_7, tensor var_397_cast_fp16_8, tensor var_397_cast_fp16_9, tensor var_397_cast_fp16_10, tensor var_397_cast_fp16_11, tensor var_397_cast_fp16_12, tensor var_397_cast_fp16_13, tensor var_397_cast_fp16_14, tensor var_397_cast_fp16_15 = split(axis = var_397_axis_0, split_sizes = tile_3, x = var_396_cast_fp16)[name = tensor("op_397_cast_fp16")]; + tensor var_414_perm_0 = const()[name = tensor("op_414_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_415_axis_0 = const()[name = tensor("op_415_axis_0"), val = tensor(3)]; + tensor var_414_cast_fp16 = transpose(perm = var_414_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_23")]; + tensor var_415_cast_fp16_0, tensor var_415_cast_fp16_1, tensor var_415_cast_fp16_2, tensor var_415_cast_fp16_3, tensor var_415_cast_fp16_4, tensor var_415_cast_fp16_5, tensor var_415_cast_fp16_6, tensor var_415_cast_fp16_7, tensor var_415_cast_fp16_8, tensor var_415_cast_fp16_9, tensor var_415_cast_fp16_10, tensor var_415_cast_fp16_11, tensor var_415_cast_fp16_12, tensor var_415_cast_fp16_13, tensor var_415_cast_fp16_14, tensor var_415_cast_fp16_15 = split(axis = var_415_axis_0, split_sizes = tile_4, x = var_414_cast_fp16)[name = tensor("op_415_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_432_axis_0 = const()[name = tensor("op_432_axis_0"), val = tensor(1)]; + tensor var_432_cast_fp16_0, tensor var_432_cast_fp16_1, tensor var_432_cast_fp16_2, tensor var_432_cast_fp16_3, tensor var_432_cast_fp16_4, tensor var_432_cast_fp16_5, tensor var_432_cast_fp16_6, tensor var_432_cast_fp16_7, tensor var_432_cast_fp16_8, tensor var_432_cast_fp16_9, tensor var_432_cast_fp16_10, tensor var_432_cast_fp16_11, tensor var_432_cast_fp16_12, tensor var_432_cast_fp16_13, tensor var_432_cast_fp16_14, tensor var_432_cast_fp16_15 = split(axis = var_432_axis_0, split_sizes = tile_5, x = var_394_cast_fp16)[name = tensor("op_432_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_415_cast_fp16_0, var_397_cast_fp16_0))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_415_cast_fp16_1, var_397_cast_fp16_1))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_415_cast_fp16_2, var_397_cast_fp16_2))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_415_cast_fp16_3, var_397_cast_fp16_3))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_415_cast_fp16_4, var_397_cast_fp16_4))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_415_cast_fp16_5, var_397_cast_fp16_5))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_415_cast_fp16_6, var_397_cast_fp16_6))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_415_cast_fp16_7, var_397_cast_fp16_7))[name = tensor("aw_47_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_415_cast_fp16_8, var_397_cast_fp16_8))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_415_cast_fp16_9, var_397_cast_fp16_9))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_415_cast_fp16_10, var_397_cast_fp16_10))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_415_cast_fp16_11, var_397_cast_fp16_11))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_415_cast_fp16_12, var_397_cast_fp16_12))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_415_cast_fp16_13, var_397_cast_fp16_13))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_415_cast_fp16_14, var_397_cast_fp16_14))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_415_cast_fp16_15, var_397_cast_fp16_15))[name = tensor("aw_63_cast_fp16")]; + tensor var_481_cast_fp16 = softmax(axis = var_345, x = aw_33_cast_fp16)[name = tensor("op_481_cast_fp16")]; + tensor var_482_cast_fp16 = softmax(axis = var_345, x = aw_35_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor var_483_cast_fp16 = softmax(axis = var_345, x = aw_37_cast_fp16)[name = tensor("op_483_cast_fp16")]; + tensor var_484_cast_fp16 = softmax(axis = var_345, x = aw_39_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_485_cast_fp16 = softmax(axis = var_345, x = aw_41_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor var_486_cast_fp16 = softmax(axis = var_345, x = aw_43_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_487_cast_fp16 = softmax(axis = var_345, x = aw_45_cast_fp16)[name = tensor("op_487_cast_fp16")]; + tensor var_488_cast_fp16 = softmax(axis = var_345, x = aw_47_cast_fp16)[name = tensor("op_488_cast_fp16")]; + tensor var_489_cast_fp16 = softmax(axis = var_345, x = aw_49_cast_fp16)[name = tensor("op_489_cast_fp16")]; + tensor var_490_cast_fp16 = softmax(axis = var_345, x = aw_51_cast_fp16)[name = tensor("op_490_cast_fp16")]; + tensor var_491_cast_fp16 = softmax(axis = var_345, x = aw_53_cast_fp16)[name = tensor("op_491_cast_fp16")]; + tensor var_492_cast_fp16 = softmax(axis = var_345, x = aw_55_cast_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_493_cast_fp16 = softmax(axis = var_345, x = aw_57_cast_fp16)[name = tensor("op_493_cast_fp16")]; + tensor var_494_cast_fp16 = softmax(axis = var_345, x = aw_59_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor var_495_cast_fp16 = softmax(axis = var_345, x = aw_61_cast_fp16)[name = tensor("op_495_cast_fp16")]; + tensor var_496_cast_fp16 = softmax(axis = var_345, x = aw_63_cast_fp16)[name = tensor("op_496_cast_fp16")]; + tensor var_498_equation_0 = const()[name = tensor("op_498_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_498_cast_fp16 = einsum(equation = var_498_equation_0, values = (var_432_cast_fp16_0, var_481_cast_fp16))[name = tensor("op_498_cast_fp16")]; + tensor var_500_equation_0 = const()[name = tensor("op_500_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_500_cast_fp16 = einsum(equation = var_500_equation_0, values = (var_432_cast_fp16_1, var_482_cast_fp16))[name = tensor("op_500_cast_fp16")]; + tensor var_502_equation_0 = const()[name = tensor("op_502_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_502_cast_fp16 = einsum(equation = var_502_equation_0, values = (var_432_cast_fp16_2, var_483_cast_fp16))[name = tensor("op_502_cast_fp16")]; + tensor var_504_equation_0 = const()[name = tensor("op_504_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_504_cast_fp16 = einsum(equation = var_504_equation_0, values = (var_432_cast_fp16_3, var_484_cast_fp16))[name = tensor("op_504_cast_fp16")]; + tensor var_506_equation_0 = const()[name = tensor("op_506_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_506_cast_fp16 = einsum(equation = var_506_equation_0, values = (var_432_cast_fp16_4, var_485_cast_fp16))[name = tensor("op_506_cast_fp16")]; + tensor var_508_equation_0 = const()[name = tensor("op_508_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_508_cast_fp16 = einsum(equation = var_508_equation_0, values = (var_432_cast_fp16_5, var_486_cast_fp16))[name = tensor("op_508_cast_fp16")]; + tensor var_510_equation_0 = const()[name = tensor("op_510_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_510_cast_fp16 = einsum(equation = var_510_equation_0, values = (var_432_cast_fp16_6, var_487_cast_fp16))[name = tensor("op_510_cast_fp16")]; + tensor var_512_equation_0 = const()[name = tensor("op_512_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_512_cast_fp16 = einsum(equation = var_512_equation_0, values = (var_432_cast_fp16_7, var_488_cast_fp16))[name = tensor("op_512_cast_fp16")]; + tensor var_514_equation_0 = const()[name = tensor("op_514_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_514_cast_fp16 = einsum(equation = var_514_equation_0, values = (var_432_cast_fp16_8, var_489_cast_fp16))[name = tensor("op_514_cast_fp16")]; + tensor var_516_equation_0 = const()[name = tensor("op_516_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_516_cast_fp16 = einsum(equation = var_516_equation_0, values = (var_432_cast_fp16_9, var_490_cast_fp16))[name = tensor("op_516_cast_fp16")]; + tensor var_518_equation_0 = const()[name = tensor("op_518_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_518_cast_fp16 = einsum(equation = var_518_equation_0, values = (var_432_cast_fp16_10, var_491_cast_fp16))[name = tensor("op_518_cast_fp16")]; + tensor var_520_equation_0 = const()[name = tensor("op_520_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_520_cast_fp16 = einsum(equation = var_520_equation_0, values = (var_432_cast_fp16_11, var_492_cast_fp16))[name = tensor("op_520_cast_fp16")]; + tensor var_522_equation_0 = const()[name = tensor("op_522_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_522_cast_fp16 = einsum(equation = var_522_equation_0, values = (var_432_cast_fp16_12, var_493_cast_fp16))[name = tensor("op_522_cast_fp16")]; + tensor var_524_equation_0 = const()[name = tensor("op_524_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_524_cast_fp16 = einsum(equation = var_524_equation_0, values = (var_432_cast_fp16_13, var_494_cast_fp16))[name = tensor("op_524_cast_fp16")]; + tensor var_526_equation_0 = const()[name = tensor("op_526_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_526_cast_fp16 = einsum(equation = var_526_equation_0, values = (var_432_cast_fp16_14, var_495_cast_fp16))[name = tensor("op_526_cast_fp16")]; + tensor var_528_equation_0 = const()[name = tensor("op_528_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_528_cast_fp16 = einsum(equation = var_528_equation_0, values = (var_432_cast_fp16_15, var_496_cast_fp16))[name = tensor("op_528_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_345, interleave = input_15_interleave_0, values = (var_498_cast_fp16, var_500_cast_fp16, var_502_cast_fp16, var_504_cast_fp16, var_506_cast_fp16, var_508_cast_fp16, var_510_cast_fp16, var_512_cast_fp16, var_514_cast_fp16, var_516_cast_fp16, var_518_cast_fp16, var_520_cast_fp16, var_522_cast_fp16, var_524_cast_fp16, var_526_cast_fp16, var_528_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_537_pad_type_0 = const()[name = tensor("op_537_pad_type_0"), val = tensor("valid")]; + tensor var_537_strides_0 = const()[name = tensor("op_537_strides_0"), val = tensor([1, 1])]; + tensor var_537_pad_0 = const()[name = tensor("op_537_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_537_dilations_0 = const()[name = tensor("op_537_dilations_0"), val = tensor([1, 1])]; + tensor var_537_groups_0 = const()[name = tensor("op_537_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41350912)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43448128)))]; + tensor var_537_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_537_dilations_0, groups = var_537_groups_0, pad = var_537_pad_0, pad_type = var_537_pad_type_0, strides = var_537_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_537_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_537_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43450240)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43452352)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_547_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43454464)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51843136)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_573_pad_type_0 = const()[name = tensor("op_573_pad_type_0"), val = tensor("valid")]; + tensor var_573_strides_0 = const()[name = tensor("op_573_strides_0"), val = tensor([1, 1])]; + tensor var_573_pad_0 = const()[name = tensor("op_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_573_dilations_0 = const()[name = tensor("op_573_dilations_0"), val = tensor([1, 1])]; + tensor var_573_groups_0 = const()[name = tensor("op_573_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51851392)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60240064)))]; + tensor var_573_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_573_dilations_0, groups = var_573_groups_0, pad = var_573_pad_0, pad_type = var_573_pad_type_0, strides = var_573_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_573_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_582 = const()[name = tensor("op_582"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60242176)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60244288)))]; + tensor var_598_to_fp16 = const()[name = tensor("op_598_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_598_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_633_weight_0_to_fp16 = const()[name = tensor("op_633_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60246400)))]; + tensor var_633_bias_0_to_fp16 = const()[name = tensor("op_633_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62343616)))]; + tensor var_633_cast_fp16 = conv(bias = var_633_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_633_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_633_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62345728)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_631_pad_type_0 = const()[name = tensor("op_631_pad_type_0"), val = tensor("valid")]; + tensor var_631_strides_0 = const()[name = tensor("op_631_strides_0"), val = tensor([1, 1])]; + tensor var_631_pad_0 = const()[name = tensor("op_631_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_631_dilations_0 = const()[name = tensor("op_631_dilations_0"), val = tensor([1, 1])]; + tensor var_631_groups_0 = const()[name = tensor("op_631_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64442944)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66540160)))]; + tensor var_631_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_631_dilations_0, groups = var_631_groups_0, pad = var_631_pad_0, pad_type = var_631_pad_type_0, strides = var_631_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_634_axis_0 = const()[name = tensor("op_634_axis_0"), val = tensor(1)]; + tensor var_634_cast_fp16_0, tensor var_634_cast_fp16_1, tensor var_634_cast_fp16_2, tensor var_634_cast_fp16_3, tensor var_634_cast_fp16_4, tensor var_634_cast_fp16_5, tensor var_634_cast_fp16_6, tensor var_634_cast_fp16_7, tensor var_634_cast_fp16_8, tensor var_634_cast_fp16_9, tensor var_634_cast_fp16_10, tensor var_634_cast_fp16_11, tensor var_634_cast_fp16_12, tensor var_634_cast_fp16_13, tensor var_634_cast_fp16_14, tensor var_634_cast_fp16_15 = split(axis = var_634_axis_0, split_sizes = tile_6, x = var_633_cast_fp16)[name = tensor("op_634_cast_fp16")]; + tensor var_651_perm_0 = const()[name = tensor("op_651_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(3)]; + tensor var_651_cast_fp16 = transpose(perm = var_651_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_22")]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2, tensor var_652_cast_fp16_3, tensor var_652_cast_fp16_4, tensor var_652_cast_fp16_5, tensor var_652_cast_fp16_6, tensor var_652_cast_fp16_7, tensor var_652_cast_fp16_8, tensor var_652_cast_fp16_9, tensor var_652_cast_fp16_10, tensor var_652_cast_fp16_11, tensor var_652_cast_fp16_12, tensor var_652_cast_fp16_13, tensor var_652_cast_fp16_14, tensor var_652_cast_fp16_15 = split(axis = var_652_axis_0, split_sizes = tile_7, x = var_651_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_669_axis_0 = const()[name = tensor("op_669_axis_0"), val = tensor(1)]; + tensor var_669_cast_fp16_0, tensor var_669_cast_fp16_1, tensor var_669_cast_fp16_2, tensor var_669_cast_fp16_3, tensor var_669_cast_fp16_4, tensor var_669_cast_fp16_5, tensor var_669_cast_fp16_6, tensor var_669_cast_fp16_7, tensor var_669_cast_fp16_8, tensor var_669_cast_fp16_9, tensor var_669_cast_fp16_10, tensor var_669_cast_fp16_11, tensor var_669_cast_fp16_12, tensor var_669_cast_fp16_13, tensor var_669_cast_fp16_14, tensor var_669_cast_fp16_15 = split(axis = var_669_axis_0, split_sizes = tile_8, x = var_631_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_652_cast_fp16_0, var_634_cast_fp16_0))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_652_cast_fp16_1, var_634_cast_fp16_1))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_652_cast_fp16_2, var_634_cast_fp16_2))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_652_cast_fp16_3, var_634_cast_fp16_3))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_652_cast_fp16_4, var_634_cast_fp16_4))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_652_cast_fp16_5, var_634_cast_fp16_5))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_652_cast_fp16_6, var_634_cast_fp16_6))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_652_cast_fp16_7, var_634_cast_fp16_7))[name = tensor("aw_79_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_652_cast_fp16_8, var_634_cast_fp16_8))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_652_cast_fp16_9, var_634_cast_fp16_9))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_652_cast_fp16_10, var_634_cast_fp16_10))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_652_cast_fp16_11, var_634_cast_fp16_11))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_652_cast_fp16_12, var_634_cast_fp16_12))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_652_cast_fp16_13, var_634_cast_fp16_13))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_652_cast_fp16_14, var_634_cast_fp16_14))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_652_cast_fp16_15, var_634_cast_fp16_15))[name = tensor("aw_95_cast_fp16")]; + tensor var_718_cast_fp16 = softmax(axis = var_582, x = aw_65_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor var_719_cast_fp16 = softmax(axis = var_582, x = aw_67_cast_fp16)[name = tensor("op_719_cast_fp16")]; + tensor var_720_cast_fp16 = softmax(axis = var_582, x = aw_69_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor var_721_cast_fp16 = softmax(axis = var_582, x = aw_71_cast_fp16)[name = tensor("op_721_cast_fp16")]; + tensor var_722_cast_fp16 = softmax(axis = var_582, x = aw_73_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = softmax(axis = var_582, x = aw_75_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor var_724_cast_fp16 = softmax(axis = var_582, x = aw_77_cast_fp16)[name = tensor("op_724_cast_fp16")]; + tensor var_725_cast_fp16 = softmax(axis = var_582, x = aw_79_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor var_726_cast_fp16 = softmax(axis = var_582, x = aw_81_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = softmax(axis = var_582, x = aw_83_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor var_728_cast_fp16 = softmax(axis = var_582, x = aw_85_cast_fp16)[name = tensor("op_728_cast_fp16")]; + tensor var_729_cast_fp16 = softmax(axis = var_582, x = aw_87_cast_fp16)[name = tensor("op_729_cast_fp16")]; + tensor var_730_cast_fp16 = softmax(axis = var_582, x = aw_89_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = softmax(axis = var_582, x = aw_91_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor var_732_cast_fp16 = softmax(axis = var_582, x = aw_93_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_cast_fp16 = softmax(axis = var_582, x = aw_95_cast_fp16)[name = tensor("op_733_cast_fp16")]; + tensor var_735_equation_0 = const()[name = tensor("op_735_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_735_cast_fp16 = einsum(equation = var_735_equation_0, values = (var_669_cast_fp16_0, var_718_cast_fp16))[name = tensor("op_735_cast_fp16")]; + tensor var_737_equation_0 = const()[name = tensor("op_737_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_737_cast_fp16 = einsum(equation = var_737_equation_0, values = (var_669_cast_fp16_1, var_719_cast_fp16))[name = tensor("op_737_cast_fp16")]; + tensor var_739_equation_0 = const()[name = tensor("op_739_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_739_cast_fp16 = einsum(equation = var_739_equation_0, values = (var_669_cast_fp16_2, var_720_cast_fp16))[name = tensor("op_739_cast_fp16")]; + tensor var_741_equation_0 = const()[name = tensor("op_741_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_741_cast_fp16 = einsum(equation = var_741_equation_0, values = (var_669_cast_fp16_3, var_721_cast_fp16))[name = tensor("op_741_cast_fp16")]; + tensor var_743_equation_0 = const()[name = tensor("op_743_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_743_cast_fp16 = einsum(equation = var_743_equation_0, values = (var_669_cast_fp16_4, var_722_cast_fp16))[name = tensor("op_743_cast_fp16")]; + tensor var_745_equation_0 = const()[name = tensor("op_745_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_745_cast_fp16 = einsum(equation = var_745_equation_0, values = (var_669_cast_fp16_5, var_723_cast_fp16))[name = tensor("op_745_cast_fp16")]; + tensor var_747_equation_0 = const()[name = tensor("op_747_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_747_cast_fp16 = einsum(equation = var_747_equation_0, values = (var_669_cast_fp16_6, var_724_cast_fp16))[name = tensor("op_747_cast_fp16")]; + tensor var_749_equation_0 = const()[name = tensor("op_749_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_749_cast_fp16 = einsum(equation = var_749_equation_0, values = (var_669_cast_fp16_7, var_725_cast_fp16))[name = tensor("op_749_cast_fp16")]; + tensor var_751_equation_0 = const()[name = tensor("op_751_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_751_cast_fp16 = einsum(equation = var_751_equation_0, values = (var_669_cast_fp16_8, var_726_cast_fp16))[name = tensor("op_751_cast_fp16")]; + tensor var_753_equation_0 = const()[name = tensor("op_753_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_753_cast_fp16 = einsum(equation = var_753_equation_0, values = (var_669_cast_fp16_9, var_727_cast_fp16))[name = tensor("op_753_cast_fp16")]; + tensor var_755_equation_0 = const()[name = tensor("op_755_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_755_cast_fp16 = einsum(equation = var_755_equation_0, values = (var_669_cast_fp16_10, var_728_cast_fp16))[name = tensor("op_755_cast_fp16")]; + tensor var_757_equation_0 = const()[name = tensor("op_757_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_757_cast_fp16 = einsum(equation = var_757_equation_0, values = (var_669_cast_fp16_11, var_729_cast_fp16))[name = tensor("op_757_cast_fp16")]; + tensor var_759_equation_0 = const()[name = tensor("op_759_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_759_cast_fp16 = einsum(equation = var_759_equation_0, values = (var_669_cast_fp16_12, var_730_cast_fp16))[name = tensor("op_759_cast_fp16")]; + tensor var_761_equation_0 = const()[name = tensor("op_761_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_761_cast_fp16 = einsum(equation = var_761_equation_0, values = (var_669_cast_fp16_13, var_731_cast_fp16))[name = tensor("op_761_cast_fp16")]; + tensor var_763_equation_0 = const()[name = tensor("op_763_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_763_cast_fp16 = einsum(equation = var_763_equation_0, values = (var_669_cast_fp16_14, var_732_cast_fp16))[name = tensor("op_763_cast_fp16")]; + tensor var_765_equation_0 = const()[name = tensor("op_765_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_765_cast_fp16 = einsum(equation = var_765_equation_0, values = (var_669_cast_fp16_15, var_733_cast_fp16))[name = tensor("op_765_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_582, interleave = input_25_interleave_0, values = (var_735_cast_fp16, var_737_cast_fp16, var_739_cast_fp16, var_741_cast_fp16, var_743_cast_fp16, var_745_cast_fp16, var_747_cast_fp16, var_749_cast_fp16, var_751_cast_fp16, var_753_cast_fp16, var_755_cast_fp16, var_757_cast_fp16, var_759_cast_fp16, var_761_cast_fp16, var_763_cast_fp16, var_765_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_774_pad_type_0 = const()[name = tensor("op_774_pad_type_0"), val = tensor("valid")]; + tensor var_774_strides_0 = const()[name = tensor("op_774_strides_0"), val = tensor([1, 1])]; + tensor var_774_pad_0 = const()[name = tensor("op_774_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_774_dilations_0 = const()[name = tensor("op_774_dilations_0"), val = tensor([1, 1])]; + tensor var_774_groups_0 = const()[name = tensor("op_774_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66542272)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68639488)))]; + tensor var_774_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_774_dilations_0, groups = var_774_groups_0, pad = var_774_pad_0, pad_type = var_774_pad_type_0, strides = var_774_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_774_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68641600)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68643712)))]; + tensor var_784_to_fp16 = const()[name = tensor("op_784_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_784_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68645824)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77034496)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_810_pad_type_0 = const()[name = tensor("op_810_pad_type_0"), val = tensor("valid")]; + tensor var_810_strides_0 = const()[name = tensor("op_810_strides_0"), val = tensor([1, 1])]; + tensor var_810_pad_0 = const()[name = tensor("op_810_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_810_dilations_0 = const()[name = tensor("op_810_dilations_0"), val = tensor([1, 1])]; + tensor var_810_groups_0 = const()[name = tensor("op_810_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77042752)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85431424)))]; + tensor var_810_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_810_dilations_0, groups = var_810_groups_0, pad = var_810_pad_0, pad_type = var_810_pad_type_0, strides = var_810_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_810_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_819 = const()[name = tensor("op_819"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85433536)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85435648)))]; + tensor var_835_to_fp16 = const()[name = tensor("op_835_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_835_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_870_weight_0_to_fp16 = const()[name = tensor("op_870_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85437760)))]; + tensor var_870_bias_0_to_fp16 = const()[name = tensor("op_870_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87534976)))]; + tensor var_870_cast_fp16 = conv(bias = var_870_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_870_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_870_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87537088)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_868_pad_type_0 = const()[name = tensor("op_868_pad_type_0"), val = tensor("valid")]; + tensor var_868_strides_0 = const()[name = tensor("op_868_strides_0"), val = tensor([1, 1])]; + tensor var_868_pad_0 = const()[name = tensor("op_868_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_868_dilations_0 = const()[name = tensor("op_868_dilations_0"), val = tensor([1, 1])]; + tensor var_868_groups_0 = const()[name = tensor("op_868_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89634304)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91731520)))]; + tensor var_868_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_868_dilations_0, groups = var_868_groups_0, pad = var_868_pad_0, pad_type = var_868_pad_type_0, strides = var_868_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_871_axis_0 = const()[name = tensor("op_871_axis_0"), val = tensor(1)]; + tensor var_871_cast_fp16_0, tensor var_871_cast_fp16_1, tensor var_871_cast_fp16_2, tensor var_871_cast_fp16_3, tensor var_871_cast_fp16_4, tensor var_871_cast_fp16_5, tensor var_871_cast_fp16_6, tensor var_871_cast_fp16_7, tensor var_871_cast_fp16_8, tensor var_871_cast_fp16_9, tensor var_871_cast_fp16_10, tensor var_871_cast_fp16_11, tensor var_871_cast_fp16_12, tensor var_871_cast_fp16_13, tensor var_871_cast_fp16_14, tensor var_871_cast_fp16_15 = split(axis = var_871_axis_0, split_sizes = tile_9, x = var_870_cast_fp16)[name = tensor("op_871_cast_fp16")]; + tensor var_888_perm_0 = const()[name = tensor("op_888_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_889_axis_0 = const()[name = tensor("op_889_axis_0"), val = tensor(3)]; + tensor var_888_cast_fp16 = transpose(perm = var_888_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_21")]; + tensor var_889_cast_fp16_0, tensor var_889_cast_fp16_1, tensor var_889_cast_fp16_2, tensor var_889_cast_fp16_3, tensor var_889_cast_fp16_4, tensor var_889_cast_fp16_5, tensor var_889_cast_fp16_6, tensor var_889_cast_fp16_7, tensor var_889_cast_fp16_8, tensor var_889_cast_fp16_9, tensor var_889_cast_fp16_10, tensor var_889_cast_fp16_11, tensor var_889_cast_fp16_12, tensor var_889_cast_fp16_13, tensor var_889_cast_fp16_14, tensor var_889_cast_fp16_15 = split(axis = var_889_axis_0, split_sizes = tile_10, x = var_888_cast_fp16)[name = tensor("op_889_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_906_axis_0 = const()[name = tensor("op_906_axis_0"), val = tensor(1)]; + tensor var_906_cast_fp16_0, tensor var_906_cast_fp16_1, tensor var_906_cast_fp16_2, tensor var_906_cast_fp16_3, tensor var_906_cast_fp16_4, tensor var_906_cast_fp16_5, tensor var_906_cast_fp16_6, tensor var_906_cast_fp16_7, tensor var_906_cast_fp16_8, tensor var_906_cast_fp16_9, tensor var_906_cast_fp16_10, tensor var_906_cast_fp16_11, tensor var_906_cast_fp16_12, tensor var_906_cast_fp16_13, tensor var_906_cast_fp16_14, tensor var_906_cast_fp16_15 = split(axis = var_906_axis_0, split_sizes = tile_11, x = var_868_cast_fp16)[name = tensor("op_906_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_889_cast_fp16_0, var_871_cast_fp16_0))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_889_cast_fp16_1, var_871_cast_fp16_1))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_889_cast_fp16_2, var_871_cast_fp16_2))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_889_cast_fp16_3, var_871_cast_fp16_3))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_889_cast_fp16_4, var_871_cast_fp16_4))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_889_cast_fp16_5, var_871_cast_fp16_5))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_889_cast_fp16_6, var_871_cast_fp16_6))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_889_cast_fp16_7, var_871_cast_fp16_7))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_889_cast_fp16_8, var_871_cast_fp16_8))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_889_cast_fp16_9, var_871_cast_fp16_9))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_889_cast_fp16_10, var_871_cast_fp16_10))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_889_cast_fp16_11, var_871_cast_fp16_11))[name = tensor("aw_119_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_889_cast_fp16_12, var_871_cast_fp16_12))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_889_cast_fp16_13, var_871_cast_fp16_13))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_889_cast_fp16_14, var_871_cast_fp16_14))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_889_cast_fp16_15, var_871_cast_fp16_15))[name = tensor("aw_127_cast_fp16")]; + tensor var_955_cast_fp16 = softmax(axis = var_819, x = aw_97_cast_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_956_cast_fp16 = softmax(axis = var_819, x = aw_99_cast_fp16)[name = tensor("op_956_cast_fp16")]; + tensor var_957_cast_fp16 = softmax(axis = var_819, x = aw_101_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor var_958_cast_fp16 = softmax(axis = var_819, x = aw_103_cast_fp16)[name = tensor("op_958_cast_fp16")]; + tensor var_959_cast_fp16 = softmax(axis = var_819, x = aw_105_cast_fp16)[name = tensor("op_959_cast_fp16")]; + tensor var_960_cast_fp16 = softmax(axis = var_819, x = aw_107_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor var_961_cast_fp16 = softmax(axis = var_819, x = aw_109_cast_fp16)[name = tensor("op_961_cast_fp16")]; + tensor var_962_cast_fp16 = softmax(axis = var_819, x = aw_111_cast_fp16)[name = tensor("op_962_cast_fp16")]; + tensor var_963_cast_fp16 = softmax(axis = var_819, x = aw_113_cast_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_964_cast_fp16 = softmax(axis = var_819, x = aw_115_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = softmax(axis = var_819, x = aw_117_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor var_966_cast_fp16 = softmax(axis = var_819, x = aw_119_cast_fp16)[name = tensor("op_966_cast_fp16")]; + tensor var_967_cast_fp16 = softmax(axis = var_819, x = aw_121_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor var_968_cast_fp16 = softmax(axis = var_819, x = aw_123_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = softmax(axis = var_819, x = aw_125_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor var_970_cast_fp16 = softmax(axis = var_819, x = aw_127_cast_fp16)[name = tensor("op_970_cast_fp16")]; + tensor var_972_equation_0 = const()[name = tensor("op_972_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_972_cast_fp16 = einsum(equation = var_972_equation_0, values = (var_906_cast_fp16_0, var_955_cast_fp16))[name = tensor("op_972_cast_fp16")]; + tensor var_974_equation_0 = const()[name = tensor("op_974_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_974_cast_fp16 = einsum(equation = var_974_equation_0, values = (var_906_cast_fp16_1, var_956_cast_fp16))[name = tensor("op_974_cast_fp16")]; + tensor var_976_equation_0 = const()[name = tensor("op_976_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_976_cast_fp16 = einsum(equation = var_976_equation_0, values = (var_906_cast_fp16_2, var_957_cast_fp16))[name = tensor("op_976_cast_fp16")]; + tensor var_978_equation_0 = const()[name = tensor("op_978_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_978_cast_fp16 = einsum(equation = var_978_equation_0, values = (var_906_cast_fp16_3, var_958_cast_fp16))[name = tensor("op_978_cast_fp16")]; + tensor var_980_equation_0 = const()[name = tensor("op_980_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_980_cast_fp16 = einsum(equation = var_980_equation_0, values = (var_906_cast_fp16_4, var_959_cast_fp16))[name = tensor("op_980_cast_fp16")]; + tensor var_982_equation_0 = const()[name = tensor("op_982_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_982_cast_fp16 = einsum(equation = var_982_equation_0, values = (var_906_cast_fp16_5, var_960_cast_fp16))[name = tensor("op_982_cast_fp16")]; + tensor var_984_equation_0 = const()[name = tensor("op_984_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_984_cast_fp16 = einsum(equation = var_984_equation_0, values = (var_906_cast_fp16_6, var_961_cast_fp16))[name = tensor("op_984_cast_fp16")]; + tensor var_986_equation_0 = const()[name = tensor("op_986_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_986_cast_fp16 = einsum(equation = var_986_equation_0, values = (var_906_cast_fp16_7, var_962_cast_fp16))[name = tensor("op_986_cast_fp16")]; + tensor var_988_equation_0 = const()[name = tensor("op_988_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_988_cast_fp16 = einsum(equation = var_988_equation_0, values = (var_906_cast_fp16_8, var_963_cast_fp16))[name = tensor("op_988_cast_fp16")]; + tensor var_990_equation_0 = const()[name = tensor("op_990_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_990_cast_fp16 = einsum(equation = var_990_equation_0, values = (var_906_cast_fp16_9, var_964_cast_fp16))[name = tensor("op_990_cast_fp16")]; + tensor var_992_equation_0 = const()[name = tensor("op_992_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_992_cast_fp16 = einsum(equation = var_992_equation_0, values = (var_906_cast_fp16_10, var_965_cast_fp16))[name = tensor("op_992_cast_fp16")]; + tensor var_994_equation_0 = const()[name = tensor("op_994_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_994_cast_fp16 = einsum(equation = var_994_equation_0, values = (var_906_cast_fp16_11, var_966_cast_fp16))[name = tensor("op_994_cast_fp16")]; + tensor var_996_equation_0 = const()[name = tensor("op_996_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_996_cast_fp16 = einsum(equation = var_996_equation_0, values = (var_906_cast_fp16_12, var_967_cast_fp16))[name = tensor("op_996_cast_fp16")]; + tensor var_998_equation_0 = const()[name = tensor("op_998_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_998_cast_fp16 = einsum(equation = var_998_equation_0, values = (var_906_cast_fp16_13, var_968_cast_fp16))[name = tensor("op_998_cast_fp16")]; + tensor var_1000_equation_0 = const()[name = tensor("op_1000_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1000_cast_fp16 = einsum(equation = var_1000_equation_0, values = (var_906_cast_fp16_14, var_969_cast_fp16))[name = tensor("op_1000_cast_fp16")]; + tensor var_1002_equation_0 = const()[name = tensor("op_1002_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1002_cast_fp16 = einsum(equation = var_1002_equation_0, values = (var_906_cast_fp16_15, var_970_cast_fp16))[name = tensor("op_1002_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_819, interleave = input_35_interleave_0, values = (var_972_cast_fp16, var_974_cast_fp16, var_976_cast_fp16, var_978_cast_fp16, var_980_cast_fp16, var_982_cast_fp16, var_984_cast_fp16, var_986_cast_fp16, var_988_cast_fp16, var_990_cast_fp16, var_992_cast_fp16, var_994_cast_fp16, var_996_cast_fp16, var_998_cast_fp16, var_1000_cast_fp16, var_1002_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_1011_pad_type_0 = const()[name = tensor("op_1011_pad_type_0"), val = tensor("valid")]; + tensor var_1011_strides_0 = const()[name = tensor("op_1011_strides_0"), val = tensor([1, 1])]; + tensor var_1011_pad_0 = const()[name = tensor("op_1011_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1011_dilations_0 = const()[name = tensor("op_1011_dilations_0"), val = tensor([1, 1])]; + tensor var_1011_groups_0 = const()[name = tensor("op_1011_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91733632)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93830848)))]; + tensor var_1011_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_1011_dilations_0, groups = var_1011_groups_0, pad = var_1011_pad_0, pad_type = var_1011_pad_type_0, strides = var_1011_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_1011_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93832960)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93835072)))]; + tensor var_1021_to_fp16 = const()[name = tensor("op_1021_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_1021_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93837184)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102225856)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1047_pad_type_0 = const()[name = tensor("op_1047_pad_type_0"), val = tensor("valid")]; + tensor var_1047_strides_0 = const()[name = tensor("op_1047_strides_0"), val = tensor([1, 1])]; + tensor var_1047_pad_0 = const()[name = tensor("op_1047_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1047_dilations_0 = const()[name = tensor("op_1047_dilations_0"), val = tensor([1, 1])]; + tensor var_1047_groups_0 = const()[name = tensor("op_1047_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102234112)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110622784)))]; + tensor var_1047_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_1047_dilations_0, groups = var_1047_groups_0, pad = var_1047_pad_0, pad_type = var_1047_pad_type_0, strides = var_1047_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_1047_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_1047_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110624896)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110627008)))]; + tensor var_1072_to_fp16 = const()[name = tensor("op_1072_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_1072_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_1107_weight_0_to_fp16 = const()[name = tensor("op_1107_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110629120)))]; + tensor var_1107_bias_0_to_fp16 = const()[name = tensor("op_1107_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112726336)))]; + tensor var_1107_cast_fp16 = conv(bias = var_1107_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_1107_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1107_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112728448)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_1105_pad_type_0 = const()[name = tensor("op_1105_pad_type_0"), val = tensor("valid")]; + tensor var_1105_strides_0 = const()[name = tensor("op_1105_strides_0"), val = tensor([1, 1])]; + tensor var_1105_pad_0 = const()[name = tensor("op_1105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1105_dilations_0 = const()[name = tensor("op_1105_dilations_0"), val = tensor([1, 1])]; + tensor var_1105_groups_0 = const()[name = tensor("op_1105_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114825664)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116922880)))]; + tensor var_1105_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_1105_dilations_0, groups = var_1105_groups_0, pad = var_1105_pad_0, pad_type = var_1105_pad_type_0, strides = var_1105_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1108_axis_0 = const()[name = tensor("op_1108_axis_0"), val = tensor(1)]; + tensor var_1108_cast_fp16_0, tensor var_1108_cast_fp16_1, tensor var_1108_cast_fp16_2, tensor var_1108_cast_fp16_3, tensor var_1108_cast_fp16_4, tensor var_1108_cast_fp16_5, tensor var_1108_cast_fp16_6, tensor var_1108_cast_fp16_7, tensor var_1108_cast_fp16_8, tensor var_1108_cast_fp16_9, tensor var_1108_cast_fp16_10, tensor var_1108_cast_fp16_11, tensor var_1108_cast_fp16_12, tensor var_1108_cast_fp16_13, tensor var_1108_cast_fp16_14, tensor var_1108_cast_fp16_15 = split(axis = var_1108_axis_0, split_sizes = tile_12, x = var_1107_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1126_axis_0 = const()[name = tensor("op_1126_axis_0"), val = tensor(3)]; + tensor var_1125_cast_fp16 = transpose(perm = var_1125_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_20")]; + tensor var_1126_cast_fp16_0, tensor var_1126_cast_fp16_1, tensor var_1126_cast_fp16_2, tensor var_1126_cast_fp16_3, tensor var_1126_cast_fp16_4, tensor var_1126_cast_fp16_5, tensor var_1126_cast_fp16_6, tensor var_1126_cast_fp16_7, tensor var_1126_cast_fp16_8, tensor var_1126_cast_fp16_9, tensor var_1126_cast_fp16_10, tensor var_1126_cast_fp16_11, tensor var_1126_cast_fp16_12, tensor var_1126_cast_fp16_13, tensor var_1126_cast_fp16_14, tensor var_1126_cast_fp16_15 = split(axis = var_1126_axis_0, split_sizes = tile_13, x = var_1125_cast_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1143_axis_0 = const()[name = tensor("op_1143_axis_0"), val = tensor(1)]; + tensor var_1143_cast_fp16_0, tensor var_1143_cast_fp16_1, tensor var_1143_cast_fp16_2, tensor var_1143_cast_fp16_3, tensor var_1143_cast_fp16_4, tensor var_1143_cast_fp16_5, tensor var_1143_cast_fp16_6, tensor var_1143_cast_fp16_7, tensor var_1143_cast_fp16_8, tensor var_1143_cast_fp16_9, tensor var_1143_cast_fp16_10, tensor var_1143_cast_fp16_11, tensor var_1143_cast_fp16_12, tensor var_1143_cast_fp16_13, tensor var_1143_cast_fp16_14, tensor var_1143_cast_fp16_15 = split(axis = var_1143_axis_0, split_sizes = tile_14, x = var_1105_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1126_cast_fp16_0, var_1108_cast_fp16_0))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1126_cast_fp16_1, var_1108_cast_fp16_1))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1126_cast_fp16_2, var_1108_cast_fp16_2))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1126_cast_fp16_3, var_1108_cast_fp16_3))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1126_cast_fp16_4, var_1108_cast_fp16_4))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1126_cast_fp16_5, var_1108_cast_fp16_5))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1126_cast_fp16_6, var_1108_cast_fp16_6))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1126_cast_fp16_7, var_1108_cast_fp16_7))[name = tensor("aw_143_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1126_cast_fp16_8, var_1108_cast_fp16_8))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1126_cast_fp16_9, var_1108_cast_fp16_9))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1126_cast_fp16_10, var_1108_cast_fp16_10))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1126_cast_fp16_11, var_1108_cast_fp16_11))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1126_cast_fp16_12, var_1108_cast_fp16_12))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1126_cast_fp16_13, var_1108_cast_fp16_13))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1126_cast_fp16_14, var_1108_cast_fp16_14))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1126_cast_fp16_15, var_1108_cast_fp16_15))[name = tensor("aw_159_cast_fp16")]; + tensor var_1192_cast_fp16 = softmax(axis = var_1056, x = aw_129_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor var_1193_cast_fp16 = softmax(axis = var_1056, x = aw_131_cast_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194_cast_fp16 = softmax(axis = var_1056, x = aw_133_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1195_cast_fp16 = softmax(axis = var_1056, x = aw_135_cast_fp16)[name = tensor("op_1195_cast_fp16")]; + tensor var_1196_cast_fp16 = softmax(axis = var_1056, x = aw_137_cast_fp16)[name = tensor("op_1196_cast_fp16")]; + tensor var_1197_cast_fp16 = softmax(axis = var_1056, x = aw_139_cast_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1198_cast_fp16 = softmax(axis = var_1056, x = aw_141_cast_fp16)[name = tensor("op_1198_cast_fp16")]; + tensor var_1199_cast_fp16 = softmax(axis = var_1056, x = aw_143_cast_fp16)[name = tensor("op_1199_cast_fp16")]; + tensor var_1200_cast_fp16 = softmax(axis = var_1056, x = aw_145_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor var_1201_cast_fp16 = softmax(axis = var_1056, x = aw_147_cast_fp16)[name = tensor("op_1201_cast_fp16")]; + tensor var_1202_cast_fp16 = softmax(axis = var_1056, x = aw_149_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1203_cast_fp16 = softmax(axis = var_1056, x = aw_151_cast_fp16)[name = tensor("op_1203_cast_fp16")]; + tensor var_1204_cast_fp16 = softmax(axis = var_1056, x = aw_153_cast_fp16)[name = tensor("op_1204_cast_fp16")]; + tensor var_1205_cast_fp16 = softmax(axis = var_1056, x = aw_155_cast_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1206_cast_fp16 = softmax(axis = var_1056, x = aw_157_cast_fp16)[name = tensor("op_1206_cast_fp16")]; + tensor var_1207_cast_fp16 = softmax(axis = var_1056, x = aw_159_cast_fp16)[name = tensor("op_1207_cast_fp16")]; + tensor var_1209_equation_0 = const()[name = tensor("op_1209_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1209_cast_fp16 = einsum(equation = var_1209_equation_0, values = (var_1143_cast_fp16_0, var_1192_cast_fp16))[name = tensor("op_1209_cast_fp16")]; + tensor var_1211_equation_0 = const()[name = tensor("op_1211_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1211_cast_fp16 = einsum(equation = var_1211_equation_0, values = (var_1143_cast_fp16_1, var_1193_cast_fp16))[name = tensor("op_1211_cast_fp16")]; + tensor var_1213_equation_0 = const()[name = tensor("op_1213_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1213_cast_fp16 = einsum(equation = var_1213_equation_0, values = (var_1143_cast_fp16_2, var_1194_cast_fp16))[name = tensor("op_1213_cast_fp16")]; + tensor var_1215_equation_0 = const()[name = tensor("op_1215_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1215_cast_fp16 = einsum(equation = var_1215_equation_0, values = (var_1143_cast_fp16_3, var_1195_cast_fp16))[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_equation_0 = const()[name = tensor("op_1217_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1217_cast_fp16 = einsum(equation = var_1217_equation_0, values = (var_1143_cast_fp16_4, var_1196_cast_fp16))[name = tensor("op_1217_cast_fp16")]; + tensor var_1219_equation_0 = const()[name = tensor("op_1219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1219_cast_fp16 = einsum(equation = var_1219_equation_0, values = (var_1143_cast_fp16_5, var_1197_cast_fp16))[name = tensor("op_1219_cast_fp16")]; + tensor var_1221_equation_0 = const()[name = tensor("op_1221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1221_cast_fp16 = einsum(equation = var_1221_equation_0, values = (var_1143_cast_fp16_6, var_1198_cast_fp16))[name = tensor("op_1221_cast_fp16")]; + tensor var_1223_equation_0 = const()[name = tensor("op_1223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1223_cast_fp16 = einsum(equation = var_1223_equation_0, values = (var_1143_cast_fp16_7, var_1199_cast_fp16))[name = tensor("op_1223_cast_fp16")]; + tensor var_1225_equation_0 = const()[name = tensor("op_1225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1225_cast_fp16 = einsum(equation = var_1225_equation_0, values = (var_1143_cast_fp16_8, var_1200_cast_fp16))[name = tensor("op_1225_cast_fp16")]; + tensor var_1227_equation_0 = const()[name = tensor("op_1227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1227_cast_fp16 = einsum(equation = var_1227_equation_0, values = (var_1143_cast_fp16_9, var_1201_cast_fp16))[name = tensor("op_1227_cast_fp16")]; + tensor var_1229_equation_0 = const()[name = tensor("op_1229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1229_cast_fp16 = einsum(equation = var_1229_equation_0, values = (var_1143_cast_fp16_10, var_1202_cast_fp16))[name = tensor("op_1229_cast_fp16")]; + tensor var_1231_equation_0 = const()[name = tensor("op_1231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1231_cast_fp16 = einsum(equation = var_1231_equation_0, values = (var_1143_cast_fp16_11, var_1203_cast_fp16))[name = tensor("op_1231_cast_fp16")]; + tensor var_1233_equation_0 = const()[name = tensor("op_1233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1233_cast_fp16 = einsum(equation = var_1233_equation_0, values = (var_1143_cast_fp16_12, var_1204_cast_fp16))[name = tensor("op_1233_cast_fp16")]; + tensor var_1235_equation_0 = const()[name = tensor("op_1235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1235_cast_fp16 = einsum(equation = var_1235_equation_0, values = (var_1143_cast_fp16_13, var_1205_cast_fp16))[name = tensor("op_1235_cast_fp16")]; + tensor var_1237_equation_0 = const()[name = tensor("op_1237_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1237_cast_fp16 = einsum(equation = var_1237_equation_0, values = (var_1143_cast_fp16_14, var_1206_cast_fp16))[name = tensor("op_1237_cast_fp16")]; + tensor var_1239_equation_0 = const()[name = tensor("op_1239_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1239_cast_fp16 = einsum(equation = var_1239_equation_0, values = (var_1143_cast_fp16_15, var_1207_cast_fp16))[name = tensor("op_1239_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_1056, interleave = input_45_interleave_0, values = (var_1209_cast_fp16, var_1211_cast_fp16, var_1213_cast_fp16, var_1215_cast_fp16, var_1217_cast_fp16, var_1219_cast_fp16, var_1221_cast_fp16, var_1223_cast_fp16, var_1225_cast_fp16, var_1227_cast_fp16, var_1229_cast_fp16, var_1231_cast_fp16, var_1233_cast_fp16, var_1235_cast_fp16, var_1237_cast_fp16, var_1239_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1248_pad_type_0 = const()[name = tensor("op_1248_pad_type_0"), val = tensor("valid")]; + tensor var_1248_strides_0 = const()[name = tensor("op_1248_strides_0"), val = tensor([1, 1])]; + tensor var_1248_pad_0 = const()[name = tensor("op_1248_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1248_dilations_0 = const()[name = tensor("op_1248_dilations_0"), val = tensor([1, 1])]; + tensor var_1248_groups_0 = const()[name = tensor("op_1248_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116924992)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119022208)))]; + tensor var_1248_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1248_dilations_0, groups = var_1248_groups_0, pad = var_1248_pad_0, pad_type = var_1248_pad_type_0, strides = var_1248_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1248_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1248_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119024320)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119026432)))]; + tensor var_1258_to_fp16 = const()[name = tensor("op_1258_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1258_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119028544)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127417216)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1284_pad_type_0 = const()[name = tensor("op_1284_pad_type_0"), val = tensor("valid")]; + tensor var_1284_strides_0 = const()[name = tensor("op_1284_strides_0"), val = tensor([1, 1])]; + tensor var_1284_pad_0 = const()[name = tensor("op_1284_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1284_dilations_0 = const()[name = tensor("op_1284_dilations_0"), val = tensor([1, 1])]; + tensor var_1284_groups_0 = const()[name = tensor("op_1284_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127425472)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135814144)))]; + tensor var_1284_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1284_dilations_0, groups = var_1284_groups_0, pad = var_1284_pad_0, pad_type = var_1284_pad_type_0, strides = var_1284_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1284_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1284_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135816256)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135818368)))]; + tensor var_1309_to_fp16 = const()[name = tensor("op_1309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1309_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1344_weight_0_to_fp16 = const()[name = tensor("op_1344_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135820480)))]; + tensor var_1344_bias_0_to_fp16 = const()[name = tensor("op_1344_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137917696)))]; + tensor var_1344_cast_fp16 = conv(bias = var_1344_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1344_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137919808)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1342_pad_type_0 = const()[name = tensor("op_1342_pad_type_0"), val = tensor("valid")]; + tensor var_1342_strides_0 = const()[name = tensor("op_1342_strides_0"), val = tensor([1, 1])]; + tensor var_1342_pad_0 = const()[name = tensor("op_1342_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1342_dilations_0 = const()[name = tensor("op_1342_dilations_0"), val = tensor([1, 1])]; + tensor var_1342_groups_0 = const()[name = tensor("op_1342_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140017024)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142114240)))]; + tensor var_1342_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1342_dilations_0, groups = var_1342_groups_0, pad = var_1342_pad_0, pad_type = var_1342_pad_type_0, strides = var_1342_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1345_axis_0 = const()[name = tensor("op_1345_axis_0"), val = tensor(1)]; + tensor var_1345_cast_fp16_0, tensor var_1345_cast_fp16_1, tensor var_1345_cast_fp16_2, tensor var_1345_cast_fp16_3, tensor var_1345_cast_fp16_4, tensor var_1345_cast_fp16_5, tensor var_1345_cast_fp16_6, tensor var_1345_cast_fp16_7, tensor var_1345_cast_fp16_8, tensor var_1345_cast_fp16_9, tensor var_1345_cast_fp16_10, tensor var_1345_cast_fp16_11, tensor var_1345_cast_fp16_12, tensor var_1345_cast_fp16_13, tensor var_1345_cast_fp16_14, tensor var_1345_cast_fp16_15 = split(axis = var_1345_axis_0, split_sizes = tile_15, x = var_1344_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor var_1362_perm_0 = const()[name = tensor("op_1362_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1363_axis_0 = const()[name = tensor("op_1363_axis_0"), val = tensor(3)]; + tensor var_1362_cast_fp16 = transpose(perm = var_1362_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_19")]; + tensor var_1363_cast_fp16_0, tensor var_1363_cast_fp16_1, tensor var_1363_cast_fp16_2, tensor var_1363_cast_fp16_3, tensor var_1363_cast_fp16_4, tensor var_1363_cast_fp16_5, tensor var_1363_cast_fp16_6, tensor var_1363_cast_fp16_7, tensor var_1363_cast_fp16_8, tensor var_1363_cast_fp16_9, tensor var_1363_cast_fp16_10, tensor var_1363_cast_fp16_11, tensor var_1363_cast_fp16_12, tensor var_1363_cast_fp16_13, tensor var_1363_cast_fp16_14, tensor var_1363_cast_fp16_15 = split(axis = var_1363_axis_0, split_sizes = tile_16, x = var_1362_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(1)]; + tensor var_1380_cast_fp16_0, tensor var_1380_cast_fp16_1, tensor var_1380_cast_fp16_2, tensor var_1380_cast_fp16_3, tensor var_1380_cast_fp16_4, tensor var_1380_cast_fp16_5, tensor var_1380_cast_fp16_6, tensor var_1380_cast_fp16_7, tensor var_1380_cast_fp16_8, tensor var_1380_cast_fp16_9, tensor var_1380_cast_fp16_10, tensor var_1380_cast_fp16_11, tensor var_1380_cast_fp16_12, tensor var_1380_cast_fp16_13, tensor var_1380_cast_fp16_14, tensor var_1380_cast_fp16_15 = split(axis = var_1380_axis_0, split_sizes = tile_17, x = var_1342_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1363_cast_fp16_0, var_1345_cast_fp16_0))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1363_cast_fp16_1, var_1345_cast_fp16_1))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1363_cast_fp16_2, var_1345_cast_fp16_2))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1363_cast_fp16_3, var_1345_cast_fp16_3))[name = tensor("aw_167_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1363_cast_fp16_4, var_1345_cast_fp16_4))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1363_cast_fp16_5, var_1345_cast_fp16_5))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1363_cast_fp16_6, var_1345_cast_fp16_6))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1363_cast_fp16_7, var_1345_cast_fp16_7))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1363_cast_fp16_8, var_1345_cast_fp16_8))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1363_cast_fp16_9, var_1345_cast_fp16_9))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1363_cast_fp16_10, var_1345_cast_fp16_10))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1363_cast_fp16_11, var_1345_cast_fp16_11))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1363_cast_fp16_12, var_1345_cast_fp16_12))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1363_cast_fp16_13, var_1345_cast_fp16_13))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1363_cast_fp16_14, var_1345_cast_fp16_14))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1363_cast_fp16_15, var_1345_cast_fp16_15))[name = tensor("aw_191_cast_fp16")]; + tensor var_1429_cast_fp16 = softmax(axis = var_1293, x = aw_161_cast_fp16)[name = tensor("op_1429_cast_fp16")]; + tensor var_1430_cast_fp16 = softmax(axis = var_1293, x = aw_163_cast_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor var_1431_cast_fp16 = softmax(axis = var_1293, x = aw_165_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1432_cast_fp16 = softmax(axis = var_1293, x = aw_167_cast_fp16)[name = tensor("op_1432_cast_fp16")]; + tensor var_1433_cast_fp16 = softmax(axis = var_1293, x = aw_169_cast_fp16)[name = tensor("op_1433_cast_fp16")]; + tensor var_1434_cast_fp16 = softmax(axis = var_1293, x = aw_171_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor var_1435_cast_fp16 = softmax(axis = var_1293, x = aw_173_cast_fp16)[name = tensor("op_1435_cast_fp16")]; + tensor var_1436_cast_fp16 = softmax(axis = var_1293, x = aw_175_cast_fp16)[name = tensor("op_1436_cast_fp16")]; + tensor var_1437_cast_fp16 = softmax(axis = var_1293, x = aw_177_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = softmax(axis = var_1293, x = aw_179_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor var_1439_cast_fp16 = softmax(axis = var_1293, x = aw_181_cast_fp16)[name = tensor("op_1439_cast_fp16")]; + tensor var_1440_cast_fp16 = softmax(axis = var_1293, x = aw_183_cast_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1441_cast_fp16 = softmax(axis = var_1293, x = aw_185_cast_fp16)[name = tensor("op_1441_cast_fp16")]; + tensor var_1442_cast_fp16 = softmax(axis = var_1293, x = aw_187_cast_fp16)[name = tensor("op_1442_cast_fp16")]; + tensor var_1443_cast_fp16 = softmax(axis = var_1293, x = aw_189_cast_fp16)[name = tensor("op_1443_cast_fp16")]; + tensor var_1444_cast_fp16 = softmax(axis = var_1293, x = aw_191_cast_fp16)[name = tensor("op_1444_cast_fp16")]; + tensor var_1446_equation_0 = const()[name = tensor("op_1446_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1446_cast_fp16 = einsum(equation = var_1446_equation_0, values = (var_1380_cast_fp16_0, var_1429_cast_fp16))[name = tensor("op_1446_cast_fp16")]; + tensor var_1448_equation_0 = const()[name = tensor("op_1448_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1448_cast_fp16 = einsum(equation = var_1448_equation_0, values = (var_1380_cast_fp16_1, var_1430_cast_fp16))[name = tensor("op_1448_cast_fp16")]; + tensor var_1450_equation_0 = const()[name = tensor("op_1450_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1450_cast_fp16 = einsum(equation = var_1450_equation_0, values = (var_1380_cast_fp16_2, var_1431_cast_fp16))[name = tensor("op_1450_cast_fp16")]; + tensor var_1452_equation_0 = const()[name = tensor("op_1452_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1452_cast_fp16 = einsum(equation = var_1452_equation_0, values = (var_1380_cast_fp16_3, var_1432_cast_fp16))[name = tensor("op_1452_cast_fp16")]; + tensor var_1454_equation_0 = const()[name = tensor("op_1454_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1454_cast_fp16 = einsum(equation = var_1454_equation_0, values = (var_1380_cast_fp16_4, var_1433_cast_fp16))[name = tensor("op_1454_cast_fp16")]; + tensor var_1456_equation_0 = const()[name = tensor("op_1456_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1456_cast_fp16 = einsum(equation = var_1456_equation_0, values = (var_1380_cast_fp16_5, var_1434_cast_fp16))[name = tensor("op_1456_cast_fp16")]; + tensor var_1458_equation_0 = const()[name = tensor("op_1458_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1458_cast_fp16 = einsum(equation = var_1458_equation_0, values = (var_1380_cast_fp16_6, var_1435_cast_fp16))[name = tensor("op_1458_cast_fp16")]; + tensor var_1460_equation_0 = const()[name = tensor("op_1460_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1460_cast_fp16 = einsum(equation = var_1460_equation_0, values = (var_1380_cast_fp16_7, var_1436_cast_fp16))[name = tensor("op_1460_cast_fp16")]; + tensor var_1462_equation_0 = const()[name = tensor("op_1462_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1462_cast_fp16 = einsum(equation = var_1462_equation_0, values = (var_1380_cast_fp16_8, var_1437_cast_fp16))[name = tensor("op_1462_cast_fp16")]; + tensor var_1464_equation_0 = const()[name = tensor("op_1464_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1464_cast_fp16 = einsum(equation = var_1464_equation_0, values = (var_1380_cast_fp16_9, var_1438_cast_fp16))[name = tensor("op_1464_cast_fp16")]; + tensor var_1466_equation_0 = const()[name = tensor("op_1466_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1466_cast_fp16 = einsum(equation = var_1466_equation_0, values = (var_1380_cast_fp16_10, var_1439_cast_fp16))[name = tensor("op_1466_cast_fp16")]; + tensor var_1468_equation_0 = const()[name = tensor("op_1468_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1468_cast_fp16 = einsum(equation = var_1468_equation_0, values = (var_1380_cast_fp16_11, var_1440_cast_fp16))[name = tensor("op_1468_cast_fp16")]; + tensor var_1470_equation_0 = const()[name = tensor("op_1470_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1470_cast_fp16 = einsum(equation = var_1470_equation_0, values = (var_1380_cast_fp16_12, var_1441_cast_fp16))[name = tensor("op_1470_cast_fp16")]; + tensor var_1472_equation_0 = const()[name = tensor("op_1472_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1472_cast_fp16 = einsum(equation = var_1472_equation_0, values = (var_1380_cast_fp16_13, var_1442_cast_fp16))[name = tensor("op_1472_cast_fp16")]; + tensor var_1474_equation_0 = const()[name = tensor("op_1474_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1474_cast_fp16 = einsum(equation = var_1474_equation_0, values = (var_1380_cast_fp16_14, var_1443_cast_fp16))[name = tensor("op_1474_cast_fp16")]; + tensor var_1476_equation_0 = const()[name = tensor("op_1476_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1476_cast_fp16 = einsum(equation = var_1476_equation_0, values = (var_1380_cast_fp16_15, var_1444_cast_fp16))[name = tensor("op_1476_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1293, interleave = input_55_interleave_0, values = (var_1446_cast_fp16, var_1448_cast_fp16, var_1450_cast_fp16, var_1452_cast_fp16, var_1454_cast_fp16, var_1456_cast_fp16, var_1458_cast_fp16, var_1460_cast_fp16, var_1462_cast_fp16, var_1464_cast_fp16, var_1466_cast_fp16, var_1468_cast_fp16, var_1470_cast_fp16, var_1472_cast_fp16, var_1474_cast_fp16, var_1476_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1485_pad_type_0 = const()[name = tensor("op_1485_pad_type_0"), val = tensor("valid")]; + tensor var_1485_strides_0 = const()[name = tensor("op_1485_strides_0"), val = tensor([1, 1])]; + tensor var_1485_pad_0 = const()[name = tensor("op_1485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1485_dilations_0 = const()[name = tensor("op_1485_dilations_0"), val = tensor([1, 1])]; + tensor var_1485_groups_0 = const()[name = tensor("op_1485_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142116352)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144213568)))]; + tensor var_1485_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1485_dilations_0, groups = var_1485_groups_0, pad = var_1485_pad_0, pad_type = var_1485_pad_type_0, strides = var_1485_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1485_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144215680)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144217792)))]; + tensor var_1495_to_fp16 = const()[name = tensor("op_1495_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1495_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144219904)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152608576)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1521_pad_type_0 = const()[name = tensor("op_1521_pad_type_0"), val = tensor("valid")]; + tensor var_1521_strides_0 = const()[name = tensor("op_1521_strides_0"), val = tensor([1, 1])]; + tensor var_1521_pad_0 = const()[name = tensor("op_1521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1521_dilations_0 = const()[name = tensor("op_1521_dilations_0"), val = tensor([1, 1])]; + tensor var_1521_groups_0 = const()[name = tensor("op_1521_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152616832)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161005504)))]; + tensor var_1521_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1521_dilations_0, groups = var_1521_groups_0, pad = var_1521_pad_0, pad_type = var_1521_pad_type_0, strides = var_1521_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1521_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1521_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161007616)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161009728)))]; + tensor var_1546_to_fp16 = const()[name = tensor("op_1546_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1546_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1581_weight_0_to_fp16 = const()[name = tensor("op_1581_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161011840)))]; + tensor var_1581_bias_0_to_fp16 = const()[name = tensor("op_1581_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163109056)))]; + tensor var_1581_cast_fp16 = conv(bias = var_1581_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1581_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1581_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163111168)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1579_pad_type_0 = const()[name = tensor("op_1579_pad_type_0"), val = tensor("valid")]; + tensor var_1579_strides_0 = const()[name = tensor("op_1579_strides_0"), val = tensor([1, 1])]; + tensor var_1579_pad_0 = const()[name = tensor("op_1579_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1579_dilations_0 = const()[name = tensor("op_1579_dilations_0"), val = tensor([1, 1])]; + tensor var_1579_groups_0 = const()[name = tensor("op_1579_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165208384)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167305600)))]; + tensor var_1579_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1579_dilations_0, groups = var_1579_groups_0, pad = var_1579_pad_0, pad_type = var_1579_pad_type_0, strides = var_1579_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1582_axis_0 = const()[name = tensor("op_1582_axis_0"), val = tensor(1)]; + tensor var_1582_cast_fp16_0, tensor var_1582_cast_fp16_1, tensor var_1582_cast_fp16_2, tensor var_1582_cast_fp16_3, tensor var_1582_cast_fp16_4, tensor var_1582_cast_fp16_5, tensor var_1582_cast_fp16_6, tensor var_1582_cast_fp16_7, tensor var_1582_cast_fp16_8, tensor var_1582_cast_fp16_9, tensor var_1582_cast_fp16_10, tensor var_1582_cast_fp16_11, tensor var_1582_cast_fp16_12, tensor var_1582_cast_fp16_13, tensor var_1582_cast_fp16_14, tensor var_1582_cast_fp16_15 = split(axis = var_1582_axis_0, split_sizes = tile_18, x = var_1581_cast_fp16)[name = tensor("op_1582_cast_fp16")]; + tensor var_1599_perm_0 = const()[name = tensor("op_1599_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1600_axis_0 = const()[name = tensor("op_1600_axis_0"), val = tensor(3)]; + tensor var_1599_cast_fp16 = transpose(perm = var_1599_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_18")]; + tensor var_1600_cast_fp16_0, tensor var_1600_cast_fp16_1, tensor var_1600_cast_fp16_2, tensor var_1600_cast_fp16_3, tensor var_1600_cast_fp16_4, tensor var_1600_cast_fp16_5, tensor var_1600_cast_fp16_6, tensor var_1600_cast_fp16_7, tensor var_1600_cast_fp16_8, tensor var_1600_cast_fp16_9, tensor var_1600_cast_fp16_10, tensor var_1600_cast_fp16_11, tensor var_1600_cast_fp16_12, tensor var_1600_cast_fp16_13, tensor var_1600_cast_fp16_14, tensor var_1600_cast_fp16_15 = split(axis = var_1600_axis_0, split_sizes = tile_19, x = var_1599_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1617_axis_0 = const()[name = tensor("op_1617_axis_0"), val = tensor(1)]; + tensor var_1617_cast_fp16_0, tensor var_1617_cast_fp16_1, tensor var_1617_cast_fp16_2, tensor var_1617_cast_fp16_3, tensor var_1617_cast_fp16_4, tensor var_1617_cast_fp16_5, tensor var_1617_cast_fp16_6, tensor var_1617_cast_fp16_7, tensor var_1617_cast_fp16_8, tensor var_1617_cast_fp16_9, tensor var_1617_cast_fp16_10, tensor var_1617_cast_fp16_11, tensor var_1617_cast_fp16_12, tensor var_1617_cast_fp16_13, tensor var_1617_cast_fp16_14, tensor var_1617_cast_fp16_15 = split(axis = var_1617_axis_0, split_sizes = tile_20, x = var_1579_cast_fp16)[name = tensor("op_1617_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1600_cast_fp16_0, var_1582_cast_fp16_0))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1600_cast_fp16_1, var_1582_cast_fp16_1))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1600_cast_fp16_2, var_1582_cast_fp16_2))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1600_cast_fp16_3, var_1582_cast_fp16_3))[name = tensor("aw_199_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1600_cast_fp16_4, var_1582_cast_fp16_4))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1600_cast_fp16_5, var_1582_cast_fp16_5))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1600_cast_fp16_6, var_1582_cast_fp16_6))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1600_cast_fp16_7, var_1582_cast_fp16_7))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1600_cast_fp16_8, var_1582_cast_fp16_8))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1600_cast_fp16_9, var_1582_cast_fp16_9))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1600_cast_fp16_10, var_1582_cast_fp16_10))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1600_cast_fp16_11, var_1582_cast_fp16_11))[name = tensor("aw_215_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1600_cast_fp16_12, var_1582_cast_fp16_12))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1600_cast_fp16_13, var_1582_cast_fp16_13))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1600_cast_fp16_14, var_1582_cast_fp16_14))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1600_cast_fp16_15, var_1582_cast_fp16_15))[name = tensor("aw_223_cast_fp16")]; + tensor var_1666_cast_fp16 = softmax(axis = var_1530, x = aw_193_cast_fp16)[name = tensor("op_1666_cast_fp16")]; + tensor var_1667_cast_fp16 = softmax(axis = var_1530, x = aw_195_cast_fp16)[name = tensor("op_1667_cast_fp16")]; + tensor var_1668_cast_fp16 = softmax(axis = var_1530, x = aw_197_cast_fp16)[name = tensor("op_1668_cast_fp16")]; + tensor var_1669_cast_fp16 = softmax(axis = var_1530, x = aw_199_cast_fp16)[name = tensor("op_1669_cast_fp16")]; + tensor var_1670_cast_fp16 = softmax(axis = var_1530, x = aw_201_cast_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1671_cast_fp16 = softmax(axis = var_1530, x = aw_203_cast_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1672_cast_fp16 = softmax(axis = var_1530, x = aw_205_cast_fp16)[name = tensor("op_1672_cast_fp16")]; + tensor var_1673_cast_fp16 = softmax(axis = var_1530, x = aw_207_cast_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1674_cast_fp16 = softmax(axis = var_1530, x = aw_209_cast_fp16)[name = tensor("op_1674_cast_fp16")]; + tensor var_1675_cast_fp16 = softmax(axis = var_1530, x = aw_211_cast_fp16)[name = tensor("op_1675_cast_fp16")]; + tensor var_1676_cast_fp16 = softmax(axis = var_1530, x = aw_213_cast_fp16)[name = tensor("op_1676_cast_fp16")]; + tensor var_1677_cast_fp16 = softmax(axis = var_1530, x = aw_215_cast_fp16)[name = tensor("op_1677_cast_fp16")]; + tensor var_1678_cast_fp16 = softmax(axis = var_1530, x = aw_217_cast_fp16)[name = tensor("op_1678_cast_fp16")]; + tensor var_1679_cast_fp16 = softmax(axis = var_1530, x = aw_219_cast_fp16)[name = tensor("op_1679_cast_fp16")]; + tensor var_1680_cast_fp16 = softmax(axis = var_1530, x = aw_221_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor var_1681_cast_fp16 = softmax(axis = var_1530, x = aw_223_cast_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1683_equation_0 = const()[name = tensor("op_1683_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1683_cast_fp16 = einsum(equation = var_1683_equation_0, values = (var_1617_cast_fp16_0, var_1666_cast_fp16))[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_equation_0 = const()[name = tensor("op_1685_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1685_cast_fp16 = einsum(equation = var_1685_equation_0, values = (var_1617_cast_fp16_1, var_1667_cast_fp16))[name = tensor("op_1685_cast_fp16")]; + tensor var_1687_equation_0 = const()[name = tensor("op_1687_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1687_cast_fp16 = einsum(equation = var_1687_equation_0, values = (var_1617_cast_fp16_2, var_1668_cast_fp16))[name = tensor("op_1687_cast_fp16")]; + tensor var_1689_equation_0 = const()[name = tensor("op_1689_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1689_cast_fp16 = einsum(equation = var_1689_equation_0, values = (var_1617_cast_fp16_3, var_1669_cast_fp16))[name = tensor("op_1689_cast_fp16")]; + tensor var_1691_equation_0 = const()[name = tensor("op_1691_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1691_cast_fp16 = einsum(equation = var_1691_equation_0, values = (var_1617_cast_fp16_4, var_1670_cast_fp16))[name = tensor("op_1691_cast_fp16")]; + tensor var_1693_equation_0 = const()[name = tensor("op_1693_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1693_cast_fp16 = einsum(equation = var_1693_equation_0, values = (var_1617_cast_fp16_5, var_1671_cast_fp16))[name = tensor("op_1693_cast_fp16")]; + tensor var_1695_equation_0 = const()[name = tensor("op_1695_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1695_cast_fp16 = einsum(equation = var_1695_equation_0, values = (var_1617_cast_fp16_6, var_1672_cast_fp16))[name = tensor("op_1695_cast_fp16")]; + tensor var_1697_equation_0 = const()[name = tensor("op_1697_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1697_cast_fp16 = einsum(equation = var_1697_equation_0, values = (var_1617_cast_fp16_7, var_1673_cast_fp16))[name = tensor("op_1697_cast_fp16")]; + tensor var_1699_equation_0 = const()[name = tensor("op_1699_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1699_cast_fp16 = einsum(equation = var_1699_equation_0, values = (var_1617_cast_fp16_8, var_1674_cast_fp16))[name = tensor("op_1699_cast_fp16")]; + tensor var_1701_equation_0 = const()[name = tensor("op_1701_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1701_cast_fp16 = einsum(equation = var_1701_equation_0, values = (var_1617_cast_fp16_9, var_1675_cast_fp16))[name = tensor("op_1701_cast_fp16")]; + tensor var_1703_equation_0 = const()[name = tensor("op_1703_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1703_cast_fp16 = einsum(equation = var_1703_equation_0, values = (var_1617_cast_fp16_10, var_1676_cast_fp16))[name = tensor("op_1703_cast_fp16")]; + tensor var_1705_equation_0 = const()[name = tensor("op_1705_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1705_cast_fp16 = einsum(equation = var_1705_equation_0, values = (var_1617_cast_fp16_11, var_1677_cast_fp16))[name = tensor("op_1705_cast_fp16")]; + tensor var_1707_equation_0 = const()[name = tensor("op_1707_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1707_cast_fp16 = einsum(equation = var_1707_equation_0, values = (var_1617_cast_fp16_12, var_1678_cast_fp16))[name = tensor("op_1707_cast_fp16")]; + tensor var_1709_equation_0 = const()[name = tensor("op_1709_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1709_cast_fp16 = einsum(equation = var_1709_equation_0, values = (var_1617_cast_fp16_13, var_1679_cast_fp16))[name = tensor("op_1709_cast_fp16")]; + tensor var_1711_equation_0 = const()[name = tensor("op_1711_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1711_cast_fp16 = einsum(equation = var_1711_equation_0, values = (var_1617_cast_fp16_14, var_1680_cast_fp16))[name = tensor("op_1711_cast_fp16")]; + tensor var_1713_equation_0 = const()[name = tensor("op_1713_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1713_cast_fp16 = einsum(equation = var_1713_equation_0, values = (var_1617_cast_fp16_15, var_1681_cast_fp16))[name = tensor("op_1713_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1530, interleave = input_65_interleave_0, values = (var_1683_cast_fp16, var_1685_cast_fp16, var_1687_cast_fp16, var_1689_cast_fp16, var_1691_cast_fp16, var_1693_cast_fp16, var_1695_cast_fp16, var_1697_cast_fp16, var_1699_cast_fp16, var_1701_cast_fp16, var_1703_cast_fp16, var_1705_cast_fp16, var_1707_cast_fp16, var_1709_cast_fp16, var_1711_cast_fp16, var_1713_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1722_pad_type_0 = const()[name = tensor("op_1722_pad_type_0"), val = tensor("valid")]; + tensor var_1722_strides_0 = const()[name = tensor("op_1722_strides_0"), val = tensor([1, 1])]; + tensor var_1722_pad_0 = const()[name = tensor("op_1722_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1722_dilations_0 = const()[name = tensor("op_1722_dilations_0"), val = tensor([1, 1])]; + tensor var_1722_groups_0 = const()[name = tensor("op_1722_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167307712)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169404928)))]; + tensor var_1722_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1722_dilations_0, groups = var_1722_groups_0, pad = var_1722_pad_0, pad_type = var_1722_pad_type_0, strides = var_1722_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1722_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1722_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169407040)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169409152)))]; + tensor var_1732_to_fp16 = const()[name = tensor("op_1732_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1732_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169411264)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177799936)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1758_pad_type_0 = const()[name = tensor("op_1758_pad_type_0"), val = tensor("valid")]; + tensor var_1758_strides_0 = const()[name = tensor("op_1758_strides_0"), val = tensor([1, 1])]; + tensor var_1758_pad_0 = const()[name = tensor("op_1758_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1758_dilations_0 = const()[name = tensor("op_1758_dilations_0"), val = tensor([1, 1])]; + tensor var_1758_groups_0 = const()[name = tensor("op_1758_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177808192)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186196864)))]; + tensor var_1758_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1758_dilations_0, groups = var_1758_groups_0, pad = var_1758_pad_0, pad_type = var_1758_pad_type_0, strides = var_1758_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1758_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1758_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1767 = const()[name = tensor("op_1767"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186198976)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186201088)))]; + tensor var_1783_to_fp16 = const()[name = tensor("op_1783_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_1783_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_1818_weight_0_to_fp16 = const()[name = tensor("op_1818_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186203200)))]; + tensor var_1818_bias_0_to_fp16 = const()[name = tensor("op_1818_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188300416)))]; + tensor var_1818_cast_fp16 = conv(bias = var_1818_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_1818_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1818_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188302528)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_1816_pad_type_0 = const()[name = tensor("op_1816_pad_type_0"), val = tensor("valid")]; + tensor var_1816_strides_0 = const()[name = tensor("op_1816_strides_0"), val = tensor([1, 1])]; + tensor var_1816_pad_0 = const()[name = tensor("op_1816_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1816_dilations_0 = const()[name = tensor("op_1816_dilations_0"), val = tensor([1, 1])]; + tensor var_1816_groups_0 = const()[name = tensor("op_1816_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190399744)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192496960)))]; + tensor var_1816_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_1816_dilations_0, groups = var_1816_groups_0, pad = var_1816_pad_0, pad_type = var_1816_pad_type_0, strides = var_1816_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1819_axis_0 = const()[name = tensor("op_1819_axis_0"), val = tensor(1)]; + tensor var_1819_cast_fp16_0, tensor var_1819_cast_fp16_1, tensor var_1819_cast_fp16_2, tensor var_1819_cast_fp16_3, tensor var_1819_cast_fp16_4, tensor var_1819_cast_fp16_5, tensor var_1819_cast_fp16_6, tensor var_1819_cast_fp16_7, tensor var_1819_cast_fp16_8, tensor var_1819_cast_fp16_9, tensor var_1819_cast_fp16_10, tensor var_1819_cast_fp16_11, tensor var_1819_cast_fp16_12, tensor var_1819_cast_fp16_13, tensor var_1819_cast_fp16_14, tensor var_1819_cast_fp16_15 = split(axis = var_1819_axis_0, split_sizes = tile_21, x = var_1818_cast_fp16)[name = tensor("op_1819_cast_fp16")]; + tensor var_1836_perm_0 = const()[name = tensor("op_1836_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1837_axis_0 = const()[name = tensor("op_1837_axis_0"), val = tensor(3)]; + tensor var_1836_cast_fp16 = transpose(perm = var_1836_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_17")]; + tensor var_1837_cast_fp16_0, tensor var_1837_cast_fp16_1, tensor var_1837_cast_fp16_2, tensor var_1837_cast_fp16_3, tensor var_1837_cast_fp16_4, tensor var_1837_cast_fp16_5, tensor var_1837_cast_fp16_6, tensor var_1837_cast_fp16_7, tensor var_1837_cast_fp16_8, tensor var_1837_cast_fp16_9, tensor var_1837_cast_fp16_10, tensor var_1837_cast_fp16_11, tensor var_1837_cast_fp16_12, tensor var_1837_cast_fp16_13, tensor var_1837_cast_fp16_14, tensor var_1837_cast_fp16_15 = split(axis = var_1837_axis_0, split_sizes = tile_22, x = var_1836_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1854_axis_0 = const()[name = tensor("op_1854_axis_0"), val = tensor(1)]; + tensor var_1854_cast_fp16_0, tensor var_1854_cast_fp16_1, tensor var_1854_cast_fp16_2, tensor var_1854_cast_fp16_3, tensor var_1854_cast_fp16_4, tensor var_1854_cast_fp16_5, tensor var_1854_cast_fp16_6, tensor var_1854_cast_fp16_7, tensor var_1854_cast_fp16_8, tensor var_1854_cast_fp16_9, tensor var_1854_cast_fp16_10, tensor var_1854_cast_fp16_11, tensor var_1854_cast_fp16_12, tensor var_1854_cast_fp16_13, tensor var_1854_cast_fp16_14, tensor var_1854_cast_fp16_15 = split(axis = var_1854_axis_0, split_sizes = tile_23, x = var_1816_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1837_cast_fp16_0, var_1819_cast_fp16_0))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1837_cast_fp16_1, var_1819_cast_fp16_1))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1837_cast_fp16_2, var_1819_cast_fp16_2))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1837_cast_fp16_3, var_1819_cast_fp16_3))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1837_cast_fp16_4, var_1819_cast_fp16_4))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1837_cast_fp16_5, var_1819_cast_fp16_5))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1837_cast_fp16_6, var_1819_cast_fp16_6))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1837_cast_fp16_7, var_1819_cast_fp16_7))[name = tensor("aw_239_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_1837_cast_fp16_8, var_1819_cast_fp16_8))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_1837_cast_fp16_9, var_1819_cast_fp16_9))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_1837_cast_fp16_10, var_1819_cast_fp16_10))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_1837_cast_fp16_11, var_1819_cast_fp16_11))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_1837_cast_fp16_12, var_1819_cast_fp16_12))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_1837_cast_fp16_13, var_1819_cast_fp16_13))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_1837_cast_fp16_14, var_1819_cast_fp16_14))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_1837_cast_fp16_15, var_1819_cast_fp16_15))[name = tensor("aw_255_cast_fp16")]; + tensor var_1903_cast_fp16 = softmax(axis = var_1767, x = aw_225_cast_fp16)[name = tensor("op_1903_cast_fp16")]; + tensor var_1904_cast_fp16 = softmax(axis = var_1767, x = aw_227_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1905_cast_fp16 = softmax(axis = var_1767, x = aw_229_cast_fp16)[name = tensor("op_1905_cast_fp16")]; + tensor var_1906_cast_fp16 = softmax(axis = var_1767, x = aw_231_cast_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1907_cast_fp16 = softmax(axis = var_1767, x = aw_233_cast_fp16)[name = tensor("op_1907_cast_fp16")]; + tensor var_1908_cast_fp16 = softmax(axis = var_1767, x = aw_235_cast_fp16)[name = tensor("op_1908_cast_fp16")]; + tensor var_1909_cast_fp16 = softmax(axis = var_1767, x = aw_237_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor var_1910_cast_fp16 = softmax(axis = var_1767, x = aw_239_cast_fp16)[name = tensor("op_1910_cast_fp16")]; + tensor var_1911_cast_fp16 = softmax(axis = var_1767, x = aw_241_cast_fp16)[name = tensor("op_1911_cast_fp16")]; + tensor var_1912_cast_fp16 = softmax(axis = var_1767, x = aw_243_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1913_cast_fp16 = softmax(axis = var_1767, x = aw_245_cast_fp16)[name = tensor("op_1913_cast_fp16")]; + tensor var_1914_cast_fp16 = softmax(axis = var_1767, x = aw_247_cast_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1915_cast_fp16 = softmax(axis = var_1767, x = aw_249_cast_fp16)[name = tensor("op_1915_cast_fp16")]; + tensor var_1916_cast_fp16 = softmax(axis = var_1767, x = aw_251_cast_fp16)[name = tensor("op_1916_cast_fp16")]; + tensor var_1917_cast_fp16 = softmax(axis = var_1767, x = aw_253_cast_fp16)[name = tensor("op_1917_cast_fp16")]; + tensor var_1918_cast_fp16 = softmax(axis = var_1767, x = aw_255_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor var_1920_equation_0 = const()[name = tensor("op_1920_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1920_cast_fp16 = einsum(equation = var_1920_equation_0, values = (var_1854_cast_fp16_0, var_1903_cast_fp16))[name = tensor("op_1920_cast_fp16")]; + tensor var_1922_equation_0 = const()[name = tensor("op_1922_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1922_cast_fp16 = einsum(equation = var_1922_equation_0, values = (var_1854_cast_fp16_1, var_1904_cast_fp16))[name = tensor("op_1922_cast_fp16")]; + tensor var_1924_equation_0 = const()[name = tensor("op_1924_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1924_cast_fp16 = einsum(equation = var_1924_equation_0, values = (var_1854_cast_fp16_2, var_1905_cast_fp16))[name = tensor("op_1924_cast_fp16")]; + tensor var_1926_equation_0 = const()[name = tensor("op_1926_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1926_cast_fp16 = einsum(equation = var_1926_equation_0, values = (var_1854_cast_fp16_3, var_1906_cast_fp16))[name = tensor("op_1926_cast_fp16")]; + tensor var_1928_equation_0 = const()[name = tensor("op_1928_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1928_cast_fp16 = einsum(equation = var_1928_equation_0, values = (var_1854_cast_fp16_4, var_1907_cast_fp16))[name = tensor("op_1928_cast_fp16")]; + tensor var_1930_equation_0 = const()[name = tensor("op_1930_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1930_cast_fp16 = einsum(equation = var_1930_equation_0, values = (var_1854_cast_fp16_5, var_1908_cast_fp16))[name = tensor("op_1930_cast_fp16")]; + tensor var_1932_equation_0 = const()[name = tensor("op_1932_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1932_cast_fp16 = einsum(equation = var_1932_equation_0, values = (var_1854_cast_fp16_6, var_1909_cast_fp16))[name = tensor("op_1932_cast_fp16")]; + tensor var_1934_equation_0 = const()[name = tensor("op_1934_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1934_cast_fp16 = einsum(equation = var_1934_equation_0, values = (var_1854_cast_fp16_7, var_1910_cast_fp16))[name = tensor("op_1934_cast_fp16")]; + tensor var_1936_equation_0 = const()[name = tensor("op_1936_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1936_cast_fp16 = einsum(equation = var_1936_equation_0, values = (var_1854_cast_fp16_8, var_1911_cast_fp16))[name = tensor("op_1936_cast_fp16")]; + tensor var_1938_equation_0 = const()[name = tensor("op_1938_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1938_cast_fp16 = einsum(equation = var_1938_equation_0, values = (var_1854_cast_fp16_9, var_1912_cast_fp16))[name = tensor("op_1938_cast_fp16")]; + tensor var_1940_equation_0 = const()[name = tensor("op_1940_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1940_cast_fp16 = einsum(equation = var_1940_equation_0, values = (var_1854_cast_fp16_10, var_1913_cast_fp16))[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_equation_0 = const()[name = tensor("op_1942_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1942_cast_fp16 = einsum(equation = var_1942_equation_0, values = (var_1854_cast_fp16_11, var_1914_cast_fp16))[name = tensor("op_1942_cast_fp16")]; + tensor var_1944_equation_0 = const()[name = tensor("op_1944_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1944_cast_fp16 = einsum(equation = var_1944_equation_0, values = (var_1854_cast_fp16_12, var_1915_cast_fp16))[name = tensor("op_1944_cast_fp16")]; + tensor var_1946_equation_0 = const()[name = tensor("op_1946_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1946_cast_fp16 = einsum(equation = var_1946_equation_0, values = (var_1854_cast_fp16_13, var_1916_cast_fp16))[name = tensor("op_1946_cast_fp16")]; + tensor var_1948_equation_0 = const()[name = tensor("op_1948_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1948_cast_fp16 = einsum(equation = var_1948_equation_0, values = (var_1854_cast_fp16_14, var_1917_cast_fp16))[name = tensor("op_1948_cast_fp16")]; + tensor var_1950_equation_0 = const()[name = tensor("op_1950_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1950_cast_fp16 = einsum(equation = var_1950_equation_0, values = (var_1854_cast_fp16_15, var_1918_cast_fp16))[name = tensor("op_1950_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_1767, interleave = input_75_interleave_0, values = (var_1920_cast_fp16, var_1922_cast_fp16, var_1924_cast_fp16, var_1926_cast_fp16, var_1928_cast_fp16, var_1930_cast_fp16, var_1932_cast_fp16, var_1934_cast_fp16, var_1936_cast_fp16, var_1938_cast_fp16, var_1940_cast_fp16, var_1942_cast_fp16, var_1944_cast_fp16, var_1946_cast_fp16, var_1948_cast_fp16, var_1950_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_1959_pad_type_0 = const()[name = tensor("op_1959_pad_type_0"), val = tensor("valid")]; + tensor var_1959_strides_0 = const()[name = tensor("op_1959_strides_0"), val = tensor([1, 1])]; + tensor var_1959_pad_0 = const()[name = tensor("op_1959_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1959_dilations_0 = const()[name = tensor("op_1959_dilations_0"), val = tensor([1, 1])]; + tensor var_1959_groups_0 = const()[name = tensor("op_1959_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192499072)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194596288)))]; + tensor var_1959_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_1959_dilations_0, groups = var_1959_groups_0, pad = var_1959_pad_0, pad_type = var_1959_pad_type_0, strides = var_1959_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_1959_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194598400)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194600512)))]; + tensor var_1969_to_fp16 = const()[name = tensor("op_1969_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_1969_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194602624)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202991296)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1995_pad_type_0 = const()[name = tensor("op_1995_pad_type_0"), val = tensor("valid")]; + tensor var_1995_strides_0 = const()[name = tensor("op_1995_strides_0"), val = tensor([1, 1])]; + tensor var_1995_pad_0 = const()[name = tensor("op_1995_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1995_dilations_0 = const()[name = tensor("op_1995_dilations_0"), val = tensor([1, 1])]; + tensor var_1995_groups_0 = const()[name = tensor("op_1995_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202999552)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211388224)))]; + tensor var_1995_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_1995_dilations_0, groups = var_1995_groups_0, pad = var_1995_pad_0, pad_type = var_1995_pad_type_0, strides = var_1995_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1995_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211390336)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211392448)))]; + tensor var_2020_to_fp16 = const()[name = tensor("op_2020_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_2020_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_2055_weight_0_to_fp16 = const()[name = tensor("op_2055_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211394560)))]; + tensor var_2055_bias_0_to_fp16 = const()[name = tensor("op_2055_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213491776)))]; + tensor var_2055_cast_fp16 = conv(bias = var_2055_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_2055_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2055_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213493888)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_2053_pad_type_0 = const()[name = tensor("op_2053_pad_type_0"), val = tensor("valid")]; + tensor var_2053_strides_0 = const()[name = tensor("op_2053_strides_0"), val = tensor([1, 1])]; + tensor var_2053_pad_0 = const()[name = tensor("op_2053_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2053_dilations_0 = const()[name = tensor("op_2053_dilations_0"), val = tensor([1, 1])]; + tensor var_2053_groups_0 = const()[name = tensor("op_2053_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215591104)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217688320)))]; + tensor var_2053_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_2053_dilations_0, groups = var_2053_groups_0, pad = var_2053_pad_0, pad_type = var_2053_pad_type_0, strides = var_2053_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2056_axis_0 = const()[name = tensor("op_2056_axis_0"), val = tensor(1)]; + tensor var_2056_cast_fp16_0, tensor var_2056_cast_fp16_1, tensor var_2056_cast_fp16_2, tensor var_2056_cast_fp16_3, tensor var_2056_cast_fp16_4, tensor var_2056_cast_fp16_5, tensor var_2056_cast_fp16_6, tensor var_2056_cast_fp16_7, tensor var_2056_cast_fp16_8, tensor var_2056_cast_fp16_9, tensor var_2056_cast_fp16_10, tensor var_2056_cast_fp16_11, tensor var_2056_cast_fp16_12, tensor var_2056_cast_fp16_13, tensor var_2056_cast_fp16_14, tensor var_2056_cast_fp16_15 = split(axis = var_2056_axis_0, split_sizes = tile_24, x = var_2055_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor var_2073_perm_0 = const()[name = tensor("op_2073_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2074_axis_0 = const()[name = tensor("op_2074_axis_0"), val = tensor(3)]; + tensor var_2073_cast_fp16 = transpose(perm = var_2073_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_16")]; + tensor var_2074_cast_fp16_0, tensor var_2074_cast_fp16_1, tensor var_2074_cast_fp16_2, tensor var_2074_cast_fp16_3, tensor var_2074_cast_fp16_4, tensor var_2074_cast_fp16_5, tensor var_2074_cast_fp16_6, tensor var_2074_cast_fp16_7, tensor var_2074_cast_fp16_8, tensor var_2074_cast_fp16_9, tensor var_2074_cast_fp16_10, tensor var_2074_cast_fp16_11, tensor var_2074_cast_fp16_12, tensor var_2074_cast_fp16_13, tensor var_2074_cast_fp16_14, tensor var_2074_cast_fp16_15 = split(axis = var_2074_axis_0, split_sizes = tile_25, x = var_2073_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2091_axis_0 = const()[name = tensor("op_2091_axis_0"), val = tensor(1)]; + tensor var_2091_cast_fp16_0, tensor var_2091_cast_fp16_1, tensor var_2091_cast_fp16_2, tensor var_2091_cast_fp16_3, tensor var_2091_cast_fp16_4, tensor var_2091_cast_fp16_5, tensor var_2091_cast_fp16_6, tensor var_2091_cast_fp16_7, tensor var_2091_cast_fp16_8, tensor var_2091_cast_fp16_9, tensor var_2091_cast_fp16_10, tensor var_2091_cast_fp16_11, tensor var_2091_cast_fp16_12, tensor var_2091_cast_fp16_13, tensor var_2091_cast_fp16_14, tensor var_2091_cast_fp16_15 = split(axis = var_2091_axis_0, split_sizes = tile_26, x = var_2053_cast_fp16)[name = tensor("op_2091_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_2074_cast_fp16_0, var_2056_cast_fp16_0))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_2074_cast_fp16_1, var_2056_cast_fp16_1))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_2074_cast_fp16_2, var_2056_cast_fp16_2))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_2074_cast_fp16_3, var_2056_cast_fp16_3))[name = tensor("aw_263_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_2074_cast_fp16_4, var_2056_cast_fp16_4))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_2074_cast_fp16_5, var_2056_cast_fp16_5))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_2074_cast_fp16_6, var_2056_cast_fp16_6))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_2074_cast_fp16_7, var_2056_cast_fp16_7))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_2074_cast_fp16_8, var_2056_cast_fp16_8))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_2074_cast_fp16_9, var_2056_cast_fp16_9))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_2074_cast_fp16_10, var_2056_cast_fp16_10))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_2074_cast_fp16_11, var_2056_cast_fp16_11))[name = tensor("aw_279_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2074_cast_fp16_12, var_2056_cast_fp16_12))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2074_cast_fp16_13, var_2056_cast_fp16_13))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2074_cast_fp16_14, var_2056_cast_fp16_14))[name = tensor("aw_285_cast_fp16")]; + tensor aw_287_equation_0 = const()[name = tensor("aw_287_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_287_cast_fp16 = einsum(equation = aw_287_equation_0, values = (var_2074_cast_fp16_15, var_2056_cast_fp16_15))[name = tensor("aw_287_cast_fp16")]; + tensor var_2140_cast_fp16 = softmax(axis = var_2004, x = aw_257_cast_fp16)[name = tensor("op_2140_cast_fp16")]; + tensor var_2141_cast_fp16 = softmax(axis = var_2004, x = aw_259_cast_fp16)[name = tensor("op_2141_cast_fp16")]; + tensor var_2142_cast_fp16 = softmax(axis = var_2004, x = aw_261_cast_fp16)[name = tensor("op_2142_cast_fp16")]; + tensor var_2143_cast_fp16 = softmax(axis = var_2004, x = aw_263_cast_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2144_cast_fp16 = softmax(axis = var_2004, x = aw_265_cast_fp16)[name = tensor("op_2144_cast_fp16")]; + tensor var_2145_cast_fp16 = softmax(axis = var_2004, x = aw_267_cast_fp16)[name = tensor("op_2145_cast_fp16")]; + tensor var_2146_cast_fp16 = softmax(axis = var_2004, x = aw_269_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor var_2147_cast_fp16 = softmax(axis = var_2004, x = aw_271_cast_fp16)[name = tensor("op_2147_cast_fp16")]; + tensor var_2148_cast_fp16 = softmax(axis = var_2004, x = aw_273_cast_fp16)[name = tensor("op_2148_cast_fp16")]; + tensor var_2149_cast_fp16 = softmax(axis = var_2004, x = aw_275_cast_fp16)[name = tensor("op_2149_cast_fp16")]; + tensor var_2150_cast_fp16 = softmax(axis = var_2004, x = aw_277_cast_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2151_cast_fp16 = softmax(axis = var_2004, x = aw_279_cast_fp16)[name = tensor("op_2151_cast_fp16")]; + tensor var_2152_cast_fp16 = softmax(axis = var_2004, x = aw_281_cast_fp16)[name = tensor("op_2152_cast_fp16")]; + tensor var_2153_cast_fp16 = softmax(axis = var_2004, x = aw_283_cast_fp16)[name = tensor("op_2153_cast_fp16")]; + tensor var_2154_cast_fp16 = softmax(axis = var_2004, x = aw_285_cast_fp16)[name = tensor("op_2154_cast_fp16")]; + tensor var_2155_cast_fp16 = softmax(axis = var_2004, x = aw_287_cast_fp16)[name = tensor("op_2155_cast_fp16")]; + tensor var_2157_equation_0 = const()[name = tensor("op_2157_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2157_cast_fp16 = einsum(equation = var_2157_equation_0, values = (var_2091_cast_fp16_0, var_2140_cast_fp16))[name = tensor("op_2157_cast_fp16")]; + tensor var_2159_equation_0 = const()[name = tensor("op_2159_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2159_cast_fp16 = einsum(equation = var_2159_equation_0, values = (var_2091_cast_fp16_1, var_2141_cast_fp16))[name = tensor("op_2159_cast_fp16")]; + tensor var_2161_equation_0 = const()[name = tensor("op_2161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2161_cast_fp16 = einsum(equation = var_2161_equation_0, values = (var_2091_cast_fp16_2, var_2142_cast_fp16))[name = tensor("op_2161_cast_fp16")]; + tensor var_2163_equation_0 = const()[name = tensor("op_2163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2163_cast_fp16 = einsum(equation = var_2163_equation_0, values = (var_2091_cast_fp16_3, var_2143_cast_fp16))[name = tensor("op_2163_cast_fp16")]; + tensor var_2165_equation_0 = const()[name = tensor("op_2165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2165_cast_fp16 = einsum(equation = var_2165_equation_0, values = (var_2091_cast_fp16_4, var_2144_cast_fp16))[name = tensor("op_2165_cast_fp16")]; + tensor var_2167_equation_0 = const()[name = tensor("op_2167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2167_cast_fp16 = einsum(equation = var_2167_equation_0, values = (var_2091_cast_fp16_5, var_2145_cast_fp16))[name = tensor("op_2167_cast_fp16")]; + tensor var_2169_equation_0 = const()[name = tensor("op_2169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2169_cast_fp16 = einsum(equation = var_2169_equation_0, values = (var_2091_cast_fp16_6, var_2146_cast_fp16))[name = tensor("op_2169_cast_fp16")]; + tensor var_2171_equation_0 = const()[name = tensor("op_2171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2171_cast_fp16 = einsum(equation = var_2171_equation_0, values = (var_2091_cast_fp16_7, var_2147_cast_fp16))[name = tensor("op_2171_cast_fp16")]; + tensor var_2173_equation_0 = const()[name = tensor("op_2173_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2173_cast_fp16 = einsum(equation = var_2173_equation_0, values = (var_2091_cast_fp16_8, var_2148_cast_fp16))[name = tensor("op_2173_cast_fp16")]; + tensor var_2175_equation_0 = const()[name = tensor("op_2175_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2175_cast_fp16 = einsum(equation = var_2175_equation_0, values = (var_2091_cast_fp16_9, var_2149_cast_fp16))[name = tensor("op_2175_cast_fp16")]; + tensor var_2177_equation_0 = const()[name = tensor("op_2177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2177_cast_fp16 = einsum(equation = var_2177_equation_0, values = (var_2091_cast_fp16_10, var_2150_cast_fp16))[name = tensor("op_2177_cast_fp16")]; + tensor var_2179_equation_0 = const()[name = tensor("op_2179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2179_cast_fp16 = einsum(equation = var_2179_equation_0, values = (var_2091_cast_fp16_11, var_2151_cast_fp16))[name = tensor("op_2179_cast_fp16")]; + tensor var_2181_equation_0 = const()[name = tensor("op_2181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2181_cast_fp16 = einsum(equation = var_2181_equation_0, values = (var_2091_cast_fp16_12, var_2152_cast_fp16))[name = tensor("op_2181_cast_fp16")]; + tensor var_2183_equation_0 = const()[name = tensor("op_2183_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2183_cast_fp16 = einsum(equation = var_2183_equation_0, values = (var_2091_cast_fp16_13, var_2153_cast_fp16))[name = tensor("op_2183_cast_fp16")]; + tensor var_2185_equation_0 = const()[name = tensor("op_2185_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2185_cast_fp16 = einsum(equation = var_2185_equation_0, values = (var_2091_cast_fp16_14, var_2154_cast_fp16))[name = tensor("op_2185_cast_fp16")]; + tensor var_2187_equation_0 = const()[name = tensor("op_2187_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2187_cast_fp16 = einsum(equation = var_2187_equation_0, values = (var_2091_cast_fp16_15, var_2155_cast_fp16))[name = tensor("op_2187_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_2004, interleave = input_85_interleave_0, values = (var_2157_cast_fp16, var_2159_cast_fp16, var_2161_cast_fp16, var_2163_cast_fp16, var_2165_cast_fp16, var_2167_cast_fp16, var_2169_cast_fp16, var_2171_cast_fp16, var_2173_cast_fp16, var_2175_cast_fp16, var_2177_cast_fp16, var_2179_cast_fp16, var_2181_cast_fp16, var_2183_cast_fp16, var_2185_cast_fp16, var_2187_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_2196_pad_type_0 = const()[name = tensor("op_2196_pad_type_0"), val = tensor("valid")]; + tensor var_2196_strides_0 = const()[name = tensor("op_2196_strides_0"), val = tensor([1, 1])]; + tensor var_2196_pad_0 = const()[name = tensor("op_2196_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2196_dilations_0 = const()[name = tensor("op_2196_dilations_0"), val = tensor([1, 1])]; + tensor var_2196_groups_0 = const()[name = tensor("op_2196_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217690432)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219787648)))]; + tensor var_2196_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_2196_dilations_0, groups = var_2196_groups_0, pad = var_2196_pad_0, pad_type = var_2196_pad_type_0, strides = var_2196_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_2196_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219789760)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219791872)))]; + tensor var_2206_to_fp16 = const()[name = tensor("op_2206_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_2206_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219793984)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228182656)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2232_pad_type_0 = const()[name = tensor("op_2232_pad_type_0"), val = tensor("valid")]; + tensor var_2232_strides_0 = const()[name = tensor("op_2232_strides_0"), val = tensor([1, 1])]; + tensor var_2232_pad_0 = const()[name = tensor("op_2232_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2232_dilations_0 = const()[name = tensor("op_2232_dilations_0"), val = tensor([1, 1])]; + tensor var_2232_groups_0 = const()[name = tensor("op_2232_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228190912)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236579584)))]; + tensor var_2232_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_2232_dilations_0, groups = var_2232_groups_0, pad = var_2232_pad_0, pad_type = var_2232_pad_type_0, strides = var_2232_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_2232_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2241 = const()[name = tensor("op_2241"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236581696)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236583808)))]; + tensor var_2257_to_fp16 = const()[name = tensor("op_2257_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_2257_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_2292_weight_0_to_fp16 = const()[name = tensor("op_2292_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236585920)))]; + tensor var_2292_bias_0_to_fp16 = const()[name = tensor("op_2292_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238683136)))]; + tensor var_2292_cast_fp16 = conv(bias = var_2292_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_2292_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238685248)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_2290_pad_type_0 = const()[name = tensor("op_2290_pad_type_0"), val = tensor("valid")]; + tensor var_2290_strides_0 = const()[name = tensor("op_2290_strides_0"), val = tensor([1, 1])]; + tensor var_2290_pad_0 = const()[name = tensor("op_2290_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2290_dilations_0 = const()[name = tensor("op_2290_dilations_0"), val = tensor([1, 1])]; + tensor var_2290_groups_0 = const()[name = tensor("op_2290_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240782464)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242879680)))]; + tensor var_2290_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_2290_dilations_0, groups = var_2290_groups_0, pad = var_2290_pad_0, pad_type = var_2290_pad_type_0, strides = var_2290_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2290_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2293_axis_0 = const()[name = tensor("op_2293_axis_0"), val = tensor(1)]; + tensor var_2293_cast_fp16_0, tensor var_2293_cast_fp16_1, tensor var_2293_cast_fp16_2, tensor var_2293_cast_fp16_3, tensor var_2293_cast_fp16_4, tensor var_2293_cast_fp16_5, tensor var_2293_cast_fp16_6, tensor var_2293_cast_fp16_7, tensor var_2293_cast_fp16_8, tensor var_2293_cast_fp16_9, tensor var_2293_cast_fp16_10, tensor var_2293_cast_fp16_11, tensor var_2293_cast_fp16_12, tensor var_2293_cast_fp16_13, tensor var_2293_cast_fp16_14, tensor var_2293_cast_fp16_15 = split(axis = var_2293_axis_0, split_sizes = tile_27, x = var_2292_cast_fp16)[name = tensor("op_2293_cast_fp16")]; + tensor var_2310_perm_0 = const()[name = tensor("op_2310_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2311_axis_0 = const()[name = tensor("op_2311_axis_0"), val = tensor(3)]; + tensor var_2310_cast_fp16 = transpose(perm = var_2310_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_15")]; + tensor var_2311_cast_fp16_0, tensor var_2311_cast_fp16_1, tensor var_2311_cast_fp16_2, tensor var_2311_cast_fp16_3, tensor var_2311_cast_fp16_4, tensor var_2311_cast_fp16_5, tensor var_2311_cast_fp16_6, tensor var_2311_cast_fp16_7, tensor var_2311_cast_fp16_8, tensor var_2311_cast_fp16_9, tensor var_2311_cast_fp16_10, tensor var_2311_cast_fp16_11, tensor var_2311_cast_fp16_12, tensor var_2311_cast_fp16_13, tensor var_2311_cast_fp16_14, tensor var_2311_cast_fp16_15 = split(axis = var_2311_axis_0, split_sizes = tile_28, x = var_2310_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2328_axis_0 = const()[name = tensor("op_2328_axis_0"), val = tensor(1)]; + tensor var_2328_cast_fp16_0, tensor var_2328_cast_fp16_1, tensor var_2328_cast_fp16_2, tensor var_2328_cast_fp16_3, tensor var_2328_cast_fp16_4, tensor var_2328_cast_fp16_5, tensor var_2328_cast_fp16_6, tensor var_2328_cast_fp16_7, tensor var_2328_cast_fp16_8, tensor var_2328_cast_fp16_9, tensor var_2328_cast_fp16_10, tensor var_2328_cast_fp16_11, tensor var_2328_cast_fp16_12, tensor var_2328_cast_fp16_13, tensor var_2328_cast_fp16_14, tensor var_2328_cast_fp16_15 = split(axis = var_2328_axis_0, split_sizes = tile_29, x = var_2290_cast_fp16)[name = tensor("op_2328_cast_fp16")]; + tensor aw_289_equation_0 = const()[name = tensor("aw_289_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_289_cast_fp16 = einsum(equation = aw_289_equation_0, values = (var_2311_cast_fp16_0, var_2293_cast_fp16_0))[name = tensor("aw_289_cast_fp16")]; + tensor aw_291_equation_0 = const()[name = tensor("aw_291_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_291_cast_fp16 = einsum(equation = aw_291_equation_0, values = (var_2311_cast_fp16_1, var_2293_cast_fp16_1))[name = tensor("aw_291_cast_fp16")]; + tensor aw_293_equation_0 = const()[name = tensor("aw_293_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_293_cast_fp16 = einsum(equation = aw_293_equation_0, values = (var_2311_cast_fp16_2, var_2293_cast_fp16_2))[name = tensor("aw_293_cast_fp16")]; + tensor aw_295_equation_0 = const()[name = tensor("aw_295_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_295_cast_fp16 = einsum(equation = aw_295_equation_0, values = (var_2311_cast_fp16_3, var_2293_cast_fp16_3))[name = tensor("aw_295_cast_fp16")]; + tensor aw_297_equation_0 = const()[name = tensor("aw_297_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_297_cast_fp16 = einsum(equation = aw_297_equation_0, values = (var_2311_cast_fp16_4, var_2293_cast_fp16_4))[name = tensor("aw_297_cast_fp16")]; + tensor aw_299_equation_0 = const()[name = tensor("aw_299_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_299_cast_fp16 = einsum(equation = aw_299_equation_0, values = (var_2311_cast_fp16_5, var_2293_cast_fp16_5))[name = tensor("aw_299_cast_fp16")]; + tensor aw_301_equation_0 = const()[name = tensor("aw_301_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_301_cast_fp16 = einsum(equation = aw_301_equation_0, values = (var_2311_cast_fp16_6, var_2293_cast_fp16_6))[name = tensor("aw_301_cast_fp16")]; + tensor aw_303_equation_0 = const()[name = tensor("aw_303_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_303_cast_fp16 = einsum(equation = aw_303_equation_0, values = (var_2311_cast_fp16_7, var_2293_cast_fp16_7))[name = tensor("aw_303_cast_fp16")]; + tensor aw_305_equation_0 = const()[name = tensor("aw_305_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_305_cast_fp16 = einsum(equation = aw_305_equation_0, values = (var_2311_cast_fp16_8, var_2293_cast_fp16_8))[name = tensor("aw_305_cast_fp16")]; + tensor aw_307_equation_0 = const()[name = tensor("aw_307_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_307_cast_fp16 = einsum(equation = aw_307_equation_0, values = (var_2311_cast_fp16_9, var_2293_cast_fp16_9))[name = tensor("aw_307_cast_fp16")]; + tensor aw_309_equation_0 = const()[name = tensor("aw_309_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_309_cast_fp16 = einsum(equation = aw_309_equation_0, values = (var_2311_cast_fp16_10, var_2293_cast_fp16_10))[name = tensor("aw_309_cast_fp16")]; + tensor aw_311_equation_0 = const()[name = tensor("aw_311_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_311_cast_fp16 = einsum(equation = aw_311_equation_0, values = (var_2311_cast_fp16_11, var_2293_cast_fp16_11))[name = tensor("aw_311_cast_fp16")]; + tensor aw_313_equation_0 = const()[name = tensor("aw_313_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_313_cast_fp16 = einsum(equation = aw_313_equation_0, values = (var_2311_cast_fp16_12, var_2293_cast_fp16_12))[name = tensor("aw_313_cast_fp16")]; + tensor aw_315_equation_0 = const()[name = tensor("aw_315_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_315_cast_fp16 = einsum(equation = aw_315_equation_0, values = (var_2311_cast_fp16_13, var_2293_cast_fp16_13))[name = tensor("aw_315_cast_fp16")]; + tensor aw_317_equation_0 = const()[name = tensor("aw_317_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_317_cast_fp16 = einsum(equation = aw_317_equation_0, values = (var_2311_cast_fp16_14, var_2293_cast_fp16_14))[name = tensor("aw_317_cast_fp16")]; + tensor aw_319_equation_0 = const()[name = tensor("aw_319_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_319_cast_fp16 = einsum(equation = aw_319_equation_0, values = (var_2311_cast_fp16_15, var_2293_cast_fp16_15))[name = tensor("aw_319_cast_fp16")]; + tensor var_2377_cast_fp16 = softmax(axis = var_2241, x = aw_289_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_cast_fp16 = softmax(axis = var_2241, x = aw_291_cast_fp16)[name = tensor("op_2378_cast_fp16")]; + tensor var_2379_cast_fp16 = softmax(axis = var_2241, x = aw_293_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = softmax(axis = var_2241, x = aw_295_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor var_2381_cast_fp16 = softmax(axis = var_2241, x = aw_297_cast_fp16)[name = tensor("op_2381_cast_fp16")]; + tensor var_2382_cast_fp16 = softmax(axis = var_2241, x = aw_299_cast_fp16)[name = tensor("op_2382_cast_fp16")]; + tensor var_2383_cast_fp16 = softmax(axis = var_2241, x = aw_301_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = softmax(axis = var_2241, x = aw_303_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor var_2385_cast_fp16 = softmax(axis = var_2241, x = aw_305_cast_fp16)[name = tensor("op_2385_cast_fp16")]; + tensor var_2386_cast_fp16 = softmax(axis = var_2241, x = aw_307_cast_fp16)[name = tensor("op_2386_cast_fp16")]; + tensor var_2387_cast_fp16 = softmax(axis = var_2241, x = aw_309_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = softmax(axis = var_2241, x = aw_311_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor var_2389_cast_fp16 = softmax(axis = var_2241, x = aw_313_cast_fp16)[name = tensor("op_2389_cast_fp16")]; + tensor var_2390_cast_fp16 = softmax(axis = var_2241, x = aw_315_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor var_2391_cast_fp16 = softmax(axis = var_2241, x = aw_317_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = softmax(axis = var_2241, x = aw_319_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor var_2394_equation_0 = const()[name = tensor("op_2394_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2394_cast_fp16 = einsum(equation = var_2394_equation_0, values = (var_2328_cast_fp16_0, var_2377_cast_fp16))[name = tensor("op_2394_cast_fp16")]; + tensor var_2396_equation_0 = const()[name = tensor("op_2396_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2396_cast_fp16 = einsum(equation = var_2396_equation_0, values = (var_2328_cast_fp16_1, var_2378_cast_fp16))[name = tensor("op_2396_cast_fp16")]; + tensor var_2398_equation_0 = const()[name = tensor("op_2398_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2398_cast_fp16 = einsum(equation = var_2398_equation_0, values = (var_2328_cast_fp16_2, var_2379_cast_fp16))[name = tensor("op_2398_cast_fp16")]; + tensor var_2400_equation_0 = const()[name = tensor("op_2400_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2400_cast_fp16 = einsum(equation = var_2400_equation_0, values = (var_2328_cast_fp16_3, var_2380_cast_fp16))[name = tensor("op_2400_cast_fp16")]; + tensor var_2402_equation_0 = const()[name = tensor("op_2402_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2402_cast_fp16 = einsum(equation = var_2402_equation_0, values = (var_2328_cast_fp16_4, var_2381_cast_fp16))[name = tensor("op_2402_cast_fp16")]; + tensor var_2404_equation_0 = const()[name = tensor("op_2404_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2404_cast_fp16 = einsum(equation = var_2404_equation_0, values = (var_2328_cast_fp16_5, var_2382_cast_fp16))[name = tensor("op_2404_cast_fp16")]; + tensor var_2406_equation_0 = const()[name = tensor("op_2406_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2406_cast_fp16 = einsum(equation = var_2406_equation_0, values = (var_2328_cast_fp16_6, var_2383_cast_fp16))[name = tensor("op_2406_cast_fp16")]; + tensor var_2408_equation_0 = const()[name = tensor("op_2408_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2408_cast_fp16 = einsum(equation = var_2408_equation_0, values = (var_2328_cast_fp16_7, var_2384_cast_fp16))[name = tensor("op_2408_cast_fp16")]; + tensor var_2410_equation_0 = const()[name = tensor("op_2410_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2410_cast_fp16 = einsum(equation = var_2410_equation_0, values = (var_2328_cast_fp16_8, var_2385_cast_fp16))[name = tensor("op_2410_cast_fp16")]; + tensor var_2412_equation_0 = const()[name = tensor("op_2412_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2412_cast_fp16 = einsum(equation = var_2412_equation_0, values = (var_2328_cast_fp16_9, var_2386_cast_fp16))[name = tensor("op_2412_cast_fp16")]; + tensor var_2414_equation_0 = const()[name = tensor("op_2414_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2414_cast_fp16 = einsum(equation = var_2414_equation_0, values = (var_2328_cast_fp16_10, var_2387_cast_fp16))[name = tensor("op_2414_cast_fp16")]; + tensor var_2416_equation_0 = const()[name = tensor("op_2416_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2416_cast_fp16 = einsum(equation = var_2416_equation_0, values = (var_2328_cast_fp16_11, var_2388_cast_fp16))[name = tensor("op_2416_cast_fp16")]; + tensor var_2418_equation_0 = const()[name = tensor("op_2418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2418_cast_fp16 = einsum(equation = var_2418_equation_0, values = (var_2328_cast_fp16_12, var_2389_cast_fp16))[name = tensor("op_2418_cast_fp16")]; + tensor var_2420_equation_0 = const()[name = tensor("op_2420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2420_cast_fp16 = einsum(equation = var_2420_equation_0, values = (var_2328_cast_fp16_13, var_2390_cast_fp16))[name = tensor("op_2420_cast_fp16")]; + tensor var_2422_equation_0 = const()[name = tensor("op_2422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2422_cast_fp16 = einsum(equation = var_2422_equation_0, values = (var_2328_cast_fp16_14, var_2391_cast_fp16))[name = tensor("op_2422_cast_fp16")]; + tensor var_2424_equation_0 = const()[name = tensor("op_2424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2424_cast_fp16 = einsum(equation = var_2424_equation_0, values = (var_2328_cast_fp16_15, var_2392_cast_fp16))[name = tensor("op_2424_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_2241, interleave = input_95_interleave_0, values = (var_2394_cast_fp16, var_2396_cast_fp16, var_2398_cast_fp16, var_2400_cast_fp16, var_2402_cast_fp16, var_2404_cast_fp16, var_2406_cast_fp16, var_2408_cast_fp16, var_2410_cast_fp16, var_2412_cast_fp16, var_2414_cast_fp16, var_2416_cast_fp16, var_2418_cast_fp16, var_2420_cast_fp16, var_2422_cast_fp16, var_2424_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2433_pad_type_0 = const()[name = tensor("op_2433_pad_type_0"), val = tensor("valid")]; + tensor var_2433_strides_0 = const()[name = tensor("op_2433_strides_0"), val = tensor([1, 1])]; + tensor var_2433_pad_0 = const()[name = tensor("op_2433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2433_dilations_0 = const()[name = tensor("op_2433_dilations_0"), val = tensor([1, 1])]; + tensor var_2433_groups_0 = const()[name = tensor("op_2433_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242881792)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244979008)))]; + tensor var_2433_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2433_dilations_0, groups = var_2433_groups_0, pad = var_2433_pad_0, pad_type = var_2433_pad_type_0, strides = var_2433_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2433_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2433_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244981120)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244983232)))]; + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2443_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244985344)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253374016)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2469_pad_type_0 = const()[name = tensor("op_2469_pad_type_0"), val = tensor("valid")]; + tensor var_2469_strides_0 = const()[name = tensor("op_2469_strides_0"), val = tensor([1, 1])]; + tensor var_2469_pad_0 = const()[name = tensor("op_2469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2469_dilations_0 = const()[name = tensor("op_2469_dilations_0"), val = tensor([1, 1])]; + tensor var_2469_groups_0 = const()[name = tensor("op_2469_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253382272)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261770944)))]; + tensor var_2469_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2469_dilations_0, groups = var_2469_groups_0, pad = var_2469_pad_0, pad_type = var_2469_pad_type_0, strides = var_2469_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2469_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2469_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2478 = const()[name = tensor("op_2478"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261773056)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261775168)))]; + tensor var_2494_to_fp16 = const()[name = tensor("op_2494_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2494_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2529_weight_0_to_fp16 = const()[name = tensor("op_2529_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261777280)))]; + tensor var_2529_bias_0_to_fp16 = const()[name = tensor("op_2529_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263874496)))]; + tensor var_2529_cast_fp16 = conv(bias = var_2529_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2529_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263876608)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2527_pad_type_0 = const()[name = tensor("op_2527_pad_type_0"), val = tensor("valid")]; + tensor var_2527_strides_0 = const()[name = tensor("op_2527_strides_0"), val = tensor([1, 1])]; + tensor var_2527_pad_0 = const()[name = tensor("op_2527_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2527_dilations_0 = const()[name = tensor("op_2527_dilations_0"), val = tensor([1, 1])]; + tensor var_2527_groups_0 = const()[name = tensor("op_2527_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265973824)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268071040)))]; + tensor var_2527_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2527_dilations_0, groups = var_2527_groups_0, pad = var_2527_pad_0, pad_type = var_2527_pad_type_0, strides = var_2527_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2530_axis_0 = const()[name = tensor("op_2530_axis_0"), val = tensor(1)]; + tensor var_2530_cast_fp16_0, tensor var_2530_cast_fp16_1, tensor var_2530_cast_fp16_2, tensor var_2530_cast_fp16_3, tensor var_2530_cast_fp16_4, tensor var_2530_cast_fp16_5, tensor var_2530_cast_fp16_6, tensor var_2530_cast_fp16_7, tensor var_2530_cast_fp16_8, tensor var_2530_cast_fp16_9, tensor var_2530_cast_fp16_10, tensor var_2530_cast_fp16_11, tensor var_2530_cast_fp16_12, tensor var_2530_cast_fp16_13, tensor var_2530_cast_fp16_14, tensor var_2530_cast_fp16_15 = split(axis = var_2530_axis_0, split_sizes = tile_30, x = var_2529_cast_fp16)[name = tensor("op_2530_cast_fp16")]; + tensor var_2547_perm_0 = const()[name = tensor("op_2547_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2548_axis_0 = const()[name = tensor("op_2548_axis_0"), val = tensor(3)]; + tensor var_2547_cast_fp16 = transpose(perm = var_2547_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_14")]; + tensor var_2548_cast_fp16_0, tensor var_2548_cast_fp16_1, tensor var_2548_cast_fp16_2, tensor var_2548_cast_fp16_3, tensor var_2548_cast_fp16_4, tensor var_2548_cast_fp16_5, tensor var_2548_cast_fp16_6, tensor var_2548_cast_fp16_7, tensor var_2548_cast_fp16_8, tensor var_2548_cast_fp16_9, tensor var_2548_cast_fp16_10, tensor var_2548_cast_fp16_11, tensor var_2548_cast_fp16_12, tensor var_2548_cast_fp16_13, tensor var_2548_cast_fp16_14, tensor var_2548_cast_fp16_15 = split(axis = var_2548_axis_0, split_sizes = tile_31, x = var_2547_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2565_axis_0 = const()[name = tensor("op_2565_axis_0"), val = tensor(1)]; + tensor var_2565_cast_fp16_0, tensor var_2565_cast_fp16_1, tensor var_2565_cast_fp16_2, tensor var_2565_cast_fp16_3, tensor var_2565_cast_fp16_4, tensor var_2565_cast_fp16_5, tensor var_2565_cast_fp16_6, tensor var_2565_cast_fp16_7, tensor var_2565_cast_fp16_8, tensor var_2565_cast_fp16_9, tensor var_2565_cast_fp16_10, tensor var_2565_cast_fp16_11, tensor var_2565_cast_fp16_12, tensor var_2565_cast_fp16_13, tensor var_2565_cast_fp16_14, tensor var_2565_cast_fp16_15 = split(axis = var_2565_axis_0, split_sizes = tile_32, x = var_2527_cast_fp16)[name = tensor("op_2565_cast_fp16")]; + tensor aw_321_equation_0 = const()[name = tensor("aw_321_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_321_cast_fp16 = einsum(equation = aw_321_equation_0, values = (var_2548_cast_fp16_0, var_2530_cast_fp16_0))[name = tensor("aw_321_cast_fp16")]; + tensor aw_323_equation_0 = const()[name = tensor("aw_323_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_323_cast_fp16 = einsum(equation = aw_323_equation_0, values = (var_2548_cast_fp16_1, var_2530_cast_fp16_1))[name = tensor("aw_323_cast_fp16")]; + tensor aw_325_equation_0 = const()[name = tensor("aw_325_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_325_cast_fp16 = einsum(equation = aw_325_equation_0, values = (var_2548_cast_fp16_2, var_2530_cast_fp16_2))[name = tensor("aw_325_cast_fp16")]; + tensor aw_327_equation_0 = const()[name = tensor("aw_327_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_327_cast_fp16 = einsum(equation = aw_327_equation_0, values = (var_2548_cast_fp16_3, var_2530_cast_fp16_3))[name = tensor("aw_327_cast_fp16")]; + tensor aw_329_equation_0 = const()[name = tensor("aw_329_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_329_cast_fp16 = einsum(equation = aw_329_equation_0, values = (var_2548_cast_fp16_4, var_2530_cast_fp16_4))[name = tensor("aw_329_cast_fp16")]; + tensor aw_331_equation_0 = const()[name = tensor("aw_331_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_331_cast_fp16 = einsum(equation = aw_331_equation_0, values = (var_2548_cast_fp16_5, var_2530_cast_fp16_5))[name = tensor("aw_331_cast_fp16")]; + tensor aw_333_equation_0 = const()[name = tensor("aw_333_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_333_cast_fp16 = einsum(equation = aw_333_equation_0, values = (var_2548_cast_fp16_6, var_2530_cast_fp16_6))[name = tensor("aw_333_cast_fp16")]; + tensor aw_335_equation_0 = const()[name = tensor("aw_335_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_335_cast_fp16 = einsum(equation = aw_335_equation_0, values = (var_2548_cast_fp16_7, var_2530_cast_fp16_7))[name = tensor("aw_335_cast_fp16")]; + tensor aw_337_equation_0 = const()[name = tensor("aw_337_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_337_cast_fp16 = einsum(equation = aw_337_equation_0, values = (var_2548_cast_fp16_8, var_2530_cast_fp16_8))[name = tensor("aw_337_cast_fp16")]; + tensor aw_339_equation_0 = const()[name = tensor("aw_339_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_339_cast_fp16 = einsum(equation = aw_339_equation_0, values = (var_2548_cast_fp16_9, var_2530_cast_fp16_9))[name = tensor("aw_339_cast_fp16")]; + tensor aw_341_equation_0 = const()[name = tensor("aw_341_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_341_cast_fp16 = einsum(equation = aw_341_equation_0, values = (var_2548_cast_fp16_10, var_2530_cast_fp16_10))[name = tensor("aw_341_cast_fp16")]; + tensor aw_343_equation_0 = const()[name = tensor("aw_343_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_343_cast_fp16 = einsum(equation = aw_343_equation_0, values = (var_2548_cast_fp16_11, var_2530_cast_fp16_11))[name = tensor("aw_343_cast_fp16")]; + tensor aw_345_equation_0 = const()[name = tensor("aw_345_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_345_cast_fp16 = einsum(equation = aw_345_equation_0, values = (var_2548_cast_fp16_12, var_2530_cast_fp16_12))[name = tensor("aw_345_cast_fp16")]; + tensor aw_347_equation_0 = const()[name = tensor("aw_347_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_347_cast_fp16 = einsum(equation = aw_347_equation_0, values = (var_2548_cast_fp16_13, var_2530_cast_fp16_13))[name = tensor("aw_347_cast_fp16")]; + tensor aw_349_equation_0 = const()[name = tensor("aw_349_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_349_cast_fp16 = einsum(equation = aw_349_equation_0, values = (var_2548_cast_fp16_14, var_2530_cast_fp16_14))[name = tensor("aw_349_cast_fp16")]; + tensor aw_351_equation_0 = const()[name = tensor("aw_351_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_351_cast_fp16 = einsum(equation = aw_351_equation_0, values = (var_2548_cast_fp16_15, var_2530_cast_fp16_15))[name = tensor("aw_351_cast_fp16")]; + tensor var_2614_cast_fp16 = softmax(axis = var_2478, x = aw_321_cast_fp16)[name = tensor("op_2614_cast_fp16")]; + tensor var_2615_cast_fp16 = softmax(axis = var_2478, x = aw_323_cast_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2616_cast_fp16 = softmax(axis = var_2478, x = aw_325_cast_fp16)[name = tensor("op_2616_cast_fp16")]; + tensor var_2617_cast_fp16 = softmax(axis = var_2478, x = aw_327_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor var_2618_cast_fp16 = softmax(axis = var_2478, x = aw_329_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2619_cast_fp16 = softmax(axis = var_2478, x = aw_331_cast_fp16)[name = tensor("op_2619_cast_fp16")]; + tensor var_2620_cast_fp16 = softmax(axis = var_2478, x = aw_333_cast_fp16)[name = tensor("op_2620_cast_fp16")]; + tensor var_2621_cast_fp16 = softmax(axis = var_2478, x = aw_335_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_cast_fp16 = softmax(axis = var_2478, x = aw_337_cast_fp16)[name = tensor("op_2622_cast_fp16")]; + tensor var_2623_cast_fp16 = softmax(axis = var_2478, x = aw_339_cast_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor var_2624_cast_fp16 = softmax(axis = var_2478, x = aw_341_cast_fp16)[name = tensor("op_2624_cast_fp16")]; + tensor var_2625_cast_fp16 = softmax(axis = var_2478, x = aw_343_cast_fp16)[name = tensor("op_2625_cast_fp16")]; + tensor var_2626_cast_fp16 = softmax(axis = var_2478, x = aw_345_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2627_cast_fp16 = softmax(axis = var_2478, x = aw_347_cast_fp16)[name = tensor("op_2627_cast_fp16")]; + tensor var_2628_cast_fp16 = softmax(axis = var_2478, x = aw_349_cast_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = softmax(axis = var_2478, x = aw_351_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor var_2631_equation_0 = const()[name = tensor("op_2631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2631_cast_fp16 = einsum(equation = var_2631_equation_0, values = (var_2565_cast_fp16_0, var_2614_cast_fp16))[name = tensor("op_2631_cast_fp16")]; + tensor var_2633_equation_0 = const()[name = tensor("op_2633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2633_cast_fp16 = einsum(equation = var_2633_equation_0, values = (var_2565_cast_fp16_1, var_2615_cast_fp16))[name = tensor("op_2633_cast_fp16")]; + tensor var_2635_equation_0 = const()[name = tensor("op_2635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2635_cast_fp16 = einsum(equation = var_2635_equation_0, values = (var_2565_cast_fp16_2, var_2616_cast_fp16))[name = tensor("op_2635_cast_fp16")]; + tensor var_2637_equation_0 = const()[name = tensor("op_2637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2637_cast_fp16 = einsum(equation = var_2637_equation_0, values = (var_2565_cast_fp16_3, var_2617_cast_fp16))[name = tensor("op_2637_cast_fp16")]; + tensor var_2639_equation_0 = const()[name = tensor("op_2639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2639_cast_fp16 = einsum(equation = var_2639_equation_0, values = (var_2565_cast_fp16_4, var_2618_cast_fp16))[name = tensor("op_2639_cast_fp16")]; + tensor var_2641_equation_0 = const()[name = tensor("op_2641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2641_cast_fp16 = einsum(equation = var_2641_equation_0, values = (var_2565_cast_fp16_5, var_2619_cast_fp16))[name = tensor("op_2641_cast_fp16")]; + tensor var_2643_equation_0 = const()[name = tensor("op_2643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2643_cast_fp16 = einsum(equation = var_2643_equation_0, values = (var_2565_cast_fp16_6, var_2620_cast_fp16))[name = tensor("op_2643_cast_fp16")]; + tensor var_2645_equation_0 = const()[name = tensor("op_2645_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2645_cast_fp16 = einsum(equation = var_2645_equation_0, values = (var_2565_cast_fp16_7, var_2621_cast_fp16))[name = tensor("op_2645_cast_fp16")]; + tensor var_2647_equation_0 = const()[name = tensor("op_2647_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2647_cast_fp16 = einsum(equation = var_2647_equation_0, values = (var_2565_cast_fp16_8, var_2622_cast_fp16))[name = tensor("op_2647_cast_fp16")]; + tensor var_2649_equation_0 = const()[name = tensor("op_2649_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2649_cast_fp16 = einsum(equation = var_2649_equation_0, values = (var_2565_cast_fp16_9, var_2623_cast_fp16))[name = tensor("op_2649_cast_fp16")]; + tensor var_2651_equation_0 = const()[name = tensor("op_2651_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2651_cast_fp16 = einsum(equation = var_2651_equation_0, values = (var_2565_cast_fp16_10, var_2624_cast_fp16))[name = tensor("op_2651_cast_fp16")]; + tensor var_2653_equation_0 = const()[name = tensor("op_2653_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2653_cast_fp16 = einsum(equation = var_2653_equation_0, values = (var_2565_cast_fp16_11, var_2625_cast_fp16))[name = tensor("op_2653_cast_fp16")]; + tensor var_2655_equation_0 = const()[name = tensor("op_2655_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2655_cast_fp16 = einsum(equation = var_2655_equation_0, values = (var_2565_cast_fp16_12, var_2626_cast_fp16))[name = tensor("op_2655_cast_fp16")]; + tensor var_2657_equation_0 = const()[name = tensor("op_2657_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2657_cast_fp16 = einsum(equation = var_2657_equation_0, values = (var_2565_cast_fp16_13, var_2627_cast_fp16))[name = tensor("op_2657_cast_fp16")]; + tensor var_2659_equation_0 = const()[name = tensor("op_2659_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2659_cast_fp16 = einsum(equation = var_2659_equation_0, values = (var_2565_cast_fp16_14, var_2628_cast_fp16))[name = tensor("op_2659_cast_fp16")]; + tensor var_2661_equation_0 = const()[name = tensor("op_2661_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2661_cast_fp16 = einsum(equation = var_2661_equation_0, values = (var_2565_cast_fp16_15, var_2629_cast_fp16))[name = tensor("op_2661_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2478, interleave = input_105_interleave_0, values = (var_2631_cast_fp16, var_2633_cast_fp16, var_2635_cast_fp16, var_2637_cast_fp16, var_2639_cast_fp16, var_2641_cast_fp16, var_2643_cast_fp16, var_2645_cast_fp16, var_2647_cast_fp16, var_2649_cast_fp16, var_2651_cast_fp16, var_2653_cast_fp16, var_2655_cast_fp16, var_2657_cast_fp16, var_2659_cast_fp16, var_2661_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_2670_pad_type_0 = const()[name = tensor("op_2670_pad_type_0"), val = tensor("valid")]; + tensor var_2670_strides_0 = const()[name = tensor("op_2670_strides_0"), val = tensor([1, 1])]; + tensor var_2670_pad_0 = const()[name = tensor("op_2670_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2670_dilations_0 = const()[name = tensor("op_2670_dilations_0"), val = tensor([1, 1])]; + tensor var_2670_groups_0 = const()[name = tensor("op_2670_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268073152)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270170368)))]; + tensor var_2670_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_2670_dilations_0, groups = var_2670_groups_0, pad = var_2670_pad_0, pad_type = var_2670_pad_type_0, strides = var_2670_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_2670_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270172480)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270174592)))]; + tensor var_2680_to_fp16 = const()[name = tensor("op_2680_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_2680_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270176704)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278565376)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_2706_pad_type_0 = const()[name = tensor("op_2706_pad_type_0"), val = tensor("valid")]; + tensor var_2706_strides_0 = const()[name = tensor("op_2706_strides_0"), val = tensor([1, 1])]; + tensor var_2706_pad_0 = const()[name = tensor("op_2706_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2706_dilations_0 = const()[name = tensor("op_2706_dilations_0"), val = tensor([1, 1])]; + tensor var_2706_groups_0 = const()[name = tensor("op_2706_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278573632)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286962304)))]; + tensor var_2706_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_2706_dilations_0, groups = var_2706_groups_0, pad = var_2706_pad_0, pad_type = var_2706_pad_type_0, strides = var_2706_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_2706_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_2715 = const()[name = tensor("op_2715"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286964416)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286966528)))]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_2731_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_23_pad_type_0 = const()[name = tensor("q_23_pad_type_0"), val = tensor("valid")]; + tensor q_23_strides_0 = const()[name = tensor("q_23_strides_0"), val = tensor([1, 1])]; + tensor q_23_pad_0 = const()[name = tensor("q_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_23_dilations_0 = const()[name = tensor("q_23_dilations_0"), val = tensor([1, 1])]; + tensor q_23_groups_0 = const()[name = tensor("q_23_groups_0"), val = tensor(1)]; + tensor var_2766_weight_0_to_fp16 = const()[name = tensor("op_2766_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286968640)))]; + tensor var_2766_bias_0_to_fp16 = const()[name = tensor("op_2766_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289065856)))]; + tensor var_2766_cast_fp16 = conv(bias = var_2766_bias_0_to_fp16, dilations = q_23_dilations_0, groups = q_23_groups_0, pad = q_23_pad_0, pad_type = q_23_pad_type_0, strides = q_23_strides_0, weight = var_2766_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor k_23_pad_type_0 = const()[name = tensor("k_23_pad_type_0"), val = tensor("valid")]; + tensor k_23_strides_0 = const()[name = tensor("k_23_strides_0"), val = tensor([1, 1])]; + tensor k_23_pad_0 = const()[name = tensor("k_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_23_dilations_0 = const()[name = tensor("k_23_dilations_0"), val = tensor([1, 1])]; + tensor k_23_groups_0 = const()[name = tensor("k_23_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289067968)))]; + tensor k_23_cast_fp16 = conv(dilations = k_23_dilations_0, groups = k_23_groups_0, pad = k_23_pad_0, pad_type = k_23_pad_type_0, strides = k_23_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_2764_pad_type_0 = const()[name = tensor("op_2764_pad_type_0"), val = tensor("valid")]; + tensor var_2764_strides_0 = const()[name = tensor("op_2764_strides_0"), val = tensor([1, 1])]; + tensor var_2764_pad_0 = const()[name = tensor("op_2764_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2764_dilations_0 = const()[name = tensor("op_2764_dilations_0"), val = tensor([1, 1])]; + tensor var_2764_groups_0 = const()[name = tensor("op_2764_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291165184)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293262400)))]; + tensor var_2764_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_2764_dilations_0, groups = var_2764_groups_0, pad = var_2764_pad_0, pad_type = var_2764_pad_type_0, strides = var_2764_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2767_axis_0 = const()[name = tensor("op_2767_axis_0"), val = tensor(1)]; + tensor var_2767_cast_fp16_0, tensor var_2767_cast_fp16_1, tensor var_2767_cast_fp16_2, tensor var_2767_cast_fp16_3, tensor var_2767_cast_fp16_4, tensor var_2767_cast_fp16_5, tensor var_2767_cast_fp16_6, tensor var_2767_cast_fp16_7, tensor var_2767_cast_fp16_8, tensor var_2767_cast_fp16_9, tensor var_2767_cast_fp16_10, tensor var_2767_cast_fp16_11, tensor var_2767_cast_fp16_12, tensor var_2767_cast_fp16_13, tensor var_2767_cast_fp16_14, tensor var_2767_cast_fp16_15 = split(axis = var_2767_axis_0, split_sizes = tile_33, x = var_2766_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor var_2784_perm_0 = const()[name = tensor("op_2784_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2785_axis_0 = const()[name = tensor("op_2785_axis_0"), val = tensor(3)]; + tensor var_2784_cast_fp16 = transpose(perm = var_2784_perm_0, x = k_23_cast_fp16)[name = tensor("transpose_13")]; + tensor var_2785_cast_fp16_0, tensor var_2785_cast_fp16_1, tensor var_2785_cast_fp16_2, tensor var_2785_cast_fp16_3, tensor var_2785_cast_fp16_4, tensor var_2785_cast_fp16_5, tensor var_2785_cast_fp16_6, tensor var_2785_cast_fp16_7, tensor var_2785_cast_fp16_8, tensor var_2785_cast_fp16_9, tensor var_2785_cast_fp16_10, tensor var_2785_cast_fp16_11, tensor var_2785_cast_fp16_12, tensor var_2785_cast_fp16_13, tensor var_2785_cast_fp16_14, tensor var_2785_cast_fp16_15 = split(axis = var_2785_axis_0, split_sizes = tile_34, x = var_2784_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2802_axis_0 = const()[name = tensor("op_2802_axis_0"), val = tensor(1)]; + tensor var_2802_cast_fp16_0, tensor var_2802_cast_fp16_1, tensor var_2802_cast_fp16_2, tensor var_2802_cast_fp16_3, tensor var_2802_cast_fp16_4, tensor var_2802_cast_fp16_5, tensor var_2802_cast_fp16_6, tensor var_2802_cast_fp16_7, tensor var_2802_cast_fp16_8, tensor var_2802_cast_fp16_9, tensor var_2802_cast_fp16_10, tensor var_2802_cast_fp16_11, tensor var_2802_cast_fp16_12, tensor var_2802_cast_fp16_13, tensor var_2802_cast_fp16_14, tensor var_2802_cast_fp16_15 = split(axis = var_2802_axis_0, split_sizes = tile_35, x = var_2764_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor aw_353_equation_0 = const()[name = tensor("aw_353_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_353_cast_fp16 = einsum(equation = aw_353_equation_0, values = (var_2785_cast_fp16_0, var_2767_cast_fp16_0))[name = tensor("aw_353_cast_fp16")]; + tensor aw_355_equation_0 = const()[name = tensor("aw_355_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_355_cast_fp16 = einsum(equation = aw_355_equation_0, values = (var_2785_cast_fp16_1, var_2767_cast_fp16_1))[name = tensor("aw_355_cast_fp16")]; + tensor aw_357_equation_0 = const()[name = tensor("aw_357_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_357_cast_fp16 = einsum(equation = aw_357_equation_0, values = (var_2785_cast_fp16_2, var_2767_cast_fp16_2))[name = tensor("aw_357_cast_fp16")]; + tensor aw_359_equation_0 = const()[name = tensor("aw_359_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_359_cast_fp16 = einsum(equation = aw_359_equation_0, values = (var_2785_cast_fp16_3, var_2767_cast_fp16_3))[name = tensor("aw_359_cast_fp16")]; + tensor aw_361_equation_0 = const()[name = tensor("aw_361_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_361_cast_fp16 = einsum(equation = aw_361_equation_0, values = (var_2785_cast_fp16_4, var_2767_cast_fp16_4))[name = tensor("aw_361_cast_fp16")]; + tensor aw_363_equation_0 = const()[name = tensor("aw_363_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_363_cast_fp16 = einsum(equation = aw_363_equation_0, values = (var_2785_cast_fp16_5, var_2767_cast_fp16_5))[name = tensor("aw_363_cast_fp16")]; + tensor aw_365_equation_0 = const()[name = tensor("aw_365_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_365_cast_fp16 = einsum(equation = aw_365_equation_0, values = (var_2785_cast_fp16_6, var_2767_cast_fp16_6))[name = tensor("aw_365_cast_fp16")]; + tensor aw_367_equation_0 = const()[name = tensor("aw_367_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_367_cast_fp16 = einsum(equation = aw_367_equation_0, values = (var_2785_cast_fp16_7, var_2767_cast_fp16_7))[name = tensor("aw_367_cast_fp16")]; + tensor aw_369_equation_0 = const()[name = tensor("aw_369_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_369_cast_fp16 = einsum(equation = aw_369_equation_0, values = (var_2785_cast_fp16_8, var_2767_cast_fp16_8))[name = tensor("aw_369_cast_fp16")]; + tensor aw_371_equation_0 = const()[name = tensor("aw_371_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_371_cast_fp16 = einsum(equation = aw_371_equation_0, values = (var_2785_cast_fp16_9, var_2767_cast_fp16_9))[name = tensor("aw_371_cast_fp16")]; + tensor aw_373_equation_0 = const()[name = tensor("aw_373_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_373_cast_fp16 = einsum(equation = aw_373_equation_0, values = (var_2785_cast_fp16_10, var_2767_cast_fp16_10))[name = tensor("aw_373_cast_fp16")]; + tensor aw_375_equation_0 = const()[name = tensor("aw_375_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_375_cast_fp16 = einsum(equation = aw_375_equation_0, values = (var_2785_cast_fp16_11, var_2767_cast_fp16_11))[name = tensor("aw_375_cast_fp16")]; + tensor aw_377_equation_0 = const()[name = tensor("aw_377_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_377_cast_fp16 = einsum(equation = aw_377_equation_0, values = (var_2785_cast_fp16_12, var_2767_cast_fp16_12))[name = tensor("aw_377_cast_fp16")]; + tensor aw_379_equation_0 = const()[name = tensor("aw_379_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_379_cast_fp16 = einsum(equation = aw_379_equation_0, values = (var_2785_cast_fp16_13, var_2767_cast_fp16_13))[name = tensor("aw_379_cast_fp16")]; + tensor aw_381_equation_0 = const()[name = tensor("aw_381_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_381_cast_fp16 = einsum(equation = aw_381_equation_0, values = (var_2785_cast_fp16_14, var_2767_cast_fp16_14))[name = tensor("aw_381_cast_fp16")]; + tensor aw_383_equation_0 = const()[name = tensor("aw_383_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_383_cast_fp16 = einsum(equation = aw_383_equation_0, values = (var_2785_cast_fp16_15, var_2767_cast_fp16_15))[name = tensor("aw_383_cast_fp16")]; + tensor var_2851_cast_fp16 = softmax(axis = var_2715, x = aw_353_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor var_2852_cast_fp16 = softmax(axis = var_2715, x = aw_355_cast_fp16)[name = tensor("op_2852_cast_fp16")]; + tensor var_2853_cast_fp16 = softmax(axis = var_2715, x = aw_357_cast_fp16)[name = tensor("op_2853_cast_fp16")]; + tensor var_2854_cast_fp16 = softmax(axis = var_2715, x = aw_359_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = softmax(axis = var_2715, x = aw_361_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = softmax(axis = var_2715, x = aw_363_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2857_cast_fp16 = softmax(axis = var_2715, x = aw_365_cast_fp16)[name = tensor("op_2857_cast_fp16")]; + tensor var_2858_cast_fp16 = softmax(axis = var_2715, x = aw_367_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor var_2859_cast_fp16 = softmax(axis = var_2715, x = aw_369_cast_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2860_cast_fp16 = softmax(axis = var_2715, x = aw_371_cast_fp16)[name = tensor("op_2860_cast_fp16")]; + tensor var_2861_cast_fp16 = softmax(axis = var_2715, x = aw_373_cast_fp16)[name = tensor("op_2861_cast_fp16")]; + tensor var_2862_cast_fp16 = softmax(axis = var_2715, x = aw_375_cast_fp16)[name = tensor("op_2862_cast_fp16")]; + tensor var_2863_cast_fp16 = softmax(axis = var_2715, x = aw_377_cast_fp16)[name = tensor("op_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = softmax(axis = var_2715, x = aw_379_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_cast_fp16 = softmax(axis = var_2715, x = aw_381_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_cast_fp16 = softmax(axis = var_2715, x = aw_383_cast_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor var_2868_equation_0 = const()[name = tensor("op_2868_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2868_cast_fp16 = einsum(equation = var_2868_equation_0, values = (var_2802_cast_fp16_0, var_2851_cast_fp16))[name = tensor("op_2868_cast_fp16")]; + tensor var_2870_equation_0 = const()[name = tensor("op_2870_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2870_cast_fp16 = einsum(equation = var_2870_equation_0, values = (var_2802_cast_fp16_1, var_2852_cast_fp16))[name = tensor("op_2870_cast_fp16")]; + tensor var_2872_equation_0 = const()[name = tensor("op_2872_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2872_cast_fp16 = einsum(equation = var_2872_equation_0, values = (var_2802_cast_fp16_2, var_2853_cast_fp16))[name = tensor("op_2872_cast_fp16")]; + tensor var_2874_equation_0 = const()[name = tensor("op_2874_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2874_cast_fp16 = einsum(equation = var_2874_equation_0, values = (var_2802_cast_fp16_3, var_2854_cast_fp16))[name = tensor("op_2874_cast_fp16")]; + tensor var_2876_equation_0 = const()[name = tensor("op_2876_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2876_cast_fp16 = einsum(equation = var_2876_equation_0, values = (var_2802_cast_fp16_4, var_2855_cast_fp16))[name = tensor("op_2876_cast_fp16")]; + tensor var_2878_equation_0 = const()[name = tensor("op_2878_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2878_cast_fp16 = einsum(equation = var_2878_equation_0, values = (var_2802_cast_fp16_5, var_2856_cast_fp16))[name = tensor("op_2878_cast_fp16")]; + tensor var_2880_equation_0 = const()[name = tensor("op_2880_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2880_cast_fp16 = einsum(equation = var_2880_equation_0, values = (var_2802_cast_fp16_6, var_2857_cast_fp16))[name = tensor("op_2880_cast_fp16")]; + tensor var_2882_equation_0 = const()[name = tensor("op_2882_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2882_cast_fp16 = einsum(equation = var_2882_equation_0, values = (var_2802_cast_fp16_7, var_2858_cast_fp16))[name = tensor("op_2882_cast_fp16")]; + tensor var_2884_equation_0 = const()[name = tensor("op_2884_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2884_cast_fp16 = einsum(equation = var_2884_equation_0, values = (var_2802_cast_fp16_8, var_2859_cast_fp16))[name = tensor("op_2884_cast_fp16")]; + tensor var_2886_equation_0 = const()[name = tensor("op_2886_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2886_cast_fp16 = einsum(equation = var_2886_equation_0, values = (var_2802_cast_fp16_9, var_2860_cast_fp16))[name = tensor("op_2886_cast_fp16")]; + tensor var_2888_equation_0 = const()[name = tensor("op_2888_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2888_cast_fp16 = einsum(equation = var_2888_equation_0, values = (var_2802_cast_fp16_10, var_2861_cast_fp16))[name = tensor("op_2888_cast_fp16")]; + tensor var_2890_equation_0 = const()[name = tensor("op_2890_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2890_cast_fp16 = einsum(equation = var_2890_equation_0, values = (var_2802_cast_fp16_11, var_2862_cast_fp16))[name = tensor("op_2890_cast_fp16")]; + tensor var_2892_equation_0 = const()[name = tensor("op_2892_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2892_cast_fp16 = einsum(equation = var_2892_equation_0, values = (var_2802_cast_fp16_12, var_2863_cast_fp16))[name = tensor("op_2892_cast_fp16")]; + tensor var_2894_equation_0 = const()[name = tensor("op_2894_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2894_cast_fp16 = einsum(equation = var_2894_equation_0, values = (var_2802_cast_fp16_13, var_2864_cast_fp16))[name = tensor("op_2894_cast_fp16")]; + tensor var_2896_equation_0 = const()[name = tensor("op_2896_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2896_cast_fp16 = einsum(equation = var_2896_equation_0, values = (var_2802_cast_fp16_14, var_2865_cast_fp16))[name = tensor("op_2896_cast_fp16")]; + tensor var_2898_equation_0 = const()[name = tensor("op_2898_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2898_cast_fp16 = einsum(equation = var_2898_equation_0, values = (var_2802_cast_fp16_15, var_2866_cast_fp16))[name = tensor("op_2898_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_2715, interleave = input_115_interleave_0, values = (var_2868_cast_fp16, var_2870_cast_fp16, var_2872_cast_fp16, var_2874_cast_fp16, var_2876_cast_fp16, var_2878_cast_fp16, var_2880_cast_fp16, var_2882_cast_fp16, var_2884_cast_fp16, var_2886_cast_fp16, var_2888_cast_fp16, var_2890_cast_fp16, var_2892_cast_fp16, var_2894_cast_fp16, var_2896_cast_fp16, var_2898_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_2907_pad_type_0 = const()[name = tensor("op_2907_pad_type_0"), val = tensor("valid")]; + tensor var_2907_strides_0 = const()[name = tensor("op_2907_strides_0"), val = tensor([1, 1])]; + tensor var_2907_pad_0 = const()[name = tensor("op_2907_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2907_dilations_0 = const()[name = tensor("op_2907_dilations_0"), val = tensor([1, 1])]; + tensor var_2907_groups_0 = const()[name = tensor("op_2907_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293264512)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295361728)))]; + tensor var_2907_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_2907_dilations_0, groups = var_2907_groups_0, pad = var_2907_pad_0, pad_type = var_2907_pad_type_0, strides = var_2907_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_2907_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_2907_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295363840)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295365952)))]; + tensor var_2917_to_fp16 = const()[name = tensor("op_2917_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_2917_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295368064)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756736)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_2943_pad_type_0 = const()[name = tensor("op_2943_pad_type_0"), val = tensor("valid")]; + tensor var_2943_strides_0 = const()[name = tensor("op_2943_strides_0"), val = tensor([1, 1])]; + tensor var_2943_pad_0 = const()[name = tensor("op_2943_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2943_dilations_0 = const()[name = tensor("op_2943_dilations_0"), val = tensor([1, 1])]; + tensor var_2943_groups_0 = const()[name = tensor("op_2943_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312153664)))]; + tensor var_2943_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_2943_dilations_0, groups = var_2943_groups_0, pad = var_2943_pad_0, pad_type = var_2943_pad_type_0, strides = var_2943_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("op_2943_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_2943_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_2952 = const()[name = tensor("op_2952"), val = tensor(1)]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([1])]; + tensor input_123_gamma_0_to_fp16 = const()[name = tensor("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312155776)))]; + tensor input_123_beta_0_to_fp16 = const()[name = tensor("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312157888)))]; + tensor var_2968_to_fp16 = const()[name = tensor("op_2968_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = input_123_beta_0_to_fp16, epsilon = var_2968_to_fp16, gamma = input_123_gamma_0_to_fp16, x = inputs_49_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor q_25_pad_type_0 = const()[name = tensor("q_25_pad_type_0"), val = tensor("valid")]; + tensor q_25_strides_0 = const()[name = tensor("q_25_strides_0"), val = tensor([1, 1])]; + tensor q_25_pad_0 = const()[name = tensor("q_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_25_dilations_0 = const()[name = tensor("q_25_dilations_0"), val = tensor([1, 1])]; + tensor q_25_groups_0 = const()[name = tensor("q_25_groups_0"), val = tensor(1)]; + tensor var_3003_weight_0_to_fp16 = const()[name = tensor("op_3003_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312160000)))]; + tensor var_3003_bias_0_to_fp16 = const()[name = tensor("op_3003_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314257216)))]; + tensor var_3003_cast_fp16 = conv(bias = var_3003_bias_0_to_fp16, dilations = q_25_dilations_0, groups = q_25_groups_0, pad = q_25_pad_0, pad_type = q_25_pad_type_0, strides = q_25_strides_0, weight = var_3003_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3003_cast_fp16")]; + tensor k_25_pad_type_0 = const()[name = tensor("k_25_pad_type_0"), val = tensor("valid")]; + tensor k_25_strides_0 = const()[name = tensor("k_25_strides_0"), val = tensor([1, 1])]; + tensor k_25_pad_0 = const()[name = tensor("k_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_25_dilations_0 = const()[name = tensor("k_25_dilations_0"), val = tensor([1, 1])]; + tensor k_25_groups_0 = const()[name = tensor("k_25_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_key_weight_to_fp16 = const()[name = tensor("blocks_12_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314259328)))]; + tensor k_25_cast_fp16 = conv(dilations = k_25_dilations_0, groups = k_25_groups_0, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = k_25_strides_0, weight = blocks_12_attn_key_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor var_3001_pad_type_0 = const()[name = tensor("op_3001_pad_type_0"), val = tensor("valid")]; + tensor var_3001_strides_0 = const()[name = tensor("op_3001_strides_0"), val = tensor([1, 1])]; + tensor var_3001_pad_0 = const()[name = tensor("op_3001_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3001_dilations_0 = const()[name = tensor("op_3001_dilations_0"), val = tensor([1, 1])]; + tensor var_3001_groups_0 = const()[name = tensor("op_3001_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_value_weight_to_fp16 = const()[name = tensor("blocks_12_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316356544)))]; + tensor blocks_12_attn_value_bias_to_fp16 = const()[name = tensor("blocks_12_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318453760)))]; + tensor var_3001_cast_fp16 = conv(bias = blocks_12_attn_value_bias_to_fp16, dilations = var_3001_dilations_0, groups = var_3001_groups_0, pad = var_3001_pad_0, pad_type = var_3001_pad_type_0, strides = var_3001_strides_0, weight = blocks_12_attn_value_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor tile_36 = const()[name = tensor("tile_36"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3004_axis_0 = const()[name = tensor("op_3004_axis_0"), val = tensor(1)]; + tensor var_3004_cast_fp16_0, tensor var_3004_cast_fp16_1, tensor var_3004_cast_fp16_2, tensor var_3004_cast_fp16_3, tensor var_3004_cast_fp16_4, tensor var_3004_cast_fp16_5, tensor var_3004_cast_fp16_6, tensor var_3004_cast_fp16_7, tensor var_3004_cast_fp16_8, tensor var_3004_cast_fp16_9, tensor var_3004_cast_fp16_10, tensor var_3004_cast_fp16_11, tensor var_3004_cast_fp16_12, tensor var_3004_cast_fp16_13, tensor var_3004_cast_fp16_14, tensor var_3004_cast_fp16_15 = split(axis = var_3004_axis_0, split_sizes = tile_36, x = var_3003_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3021_perm_0 = const()[name = tensor("op_3021_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_37 = const()[name = tensor("tile_37"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3022_axis_0 = const()[name = tensor("op_3022_axis_0"), val = tensor(3)]; + tensor var_3021_cast_fp16 = transpose(perm = var_3021_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_12")]; + tensor var_3022_cast_fp16_0, tensor var_3022_cast_fp16_1, tensor var_3022_cast_fp16_2, tensor var_3022_cast_fp16_3, tensor var_3022_cast_fp16_4, tensor var_3022_cast_fp16_5, tensor var_3022_cast_fp16_6, tensor var_3022_cast_fp16_7, tensor var_3022_cast_fp16_8, tensor var_3022_cast_fp16_9, tensor var_3022_cast_fp16_10, tensor var_3022_cast_fp16_11, tensor var_3022_cast_fp16_12, tensor var_3022_cast_fp16_13, tensor var_3022_cast_fp16_14, tensor var_3022_cast_fp16_15 = split(axis = var_3022_axis_0, split_sizes = tile_37, x = var_3021_cast_fp16)[name = tensor("op_3022_cast_fp16")]; + tensor tile_38 = const()[name = tensor("tile_38"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3039_axis_0 = const()[name = tensor("op_3039_axis_0"), val = tensor(1)]; + tensor var_3039_cast_fp16_0, tensor var_3039_cast_fp16_1, tensor var_3039_cast_fp16_2, tensor var_3039_cast_fp16_3, tensor var_3039_cast_fp16_4, tensor var_3039_cast_fp16_5, tensor var_3039_cast_fp16_6, tensor var_3039_cast_fp16_7, tensor var_3039_cast_fp16_8, tensor var_3039_cast_fp16_9, tensor var_3039_cast_fp16_10, tensor var_3039_cast_fp16_11, tensor var_3039_cast_fp16_12, tensor var_3039_cast_fp16_13, tensor var_3039_cast_fp16_14, tensor var_3039_cast_fp16_15 = split(axis = var_3039_axis_0, split_sizes = tile_38, x = var_3001_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor aw_385_equation_0 = const()[name = tensor("aw_385_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_385_cast_fp16 = einsum(equation = aw_385_equation_0, values = (var_3022_cast_fp16_0, var_3004_cast_fp16_0))[name = tensor("aw_385_cast_fp16")]; + tensor aw_387_equation_0 = const()[name = tensor("aw_387_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_387_cast_fp16 = einsum(equation = aw_387_equation_0, values = (var_3022_cast_fp16_1, var_3004_cast_fp16_1))[name = tensor("aw_387_cast_fp16")]; + tensor aw_389_equation_0 = const()[name = tensor("aw_389_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_389_cast_fp16 = einsum(equation = aw_389_equation_0, values = (var_3022_cast_fp16_2, var_3004_cast_fp16_2))[name = tensor("aw_389_cast_fp16")]; + tensor aw_391_equation_0 = const()[name = tensor("aw_391_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_391_cast_fp16 = einsum(equation = aw_391_equation_0, values = (var_3022_cast_fp16_3, var_3004_cast_fp16_3))[name = tensor("aw_391_cast_fp16")]; + tensor aw_393_equation_0 = const()[name = tensor("aw_393_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_393_cast_fp16 = einsum(equation = aw_393_equation_0, values = (var_3022_cast_fp16_4, var_3004_cast_fp16_4))[name = tensor("aw_393_cast_fp16")]; + tensor aw_395_equation_0 = const()[name = tensor("aw_395_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_395_cast_fp16 = einsum(equation = aw_395_equation_0, values = (var_3022_cast_fp16_5, var_3004_cast_fp16_5))[name = tensor("aw_395_cast_fp16")]; + tensor aw_397_equation_0 = const()[name = tensor("aw_397_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_397_cast_fp16 = einsum(equation = aw_397_equation_0, values = (var_3022_cast_fp16_6, var_3004_cast_fp16_6))[name = tensor("aw_397_cast_fp16")]; + tensor aw_399_equation_0 = const()[name = tensor("aw_399_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_399_cast_fp16 = einsum(equation = aw_399_equation_0, values = (var_3022_cast_fp16_7, var_3004_cast_fp16_7))[name = tensor("aw_399_cast_fp16")]; + tensor aw_401_equation_0 = const()[name = tensor("aw_401_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_401_cast_fp16 = einsum(equation = aw_401_equation_0, values = (var_3022_cast_fp16_8, var_3004_cast_fp16_8))[name = tensor("aw_401_cast_fp16")]; + tensor aw_403_equation_0 = const()[name = tensor("aw_403_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_403_cast_fp16 = einsum(equation = aw_403_equation_0, values = (var_3022_cast_fp16_9, var_3004_cast_fp16_9))[name = tensor("aw_403_cast_fp16")]; + tensor aw_405_equation_0 = const()[name = tensor("aw_405_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_405_cast_fp16 = einsum(equation = aw_405_equation_0, values = (var_3022_cast_fp16_10, var_3004_cast_fp16_10))[name = tensor("aw_405_cast_fp16")]; + tensor aw_407_equation_0 = const()[name = tensor("aw_407_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_407_cast_fp16 = einsum(equation = aw_407_equation_0, values = (var_3022_cast_fp16_11, var_3004_cast_fp16_11))[name = tensor("aw_407_cast_fp16")]; + tensor aw_409_equation_0 = const()[name = tensor("aw_409_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_409_cast_fp16 = einsum(equation = aw_409_equation_0, values = (var_3022_cast_fp16_12, var_3004_cast_fp16_12))[name = tensor("aw_409_cast_fp16")]; + tensor aw_411_equation_0 = const()[name = tensor("aw_411_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_411_cast_fp16 = einsum(equation = aw_411_equation_0, values = (var_3022_cast_fp16_13, var_3004_cast_fp16_13))[name = tensor("aw_411_cast_fp16")]; + tensor aw_413_equation_0 = const()[name = tensor("aw_413_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_413_cast_fp16 = einsum(equation = aw_413_equation_0, values = (var_3022_cast_fp16_14, var_3004_cast_fp16_14))[name = tensor("aw_413_cast_fp16")]; + tensor aw_415_equation_0 = const()[name = tensor("aw_415_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_415_cast_fp16 = einsum(equation = aw_415_equation_0, values = (var_3022_cast_fp16_15, var_3004_cast_fp16_15))[name = tensor("aw_415_cast_fp16")]; + tensor var_3088_cast_fp16 = softmax(axis = var_2952, x = aw_385_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_cast_fp16 = softmax(axis = var_2952, x = aw_387_cast_fp16)[name = tensor("op_3089_cast_fp16")]; + tensor var_3090_cast_fp16 = softmax(axis = var_2952, x = aw_389_cast_fp16)[name = tensor("op_3090_cast_fp16")]; + tensor var_3091_cast_fp16 = softmax(axis = var_2952, x = aw_391_cast_fp16)[name = tensor("op_3091_cast_fp16")]; + tensor var_3092_cast_fp16 = softmax(axis = var_2952, x = aw_393_cast_fp16)[name = tensor("op_3092_cast_fp16")]; + tensor var_3093_cast_fp16 = softmax(axis = var_2952, x = aw_395_cast_fp16)[name = tensor("op_3093_cast_fp16")]; + tensor var_3094_cast_fp16 = softmax(axis = var_2952, x = aw_397_cast_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = softmax(axis = var_2952, x = aw_399_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor var_3096_cast_fp16 = softmax(axis = var_2952, x = aw_401_cast_fp16)[name = tensor("op_3096_cast_fp16")]; + tensor var_3097_cast_fp16 = softmax(axis = var_2952, x = aw_403_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor var_3098_cast_fp16 = softmax(axis = var_2952, x = aw_405_cast_fp16)[name = tensor("op_3098_cast_fp16")]; + tensor var_3099_cast_fp16 = softmax(axis = var_2952, x = aw_407_cast_fp16)[name = tensor("op_3099_cast_fp16")]; + tensor var_3100_cast_fp16 = softmax(axis = var_2952, x = aw_409_cast_fp16)[name = tensor("op_3100_cast_fp16")]; + tensor var_3101_cast_fp16 = softmax(axis = var_2952, x = aw_411_cast_fp16)[name = tensor("op_3101_cast_fp16")]; + tensor var_3102_cast_fp16 = softmax(axis = var_2952, x = aw_413_cast_fp16)[name = tensor("op_3102_cast_fp16")]; + tensor var_3103_cast_fp16 = softmax(axis = var_2952, x = aw_415_cast_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3105_equation_0 = const()[name = tensor("op_3105_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3105_cast_fp16 = einsum(equation = var_3105_equation_0, values = (var_3039_cast_fp16_0, var_3088_cast_fp16))[name = tensor("op_3105_cast_fp16")]; + tensor var_3107_equation_0 = const()[name = tensor("op_3107_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3107_cast_fp16 = einsum(equation = var_3107_equation_0, values = (var_3039_cast_fp16_1, var_3089_cast_fp16))[name = tensor("op_3107_cast_fp16")]; + tensor var_3109_equation_0 = const()[name = tensor("op_3109_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3109_cast_fp16 = einsum(equation = var_3109_equation_0, values = (var_3039_cast_fp16_2, var_3090_cast_fp16))[name = tensor("op_3109_cast_fp16")]; + tensor var_3111_equation_0 = const()[name = tensor("op_3111_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3111_cast_fp16 = einsum(equation = var_3111_equation_0, values = (var_3039_cast_fp16_3, var_3091_cast_fp16))[name = tensor("op_3111_cast_fp16")]; + tensor var_3113_equation_0 = const()[name = tensor("op_3113_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3113_cast_fp16 = einsum(equation = var_3113_equation_0, values = (var_3039_cast_fp16_4, var_3092_cast_fp16))[name = tensor("op_3113_cast_fp16")]; + tensor var_3115_equation_0 = const()[name = tensor("op_3115_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3115_cast_fp16 = einsum(equation = var_3115_equation_0, values = (var_3039_cast_fp16_5, var_3093_cast_fp16))[name = tensor("op_3115_cast_fp16")]; + tensor var_3117_equation_0 = const()[name = tensor("op_3117_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3117_cast_fp16 = einsum(equation = var_3117_equation_0, values = (var_3039_cast_fp16_6, var_3094_cast_fp16))[name = tensor("op_3117_cast_fp16")]; + tensor var_3119_equation_0 = const()[name = tensor("op_3119_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3119_cast_fp16 = einsum(equation = var_3119_equation_0, values = (var_3039_cast_fp16_7, var_3095_cast_fp16))[name = tensor("op_3119_cast_fp16")]; + tensor var_3121_equation_0 = const()[name = tensor("op_3121_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3121_cast_fp16 = einsum(equation = var_3121_equation_0, values = (var_3039_cast_fp16_8, var_3096_cast_fp16))[name = tensor("op_3121_cast_fp16")]; + tensor var_3123_equation_0 = const()[name = tensor("op_3123_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3123_cast_fp16 = einsum(equation = var_3123_equation_0, values = (var_3039_cast_fp16_9, var_3097_cast_fp16))[name = tensor("op_3123_cast_fp16")]; + tensor var_3125_equation_0 = const()[name = tensor("op_3125_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3125_cast_fp16 = einsum(equation = var_3125_equation_0, values = (var_3039_cast_fp16_10, var_3098_cast_fp16))[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_equation_0 = const()[name = tensor("op_3127_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3127_cast_fp16 = einsum(equation = var_3127_equation_0, values = (var_3039_cast_fp16_11, var_3099_cast_fp16))[name = tensor("op_3127_cast_fp16")]; + tensor var_3129_equation_0 = const()[name = tensor("op_3129_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3129_cast_fp16 = einsum(equation = var_3129_equation_0, values = (var_3039_cast_fp16_12, var_3100_cast_fp16))[name = tensor("op_3129_cast_fp16")]; + tensor var_3131_equation_0 = const()[name = tensor("op_3131_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3131_cast_fp16 = einsum(equation = var_3131_equation_0, values = (var_3039_cast_fp16_13, var_3101_cast_fp16))[name = tensor("op_3131_cast_fp16")]; + tensor var_3133_equation_0 = const()[name = tensor("op_3133_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3133_cast_fp16 = einsum(equation = var_3133_equation_0, values = (var_3039_cast_fp16_14, var_3102_cast_fp16))[name = tensor("op_3133_cast_fp16")]; + tensor var_3135_equation_0 = const()[name = tensor("op_3135_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3135_cast_fp16 = einsum(equation = var_3135_equation_0, values = (var_3039_cast_fp16_15, var_3103_cast_fp16))[name = tensor("op_3135_cast_fp16")]; + tensor input_125_interleave_0 = const()[name = tensor("input_125_interleave_0"), val = tensor(false)]; + tensor input_125_cast_fp16 = concat(axis = var_2952, interleave = input_125_interleave_0, values = (var_3105_cast_fp16, var_3107_cast_fp16, var_3109_cast_fp16, var_3111_cast_fp16, var_3113_cast_fp16, var_3115_cast_fp16, var_3117_cast_fp16, var_3119_cast_fp16, var_3121_cast_fp16, var_3123_cast_fp16, var_3125_cast_fp16, var_3127_cast_fp16, var_3129_cast_fp16, var_3131_cast_fp16, var_3133_cast_fp16, var_3135_cast_fp16))[name = tensor("input_125_cast_fp16")]; + tensor var_3144_pad_type_0 = const()[name = tensor("op_3144_pad_type_0"), val = tensor("valid")]; + tensor var_3144_strides_0 = const()[name = tensor("op_3144_strides_0"), val = tensor([1, 1])]; + tensor var_3144_pad_0 = const()[name = tensor("op_3144_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3144_dilations_0 = const()[name = tensor("op_3144_dilations_0"), val = tensor([1, 1])]; + tensor var_3144_groups_0 = const()[name = tensor("op_3144_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_out_weight_to_fp16 = const()[name = tensor("blocks_12_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318455872)))]; + tensor blocks_12_attn_out_bias_to_fp16 = const()[name = tensor("blocks_12_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320553088)))]; + tensor var_3144_cast_fp16 = conv(bias = blocks_12_attn_out_bias_to_fp16, dilations = var_3144_dilations_0, groups = var_3144_groups_0, pad = var_3144_pad_0, pad_type = var_3144_pad_type_0, strides = var_3144_strides_0, weight = blocks_12_attn_out_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("op_3144_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = var_3144_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([1])]; + tensor input_127_gamma_0_to_fp16 = const()[name = tensor("input_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320555200)))]; + tensor input_127_beta_0_to_fp16 = const()[name = tensor("input_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320557312)))]; + tensor var_3154_to_fp16 = const()[name = tensor("op_3154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = input_127_beta_0_to_fp16, epsilon = var_3154_to_fp16, gamma = input_127_gamma_0_to_fp16, x = inputs_51_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("valid")]; + tensor input_129_strides_0 = const()[name = tensor("input_129_strides_0"), val = tensor([1, 1])]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_129_dilations_0 = const()[name = tensor("input_129_dilations_0"), val = tensor([1, 1])]; + tensor input_129_groups_0 = const()[name = tensor("input_129_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320559424)))]; + tensor blocks_12_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328948096)))]; + tensor input_129_cast_fp16 = conv(bias = blocks_12_mlp_0_bias_to_fp16, dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = blocks_12_mlp_0_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3180_pad_type_0 = const()[name = tensor("op_3180_pad_type_0"), val = tensor("valid")]; + tensor var_3180_strides_0 = const()[name = tensor("op_3180_strides_0"), val = tensor([1, 1])]; + tensor var_3180_pad_0 = const()[name = tensor("op_3180_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3180_dilations_0 = const()[name = tensor("op_3180_dilations_0"), val = tensor([1, 1])]; + tensor var_3180_groups_0 = const()[name = tensor("op_3180_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328956352)))]; + tensor blocks_12_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337345024)))]; + tensor var_3180_cast_fp16 = conv(bias = blocks_12_mlp_2_bias_to_fp16, dilations = var_3180_dilations_0, groups = var_3180_groups_0, pad = var_3180_pad_0, pad_type = var_3180_pad_type_0, strides = var_3180_strides_0, weight = blocks_12_mlp_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("op_3180_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_3180_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3189 = const()[name = tensor("op_3189"), val = tensor(1)]; + tensor input_133_axes_0 = const()[name = tensor("input_133_axes_0"), val = tensor([1])]; + tensor input_133_gamma_0_to_fp16 = const()[name = tensor("input_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337347136)))]; + tensor input_133_beta_0_to_fp16 = const()[name = tensor("input_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337349248)))]; + tensor var_3205_to_fp16 = const()[name = tensor("op_3205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_133_cast_fp16 = layer_norm(axes = input_133_axes_0, beta = input_133_beta_0_to_fp16, epsilon = var_3205_to_fp16, gamma = input_133_gamma_0_to_fp16, x = inputs_53_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor q_27_pad_type_0 = const()[name = tensor("q_27_pad_type_0"), val = tensor("valid")]; + tensor q_27_strides_0 = const()[name = tensor("q_27_strides_0"), val = tensor([1, 1])]; + tensor q_27_pad_0 = const()[name = tensor("q_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_27_dilations_0 = const()[name = tensor("q_27_dilations_0"), val = tensor([1, 1])]; + tensor q_27_groups_0 = const()[name = tensor("q_27_groups_0"), val = tensor(1)]; + tensor var_3240_weight_0_to_fp16 = const()[name = tensor("op_3240_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337351360)))]; + tensor var_3240_bias_0_to_fp16 = const()[name = tensor("op_3240_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339448576)))]; + tensor var_3240_cast_fp16 = conv(bias = var_3240_bias_0_to_fp16, dilations = q_27_dilations_0, groups = q_27_groups_0, pad = q_27_pad_0, pad_type = q_27_pad_type_0, strides = q_27_strides_0, weight = var_3240_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3240_cast_fp16")]; + tensor k_27_pad_type_0 = const()[name = tensor("k_27_pad_type_0"), val = tensor("valid")]; + tensor k_27_strides_0 = const()[name = tensor("k_27_strides_0"), val = tensor([1, 1])]; + tensor k_27_pad_0 = const()[name = tensor("k_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_27_dilations_0 = const()[name = tensor("k_27_dilations_0"), val = tensor([1, 1])]; + tensor k_27_groups_0 = const()[name = tensor("k_27_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_key_weight_to_fp16 = const()[name = tensor("blocks_13_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339450688)))]; + tensor k_27_cast_fp16 = conv(dilations = k_27_dilations_0, groups = k_27_groups_0, pad = k_27_pad_0, pad_type = k_27_pad_type_0, strides = k_27_strides_0, weight = blocks_13_attn_key_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor var_3238_pad_type_0 = const()[name = tensor("op_3238_pad_type_0"), val = tensor("valid")]; + tensor var_3238_strides_0 = const()[name = tensor("op_3238_strides_0"), val = tensor([1, 1])]; + tensor var_3238_pad_0 = const()[name = tensor("op_3238_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3238_dilations_0 = const()[name = tensor("op_3238_dilations_0"), val = tensor([1, 1])]; + tensor var_3238_groups_0 = const()[name = tensor("op_3238_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_value_weight_to_fp16 = const()[name = tensor("blocks_13_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341547904)))]; + tensor blocks_13_attn_value_bias_to_fp16 = const()[name = tensor("blocks_13_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343645120)))]; + tensor var_3238_cast_fp16 = conv(bias = blocks_13_attn_value_bias_to_fp16, dilations = var_3238_dilations_0, groups = var_3238_groups_0, pad = var_3238_pad_0, pad_type = var_3238_pad_type_0, strides = var_3238_strides_0, weight = blocks_13_attn_value_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3238_cast_fp16")]; + tensor tile_39 = const()[name = tensor("tile_39"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3241_axis_0 = const()[name = tensor("op_3241_axis_0"), val = tensor(1)]; + tensor var_3241_cast_fp16_0, tensor var_3241_cast_fp16_1, tensor var_3241_cast_fp16_2, tensor var_3241_cast_fp16_3, tensor var_3241_cast_fp16_4, tensor var_3241_cast_fp16_5, tensor var_3241_cast_fp16_6, tensor var_3241_cast_fp16_7, tensor var_3241_cast_fp16_8, tensor var_3241_cast_fp16_9, tensor var_3241_cast_fp16_10, tensor var_3241_cast_fp16_11, tensor var_3241_cast_fp16_12, tensor var_3241_cast_fp16_13, tensor var_3241_cast_fp16_14, tensor var_3241_cast_fp16_15 = split(axis = var_3241_axis_0, split_sizes = tile_39, x = var_3240_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor var_3258_perm_0 = const()[name = tensor("op_3258_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_40 = const()[name = tensor("tile_40"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3259_axis_0 = const()[name = tensor("op_3259_axis_0"), val = tensor(3)]; + tensor var_3258_cast_fp16 = transpose(perm = var_3258_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_11")]; + tensor var_3259_cast_fp16_0, tensor var_3259_cast_fp16_1, tensor var_3259_cast_fp16_2, tensor var_3259_cast_fp16_3, tensor var_3259_cast_fp16_4, tensor var_3259_cast_fp16_5, tensor var_3259_cast_fp16_6, tensor var_3259_cast_fp16_7, tensor var_3259_cast_fp16_8, tensor var_3259_cast_fp16_9, tensor var_3259_cast_fp16_10, tensor var_3259_cast_fp16_11, tensor var_3259_cast_fp16_12, tensor var_3259_cast_fp16_13, tensor var_3259_cast_fp16_14, tensor var_3259_cast_fp16_15 = split(axis = var_3259_axis_0, split_sizes = tile_40, x = var_3258_cast_fp16)[name = tensor("op_3259_cast_fp16")]; + tensor tile_41 = const()[name = tensor("tile_41"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3276_axis_0 = const()[name = tensor("op_3276_axis_0"), val = tensor(1)]; + tensor var_3276_cast_fp16_0, tensor var_3276_cast_fp16_1, tensor var_3276_cast_fp16_2, tensor var_3276_cast_fp16_3, tensor var_3276_cast_fp16_4, tensor var_3276_cast_fp16_5, tensor var_3276_cast_fp16_6, tensor var_3276_cast_fp16_7, tensor var_3276_cast_fp16_8, tensor var_3276_cast_fp16_9, tensor var_3276_cast_fp16_10, tensor var_3276_cast_fp16_11, tensor var_3276_cast_fp16_12, tensor var_3276_cast_fp16_13, tensor var_3276_cast_fp16_14, tensor var_3276_cast_fp16_15 = split(axis = var_3276_axis_0, split_sizes = tile_41, x = var_3238_cast_fp16)[name = tensor("op_3276_cast_fp16")]; + tensor aw_417_equation_0 = const()[name = tensor("aw_417_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_417_cast_fp16 = einsum(equation = aw_417_equation_0, values = (var_3259_cast_fp16_0, var_3241_cast_fp16_0))[name = tensor("aw_417_cast_fp16")]; + tensor aw_419_equation_0 = const()[name = tensor("aw_419_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_419_cast_fp16 = einsum(equation = aw_419_equation_0, values = (var_3259_cast_fp16_1, var_3241_cast_fp16_1))[name = tensor("aw_419_cast_fp16")]; + tensor aw_421_equation_0 = const()[name = tensor("aw_421_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_421_cast_fp16 = einsum(equation = aw_421_equation_0, values = (var_3259_cast_fp16_2, var_3241_cast_fp16_2))[name = tensor("aw_421_cast_fp16")]; + tensor aw_423_equation_0 = const()[name = tensor("aw_423_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_423_cast_fp16 = einsum(equation = aw_423_equation_0, values = (var_3259_cast_fp16_3, var_3241_cast_fp16_3))[name = tensor("aw_423_cast_fp16")]; + tensor aw_425_equation_0 = const()[name = tensor("aw_425_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_425_cast_fp16 = einsum(equation = aw_425_equation_0, values = (var_3259_cast_fp16_4, var_3241_cast_fp16_4))[name = tensor("aw_425_cast_fp16")]; + tensor aw_427_equation_0 = const()[name = tensor("aw_427_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_427_cast_fp16 = einsum(equation = aw_427_equation_0, values = (var_3259_cast_fp16_5, var_3241_cast_fp16_5))[name = tensor("aw_427_cast_fp16")]; + tensor aw_429_equation_0 = const()[name = tensor("aw_429_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_429_cast_fp16 = einsum(equation = aw_429_equation_0, values = (var_3259_cast_fp16_6, var_3241_cast_fp16_6))[name = tensor("aw_429_cast_fp16")]; + tensor aw_431_equation_0 = const()[name = tensor("aw_431_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_431_cast_fp16 = einsum(equation = aw_431_equation_0, values = (var_3259_cast_fp16_7, var_3241_cast_fp16_7))[name = tensor("aw_431_cast_fp16")]; + tensor aw_433_equation_0 = const()[name = tensor("aw_433_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_433_cast_fp16 = einsum(equation = aw_433_equation_0, values = (var_3259_cast_fp16_8, var_3241_cast_fp16_8))[name = tensor("aw_433_cast_fp16")]; + tensor aw_435_equation_0 = const()[name = tensor("aw_435_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_435_cast_fp16 = einsum(equation = aw_435_equation_0, values = (var_3259_cast_fp16_9, var_3241_cast_fp16_9))[name = tensor("aw_435_cast_fp16")]; + tensor aw_437_equation_0 = const()[name = tensor("aw_437_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_437_cast_fp16 = einsum(equation = aw_437_equation_0, values = (var_3259_cast_fp16_10, var_3241_cast_fp16_10))[name = tensor("aw_437_cast_fp16")]; + tensor aw_439_equation_0 = const()[name = tensor("aw_439_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_439_cast_fp16 = einsum(equation = aw_439_equation_0, values = (var_3259_cast_fp16_11, var_3241_cast_fp16_11))[name = tensor("aw_439_cast_fp16")]; + tensor aw_441_equation_0 = const()[name = tensor("aw_441_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_441_cast_fp16 = einsum(equation = aw_441_equation_0, values = (var_3259_cast_fp16_12, var_3241_cast_fp16_12))[name = tensor("aw_441_cast_fp16")]; + tensor aw_443_equation_0 = const()[name = tensor("aw_443_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_443_cast_fp16 = einsum(equation = aw_443_equation_0, values = (var_3259_cast_fp16_13, var_3241_cast_fp16_13))[name = tensor("aw_443_cast_fp16")]; + tensor aw_445_equation_0 = const()[name = tensor("aw_445_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_445_cast_fp16 = einsum(equation = aw_445_equation_0, values = (var_3259_cast_fp16_14, var_3241_cast_fp16_14))[name = tensor("aw_445_cast_fp16")]; + tensor aw_447_equation_0 = const()[name = tensor("aw_447_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_447_cast_fp16 = einsum(equation = aw_447_equation_0, values = (var_3259_cast_fp16_15, var_3241_cast_fp16_15))[name = tensor("aw_447_cast_fp16")]; + tensor var_3325_cast_fp16 = softmax(axis = var_3189, x = aw_417_cast_fp16)[name = tensor("op_3325_cast_fp16")]; + tensor var_3326_cast_fp16 = softmax(axis = var_3189, x = aw_419_cast_fp16)[name = tensor("op_3326_cast_fp16")]; + tensor var_3327_cast_fp16 = softmax(axis = var_3189, x = aw_421_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = softmax(axis = var_3189, x = aw_423_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor var_3329_cast_fp16 = softmax(axis = var_3189, x = aw_425_cast_fp16)[name = tensor("op_3329_cast_fp16")]; + tensor var_3330_cast_fp16 = softmax(axis = var_3189, x = aw_427_cast_fp16)[name = tensor("op_3330_cast_fp16")]; + tensor var_3331_cast_fp16 = softmax(axis = var_3189, x = aw_429_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = softmax(axis = var_3189, x = aw_431_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_cast_fp16 = softmax(axis = var_3189, x = aw_433_cast_fp16)[name = tensor("op_3333_cast_fp16")]; + tensor var_3334_cast_fp16 = softmax(axis = var_3189, x = aw_435_cast_fp16)[name = tensor("op_3334_cast_fp16")]; + tensor var_3335_cast_fp16 = softmax(axis = var_3189, x = aw_437_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = softmax(axis = var_3189, x = aw_439_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor var_3337_cast_fp16 = softmax(axis = var_3189, x = aw_441_cast_fp16)[name = tensor("op_3337_cast_fp16")]; + tensor var_3338_cast_fp16 = softmax(axis = var_3189, x = aw_443_cast_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = softmax(axis = var_3189, x = aw_445_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor var_3340_cast_fp16 = softmax(axis = var_3189, x = aw_447_cast_fp16)[name = tensor("op_3340_cast_fp16")]; + tensor var_3342_equation_0 = const()[name = tensor("op_3342_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3342_cast_fp16 = einsum(equation = var_3342_equation_0, values = (var_3276_cast_fp16_0, var_3325_cast_fp16))[name = tensor("op_3342_cast_fp16")]; + tensor var_3344_equation_0 = const()[name = tensor("op_3344_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3344_cast_fp16 = einsum(equation = var_3344_equation_0, values = (var_3276_cast_fp16_1, var_3326_cast_fp16))[name = tensor("op_3344_cast_fp16")]; + tensor var_3346_equation_0 = const()[name = tensor("op_3346_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3346_cast_fp16 = einsum(equation = var_3346_equation_0, values = (var_3276_cast_fp16_2, var_3327_cast_fp16))[name = tensor("op_3346_cast_fp16")]; + tensor var_3348_equation_0 = const()[name = tensor("op_3348_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3348_cast_fp16 = einsum(equation = var_3348_equation_0, values = (var_3276_cast_fp16_3, var_3328_cast_fp16))[name = tensor("op_3348_cast_fp16")]; + tensor var_3350_equation_0 = const()[name = tensor("op_3350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3350_cast_fp16 = einsum(equation = var_3350_equation_0, values = (var_3276_cast_fp16_4, var_3329_cast_fp16))[name = tensor("op_3350_cast_fp16")]; + tensor var_3352_equation_0 = const()[name = tensor("op_3352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3352_cast_fp16 = einsum(equation = var_3352_equation_0, values = (var_3276_cast_fp16_5, var_3330_cast_fp16))[name = tensor("op_3352_cast_fp16")]; + tensor var_3354_equation_0 = const()[name = tensor("op_3354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3354_cast_fp16 = einsum(equation = var_3354_equation_0, values = (var_3276_cast_fp16_6, var_3331_cast_fp16))[name = tensor("op_3354_cast_fp16")]; + tensor var_3356_equation_0 = const()[name = tensor("op_3356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3356_cast_fp16 = einsum(equation = var_3356_equation_0, values = (var_3276_cast_fp16_7, var_3332_cast_fp16))[name = tensor("op_3356_cast_fp16")]; + tensor var_3358_equation_0 = const()[name = tensor("op_3358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3358_cast_fp16 = einsum(equation = var_3358_equation_0, values = (var_3276_cast_fp16_8, var_3333_cast_fp16))[name = tensor("op_3358_cast_fp16")]; + tensor var_3360_equation_0 = const()[name = tensor("op_3360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3360_cast_fp16 = einsum(equation = var_3360_equation_0, values = (var_3276_cast_fp16_9, var_3334_cast_fp16))[name = tensor("op_3360_cast_fp16")]; + tensor var_3362_equation_0 = const()[name = tensor("op_3362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3362_cast_fp16 = einsum(equation = var_3362_equation_0, values = (var_3276_cast_fp16_10, var_3335_cast_fp16))[name = tensor("op_3362_cast_fp16")]; + tensor var_3364_equation_0 = const()[name = tensor("op_3364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3364_cast_fp16 = einsum(equation = var_3364_equation_0, values = (var_3276_cast_fp16_11, var_3336_cast_fp16))[name = tensor("op_3364_cast_fp16")]; + tensor var_3366_equation_0 = const()[name = tensor("op_3366_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3366_cast_fp16 = einsum(equation = var_3366_equation_0, values = (var_3276_cast_fp16_12, var_3337_cast_fp16))[name = tensor("op_3366_cast_fp16")]; + tensor var_3368_equation_0 = const()[name = tensor("op_3368_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3368_cast_fp16 = einsum(equation = var_3368_equation_0, values = (var_3276_cast_fp16_13, var_3338_cast_fp16))[name = tensor("op_3368_cast_fp16")]; + tensor var_3370_equation_0 = const()[name = tensor("op_3370_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3370_cast_fp16 = einsum(equation = var_3370_equation_0, values = (var_3276_cast_fp16_14, var_3339_cast_fp16))[name = tensor("op_3370_cast_fp16")]; + tensor var_3372_equation_0 = const()[name = tensor("op_3372_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3372_cast_fp16 = einsum(equation = var_3372_equation_0, values = (var_3276_cast_fp16_15, var_3340_cast_fp16))[name = tensor("op_3372_cast_fp16")]; + tensor input_135_interleave_0 = const()[name = tensor("input_135_interleave_0"), val = tensor(false)]; + tensor input_135_cast_fp16 = concat(axis = var_3189, interleave = input_135_interleave_0, values = (var_3342_cast_fp16, var_3344_cast_fp16, var_3346_cast_fp16, var_3348_cast_fp16, var_3350_cast_fp16, var_3352_cast_fp16, var_3354_cast_fp16, var_3356_cast_fp16, var_3358_cast_fp16, var_3360_cast_fp16, var_3362_cast_fp16, var_3364_cast_fp16, var_3366_cast_fp16, var_3368_cast_fp16, var_3370_cast_fp16, var_3372_cast_fp16))[name = tensor("input_135_cast_fp16")]; + tensor var_3381_pad_type_0 = const()[name = tensor("op_3381_pad_type_0"), val = tensor("valid")]; + tensor var_3381_strides_0 = const()[name = tensor("op_3381_strides_0"), val = tensor([1, 1])]; + tensor var_3381_pad_0 = const()[name = tensor("op_3381_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3381_dilations_0 = const()[name = tensor("op_3381_dilations_0"), val = tensor([1, 1])]; + tensor var_3381_groups_0 = const()[name = tensor("op_3381_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_out_weight_to_fp16 = const()[name = tensor("blocks_13_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343647232)))]; + tensor blocks_13_attn_out_bias_to_fp16 = const()[name = tensor("blocks_13_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345744448)))]; + tensor var_3381_cast_fp16 = conv(bias = blocks_13_attn_out_bias_to_fp16, dilations = var_3381_dilations_0, groups = var_3381_groups_0, pad = var_3381_pad_0, pad_type = var_3381_pad_type_0, strides = var_3381_strides_0, weight = blocks_13_attn_out_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("op_3381_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = var_3381_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([1])]; + tensor input_137_gamma_0_to_fp16 = const()[name = tensor("input_137_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345746560)))]; + tensor input_137_beta_0_to_fp16 = const()[name = tensor("input_137_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345748672)))]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = input_137_beta_0_to_fp16, epsilon = var_3391_to_fp16, gamma = input_137_gamma_0_to_fp16, x = inputs_55_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("valid")]; + tensor input_139_strides_0 = const()[name = tensor("input_139_strides_0"), val = tensor([1, 1])]; + tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_139_dilations_0 = const()[name = tensor("input_139_dilations_0"), val = tensor([1, 1])]; + tensor input_139_groups_0 = const()[name = tensor("input_139_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345750784)))]; + tensor blocks_13_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354139456)))]; + tensor input_139_cast_fp16 = conv(bias = blocks_13_mlp_0_bias_to_fp16, dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = blocks_13_mlp_0_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3417_pad_type_0 = const()[name = tensor("op_3417_pad_type_0"), val = tensor("valid")]; + tensor var_3417_strides_0 = const()[name = tensor("op_3417_strides_0"), val = tensor([1, 1])]; + tensor var_3417_pad_0 = const()[name = tensor("op_3417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3417_dilations_0 = const()[name = tensor("op_3417_dilations_0"), val = tensor([1, 1])]; + tensor var_3417_groups_0 = const()[name = tensor("op_3417_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354147712)))]; + tensor blocks_13_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362536384)))]; + tensor var_3417_cast_fp16 = conv(bias = blocks_13_mlp_2_bias_to_fp16, dilations = var_3417_dilations_0, groups = var_3417_groups_0, pad = var_3417_pad_0, pad_type = var_3417_pad_type_0, strides = var_3417_strides_0, weight = blocks_13_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = var_3417_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3426 = const()[name = tensor("op_3426"), val = tensor(1)]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([1])]; + tensor input_143_gamma_0_to_fp16 = const()[name = tensor("input_143_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362538496)))]; + tensor input_143_beta_0_to_fp16 = const()[name = tensor("input_143_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362540608)))]; + tensor var_3442_to_fp16 = const()[name = tensor("op_3442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = input_143_beta_0_to_fp16, epsilon = var_3442_to_fp16, gamma = input_143_gamma_0_to_fp16, x = inputs_57_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor q_29_pad_type_0 = const()[name = tensor("q_29_pad_type_0"), val = tensor("valid")]; + tensor q_29_strides_0 = const()[name = tensor("q_29_strides_0"), val = tensor([1, 1])]; + tensor q_29_pad_0 = const()[name = tensor("q_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_29_dilations_0 = const()[name = tensor("q_29_dilations_0"), val = tensor([1, 1])]; + tensor q_29_groups_0 = const()[name = tensor("q_29_groups_0"), val = tensor(1)]; + tensor var_3477_weight_0_to_fp16 = const()[name = tensor("op_3477_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362542720)))]; + tensor var_3477_bias_0_to_fp16 = const()[name = tensor("op_3477_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364639936)))]; + tensor var_3477_cast_fp16 = conv(bias = var_3477_bias_0_to_fp16, dilations = q_29_dilations_0, groups = q_29_groups_0, pad = q_29_pad_0, pad_type = q_29_pad_type_0, strides = q_29_strides_0, weight = var_3477_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor k_29_pad_type_0 = const()[name = tensor("k_29_pad_type_0"), val = tensor("valid")]; + tensor k_29_strides_0 = const()[name = tensor("k_29_strides_0"), val = tensor([1, 1])]; + tensor k_29_pad_0 = const()[name = tensor("k_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_29_dilations_0 = const()[name = tensor("k_29_dilations_0"), val = tensor([1, 1])]; + tensor k_29_groups_0 = const()[name = tensor("k_29_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_key_weight_to_fp16 = const()[name = tensor("blocks_14_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364642048)))]; + tensor k_29_cast_fp16 = conv(dilations = k_29_dilations_0, groups = k_29_groups_0, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = k_29_strides_0, weight = blocks_14_attn_key_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor var_3475_pad_type_0 = const()[name = tensor("op_3475_pad_type_0"), val = tensor("valid")]; + tensor var_3475_strides_0 = const()[name = tensor("op_3475_strides_0"), val = tensor([1, 1])]; + tensor var_3475_pad_0 = const()[name = tensor("op_3475_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3475_dilations_0 = const()[name = tensor("op_3475_dilations_0"), val = tensor([1, 1])]; + tensor var_3475_groups_0 = const()[name = tensor("op_3475_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_value_weight_to_fp16 = const()[name = tensor("blocks_14_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366739264)))]; + tensor blocks_14_attn_value_bias_to_fp16 = const()[name = tensor("blocks_14_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368836480)))]; + tensor var_3475_cast_fp16 = conv(bias = blocks_14_attn_value_bias_to_fp16, dilations = var_3475_dilations_0, groups = var_3475_groups_0, pad = var_3475_pad_0, pad_type = var_3475_pad_type_0, strides = var_3475_strides_0, weight = blocks_14_attn_value_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3475_cast_fp16")]; + tensor tile_42 = const()[name = tensor("tile_42"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3478_axis_0 = const()[name = tensor("op_3478_axis_0"), val = tensor(1)]; + tensor var_3478_cast_fp16_0, tensor var_3478_cast_fp16_1, tensor var_3478_cast_fp16_2, tensor var_3478_cast_fp16_3, tensor var_3478_cast_fp16_4, tensor var_3478_cast_fp16_5, tensor var_3478_cast_fp16_6, tensor var_3478_cast_fp16_7, tensor var_3478_cast_fp16_8, tensor var_3478_cast_fp16_9, tensor var_3478_cast_fp16_10, tensor var_3478_cast_fp16_11, tensor var_3478_cast_fp16_12, tensor var_3478_cast_fp16_13, tensor var_3478_cast_fp16_14, tensor var_3478_cast_fp16_15 = split(axis = var_3478_axis_0, split_sizes = tile_42, x = var_3477_cast_fp16)[name = tensor("op_3478_cast_fp16")]; + tensor var_3495_perm_0 = const()[name = tensor("op_3495_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_43 = const()[name = tensor("tile_43"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3496_axis_0 = const()[name = tensor("op_3496_axis_0"), val = tensor(3)]; + tensor var_3495_cast_fp16 = transpose(perm = var_3495_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_10")]; + tensor var_3496_cast_fp16_0, tensor var_3496_cast_fp16_1, tensor var_3496_cast_fp16_2, tensor var_3496_cast_fp16_3, tensor var_3496_cast_fp16_4, tensor var_3496_cast_fp16_5, tensor var_3496_cast_fp16_6, tensor var_3496_cast_fp16_7, tensor var_3496_cast_fp16_8, tensor var_3496_cast_fp16_9, tensor var_3496_cast_fp16_10, tensor var_3496_cast_fp16_11, tensor var_3496_cast_fp16_12, tensor var_3496_cast_fp16_13, tensor var_3496_cast_fp16_14, tensor var_3496_cast_fp16_15 = split(axis = var_3496_axis_0, split_sizes = tile_43, x = var_3495_cast_fp16)[name = tensor("op_3496_cast_fp16")]; + tensor tile_44 = const()[name = tensor("tile_44"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3513_axis_0 = const()[name = tensor("op_3513_axis_0"), val = tensor(1)]; + tensor var_3513_cast_fp16_0, tensor var_3513_cast_fp16_1, tensor var_3513_cast_fp16_2, tensor var_3513_cast_fp16_3, tensor var_3513_cast_fp16_4, tensor var_3513_cast_fp16_5, tensor var_3513_cast_fp16_6, tensor var_3513_cast_fp16_7, tensor var_3513_cast_fp16_8, tensor var_3513_cast_fp16_9, tensor var_3513_cast_fp16_10, tensor var_3513_cast_fp16_11, tensor var_3513_cast_fp16_12, tensor var_3513_cast_fp16_13, tensor var_3513_cast_fp16_14, tensor var_3513_cast_fp16_15 = split(axis = var_3513_axis_0, split_sizes = tile_44, x = var_3475_cast_fp16)[name = tensor("op_3513_cast_fp16")]; + tensor aw_449_equation_0 = const()[name = tensor("aw_449_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_449_cast_fp16 = einsum(equation = aw_449_equation_0, values = (var_3496_cast_fp16_0, var_3478_cast_fp16_0))[name = tensor("aw_449_cast_fp16")]; + tensor aw_451_equation_0 = const()[name = tensor("aw_451_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_451_cast_fp16 = einsum(equation = aw_451_equation_0, values = (var_3496_cast_fp16_1, var_3478_cast_fp16_1))[name = tensor("aw_451_cast_fp16")]; + tensor aw_453_equation_0 = const()[name = tensor("aw_453_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_453_cast_fp16 = einsum(equation = aw_453_equation_0, values = (var_3496_cast_fp16_2, var_3478_cast_fp16_2))[name = tensor("aw_453_cast_fp16")]; + tensor aw_455_equation_0 = const()[name = tensor("aw_455_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_455_cast_fp16 = einsum(equation = aw_455_equation_0, values = (var_3496_cast_fp16_3, var_3478_cast_fp16_3))[name = tensor("aw_455_cast_fp16")]; + tensor aw_457_equation_0 = const()[name = tensor("aw_457_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_457_cast_fp16 = einsum(equation = aw_457_equation_0, values = (var_3496_cast_fp16_4, var_3478_cast_fp16_4))[name = tensor("aw_457_cast_fp16")]; + tensor aw_459_equation_0 = const()[name = tensor("aw_459_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_459_cast_fp16 = einsum(equation = aw_459_equation_0, values = (var_3496_cast_fp16_5, var_3478_cast_fp16_5))[name = tensor("aw_459_cast_fp16")]; + tensor aw_461_equation_0 = const()[name = tensor("aw_461_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_461_cast_fp16 = einsum(equation = aw_461_equation_0, values = (var_3496_cast_fp16_6, var_3478_cast_fp16_6))[name = tensor("aw_461_cast_fp16")]; + tensor aw_463_equation_0 = const()[name = tensor("aw_463_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_463_cast_fp16 = einsum(equation = aw_463_equation_0, values = (var_3496_cast_fp16_7, var_3478_cast_fp16_7))[name = tensor("aw_463_cast_fp16")]; + tensor aw_465_equation_0 = const()[name = tensor("aw_465_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_465_cast_fp16 = einsum(equation = aw_465_equation_0, values = (var_3496_cast_fp16_8, var_3478_cast_fp16_8))[name = tensor("aw_465_cast_fp16")]; + tensor aw_467_equation_0 = const()[name = tensor("aw_467_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_467_cast_fp16 = einsum(equation = aw_467_equation_0, values = (var_3496_cast_fp16_9, var_3478_cast_fp16_9))[name = tensor("aw_467_cast_fp16")]; + tensor aw_469_equation_0 = const()[name = tensor("aw_469_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_469_cast_fp16 = einsum(equation = aw_469_equation_0, values = (var_3496_cast_fp16_10, var_3478_cast_fp16_10))[name = tensor("aw_469_cast_fp16")]; + tensor aw_471_equation_0 = const()[name = tensor("aw_471_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_471_cast_fp16 = einsum(equation = aw_471_equation_0, values = (var_3496_cast_fp16_11, var_3478_cast_fp16_11))[name = tensor("aw_471_cast_fp16")]; + tensor aw_473_equation_0 = const()[name = tensor("aw_473_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_473_cast_fp16 = einsum(equation = aw_473_equation_0, values = (var_3496_cast_fp16_12, var_3478_cast_fp16_12))[name = tensor("aw_473_cast_fp16")]; + tensor aw_475_equation_0 = const()[name = tensor("aw_475_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_475_cast_fp16 = einsum(equation = aw_475_equation_0, values = (var_3496_cast_fp16_13, var_3478_cast_fp16_13))[name = tensor("aw_475_cast_fp16")]; + tensor aw_477_equation_0 = const()[name = tensor("aw_477_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_477_cast_fp16 = einsum(equation = aw_477_equation_0, values = (var_3496_cast_fp16_14, var_3478_cast_fp16_14))[name = tensor("aw_477_cast_fp16")]; + tensor aw_479_equation_0 = const()[name = tensor("aw_479_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_479_cast_fp16 = einsum(equation = aw_479_equation_0, values = (var_3496_cast_fp16_15, var_3478_cast_fp16_15))[name = tensor("aw_479_cast_fp16")]; + tensor var_3562_cast_fp16 = softmax(axis = var_3426, x = aw_449_cast_fp16)[name = tensor("op_3562_cast_fp16")]; + tensor var_3563_cast_fp16 = softmax(axis = var_3426, x = aw_451_cast_fp16)[name = tensor("op_3563_cast_fp16")]; + tensor var_3564_cast_fp16 = softmax(axis = var_3426, x = aw_453_cast_fp16)[name = tensor("op_3564_cast_fp16")]; + tensor var_3565_cast_fp16 = softmax(axis = var_3426, x = aw_455_cast_fp16)[name = tensor("op_3565_cast_fp16")]; + tensor var_3566_cast_fp16 = softmax(axis = var_3426, x = aw_457_cast_fp16)[name = tensor("op_3566_cast_fp16")]; + tensor var_3567_cast_fp16 = softmax(axis = var_3426, x = aw_459_cast_fp16)[name = tensor("op_3567_cast_fp16")]; + tensor var_3568_cast_fp16 = softmax(axis = var_3426, x = aw_461_cast_fp16)[name = tensor("op_3568_cast_fp16")]; + tensor var_3569_cast_fp16 = softmax(axis = var_3426, x = aw_463_cast_fp16)[name = tensor("op_3569_cast_fp16")]; + tensor var_3570_cast_fp16 = softmax(axis = var_3426, x = aw_465_cast_fp16)[name = tensor("op_3570_cast_fp16")]; + tensor var_3571_cast_fp16 = softmax(axis = var_3426, x = aw_467_cast_fp16)[name = tensor("op_3571_cast_fp16")]; + tensor var_3572_cast_fp16 = softmax(axis = var_3426, x = aw_469_cast_fp16)[name = tensor("op_3572_cast_fp16")]; + tensor var_3573_cast_fp16 = softmax(axis = var_3426, x = aw_471_cast_fp16)[name = tensor("op_3573_cast_fp16")]; + tensor var_3574_cast_fp16 = softmax(axis = var_3426, x = aw_473_cast_fp16)[name = tensor("op_3574_cast_fp16")]; + tensor var_3575_cast_fp16 = softmax(axis = var_3426, x = aw_475_cast_fp16)[name = tensor("op_3575_cast_fp16")]; + tensor var_3576_cast_fp16 = softmax(axis = var_3426, x = aw_477_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_cast_fp16 = softmax(axis = var_3426, x = aw_479_cast_fp16)[name = tensor("op_3577_cast_fp16")]; + tensor var_3579_equation_0 = const()[name = tensor("op_3579_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3579_cast_fp16 = einsum(equation = var_3579_equation_0, values = (var_3513_cast_fp16_0, var_3562_cast_fp16))[name = tensor("op_3579_cast_fp16")]; + tensor var_3581_equation_0 = const()[name = tensor("op_3581_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3581_cast_fp16 = einsum(equation = var_3581_equation_0, values = (var_3513_cast_fp16_1, var_3563_cast_fp16))[name = tensor("op_3581_cast_fp16")]; + tensor var_3583_equation_0 = const()[name = tensor("op_3583_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3583_cast_fp16 = einsum(equation = var_3583_equation_0, values = (var_3513_cast_fp16_2, var_3564_cast_fp16))[name = tensor("op_3583_cast_fp16")]; + tensor var_3585_equation_0 = const()[name = tensor("op_3585_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3585_cast_fp16 = einsum(equation = var_3585_equation_0, values = (var_3513_cast_fp16_3, var_3565_cast_fp16))[name = tensor("op_3585_cast_fp16")]; + tensor var_3587_equation_0 = const()[name = tensor("op_3587_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3587_cast_fp16 = einsum(equation = var_3587_equation_0, values = (var_3513_cast_fp16_4, var_3566_cast_fp16))[name = tensor("op_3587_cast_fp16")]; + tensor var_3589_equation_0 = const()[name = tensor("op_3589_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3589_cast_fp16 = einsum(equation = var_3589_equation_0, values = (var_3513_cast_fp16_5, var_3567_cast_fp16))[name = tensor("op_3589_cast_fp16")]; + tensor var_3591_equation_0 = const()[name = tensor("op_3591_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3591_cast_fp16 = einsum(equation = var_3591_equation_0, values = (var_3513_cast_fp16_6, var_3568_cast_fp16))[name = tensor("op_3591_cast_fp16")]; + tensor var_3593_equation_0 = const()[name = tensor("op_3593_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3593_cast_fp16 = einsum(equation = var_3593_equation_0, values = (var_3513_cast_fp16_7, var_3569_cast_fp16))[name = tensor("op_3593_cast_fp16")]; + tensor var_3595_equation_0 = const()[name = tensor("op_3595_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3595_cast_fp16 = einsum(equation = var_3595_equation_0, values = (var_3513_cast_fp16_8, var_3570_cast_fp16))[name = tensor("op_3595_cast_fp16")]; + tensor var_3597_equation_0 = const()[name = tensor("op_3597_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3597_cast_fp16 = einsum(equation = var_3597_equation_0, values = (var_3513_cast_fp16_9, var_3571_cast_fp16))[name = tensor("op_3597_cast_fp16")]; + tensor var_3599_equation_0 = const()[name = tensor("op_3599_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3599_cast_fp16 = einsum(equation = var_3599_equation_0, values = (var_3513_cast_fp16_10, var_3572_cast_fp16))[name = tensor("op_3599_cast_fp16")]; + tensor var_3601_equation_0 = const()[name = tensor("op_3601_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3601_cast_fp16 = einsum(equation = var_3601_equation_0, values = (var_3513_cast_fp16_11, var_3573_cast_fp16))[name = tensor("op_3601_cast_fp16")]; + tensor var_3603_equation_0 = const()[name = tensor("op_3603_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3603_cast_fp16 = einsum(equation = var_3603_equation_0, values = (var_3513_cast_fp16_12, var_3574_cast_fp16))[name = tensor("op_3603_cast_fp16")]; + tensor var_3605_equation_0 = const()[name = tensor("op_3605_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3605_cast_fp16 = einsum(equation = var_3605_equation_0, values = (var_3513_cast_fp16_13, var_3575_cast_fp16))[name = tensor("op_3605_cast_fp16")]; + tensor var_3607_equation_0 = const()[name = tensor("op_3607_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3607_cast_fp16 = einsum(equation = var_3607_equation_0, values = (var_3513_cast_fp16_14, var_3576_cast_fp16))[name = tensor("op_3607_cast_fp16")]; + tensor var_3609_equation_0 = const()[name = tensor("op_3609_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3609_cast_fp16 = einsum(equation = var_3609_equation_0, values = (var_3513_cast_fp16_15, var_3577_cast_fp16))[name = tensor("op_3609_cast_fp16")]; + tensor input_145_interleave_0 = const()[name = tensor("input_145_interleave_0"), val = tensor(false)]; + tensor input_145_cast_fp16 = concat(axis = var_3426, interleave = input_145_interleave_0, values = (var_3579_cast_fp16, var_3581_cast_fp16, var_3583_cast_fp16, var_3585_cast_fp16, var_3587_cast_fp16, var_3589_cast_fp16, var_3591_cast_fp16, var_3593_cast_fp16, var_3595_cast_fp16, var_3597_cast_fp16, var_3599_cast_fp16, var_3601_cast_fp16, var_3603_cast_fp16, var_3605_cast_fp16, var_3607_cast_fp16, var_3609_cast_fp16))[name = tensor("input_145_cast_fp16")]; + tensor var_3618_pad_type_0 = const()[name = tensor("op_3618_pad_type_0"), val = tensor("valid")]; + tensor var_3618_strides_0 = const()[name = tensor("op_3618_strides_0"), val = tensor([1, 1])]; + tensor var_3618_pad_0 = const()[name = tensor("op_3618_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3618_dilations_0 = const()[name = tensor("op_3618_dilations_0"), val = tensor([1, 1])]; + tensor var_3618_groups_0 = const()[name = tensor("op_3618_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_out_weight_to_fp16 = const()[name = tensor("blocks_14_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368838592)))]; + tensor blocks_14_attn_out_bias_to_fp16 = const()[name = tensor("blocks_14_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370935808)))]; + tensor var_3618_cast_fp16 = conv(bias = blocks_14_attn_out_bias_to_fp16, dilations = var_3618_dilations_0, groups = var_3618_groups_0, pad = var_3618_pad_0, pad_type = var_3618_pad_type_0, strides = var_3618_strides_0, weight = blocks_14_attn_out_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_3618_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([1])]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370937920)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370940032)))]; + tensor var_3628_to_fp16 = const()[name = tensor("op_3628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = input_147_beta_0_to_fp16, epsilon = var_3628_to_fp16, gamma = input_147_gamma_0_to_fp16, x = inputs_59_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370942144)))]; + tensor blocks_14_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379330816)))]; + tensor input_149_cast_fp16 = conv(bias = blocks_14_mlp_0_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = blocks_14_mlp_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_3654_pad_type_0 = const()[name = tensor("op_3654_pad_type_0"), val = tensor("valid")]; + tensor var_3654_strides_0 = const()[name = tensor("op_3654_strides_0"), val = tensor([1, 1])]; + tensor var_3654_pad_0 = const()[name = tensor("op_3654_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3654_dilations_0 = const()[name = tensor("op_3654_dilations_0"), val = tensor([1, 1])]; + tensor var_3654_groups_0 = const()[name = tensor("op_3654_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379339072)))]; + tensor blocks_14_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387727744)))]; + tensor var_3654_cast_fp16 = conv(bias = blocks_14_mlp_2_bias_to_fp16, dilations = var_3654_dilations_0, groups = var_3654_groups_0, pad = var_3654_pad_0, pad_type = var_3654_pad_type_0, strides = var_3654_strides_0, weight = blocks_14_mlp_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("op_3654_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = var_3654_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_3663 = const()[name = tensor("op_3663"), val = tensor(1)]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([1])]; + tensor input_153_gamma_0_to_fp16 = const()[name = tensor("input_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387729856)))]; + tensor input_153_beta_0_to_fp16 = const()[name = tensor("input_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387731968)))]; + tensor var_3679_to_fp16 = const()[name = tensor("op_3679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = input_153_beta_0_to_fp16, epsilon = var_3679_to_fp16, gamma = input_153_gamma_0_to_fp16, x = inputs_61_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor q_31_pad_type_0 = const()[name = tensor("q_31_pad_type_0"), val = tensor("valid")]; + tensor q_31_strides_0 = const()[name = tensor("q_31_strides_0"), val = tensor([1, 1])]; + tensor q_31_pad_0 = const()[name = tensor("q_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_31_dilations_0 = const()[name = tensor("q_31_dilations_0"), val = tensor([1, 1])]; + tensor q_31_groups_0 = const()[name = tensor("q_31_groups_0"), val = tensor(1)]; + tensor var_3714_weight_0_to_fp16 = const()[name = tensor("op_3714_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387734080)))]; + tensor var_3714_bias_0_to_fp16 = const()[name = tensor("op_3714_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389831296)))]; + tensor var_3714_cast_fp16 = conv(bias = var_3714_bias_0_to_fp16, dilations = q_31_dilations_0, groups = q_31_groups_0, pad = q_31_pad_0, pad_type = q_31_pad_type_0, strides = q_31_strides_0, weight = var_3714_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("op_3714_cast_fp16")]; + tensor k_31_pad_type_0 = const()[name = tensor("k_31_pad_type_0"), val = tensor("valid")]; + tensor k_31_strides_0 = const()[name = tensor("k_31_strides_0"), val = tensor([1, 1])]; + tensor k_31_pad_0 = const()[name = tensor("k_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_31_dilations_0 = const()[name = tensor("k_31_dilations_0"), val = tensor([1, 1])]; + tensor k_31_groups_0 = const()[name = tensor("k_31_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_key_weight_to_fp16 = const()[name = tensor("blocks_15_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389833408)))]; + tensor k_31_cast_fp16 = conv(dilations = k_31_dilations_0, groups = k_31_groups_0, pad = k_31_pad_0, pad_type = k_31_pad_type_0, strides = k_31_strides_0, weight = blocks_15_attn_key_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor var_3712_pad_type_0 = const()[name = tensor("op_3712_pad_type_0"), val = tensor("valid")]; + tensor var_3712_strides_0 = const()[name = tensor("op_3712_strides_0"), val = tensor([1, 1])]; + tensor var_3712_pad_0 = const()[name = tensor("op_3712_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3712_dilations_0 = const()[name = tensor("op_3712_dilations_0"), val = tensor([1, 1])]; + tensor var_3712_groups_0 = const()[name = tensor("op_3712_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_value_weight_to_fp16 = const()[name = tensor("blocks_15_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391930624)))]; + tensor blocks_15_attn_value_bias_to_fp16 = const()[name = tensor("blocks_15_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394027840)))]; + tensor var_3712_cast_fp16 = conv(bias = blocks_15_attn_value_bias_to_fp16, dilations = var_3712_dilations_0, groups = var_3712_groups_0, pad = var_3712_pad_0, pad_type = var_3712_pad_type_0, strides = var_3712_strides_0, weight = blocks_15_attn_value_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("op_3712_cast_fp16")]; + tensor tile_45 = const()[name = tensor("tile_45"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3715_axis_0 = const()[name = tensor("op_3715_axis_0"), val = tensor(1)]; + tensor var_3715_cast_fp16_0, tensor var_3715_cast_fp16_1, tensor var_3715_cast_fp16_2, tensor var_3715_cast_fp16_3, tensor var_3715_cast_fp16_4, tensor var_3715_cast_fp16_5, tensor var_3715_cast_fp16_6, tensor var_3715_cast_fp16_7, tensor var_3715_cast_fp16_8, tensor var_3715_cast_fp16_9, tensor var_3715_cast_fp16_10, tensor var_3715_cast_fp16_11, tensor var_3715_cast_fp16_12, tensor var_3715_cast_fp16_13, tensor var_3715_cast_fp16_14, tensor var_3715_cast_fp16_15 = split(axis = var_3715_axis_0, split_sizes = tile_45, x = var_3714_cast_fp16)[name = tensor("op_3715_cast_fp16")]; + tensor var_3732_perm_0 = const()[name = tensor("op_3732_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_46 = const()[name = tensor("tile_46"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3733_axis_0 = const()[name = tensor("op_3733_axis_0"), val = tensor(3)]; + tensor var_3732_cast_fp16 = transpose(perm = var_3732_perm_0, x = k_31_cast_fp16)[name = tensor("transpose_9")]; + tensor var_3733_cast_fp16_0, tensor var_3733_cast_fp16_1, tensor var_3733_cast_fp16_2, tensor var_3733_cast_fp16_3, tensor var_3733_cast_fp16_4, tensor var_3733_cast_fp16_5, tensor var_3733_cast_fp16_6, tensor var_3733_cast_fp16_7, tensor var_3733_cast_fp16_8, tensor var_3733_cast_fp16_9, tensor var_3733_cast_fp16_10, tensor var_3733_cast_fp16_11, tensor var_3733_cast_fp16_12, tensor var_3733_cast_fp16_13, tensor var_3733_cast_fp16_14, tensor var_3733_cast_fp16_15 = split(axis = var_3733_axis_0, split_sizes = tile_46, x = var_3732_cast_fp16)[name = tensor("op_3733_cast_fp16")]; + tensor tile_47 = const()[name = tensor("tile_47"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3750_axis_0 = const()[name = tensor("op_3750_axis_0"), val = tensor(1)]; + tensor var_3750_cast_fp16_0, tensor var_3750_cast_fp16_1, tensor var_3750_cast_fp16_2, tensor var_3750_cast_fp16_3, tensor var_3750_cast_fp16_4, tensor var_3750_cast_fp16_5, tensor var_3750_cast_fp16_6, tensor var_3750_cast_fp16_7, tensor var_3750_cast_fp16_8, tensor var_3750_cast_fp16_9, tensor var_3750_cast_fp16_10, tensor var_3750_cast_fp16_11, tensor var_3750_cast_fp16_12, tensor var_3750_cast_fp16_13, tensor var_3750_cast_fp16_14, tensor var_3750_cast_fp16_15 = split(axis = var_3750_axis_0, split_sizes = tile_47, x = var_3712_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor aw_481_equation_0 = const()[name = tensor("aw_481_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_481_cast_fp16 = einsum(equation = aw_481_equation_0, values = (var_3733_cast_fp16_0, var_3715_cast_fp16_0))[name = tensor("aw_481_cast_fp16")]; + tensor aw_483_equation_0 = const()[name = tensor("aw_483_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_483_cast_fp16 = einsum(equation = aw_483_equation_0, values = (var_3733_cast_fp16_1, var_3715_cast_fp16_1))[name = tensor("aw_483_cast_fp16")]; + tensor aw_485_equation_0 = const()[name = tensor("aw_485_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_485_cast_fp16 = einsum(equation = aw_485_equation_0, values = (var_3733_cast_fp16_2, var_3715_cast_fp16_2))[name = tensor("aw_485_cast_fp16")]; + tensor aw_487_equation_0 = const()[name = tensor("aw_487_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_487_cast_fp16 = einsum(equation = aw_487_equation_0, values = (var_3733_cast_fp16_3, var_3715_cast_fp16_3))[name = tensor("aw_487_cast_fp16")]; + tensor aw_489_equation_0 = const()[name = tensor("aw_489_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_489_cast_fp16 = einsum(equation = aw_489_equation_0, values = (var_3733_cast_fp16_4, var_3715_cast_fp16_4))[name = tensor("aw_489_cast_fp16")]; + tensor aw_491_equation_0 = const()[name = tensor("aw_491_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_491_cast_fp16 = einsum(equation = aw_491_equation_0, values = (var_3733_cast_fp16_5, var_3715_cast_fp16_5))[name = tensor("aw_491_cast_fp16")]; + tensor aw_493_equation_0 = const()[name = tensor("aw_493_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_493_cast_fp16 = einsum(equation = aw_493_equation_0, values = (var_3733_cast_fp16_6, var_3715_cast_fp16_6))[name = tensor("aw_493_cast_fp16")]; + tensor aw_495_equation_0 = const()[name = tensor("aw_495_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_495_cast_fp16 = einsum(equation = aw_495_equation_0, values = (var_3733_cast_fp16_7, var_3715_cast_fp16_7))[name = tensor("aw_495_cast_fp16")]; + tensor aw_497_equation_0 = const()[name = tensor("aw_497_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_497_cast_fp16 = einsum(equation = aw_497_equation_0, values = (var_3733_cast_fp16_8, var_3715_cast_fp16_8))[name = tensor("aw_497_cast_fp16")]; + tensor aw_499_equation_0 = const()[name = tensor("aw_499_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_499_cast_fp16 = einsum(equation = aw_499_equation_0, values = (var_3733_cast_fp16_9, var_3715_cast_fp16_9))[name = tensor("aw_499_cast_fp16")]; + tensor aw_501_equation_0 = const()[name = tensor("aw_501_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_501_cast_fp16 = einsum(equation = aw_501_equation_0, values = (var_3733_cast_fp16_10, var_3715_cast_fp16_10))[name = tensor("aw_501_cast_fp16")]; + tensor aw_503_equation_0 = const()[name = tensor("aw_503_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_503_cast_fp16 = einsum(equation = aw_503_equation_0, values = (var_3733_cast_fp16_11, var_3715_cast_fp16_11))[name = tensor("aw_503_cast_fp16")]; + tensor aw_505_equation_0 = const()[name = tensor("aw_505_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_505_cast_fp16 = einsum(equation = aw_505_equation_0, values = (var_3733_cast_fp16_12, var_3715_cast_fp16_12))[name = tensor("aw_505_cast_fp16")]; + tensor aw_507_equation_0 = const()[name = tensor("aw_507_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_507_cast_fp16 = einsum(equation = aw_507_equation_0, values = (var_3733_cast_fp16_13, var_3715_cast_fp16_13))[name = tensor("aw_507_cast_fp16")]; + tensor aw_509_equation_0 = const()[name = tensor("aw_509_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_509_cast_fp16 = einsum(equation = aw_509_equation_0, values = (var_3733_cast_fp16_14, var_3715_cast_fp16_14))[name = tensor("aw_509_cast_fp16")]; + tensor aw_511_equation_0 = const()[name = tensor("aw_511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_511_cast_fp16 = einsum(equation = aw_511_equation_0, values = (var_3733_cast_fp16_15, var_3715_cast_fp16_15))[name = tensor("aw_511_cast_fp16")]; + tensor var_3799_cast_fp16 = softmax(axis = var_3663, x = aw_481_cast_fp16)[name = tensor("op_3799_cast_fp16")]; + tensor var_3800_cast_fp16 = softmax(axis = var_3663, x = aw_483_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor var_3801_cast_fp16 = softmax(axis = var_3663, x = aw_485_cast_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor var_3802_cast_fp16 = softmax(axis = var_3663, x = aw_487_cast_fp16)[name = tensor("op_3802_cast_fp16")]; + tensor var_3803_cast_fp16 = softmax(axis = var_3663, x = aw_489_cast_fp16)[name = tensor("op_3803_cast_fp16")]; + tensor var_3804_cast_fp16 = softmax(axis = var_3663, x = aw_491_cast_fp16)[name = tensor("op_3804_cast_fp16")]; + tensor var_3805_cast_fp16 = softmax(axis = var_3663, x = aw_493_cast_fp16)[name = tensor("op_3805_cast_fp16")]; + tensor var_3806_cast_fp16 = softmax(axis = var_3663, x = aw_495_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor var_3807_cast_fp16 = softmax(axis = var_3663, x = aw_497_cast_fp16)[name = tensor("op_3807_cast_fp16")]; + tensor var_3808_cast_fp16 = softmax(axis = var_3663, x = aw_499_cast_fp16)[name = tensor("op_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = softmax(axis = var_3663, x = aw_501_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_cast_fp16 = softmax(axis = var_3663, x = aw_503_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = softmax(axis = var_3663, x = aw_505_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor var_3812_cast_fp16 = softmax(axis = var_3663, x = aw_507_cast_fp16)[name = tensor("op_3812_cast_fp16")]; + tensor var_3813_cast_fp16 = softmax(axis = var_3663, x = aw_509_cast_fp16)[name = tensor("op_3813_cast_fp16")]; + tensor var_3814_cast_fp16 = softmax(axis = var_3663, x = aw_511_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3816_equation_0 = const()[name = tensor("op_3816_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3816_cast_fp16 = einsum(equation = var_3816_equation_0, values = (var_3750_cast_fp16_0, var_3799_cast_fp16))[name = tensor("op_3816_cast_fp16")]; + tensor var_3818_equation_0 = const()[name = tensor("op_3818_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3818_cast_fp16 = einsum(equation = var_3818_equation_0, values = (var_3750_cast_fp16_1, var_3800_cast_fp16))[name = tensor("op_3818_cast_fp16")]; + tensor var_3820_equation_0 = const()[name = tensor("op_3820_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3820_cast_fp16 = einsum(equation = var_3820_equation_0, values = (var_3750_cast_fp16_2, var_3801_cast_fp16))[name = tensor("op_3820_cast_fp16")]; + tensor var_3822_equation_0 = const()[name = tensor("op_3822_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3822_cast_fp16 = einsum(equation = var_3822_equation_0, values = (var_3750_cast_fp16_3, var_3802_cast_fp16))[name = tensor("op_3822_cast_fp16")]; + tensor var_3824_equation_0 = const()[name = tensor("op_3824_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3824_cast_fp16 = einsum(equation = var_3824_equation_0, values = (var_3750_cast_fp16_4, var_3803_cast_fp16))[name = tensor("op_3824_cast_fp16")]; + tensor var_3826_equation_0 = const()[name = tensor("op_3826_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3826_cast_fp16 = einsum(equation = var_3826_equation_0, values = (var_3750_cast_fp16_5, var_3804_cast_fp16))[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_equation_0 = const()[name = tensor("op_3828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3828_cast_fp16 = einsum(equation = var_3828_equation_0, values = (var_3750_cast_fp16_6, var_3805_cast_fp16))[name = tensor("op_3828_cast_fp16")]; + tensor var_3830_equation_0 = const()[name = tensor("op_3830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3830_cast_fp16 = einsum(equation = var_3830_equation_0, values = (var_3750_cast_fp16_7, var_3806_cast_fp16))[name = tensor("op_3830_cast_fp16")]; + tensor var_3832_equation_0 = const()[name = tensor("op_3832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3832_cast_fp16 = einsum(equation = var_3832_equation_0, values = (var_3750_cast_fp16_8, var_3807_cast_fp16))[name = tensor("op_3832_cast_fp16")]; + tensor var_3834_equation_0 = const()[name = tensor("op_3834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3834_cast_fp16 = einsum(equation = var_3834_equation_0, values = (var_3750_cast_fp16_9, var_3808_cast_fp16))[name = tensor("op_3834_cast_fp16")]; + tensor var_3836_equation_0 = const()[name = tensor("op_3836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3836_cast_fp16 = einsum(equation = var_3836_equation_0, values = (var_3750_cast_fp16_10, var_3809_cast_fp16))[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_equation_0 = const()[name = tensor("op_3838_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3838_cast_fp16 = einsum(equation = var_3838_equation_0, values = (var_3750_cast_fp16_11, var_3810_cast_fp16))[name = tensor("op_3838_cast_fp16")]; + tensor var_3840_equation_0 = const()[name = tensor("op_3840_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3840_cast_fp16 = einsum(equation = var_3840_equation_0, values = (var_3750_cast_fp16_12, var_3811_cast_fp16))[name = tensor("op_3840_cast_fp16")]; + tensor var_3842_equation_0 = const()[name = tensor("op_3842_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3842_cast_fp16 = einsum(equation = var_3842_equation_0, values = (var_3750_cast_fp16_13, var_3812_cast_fp16))[name = tensor("op_3842_cast_fp16")]; + tensor var_3844_equation_0 = const()[name = tensor("op_3844_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3844_cast_fp16 = einsum(equation = var_3844_equation_0, values = (var_3750_cast_fp16_14, var_3813_cast_fp16))[name = tensor("op_3844_cast_fp16")]; + tensor var_3846_equation_0 = const()[name = tensor("op_3846_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3846_cast_fp16 = einsum(equation = var_3846_equation_0, values = (var_3750_cast_fp16_15, var_3814_cast_fp16))[name = tensor("op_3846_cast_fp16")]; + tensor input_155_interleave_0 = const()[name = tensor("input_155_interleave_0"), val = tensor(false)]; + tensor input_155_cast_fp16 = concat(axis = var_3663, interleave = input_155_interleave_0, values = (var_3816_cast_fp16, var_3818_cast_fp16, var_3820_cast_fp16, var_3822_cast_fp16, var_3824_cast_fp16, var_3826_cast_fp16, var_3828_cast_fp16, var_3830_cast_fp16, var_3832_cast_fp16, var_3834_cast_fp16, var_3836_cast_fp16, var_3838_cast_fp16, var_3840_cast_fp16, var_3842_cast_fp16, var_3844_cast_fp16, var_3846_cast_fp16))[name = tensor("input_155_cast_fp16")]; + tensor var_3855_pad_type_0 = const()[name = tensor("op_3855_pad_type_0"), val = tensor("valid")]; + tensor var_3855_strides_0 = const()[name = tensor("op_3855_strides_0"), val = tensor([1, 1])]; + tensor var_3855_pad_0 = const()[name = tensor("op_3855_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3855_dilations_0 = const()[name = tensor("op_3855_dilations_0"), val = tensor([1, 1])]; + tensor var_3855_groups_0 = const()[name = tensor("op_3855_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_out_weight_to_fp16 = const()[name = tensor("blocks_15_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394029952)))]; + tensor blocks_15_attn_out_bias_to_fp16 = const()[name = tensor("blocks_15_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396127168)))]; + tensor var_3855_cast_fp16 = conv(bias = blocks_15_attn_out_bias_to_fp16, dilations = var_3855_dilations_0, groups = var_3855_groups_0, pad = var_3855_pad_0, pad_type = var_3855_pad_type_0, strides = var_3855_strides_0, weight = blocks_15_attn_out_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("op_3855_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_3855_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor input_157_axes_0 = const()[name = tensor("input_157_axes_0"), val = tensor([1])]; + tensor input_157_gamma_0_to_fp16 = const()[name = tensor("input_157_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396129280)))]; + tensor input_157_beta_0_to_fp16 = const()[name = tensor("input_157_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396131392)))]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = input_157_beta_0_to_fp16, epsilon = var_3865_to_fp16, gamma = input_157_gamma_0_to_fp16, x = inputs_63_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("valid")]; + tensor input_159_strides_0 = const()[name = tensor("input_159_strides_0"), val = tensor([1, 1])]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_159_dilations_0 = const()[name = tensor("input_159_dilations_0"), val = tensor([1, 1])]; + tensor input_159_groups_0 = const()[name = tensor("input_159_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396133504)))]; + tensor blocks_15_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404522176)))]; + tensor input_159_cast_fp16 = conv(bias = blocks_15_mlp_0_bias_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = blocks_15_mlp_0_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_3891_pad_type_0 = const()[name = tensor("op_3891_pad_type_0"), val = tensor("valid")]; + tensor var_3891_strides_0 = const()[name = tensor("op_3891_strides_0"), val = tensor([1, 1])]; + tensor var_3891_pad_0 = const()[name = tensor("op_3891_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3891_dilations_0 = const()[name = tensor("op_3891_dilations_0"), val = tensor([1, 1])]; + tensor var_3891_groups_0 = const()[name = tensor("op_3891_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404530432)))]; + tensor blocks_15_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412919104)))]; + tensor var_3891_cast_fp16 = conv(bias = blocks_15_mlp_2_bias_to_fp16, dilations = var_3891_dilations_0, groups = var_3891_groups_0, pad = var_3891_pad_0, pad_type = var_3891_pad_type_0, strides = var_3891_strides_0, weight = blocks_15_mlp_2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("op_3891_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = var_3891_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor(1)]; + tensor input_163_axes_0 = const()[name = tensor("input_163_axes_0"), val = tensor([1])]; + tensor input_163_gamma_0_to_fp16 = const()[name = tensor("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412921216)))]; + tensor input_163_beta_0_to_fp16 = const()[name = tensor("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412923328)))]; + tensor var_3916_to_fp16 = const()[name = tensor("op_3916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = input_163_beta_0_to_fp16, epsilon = var_3916_to_fp16, gamma = input_163_gamma_0_to_fp16, x = inputs_65_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor q_33_pad_type_0 = const()[name = tensor("q_33_pad_type_0"), val = tensor("valid")]; + tensor q_33_strides_0 = const()[name = tensor("q_33_strides_0"), val = tensor([1, 1])]; + tensor q_33_pad_0 = const()[name = tensor("q_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_33_dilations_0 = const()[name = tensor("q_33_dilations_0"), val = tensor([1, 1])]; + tensor q_33_groups_0 = const()[name = tensor("q_33_groups_0"), val = tensor(1)]; + tensor var_3951_weight_0_to_fp16 = const()[name = tensor("op_3951_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412925440)))]; + tensor var_3951_bias_0_to_fp16 = const()[name = tensor("op_3951_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415022656)))]; + tensor var_3951_cast_fp16 = conv(bias = var_3951_bias_0_to_fp16, dilations = q_33_dilations_0, groups = q_33_groups_0, pad = q_33_pad_0, pad_type = q_33_pad_type_0, strides = q_33_strides_0, weight = var_3951_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor k_33_pad_type_0 = const()[name = tensor("k_33_pad_type_0"), val = tensor("valid")]; + tensor k_33_strides_0 = const()[name = tensor("k_33_strides_0"), val = tensor([1, 1])]; + tensor k_33_pad_0 = const()[name = tensor("k_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_33_dilations_0 = const()[name = tensor("k_33_dilations_0"), val = tensor([1, 1])]; + tensor k_33_groups_0 = const()[name = tensor("k_33_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_key_weight_to_fp16 = const()[name = tensor("blocks_16_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415024768)))]; + tensor k_33_cast_fp16 = conv(dilations = k_33_dilations_0, groups = k_33_groups_0, pad = k_33_pad_0, pad_type = k_33_pad_type_0, strides = k_33_strides_0, weight = blocks_16_attn_key_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor var_3949_pad_type_0 = const()[name = tensor("op_3949_pad_type_0"), val = tensor("valid")]; + tensor var_3949_strides_0 = const()[name = tensor("op_3949_strides_0"), val = tensor([1, 1])]; + tensor var_3949_pad_0 = const()[name = tensor("op_3949_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3949_dilations_0 = const()[name = tensor("op_3949_dilations_0"), val = tensor([1, 1])]; + tensor var_3949_groups_0 = const()[name = tensor("op_3949_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_value_weight_to_fp16 = const()[name = tensor("blocks_16_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417121984)))]; + tensor blocks_16_attn_value_bias_to_fp16 = const()[name = tensor("blocks_16_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419219200)))]; + tensor var_3949_cast_fp16 = conv(bias = blocks_16_attn_value_bias_to_fp16, dilations = var_3949_dilations_0, groups = var_3949_groups_0, pad = var_3949_pad_0, pad_type = var_3949_pad_type_0, strides = var_3949_strides_0, weight = blocks_16_attn_value_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("op_3949_cast_fp16")]; + tensor tile_48 = const()[name = tensor("tile_48"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3952_axis_0 = const()[name = tensor("op_3952_axis_0"), val = tensor(1)]; + tensor var_3952_cast_fp16_0, tensor var_3952_cast_fp16_1, tensor var_3952_cast_fp16_2, tensor var_3952_cast_fp16_3, tensor var_3952_cast_fp16_4, tensor var_3952_cast_fp16_5, tensor var_3952_cast_fp16_6, tensor var_3952_cast_fp16_7, tensor var_3952_cast_fp16_8, tensor var_3952_cast_fp16_9, tensor var_3952_cast_fp16_10, tensor var_3952_cast_fp16_11, tensor var_3952_cast_fp16_12, tensor var_3952_cast_fp16_13, tensor var_3952_cast_fp16_14, tensor var_3952_cast_fp16_15 = split(axis = var_3952_axis_0, split_sizes = tile_48, x = var_3951_cast_fp16)[name = tensor("op_3952_cast_fp16")]; + tensor var_3969_perm_0 = const()[name = tensor("op_3969_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_49 = const()[name = tensor("tile_49"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3970_axis_0 = const()[name = tensor("op_3970_axis_0"), val = tensor(3)]; + tensor var_3969_cast_fp16 = transpose(perm = var_3969_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_8")]; + tensor var_3970_cast_fp16_0, tensor var_3970_cast_fp16_1, tensor var_3970_cast_fp16_2, tensor var_3970_cast_fp16_3, tensor var_3970_cast_fp16_4, tensor var_3970_cast_fp16_5, tensor var_3970_cast_fp16_6, tensor var_3970_cast_fp16_7, tensor var_3970_cast_fp16_8, tensor var_3970_cast_fp16_9, tensor var_3970_cast_fp16_10, tensor var_3970_cast_fp16_11, tensor var_3970_cast_fp16_12, tensor var_3970_cast_fp16_13, tensor var_3970_cast_fp16_14, tensor var_3970_cast_fp16_15 = split(axis = var_3970_axis_0, split_sizes = tile_49, x = var_3969_cast_fp16)[name = tensor("op_3970_cast_fp16")]; + tensor tile_50 = const()[name = tensor("tile_50"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3987_axis_0 = const()[name = tensor("op_3987_axis_0"), val = tensor(1)]; + tensor var_3987_cast_fp16_0, tensor var_3987_cast_fp16_1, tensor var_3987_cast_fp16_2, tensor var_3987_cast_fp16_3, tensor var_3987_cast_fp16_4, tensor var_3987_cast_fp16_5, tensor var_3987_cast_fp16_6, tensor var_3987_cast_fp16_7, tensor var_3987_cast_fp16_8, tensor var_3987_cast_fp16_9, tensor var_3987_cast_fp16_10, tensor var_3987_cast_fp16_11, tensor var_3987_cast_fp16_12, tensor var_3987_cast_fp16_13, tensor var_3987_cast_fp16_14, tensor var_3987_cast_fp16_15 = split(axis = var_3987_axis_0, split_sizes = tile_50, x = var_3949_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor aw_513_equation_0 = const()[name = tensor("aw_513_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_513_cast_fp16 = einsum(equation = aw_513_equation_0, values = (var_3970_cast_fp16_0, var_3952_cast_fp16_0))[name = tensor("aw_513_cast_fp16")]; + tensor aw_515_equation_0 = const()[name = tensor("aw_515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_515_cast_fp16 = einsum(equation = aw_515_equation_0, values = (var_3970_cast_fp16_1, var_3952_cast_fp16_1))[name = tensor("aw_515_cast_fp16")]; + tensor aw_517_equation_0 = const()[name = tensor("aw_517_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_517_cast_fp16 = einsum(equation = aw_517_equation_0, values = (var_3970_cast_fp16_2, var_3952_cast_fp16_2))[name = tensor("aw_517_cast_fp16")]; + tensor aw_519_equation_0 = const()[name = tensor("aw_519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_519_cast_fp16 = einsum(equation = aw_519_equation_0, values = (var_3970_cast_fp16_3, var_3952_cast_fp16_3))[name = tensor("aw_519_cast_fp16")]; + tensor aw_521_equation_0 = const()[name = tensor("aw_521_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_521_cast_fp16 = einsum(equation = aw_521_equation_0, values = (var_3970_cast_fp16_4, var_3952_cast_fp16_4))[name = tensor("aw_521_cast_fp16")]; + tensor aw_523_equation_0 = const()[name = tensor("aw_523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_523_cast_fp16 = einsum(equation = aw_523_equation_0, values = (var_3970_cast_fp16_5, var_3952_cast_fp16_5))[name = tensor("aw_523_cast_fp16")]; + tensor aw_525_equation_0 = const()[name = tensor("aw_525_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_525_cast_fp16 = einsum(equation = aw_525_equation_0, values = (var_3970_cast_fp16_6, var_3952_cast_fp16_6))[name = tensor("aw_525_cast_fp16")]; + tensor aw_527_equation_0 = const()[name = tensor("aw_527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_527_cast_fp16 = einsum(equation = aw_527_equation_0, values = (var_3970_cast_fp16_7, var_3952_cast_fp16_7))[name = tensor("aw_527_cast_fp16")]; + tensor aw_529_equation_0 = const()[name = tensor("aw_529_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_529_cast_fp16 = einsum(equation = aw_529_equation_0, values = (var_3970_cast_fp16_8, var_3952_cast_fp16_8))[name = tensor("aw_529_cast_fp16")]; + tensor aw_531_equation_0 = const()[name = tensor("aw_531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_531_cast_fp16 = einsum(equation = aw_531_equation_0, values = (var_3970_cast_fp16_9, var_3952_cast_fp16_9))[name = tensor("aw_531_cast_fp16")]; + tensor aw_533_equation_0 = const()[name = tensor("aw_533_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_533_cast_fp16 = einsum(equation = aw_533_equation_0, values = (var_3970_cast_fp16_10, var_3952_cast_fp16_10))[name = tensor("aw_533_cast_fp16")]; + tensor aw_535_equation_0 = const()[name = tensor("aw_535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_535_cast_fp16 = einsum(equation = aw_535_equation_0, values = (var_3970_cast_fp16_11, var_3952_cast_fp16_11))[name = tensor("aw_535_cast_fp16")]; + tensor aw_537_equation_0 = const()[name = tensor("aw_537_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_537_cast_fp16 = einsum(equation = aw_537_equation_0, values = (var_3970_cast_fp16_12, var_3952_cast_fp16_12))[name = tensor("aw_537_cast_fp16")]; + tensor aw_539_equation_0 = const()[name = tensor("aw_539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_539_cast_fp16 = einsum(equation = aw_539_equation_0, values = (var_3970_cast_fp16_13, var_3952_cast_fp16_13))[name = tensor("aw_539_cast_fp16")]; + tensor aw_541_equation_0 = const()[name = tensor("aw_541_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_541_cast_fp16 = einsum(equation = aw_541_equation_0, values = (var_3970_cast_fp16_14, var_3952_cast_fp16_14))[name = tensor("aw_541_cast_fp16")]; + tensor aw_543_equation_0 = const()[name = tensor("aw_543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_543_cast_fp16 = einsum(equation = aw_543_equation_0, values = (var_3970_cast_fp16_15, var_3952_cast_fp16_15))[name = tensor("aw_543_cast_fp16")]; + tensor var_4036_cast_fp16 = softmax(axis = var_3900, x = aw_513_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor var_4037_cast_fp16 = softmax(axis = var_3900, x = aw_515_cast_fp16)[name = tensor("op_4037_cast_fp16")]; + tensor var_4038_cast_fp16 = softmax(axis = var_3900, x = aw_517_cast_fp16)[name = tensor("op_4038_cast_fp16")]; + tensor var_4039_cast_fp16 = softmax(axis = var_3900, x = aw_519_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = softmax(axis = var_3900, x = aw_521_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor var_4041_cast_fp16 = softmax(axis = var_3900, x = aw_523_cast_fp16)[name = tensor("op_4041_cast_fp16")]; + tensor var_4042_cast_fp16 = softmax(axis = var_3900, x = aw_525_cast_fp16)[name = tensor("op_4042_cast_fp16")]; + tensor var_4043_cast_fp16 = softmax(axis = var_3900, x = aw_527_cast_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4044_cast_fp16 = softmax(axis = var_3900, x = aw_529_cast_fp16)[name = tensor("op_4044_cast_fp16")]; + tensor var_4045_cast_fp16 = softmax(axis = var_3900, x = aw_531_cast_fp16)[name = tensor("op_4045_cast_fp16")]; + tensor var_4046_cast_fp16 = softmax(axis = var_3900, x = aw_533_cast_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor var_4047_cast_fp16 = softmax(axis = var_3900, x = aw_535_cast_fp16)[name = tensor("op_4047_cast_fp16")]; + tensor var_4048_cast_fp16 = softmax(axis = var_3900, x = aw_537_cast_fp16)[name = tensor("op_4048_cast_fp16")]; + tensor var_4049_cast_fp16 = softmax(axis = var_3900, x = aw_539_cast_fp16)[name = tensor("op_4049_cast_fp16")]; + tensor var_4050_cast_fp16 = softmax(axis = var_3900, x = aw_541_cast_fp16)[name = tensor("op_4050_cast_fp16")]; + tensor var_4051_cast_fp16 = softmax(axis = var_3900, x = aw_543_cast_fp16)[name = tensor("op_4051_cast_fp16")]; + tensor var_4053_equation_0 = const()[name = tensor("op_4053_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4053_cast_fp16 = einsum(equation = var_4053_equation_0, values = (var_3987_cast_fp16_0, var_4036_cast_fp16))[name = tensor("op_4053_cast_fp16")]; + tensor var_4055_equation_0 = const()[name = tensor("op_4055_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4055_cast_fp16 = einsum(equation = var_4055_equation_0, values = (var_3987_cast_fp16_1, var_4037_cast_fp16))[name = tensor("op_4055_cast_fp16")]; + tensor var_4057_equation_0 = const()[name = tensor("op_4057_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4057_cast_fp16 = einsum(equation = var_4057_equation_0, values = (var_3987_cast_fp16_2, var_4038_cast_fp16))[name = tensor("op_4057_cast_fp16")]; + tensor var_4059_equation_0 = const()[name = tensor("op_4059_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4059_cast_fp16 = einsum(equation = var_4059_equation_0, values = (var_3987_cast_fp16_3, var_4039_cast_fp16))[name = tensor("op_4059_cast_fp16")]; + tensor var_4061_equation_0 = const()[name = tensor("op_4061_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4061_cast_fp16 = einsum(equation = var_4061_equation_0, values = (var_3987_cast_fp16_4, var_4040_cast_fp16))[name = tensor("op_4061_cast_fp16")]; + tensor var_4063_equation_0 = const()[name = tensor("op_4063_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4063_cast_fp16 = einsum(equation = var_4063_equation_0, values = (var_3987_cast_fp16_5, var_4041_cast_fp16))[name = tensor("op_4063_cast_fp16")]; + tensor var_4065_equation_0 = const()[name = tensor("op_4065_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4065_cast_fp16 = einsum(equation = var_4065_equation_0, values = (var_3987_cast_fp16_6, var_4042_cast_fp16))[name = tensor("op_4065_cast_fp16")]; + tensor var_4067_equation_0 = const()[name = tensor("op_4067_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4067_cast_fp16 = einsum(equation = var_4067_equation_0, values = (var_3987_cast_fp16_7, var_4043_cast_fp16))[name = tensor("op_4067_cast_fp16")]; + tensor var_4069_equation_0 = const()[name = tensor("op_4069_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4069_cast_fp16 = einsum(equation = var_4069_equation_0, values = (var_3987_cast_fp16_8, var_4044_cast_fp16))[name = tensor("op_4069_cast_fp16")]; + tensor var_4071_equation_0 = const()[name = tensor("op_4071_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4071_cast_fp16 = einsum(equation = var_4071_equation_0, values = (var_3987_cast_fp16_9, var_4045_cast_fp16))[name = tensor("op_4071_cast_fp16")]; + tensor var_4073_equation_0 = const()[name = tensor("op_4073_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4073_cast_fp16 = einsum(equation = var_4073_equation_0, values = (var_3987_cast_fp16_10, var_4046_cast_fp16))[name = tensor("op_4073_cast_fp16")]; + tensor var_4075_equation_0 = const()[name = tensor("op_4075_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4075_cast_fp16 = einsum(equation = var_4075_equation_0, values = (var_3987_cast_fp16_11, var_4047_cast_fp16))[name = tensor("op_4075_cast_fp16")]; + tensor var_4077_equation_0 = const()[name = tensor("op_4077_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4077_cast_fp16 = einsum(equation = var_4077_equation_0, values = (var_3987_cast_fp16_12, var_4048_cast_fp16))[name = tensor("op_4077_cast_fp16")]; + tensor var_4079_equation_0 = const()[name = tensor("op_4079_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4079_cast_fp16 = einsum(equation = var_4079_equation_0, values = (var_3987_cast_fp16_13, var_4049_cast_fp16))[name = tensor("op_4079_cast_fp16")]; + tensor var_4081_equation_0 = const()[name = tensor("op_4081_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4081_cast_fp16 = einsum(equation = var_4081_equation_0, values = (var_3987_cast_fp16_14, var_4050_cast_fp16))[name = tensor("op_4081_cast_fp16")]; + tensor var_4083_equation_0 = const()[name = tensor("op_4083_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4083_cast_fp16 = einsum(equation = var_4083_equation_0, values = (var_3987_cast_fp16_15, var_4051_cast_fp16))[name = tensor("op_4083_cast_fp16")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165_cast_fp16 = concat(axis = var_3900, interleave = input_165_interleave_0, values = (var_4053_cast_fp16, var_4055_cast_fp16, var_4057_cast_fp16, var_4059_cast_fp16, var_4061_cast_fp16, var_4063_cast_fp16, var_4065_cast_fp16, var_4067_cast_fp16, var_4069_cast_fp16, var_4071_cast_fp16, var_4073_cast_fp16, var_4075_cast_fp16, var_4077_cast_fp16, var_4079_cast_fp16, var_4081_cast_fp16, var_4083_cast_fp16))[name = tensor("input_165_cast_fp16")]; + tensor var_4092_pad_type_0 = const()[name = tensor("op_4092_pad_type_0"), val = tensor("valid")]; + tensor var_4092_strides_0 = const()[name = tensor("op_4092_strides_0"), val = tensor([1, 1])]; + tensor var_4092_pad_0 = const()[name = tensor("op_4092_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4092_dilations_0 = const()[name = tensor("op_4092_dilations_0"), val = tensor([1, 1])]; + tensor var_4092_groups_0 = const()[name = tensor("op_4092_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_out_weight_to_fp16 = const()[name = tensor("blocks_16_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419221312)))]; + tensor blocks_16_attn_out_bias_to_fp16 = const()[name = tensor("blocks_16_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421318528)))]; + tensor var_4092_cast_fp16 = conv(bias = blocks_16_attn_out_bias_to_fp16, dilations = var_4092_dilations_0, groups = var_4092_groups_0, pad = var_4092_pad_0, pad_type = var_4092_pad_type_0, strides = var_4092_strides_0, weight = blocks_16_attn_out_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("op_4092_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = var_4092_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([1])]; + tensor input_167_gamma_0_to_fp16 = const()[name = tensor("input_167_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421320640)))]; + tensor input_167_beta_0_to_fp16 = const()[name = tensor("input_167_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421322752)))]; + tensor var_4102_to_fp16 = const()[name = tensor("op_4102_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = input_167_beta_0_to_fp16, epsilon = var_4102_to_fp16, gamma = input_167_gamma_0_to_fp16, x = inputs_67_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("valid")]; + tensor input_169_strides_0 = const()[name = tensor("input_169_strides_0"), val = tensor([1, 1])]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_169_dilations_0 = const()[name = tensor("input_169_dilations_0"), val = tensor([1, 1])]; + tensor input_169_groups_0 = const()[name = tensor("input_169_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421324864)))]; + tensor blocks_16_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429713536)))]; + tensor input_169_cast_fp16 = conv(bias = blocks_16_mlp_0_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = blocks_16_mlp_0_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_4128_pad_type_0 = const()[name = tensor("op_4128_pad_type_0"), val = tensor("valid")]; + tensor var_4128_strides_0 = const()[name = tensor("op_4128_strides_0"), val = tensor([1, 1])]; + tensor var_4128_pad_0 = const()[name = tensor("op_4128_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4128_dilations_0 = const()[name = tensor("op_4128_dilations_0"), val = tensor([1, 1])]; + tensor var_4128_groups_0 = const()[name = tensor("op_4128_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429721792)))]; + tensor blocks_16_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438110464)))]; + tensor var_4128_cast_fp16 = conv(bias = blocks_16_mlp_2_bias_to_fp16, dilations = var_4128_dilations_0, groups = var_4128_groups_0, pad = var_4128_pad_0, pad_type = var_4128_pad_type_0, strides = var_4128_strides_0, weight = blocks_16_mlp_2_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("op_4128_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_4128_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4137 = const()[name = tensor("op_4137"), val = tensor(1)]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([1])]; + tensor input_173_gamma_0_to_fp16 = const()[name = tensor("input_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438112576)))]; + tensor input_173_beta_0_to_fp16 = const()[name = tensor("input_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438114688)))]; + tensor var_4153_to_fp16 = const()[name = tensor("op_4153_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = input_173_beta_0_to_fp16, epsilon = var_4153_to_fp16, gamma = input_173_gamma_0_to_fp16, x = inputs_69_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor q_35_pad_type_0 = const()[name = tensor("q_35_pad_type_0"), val = tensor("valid")]; + tensor q_35_strides_0 = const()[name = tensor("q_35_strides_0"), val = tensor([1, 1])]; + tensor q_35_pad_0 = const()[name = tensor("q_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_35_dilations_0 = const()[name = tensor("q_35_dilations_0"), val = tensor([1, 1])]; + tensor q_35_groups_0 = const()[name = tensor("q_35_groups_0"), val = tensor(1)]; + tensor var_4188_weight_0_to_fp16 = const()[name = tensor("op_4188_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438116800)))]; + tensor var_4188_bias_0_to_fp16 = const()[name = tensor("op_4188_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440214016)))]; + tensor var_4188_cast_fp16 = conv(bias = var_4188_bias_0_to_fp16, dilations = q_35_dilations_0, groups = q_35_groups_0, pad = q_35_pad_0, pad_type = q_35_pad_type_0, strides = q_35_strides_0, weight = var_4188_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4188_cast_fp16")]; + tensor k_35_pad_type_0 = const()[name = tensor("k_35_pad_type_0"), val = tensor("valid")]; + tensor k_35_strides_0 = const()[name = tensor("k_35_strides_0"), val = tensor([1, 1])]; + tensor k_35_pad_0 = const()[name = tensor("k_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_35_dilations_0 = const()[name = tensor("k_35_dilations_0"), val = tensor([1, 1])]; + tensor k_35_groups_0 = const()[name = tensor("k_35_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_key_weight_to_fp16 = const()[name = tensor("blocks_17_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440216128)))]; + tensor k_35_cast_fp16 = conv(dilations = k_35_dilations_0, groups = k_35_groups_0, pad = k_35_pad_0, pad_type = k_35_pad_type_0, strides = k_35_strides_0, weight = blocks_17_attn_key_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_4186_pad_type_0 = const()[name = tensor("op_4186_pad_type_0"), val = tensor("valid")]; + tensor var_4186_strides_0 = const()[name = tensor("op_4186_strides_0"), val = tensor([1, 1])]; + tensor var_4186_pad_0 = const()[name = tensor("op_4186_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4186_dilations_0 = const()[name = tensor("op_4186_dilations_0"), val = tensor([1, 1])]; + tensor var_4186_groups_0 = const()[name = tensor("op_4186_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_value_weight_to_fp16 = const()[name = tensor("blocks_17_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442313344)))]; + tensor blocks_17_attn_value_bias_to_fp16 = const()[name = tensor("blocks_17_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444410560)))]; + tensor var_4186_cast_fp16 = conv(bias = blocks_17_attn_value_bias_to_fp16, dilations = var_4186_dilations_0, groups = var_4186_groups_0, pad = var_4186_pad_0, pad_type = var_4186_pad_type_0, strides = var_4186_strides_0, weight = blocks_17_attn_value_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor tile_51 = const()[name = tensor("tile_51"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4189_axis_0 = const()[name = tensor("op_4189_axis_0"), val = tensor(1)]; + tensor var_4189_cast_fp16_0, tensor var_4189_cast_fp16_1, tensor var_4189_cast_fp16_2, tensor var_4189_cast_fp16_3, tensor var_4189_cast_fp16_4, tensor var_4189_cast_fp16_5, tensor var_4189_cast_fp16_6, tensor var_4189_cast_fp16_7, tensor var_4189_cast_fp16_8, tensor var_4189_cast_fp16_9, tensor var_4189_cast_fp16_10, tensor var_4189_cast_fp16_11, tensor var_4189_cast_fp16_12, tensor var_4189_cast_fp16_13, tensor var_4189_cast_fp16_14, tensor var_4189_cast_fp16_15 = split(axis = var_4189_axis_0, split_sizes = tile_51, x = var_4188_cast_fp16)[name = tensor("op_4189_cast_fp16")]; + tensor var_4206_perm_0 = const()[name = tensor("op_4206_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_52 = const()[name = tensor("tile_52"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4207_axis_0 = const()[name = tensor("op_4207_axis_0"), val = tensor(3)]; + tensor var_4206_cast_fp16 = transpose(perm = var_4206_perm_0, x = k_35_cast_fp16)[name = tensor("transpose_7")]; + tensor var_4207_cast_fp16_0, tensor var_4207_cast_fp16_1, tensor var_4207_cast_fp16_2, tensor var_4207_cast_fp16_3, tensor var_4207_cast_fp16_4, tensor var_4207_cast_fp16_5, tensor var_4207_cast_fp16_6, tensor var_4207_cast_fp16_7, tensor var_4207_cast_fp16_8, tensor var_4207_cast_fp16_9, tensor var_4207_cast_fp16_10, tensor var_4207_cast_fp16_11, tensor var_4207_cast_fp16_12, tensor var_4207_cast_fp16_13, tensor var_4207_cast_fp16_14, tensor var_4207_cast_fp16_15 = split(axis = var_4207_axis_0, split_sizes = tile_52, x = var_4206_cast_fp16)[name = tensor("op_4207_cast_fp16")]; + tensor tile_53 = const()[name = tensor("tile_53"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4224_axis_0 = const()[name = tensor("op_4224_axis_0"), val = tensor(1)]; + tensor var_4224_cast_fp16_0, tensor var_4224_cast_fp16_1, tensor var_4224_cast_fp16_2, tensor var_4224_cast_fp16_3, tensor var_4224_cast_fp16_4, tensor var_4224_cast_fp16_5, tensor var_4224_cast_fp16_6, tensor var_4224_cast_fp16_7, tensor var_4224_cast_fp16_8, tensor var_4224_cast_fp16_9, tensor var_4224_cast_fp16_10, tensor var_4224_cast_fp16_11, tensor var_4224_cast_fp16_12, tensor var_4224_cast_fp16_13, tensor var_4224_cast_fp16_14, tensor var_4224_cast_fp16_15 = split(axis = var_4224_axis_0, split_sizes = tile_53, x = var_4186_cast_fp16)[name = tensor("op_4224_cast_fp16")]; + tensor aw_545_equation_0 = const()[name = tensor("aw_545_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_545_cast_fp16 = einsum(equation = aw_545_equation_0, values = (var_4207_cast_fp16_0, var_4189_cast_fp16_0))[name = tensor("aw_545_cast_fp16")]; + tensor aw_547_equation_0 = const()[name = tensor("aw_547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_547_cast_fp16 = einsum(equation = aw_547_equation_0, values = (var_4207_cast_fp16_1, var_4189_cast_fp16_1))[name = tensor("aw_547_cast_fp16")]; + tensor aw_549_equation_0 = const()[name = tensor("aw_549_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_549_cast_fp16 = einsum(equation = aw_549_equation_0, values = (var_4207_cast_fp16_2, var_4189_cast_fp16_2))[name = tensor("aw_549_cast_fp16")]; + tensor aw_551_equation_0 = const()[name = tensor("aw_551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_551_cast_fp16 = einsum(equation = aw_551_equation_0, values = (var_4207_cast_fp16_3, var_4189_cast_fp16_3))[name = tensor("aw_551_cast_fp16")]; + tensor aw_553_equation_0 = const()[name = tensor("aw_553_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_553_cast_fp16 = einsum(equation = aw_553_equation_0, values = (var_4207_cast_fp16_4, var_4189_cast_fp16_4))[name = tensor("aw_553_cast_fp16")]; + tensor aw_555_equation_0 = const()[name = tensor("aw_555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_555_cast_fp16 = einsum(equation = aw_555_equation_0, values = (var_4207_cast_fp16_5, var_4189_cast_fp16_5))[name = tensor("aw_555_cast_fp16")]; + tensor aw_557_equation_0 = const()[name = tensor("aw_557_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_557_cast_fp16 = einsum(equation = aw_557_equation_0, values = (var_4207_cast_fp16_6, var_4189_cast_fp16_6))[name = tensor("aw_557_cast_fp16")]; + tensor aw_559_equation_0 = const()[name = tensor("aw_559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_559_cast_fp16 = einsum(equation = aw_559_equation_0, values = (var_4207_cast_fp16_7, var_4189_cast_fp16_7))[name = tensor("aw_559_cast_fp16")]; + tensor aw_561_equation_0 = const()[name = tensor("aw_561_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_561_cast_fp16 = einsum(equation = aw_561_equation_0, values = (var_4207_cast_fp16_8, var_4189_cast_fp16_8))[name = tensor("aw_561_cast_fp16")]; + tensor aw_563_equation_0 = const()[name = tensor("aw_563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_563_cast_fp16 = einsum(equation = aw_563_equation_0, values = (var_4207_cast_fp16_9, var_4189_cast_fp16_9))[name = tensor("aw_563_cast_fp16")]; + tensor aw_565_equation_0 = const()[name = tensor("aw_565_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_565_cast_fp16 = einsum(equation = aw_565_equation_0, values = (var_4207_cast_fp16_10, var_4189_cast_fp16_10))[name = tensor("aw_565_cast_fp16")]; + tensor aw_567_equation_0 = const()[name = tensor("aw_567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_567_cast_fp16 = einsum(equation = aw_567_equation_0, values = (var_4207_cast_fp16_11, var_4189_cast_fp16_11))[name = tensor("aw_567_cast_fp16")]; + tensor aw_569_equation_0 = const()[name = tensor("aw_569_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_569_cast_fp16 = einsum(equation = aw_569_equation_0, values = (var_4207_cast_fp16_12, var_4189_cast_fp16_12))[name = tensor("aw_569_cast_fp16")]; + tensor aw_571_equation_0 = const()[name = tensor("aw_571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_571_cast_fp16 = einsum(equation = aw_571_equation_0, values = (var_4207_cast_fp16_13, var_4189_cast_fp16_13))[name = tensor("aw_571_cast_fp16")]; + tensor aw_573_equation_0 = const()[name = tensor("aw_573_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_573_cast_fp16 = einsum(equation = aw_573_equation_0, values = (var_4207_cast_fp16_14, var_4189_cast_fp16_14))[name = tensor("aw_573_cast_fp16")]; + tensor aw_575_equation_0 = const()[name = tensor("aw_575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_575_cast_fp16 = einsum(equation = aw_575_equation_0, values = (var_4207_cast_fp16_15, var_4189_cast_fp16_15))[name = tensor("aw_575_cast_fp16")]; + tensor var_4273_cast_fp16 = softmax(axis = var_4137, x = aw_545_cast_fp16)[name = tensor("op_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = softmax(axis = var_4137, x = aw_547_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_cast_fp16 = softmax(axis = var_4137, x = aw_549_cast_fp16)[name = tensor("op_4275_cast_fp16")]; + tensor var_4276_cast_fp16 = softmax(axis = var_4137, x = aw_551_cast_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor var_4277_cast_fp16 = softmax(axis = var_4137, x = aw_553_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4278_cast_fp16 = softmax(axis = var_4137, x = aw_555_cast_fp16)[name = tensor("op_4278_cast_fp16")]; + tensor var_4279_cast_fp16 = softmax(axis = var_4137, x = aw_557_cast_fp16)[name = tensor("op_4279_cast_fp16")]; + tensor var_4280_cast_fp16 = softmax(axis = var_4137, x = aw_559_cast_fp16)[name = tensor("op_4280_cast_fp16")]; + tensor var_4281_cast_fp16 = softmax(axis = var_4137, x = aw_561_cast_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4282_cast_fp16 = softmax(axis = var_4137, x = aw_563_cast_fp16)[name = tensor("op_4282_cast_fp16")]; + tensor var_4283_cast_fp16 = softmax(axis = var_4137, x = aw_565_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = softmax(axis = var_4137, x = aw_567_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor var_4285_cast_fp16 = softmax(axis = var_4137, x = aw_569_cast_fp16)[name = tensor("op_4285_cast_fp16")]; + tensor var_4286_cast_fp16 = softmax(axis = var_4137, x = aw_571_cast_fp16)[name = tensor("op_4286_cast_fp16")]; + tensor var_4287_cast_fp16 = softmax(axis = var_4137, x = aw_573_cast_fp16)[name = tensor("op_4287_cast_fp16")]; + tensor var_4288_cast_fp16 = softmax(axis = var_4137, x = aw_575_cast_fp16)[name = tensor("op_4288_cast_fp16")]; + tensor var_4290_equation_0 = const()[name = tensor("op_4290_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4290_cast_fp16 = einsum(equation = var_4290_equation_0, values = (var_4224_cast_fp16_0, var_4273_cast_fp16))[name = tensor("op_4290_cast_fp16")]; + tensor var_4292_equation_0 = const()[name = tensor("op_4292_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4292_cast_fp16 = einsum(equation = var_4292_equation_0, values = (var_4224_cast_fp16_1, var_4274_cast_fp16))[name = tensor("op_4292_cast_fp16")]; + tensor var_4294_equation_0 = const()[name = tensor("op_4294_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4294_cast_fp16 = einsum(equation = var_4294_equation_0, values = (var_4224_cast_fp16_2, var_4275_cast_fp16))[name = tensor("op_4294_cast_fp16")]; + tensor var_4296_equation_0 = const()[name = tensor("op_4296_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4296_cast_fp16 = einsum(equation = var_4296_equation_0, values = (var_4224_cast_fp16_3, var_4276_cast_fp16))[name = tensor("op_4296_cast_fp16")]; + tensor var_4298_equation_0 = const()[name = tensor("op_4298_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4298_cast_fp16 = einsum(equation = var_4298_equation_0, values = (var_4224_cast_fp16_4, var_4277_cast_fp16))[name = tensor("op_4298_cast_fp16")]; + tensor var_4300_equation_0 = const()[name = tensor("op_4300_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4300_cast_fp16 = einsum(equation = var_4300_equation_0, values = (var_4224_cast_fp16_5, var_4278_cast_fp16))[name = tensor("op_4300_cast_fp16")]; + tensor var_4302_equation_0 = const()[name = tensor("op_4302_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4302_cast_fp16 = einsum(equation = var_4302_equation_0, values = (var_4224_cast_fp16_6, var_4279_cast_fp16))[name = tensor("op_4302_cast_fp16")]; + tensor var_4304_equation_0 = const()[name = tensor("op_4304_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4304_cast_fp16 = einsum(equation = var_4304_equation_0, values = (var_4224_cast_fp16_7, var_4280_cast_fp16))[name = tensor("op_4304_cast_fp16")]; + tensor var_4306_equation_0 = const()[name = tensor("op_4306_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4306_cast_fp16 = einsum(equation = var_4306_equation_0, values = (var_4224_cast_fp16_8, var_4281_cast_fp16))[name = tensor("op_4306_cast_fp16")]; + tensor var_4308_equation_0 = const()[name = tensor("op_4308_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4308_cast_fp16 = einsum(equation = var_4308_equation_0, values = (var_4224_cast_fp16_9, var_4282_cast_fp16))[name = tensor("op_4308_cast_fp16")]; + tensor var_4310_equation_0 = const()[name = tensor("op_4310_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4310_cast_fp16 = einsum(equation = var_4310_equation_0, values = (var_4224_cast_fp16_10, var_4283_cast_fp16))[name = tensor("op_4310_cast_fp16")]; + tensor var_4312_equation_0 = const()[name = tensor("op_4312_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4312_cast_fp16 = einsum(equation = var_4312_equation_0, values = (var_4224_cast_fp16_11, var_4284_cast_fp16))[name = tensor("op_4312_cast_fp16")]; + tensor var_4314_equation_0 = const()[name = tensor("op_4314_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4314_cast_fp16 = einsum(equation = var_4314_equation_0, values = (var_4224_cast_fp16_12, var_4285_cast_fp16))[name = tensor("op_4314_cast_fp16")]; + tensor var_4316_equation_0 = const()[name = tensor("op_4316_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4316_cast_fp16 = einsum(equation = var_4316_equation_0, values = (var_4224_cast_fp16_13, var_4286_cast_fp16))[name = tensor("op_4316_cast_fp16")]; + tensor var_4318_equation_0 = const()[name = tensor("op_4318_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4318_cast_fp16 = einsum(equation = var_4318_equation_0, values = (var_4224_cast_fp16_14, var_4287_cast_fp16))[name = tensor("op_4318_cast_fp16")]; + tensor var_4320_equation_0 = const()[name = tensor("op_4320_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4320_cast_fp16 = einsum(equation = var_4320_equation_0, values = (var_4224_cast_fp16_15, var_4288_cast_fp16))[name = tensor("op_4320_cast_fp16")]; + tensor input_175_interleave_0 = const()[name = tensor("input_175_interleave_0"), val = tensor(false)]; + tensor input_175_cast_fp16 = concat(axis = var_4137, interleave = input_175_interleave_0, values = (var_4290_cast_fp16, var_4292_cast_fp16, var_4294_cast_fp16, var_4296_cast_fp16, var_4298_cast_fp16, var_4300_cast_fp16, var_4302_cast_fp16, var_4304_cast_fp16, var_4306_cast_fp16, var_4308_cast_fp16, var_4310_cast_fp16, var_4312_cast_fp16, var_4314_cast_fp16, var_4316_cast_fp16, var_4318_cast_fp16, var_4320_cast_fp16))[name = tensor("input_175_cast_fp16")]; + tensor var_4329_pad_type_0 = const()[name = tensor("op_4329_pad_type_0"), val = tensor("valid")]; + tensor var_4329_strides_0 = const()[name = tensor("op_4329_strides_0"), val = tensor([1, 1])]; + tensor var_4329_pad_0 = const()[name = tensor("op_4329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4329_dilations_0 = const()[name = tensor("op_4329_dilations_0"), val = tensor([1, 1])]; + tensor var_4329_groups_0 = const()[name = tensor("op_4329_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_out_weight_to_fp16 = const()[name = tensor("blocks_17_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444412672)))]; + tensor blocks_17_attn_out_bias_to_fp16 = const()[name = tensor("blocks_17_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446509888)))]; + tensor var_4329_cast_fp16 = conv(bias = blocks_17_attn_out_bias_to_fp16, dilations = var_4329_dilations_0, groups = var_4329_groups_0, pad = var_4329_pad_0, pad_type = var_4329_pad_type_0, strides = var_4329_strides_0, weight = blocks_17_attn_out_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = var_4329_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([1])]; + tensor input_177_gamma_0_to_fp16 = const()[name = tensor("input_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446512000)))]; + tensor input_177_beta_0_to_fp16 = const()[name = tensor("input_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446514112)))]; + tensor var_4339_to_fp16 = const()[name = tensor("op_4339_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = input_177_beta_0_to_fp16, epsilon = var_4339_to_fp16, gamma = input_177_gamma_0_to_fp16, x = inputs_71_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("valid")]; + tensor input_179_strides_0 = const()[name = tensor("input_179_strides_0"), val = tensor([1, 1])]; + tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_179_dilations_0 = const()[name = tensor("input_179_dilations_0"), val = tensor([1, 1])]; + tensor input_179_groups_0 = const()[name = tensor("input_179_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446516224)))]; + tensor blocks_17_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454904896)))]; + tensor input_179_cast_fp16 = conv(bias = blocks_17_mlp_0_bias_to_fp16, dilations = input_179_dilations_0, groups = input_179_groups_0, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = input_179_strides_0, weight = blocks_17_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4365_pad_type_0 = const()[name = tensor("op_4365_pad_type_0"), val = tensor("valid")]; + tensor var_4365_strides_0 = const()[name = tensor("op_4365_strides_0"), val = tensor([1, 1])]; + tensor var_4365_pad_0 = const()[name = tensor("op_4365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4365_dilations_0 = const()[name = tensor("op_4365_dilations_0"), val = tensor([1, 1])]; + tensor var_4365_groups_0 = const()[name = tensor("op_4365_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454913152)))]; + tensor blocks_17_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463301824)))]; + tensor var_4365_cast_fp16 = conv(bias = blocks_17_mlp_2_bias_to_fp16, dilations = var_4365_dilations_0, groups = var_4365_groups_0, pad = var_4365_pad_0, pad_type = var_4365_pad_type_0, strides = var_4365_strides_0, weight = blocks_17_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("op_4365_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_4365_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor(1)]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([1])]; + tensor input_183_gamma_0_to_fp16 = const()[name = tensor("input_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463303936)))]; + tensor input_183_beta_0_to_fp16 = const()[name = tensor("input_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463306048)))]; + tensor var_4390_to_fp16 = const()[name = tensor("op_4390_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = input_183_beta_0_to_fp16, epsilon = var_4390_to_fp16, gamma = input_183_gamma_0_to_fp16, x = inputs_73_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor q_37_pad_type_0 = const()[name = tensor("q_37_pad_type_0"), val = tensor("valid")]; + tensor q_37_strides_0 = const()[name = tensor("q_37_strides_0"), val = tensor([1, 1])]; + tensor q_37_pad_0 = const()[name = tensor("q_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_37_dilations_0 = const()[name = tensor("q_37_dilations_0"), val = tensor([1, 1])]; + tensor q_37_groups_0 = const()[name = tensor("q_37_groups_0"), val = tensor(1)]; + tensor var_4425_weight_0_to_fp16 = const()[name = tensor("op_4425_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463308160)))]; + tensor var_4425_bias_0_to_fp16 = const()[name = tensor("op_4425_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465405376)))]; + tensor var_4425_cast_fp16 = conv(bias = var_4425_bias_0_to_fp16, dilations = q_37_dilations_0, groups = q_37_groups_0, pad = q_37_pad_0, pad_type = q_37_pad_type_0, strides = q_37_strides_0, weight = var_4425_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("op_4425_cast_fp16")]; + tensor k_37_pad_type_0 = const()[name = tensor("k_37_pad_type_0"), val = tensor("valid")]; + tensor k_37_strides_0 = const()[name = tensor("k_37_strides_0"), val = tensor([1, 1])]; + tensor k_37_pad_0 = const()[name = tensor("k_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_37_dilations_0 = const()[name = tensor("k_37_dilations_0"), val = tensor([1, 1])]; + tensor k_37_groups_0 = const()[name = tensor("k_37_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_key_weight_to_fp16 = const()[name = tensor("blocks_18_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465407488)))]; + tensor k_37_cast_fp16 = conv(dilations = k_37_dilations_0, groups = k_37_groups_0, pad = k_37_pad_0, pad_type = k_37_pad_type_0, strides = k_37_strides_0, weight = blocks_18_attn_key_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor var_4423_pad_type_0 = const()[name = tensor("op_4423_pad_type_0"), val = tensor("valid")]; + tensor var_4423_strides_0 = const()[name = tensor("op_4423_strides_0"), val = tensor([1, 1])]; + tensor var_4423_pad_0 = const()[name = tensor("op_4423_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4423_dilations_0 = const()[name = tensor("op_4423_dilations_0"), val = tensor([1, 1])]; + tensor var_4423_groups_0 = const()[name = tensor("op_4423_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_value_weight_to_fp16 = const()[name = tensor("blocks_18_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467504704)))]; + tensor blocks_18_attn_value_bias_to_fp16 = const()[name = tensor("blocks_18_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469601920)))]; + tensor var_4423_cast_fp16 = conv(bias = blocks_18_attn_value_bias_to_fp16, dilations = var_4423_dilations_0, groups = var_4423_groups_0, pad = var_4423_pad_0, pad_type = var_4423_pad_type_0, strides = var_4423_strides_0, weight = blocks_18_attn_value_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("op_4423_cast_fp16")]; + tensor tile_54 = const()[name = tensor("tile_54"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4426_axis_0 = const()[name = tensor("op_4426_axis_0"), val = tensor(1)]; + tensor var_4426_cast_fp16_0, tensor var_4426_cast_fp16_1, tensor var_4426_cast_fp16_2, tensor var_4426_cast_fp16_3, tensor var_4426_cast_fp16_4, tensor var_4426_cast_fp16_5, tensor var_4426_cast_fp16_6, tensor var_4426_cast_fp16_7, tensor var_4426_cast_fp16_8, tensor var_4426_cast_fp16_9, tensor var_4426_cast_fp16_10, tensor var_4426_cast_fp16_11, tensor var_4426_cast_fp16_12, tensor var_4426_cast_fp16_13, tensor var_4426_cast_fp16_14, tensor var_4426_cast_fp16_15 = split(axis = var_4426_axis_0, split_sizes = tile_54, x = var_4425_cast_fp16)[name = tensor("op_4426_cast_fp16")]; + tensor var_4443_perm_0 = const()[name = tensor("op_4443_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_55 = const()[name = tensor("tile_55"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4444_axis_0 = const()[name = tensor("op_4444_axis_0"), val = tensor(3)]; + tensor var_4443_cast_fp16 = transpose(perm = var_4443_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_6")]; + tensor var_4444_cast_fp16_0, tensor var_4444_cast_fp16_1, tensor var_4444_cast_fp16_2, tensor var_4444_cast_fp16_3, tensor var_4444_cast_fp16_4, tensor var_4444_cast_fp16_5, tensor var_4444_cast_fp16_6, tensor var_4444_cast_fp16_7, tensor var_4444_cast_fp16_8, tensor var_4444_cast_fp16_9, tensor var_4444_cast_fp16_10, tensor var_4444_cast_fp16_11, tensor var_4444_cast_fp16_12, tensor var_4444_cast_fp16_13, tensor var_4444_cast_fp16_14, tensor var_4444_cast_fp16_15 = split(axis = var_4444_axis_0, split_sizes = tile_55, x = var_4443_cast_fp16)[name = tensor("op_4444_cast_fp16")]; + tensor tile_56 = const()[name = tensor("tile_56"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4461_axis_0 = const()[name = tensor("op_4461_axis_0"), val = tensor(1)]; + tensor var_4461_cast_fp16_0, tensor var_4461_cast_fp16_1, tensor var_4461_cast_fp16_2, tensor var_4461_cast_fp16_3, tensor var_4461_cast_fp16_4, tensor var_4461_cast_fp16_5, tensor var_4461_cast_fp16_6, tensor var_4461_cast_fp16_7, tensor var_4461_cast_fp16_8, tensor var_4461_cast_fp16_9, tensor var_4461_cast_fp16_10, tensor var_4461_cast_fp16_11, tensor var_4461_cast_fp16_12, tensor var_4461_cast_fp16_13, tensor var_4461_cast_fp16_14, tensor var_4461_cast_fp16_15 = split(axis = var_4461_axis_0, split_sizes = tile_56, x = var_4423_cast_fp16)[name = tensor("op_4461_cast_fp16")]; + tensor aw_577_equation_0 = const()[name = tensor("aw_577_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_577_cast_fp16 = einsum(equation = aw_577_equation_0, values = (var_4444_cast_fp16_0, var_4426_cast_fp16_0))[name = tensor("aw_577_cast_fp16")]; + tensor aw_579_equation_0 = const()[name = tensor("aw_579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_579_cast_fp16 = einsum(equation = aw_579_equation_0, values = (var_4444_cast_fp16_1, var_4426_cast_fp16_1))[name = tensor("aw_579_cast_fp16")]; + tensor aw_581_equation_0 = const()[name = tensor("aw_581_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_581_cast_fp16 = einsum(equation = aw_581_equation_0, values = (var_4444_cast_fp16_2, var_4426_cast_fp16_2))[name = tensor("aw_581_cast_fp16")]; + tensor aw_583_equation_0 = const()[name = tensor("aw_583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_583_cast_fp16 = einsum(equation = aw_583_equation_0, values = (var_4444_cast_fp16_3, var_4426_cast_fp16_3))[name = tensor("aw_583_cast_fp16")]; + tensor aw_585_equation_0 = const()[name = tensor("aw_585_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_585_cast_fp16 = einsum(equation = aw_585_equation_0, values = (var_4444_cast_fp16_4, var_4426_cast_fp16_4))[name = tensor("aw_585_cast_fp16")]; + tensor aw_587_equation_0 = const()[name = tensor("aw_587_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_587_cast_fp16 = einsum(equation = aw_587_equation_0, values = (var_4444_cast_fp16_5, var_4426_cast_fp16_5))[name = tensor("aw_587_cast_fp16")]; + tensor aw_589_equation_0 = const()[name = tensor("aw_589_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_589_cast_fp16 = einsum(equation = aw_589_equation_0, values = (var_4444_cast_fp16_6, var_4426_cast_fp16_6))[name = tensor("aw_589_cast_fp16")]; + tensor aw_591_equation_0 = const()[name = tensor("aw_591_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_591_cast_fp16 = einsum(equation = aw_591_equation_0, values = (var_4444_cast_fp16_7, var_4426_cast_fp16_7))[name = tensor("aw_591_cast_fp16")]; + tensor aw_593_equation_0 = const()[name = tensor("aw_593_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_593_cast_fp16 = einsum(equation = aw_593_equation_0, values = (var_4444_cast_fp16_8, var_4426_cast_fp16_8))[name = tensor("aw_593_cast_fp16")]; + tensor aw_595_equation_0 = const()[name = tensor("aw_595_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_595_cast_fp16 = einsum(equation = aw_595_equation_0, values = (var_4444_cast_fp16_9, var_4426_cast_fp16_9))[name = tensor("aw_595_cast_fp16")]; + tensor aw_597_equation_0 = const()[name = tensor("aw_597_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_597_cast_fp16 = einsum(equation = aw_597_equation_0, values = (var_4444_cast_fp16_10, var_4426_cast_fp16_10))[name = tensor("aw_597_cast_fp16")]; + tensor aw_599_equation_0 = const()[name = tensor("aw_599_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_599_cast_fp16 = einsum(equation = aw_599_equation_0, values = (var_4444_cast_fp16_11, var_4426_cast_fp16_11))[name = tensor("aw_599_cast_fp16")]; + tensor aw_601_equation_0 = const()[name = tensor("aw_601_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_601_cast_fp16 = einsum(equation = aw_601_equation_0, values = (var_4444_cast_fp16_12, var_4426_cast_fp16_12))[name = tensor("aw_601_cast_fp16")]; + tensor aw_603_equation_0 = const()[name = tensor("aw_603_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_603_cast_fp16 = einsum(equation = aw_603_equation_0, values = (var_4444_cast_fp16_13, var_4426_cast_fp16_13))[name = tensor("aw_603_cast_fp16")]; + tensor aw_605_equation_0 = const()[name = tensor("aw_605_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_605_cast_fp16 = einsum(equation = aw_605_equation_0, values = (var_4444_cast_fp16_14, var_4426_cast_fp16_14))[name = tensor("aw_605_cast_fp16")]; + tensor aw_607_equation_0 = const()[name = tensor("aw_607_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_607_cast_fp16 = einsum(equation = aw_607_equation_0, values = (var_4444_cast_fp16_15, var_4426_cast_fp16_15))[name = tensor("aw_607_cast_fp16")]; + tensor var_4510_cast_fp16 = softmax(axis = var_4374, x = aw_577_cast_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor var_4511_cast_fp16 = softmax(axis = var_4374, x = aw_579_cast_fp16)[name = tensor("op_4511_cast_fp16")]; + tensor var_4512_cast_fp16 = softmax(axis = var_4374, x = aw_581_cast_fp16)[name = tensor("op_4512_cast_fp16")]; + tensor var_4513_cast_fp16 = softmax(axis = var_4374, x = aw_583_cast_fp16)[name = tensor("op_4513_cast_fp16")]; + tensor var_4514_cast_fp16 = softmax(axis = var_4374, x = aw_585_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = softmax(axis = var_4374, x = aw_587_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor var_4516_cast_fp16 = softmax(axis = var_4374, x = aw_589_cast_fp16)[name = tensor("op_4516_cast_fp16")]; + tensor var_4517_cast_fp16 = softmax(axis = var_4374, x = aw_591_cast_fp16)[name = tensor("op_4517_cast_fp16")]; + tensor var_4518_cast_fp16 = softmax(axis = var_4374, x = aw_593_cast_fp16)[name = tensor("op_4518_cast_fp16")]; + tensor var_4519_cast_fp16 = softmax(axis = var_4374, x = aw_595_cast_fp16)[name = tensor("op_4519_cast_fp16")]; + tensor var_4520_cast_fp16 = softmax(axis = var_4374, x = aw_597_cast_fp16)[name = tensor("op_4520_cast_fp16")]; + tensor var_4521_cast_fp16 = softmax(axis = var_4374, x = aw_599_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4522_cast_fp16 = softmax(axis = var_4374, x = aw_601_cast_fp16)[name = tensor("op_4522_cast_fp16")]; + tensor var_4523_cast_fp16 = softmax(axis = var_4374, x = aw_603_cast_fp16)[name = tensor("op_4523_cast_fp16")]; + tensor var_4524_cast_fp16 = softmax(axis = var_4374, x = aw_605_cast_fp16)[name = tensor("op_4524_cast_fp16")]; + tensor var_4525_cast_fp16 = softmax(axis = var_4374, x = aw_607_cast_fp16)[name = tensor("op_4525_cast_fp16")]; + tensor var_4527_equation_0 = const()[name = tensor("op_4527_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4527_cast_fp16 = einsum(equation = var_4527_equation_0, values = (var_4461_cast_fp16_0, var_4510_cast_fp16))[name = tensor("op_4527_cast_fp16")]; + tensor var_4529_equation_0 = const()[name = tensor("op_4529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4529_cast_fp16 = einsum(equation = var_4529_equation_0, values = (var_4461_cast_fp16_1, var_4511_cast_fp16))[name = tensor("op_4529_cast_fp16")]; + tensor var_4531_equation_0 = const()[name = tensor("op_4531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4531_cast_fp16 = einsum(equation = var_4531_equation_0, values = (var_4461_cast_fp16_2, var_4512_cast_fp16))[name = tensor("op_4531_cast_fp16")]; + tensor var_4533_equation_0 = const()[name = tensor("op_4533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4533_cast_fp16 = einsum(equation = var_4533_equation_0, values = (var_4461_cast_fp16_3, var_4513_cast_fp16))[name = tensor("op_4533_cast_fp16")]; + tensor var_4535_equation_0 = const()[name = tensor("op_4535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4535_cast_fp16 = einsum(equation = var_4535_equation_0, values = (var_4461_cast_fp16_4, var_4514_cast_fp16))[name = tensor("op_4535_cast_fp16")]; + tensor var_4537_equation_0 = const()[name = tensor("op_4537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4537_cast_fp16 = einsum(equation = var_4537_equation_0, values = (var_4461_cast_fp16_5, var_4515_cast_fp16))[name = tensor("op_4537_cast_fp16")]; + tensor var_4539_equation_0 = const()[name = tensor("op_4539_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4539_cast_fp16 = einsum(equation = var_4539_equation_0, values = (var_4461_cast_fp16_6, var_4516_cast_fp16))[name = tensor("op_4539_cast_fp16")]; + tensor var_4541_equation_0 = const()[name = tensor("op_4541_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4541_cast_fp16 = einsum(equation = var_4541_equation_0, values = (var_4461_cast_fp16_7, var_4517_cast_fp16))[name = tensor("op_4541_cast_fp16")]; + tensor var_4543_equation_0 = const()[name = tensor("op_4543_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4543_cast_fp16 = einsum(equation = var_4543_equation_0, values = (var_4461_cast_fp16_8, var_4518_cast_fp16))[name = tensor("op_4543_cast_fp16")]; + tensor var_4545_equation_0 = const()[name = tensor("op_4545_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4545_cast_fp16 = einsum(equation = var_4545_equation_0, values = (var_4461_cast_fp16_9, var_4519_cast_fp16))[name = tensor("op_4545_cast_fp16")]; + tensor var_4547_equation_0 = const()[name = tensor("op_4547_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4547_cast_fp16 = einsum(equation = var_4547_equation_0, values = (var_4461_cast_fp16_10, var_4520_cast_fp16))[name = tensor("op_4547_cast_fp16")]; + tensor var_4549_equation_0 = const()[name = tensor("op_4549_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4549_cast_fp16 = einsum(equation = var_4549_equation_0, values = (var_4461_cast_fp16_11, var_4521_cast_fp16))[name = tensor("op_4549_cast_fp16")]; + tensor var_4551_equation_0 = const()[name = tensor("op_4551_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4551_cast_fp16 = einsum(equation = var_4551_equation_0, values = (var_4461_cast_fp16_12, var_4522_cast_fp16))[name = tensor("op_4551_cast_fp16")]; + tensor var_4553_equation_0 = const()[name = tensor("op_4553_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4553_cast_fp16 = einsum(equation = var_4553_equation_0, values = (var_4461_cast_fp16_13, var_4523_cast_fp16))[name = tensor("op_4553_cast_fp16")]; + tensor var_4555_equation_0 = const()[name = tensor("op_4555_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4555_cast_fp16 = einsum(equation = var_4555_equation_0, values = (var_4461_cast_fp16_14, var_4524_cast_fp16))[name = tensor("op_4555_cast_fp16")]; + tensor var_4557_equation_0 = const()[name = tensor("op_4557_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4557_cast_fp16 = einsum(equation = var_4557_equation_0, values = (var_4461_cast_fp16_15, var_4525_cast_fp16))[name = tensor("op_4557_cast_fp16")]; + tensor input_185_interleave_0 = const()[name = tensor("input_185_interleave_0"), val = tensor(false)]; + tensor input_185_cast_fp16 = concat(axis = var_4374, interleave = input_185_interleave_0, values = (var_4527_cast_fp16, var_4529_cast_fp16, var_4531_cast_fp16, var_4533_cast_fp16, var_4535_cast_fp16, var_4537_cast_fp16, var_4539_cast_fp16, var_4541_cast_fp16, var_4543_cast_fp16, var_4545_cast_fp16, var_4547_cast_fp16, var_4549_cast_fp16, var_4551_cast_fp16, var_4553_cast_fp16, var_4555_cast_fp16, var_4557_cast_fp16))[name = tensor("input_185_cast_fp16")]; + tensor var_4566_pad_type_0 = const()[name = tensor("op_4566_pad_type_0"), val = tensor("valid")]; + tensor var_4566_strides_0 = const()[name = tensor("op_4566_strides_0"), val = tensor([1, 1])]; + tensor var_4566_pad_0 = const()[name = tensor("op_4566_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4566_dilations_0 = const()[name = tensor("op_4566_dilations_0"), val = tensor([1, 1])]; + tensor var_4566_groups_0 = const()[name = tensor("op_4566_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_out_weight_to_fp16 = const()[name = tensor("blocks_18_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469604032)))]; + tensor blocks_18_attn_out_bias_to_fp16 = const()[name = tensor("blocks_18_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471701248)))]; + tensor var_4566_cast_fp16 = conv(bias = blocks_18_attn_out_bias_to_fp16, dilations = var_4566_dilations_0, groups = var_4566_groups_0, pad = var_4566_pad_0, pad_type = var_4566_pad_type_0, strides = var_4566_strides_0, weight = blocks_18_attn_out_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("op_4566_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = var_4566_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor input_187_axes_0 = const()[name = tensor("input_187_axes_0"), val = tensor([1])]; + tensor input_187_gamma_0_to_fp16 = const()[name = tensor("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471703360)))]; + tensor input_187_beta_0_to_fp16 = const()[name = tensor("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471705472)))]; + tensor var_4576_to_fp16 = const()[name = tensor("op_4576_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = input_187_beta_0_to_fp16, epsilon = var_4576_to_fp16, gamma = input_187_gamma_0_to_fp16, x = inputs_75_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("valid")]; + tensor input_189_strides_0 = const()[name = tensor("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = tensor("input_189_dilations_0"), val = tensor([1, 1])]; + tensor input_189_groups_0 = const()[name = tensor("input_189_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471707584)))]; + tensor blocks_18_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480096256)))]; + tensor input_189_cast_fp16 = conv(bias = blocks_18_mlp_0_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = blocks_18_mlp_0_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_4602_pad_type_0 = const()[name = tensor("op_4602_pad_type_0"), val = tensor("valid")]; + tensor var_4602_strides_0 = const()[name = tensor("op_4602_strides_0"), val = tensor([1, 1])]; + tensor var_4602_pad_0 = const()[name = tensor("op_4602_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4602_dilations_0 = const()[name = tensor("op_4602_dilations_0"), val = tensor([1, 1])]; + tensor var_4602_groups_0 = const()[name = tensor("op_4602_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480104512)))]; + tensor blocks_18_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488493184)))]; + tensor var_4602_cast_fp16 = conv(bias = blocks_18_mlp_2_bias_to_fp16, dilations = var_4602_dilations_0, groups = var_4602_groups_0, pad = var_4602_pad_0, pad_type = var_4602_pad_type_0, strides = var_4602_strides_0, weight = blocks_18_mlp_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("op_4602_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = var_4602_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_4611 = const()[name = tensor("op_4611"), val = tensor(1)]; + tensor input_193_axes_0 = const()[name = tensor("input_193_axes_0"), val = tensor([1])]; + tensor input_193_gamma_0_to_fp16 = const()[name = tensor("input_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488495296)))]; + tensor input_193_beta_0_to_fp16 = const()[name = tensor("input_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488497408)))]; + tensor var_4627_to_fp16 = const()[name = tensor("op_4627_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = input_193_beta_0_to_fp16, epsilon = var_4627_to_fp16, gamma = input_193_gamma_0_to_fp16, x = inputs_77_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor q_39_pad_type_0 = const()[name = tensor("q_39_pad_type_0"), val = tensor("valid")]; + tensor q_39_strides_0 = const()[name = tensor("q_39_strides_0"), val = tensor([1, 1])]; + tensor q_39_pad_0 = const()[name = tensor("q_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_39_dilations_0 = const()[name = tensor("q_39_dilations_0"), val = tensor([1, 1])]; + tensor q_39_groups_0 = const()[name = tensor("q_39_groups_0"), val = tensor(1)]; + tensor var_4662_weight_0_to_fp16 = const()[name = tensor("op_4662_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488499520)))]; + tensor var_4662_bias_0_to_fp16 = const()[name = tensor("op_4662_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490596736)))]; + tensor var_4662_cast_fp16 = conv(bias = var_4662_bias_0_to_fp16, dilations = q_39_dilations_0, groups = q_39_groups_0, pad = q_39_pad_0, pad_type = q_39_pad_type_0, strides = q_39_strides_0, weight = var_4662_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor k_39_pad_type_0 = const()[name = tensor("k_39_pad_type_0"), val = tensor("valid")]; + tensor k_39_strides_0 = const()[name = tensor("k_39_strides_0"), val = tensor([1, 1])]; + tensor k_39_pad_0 = const()[name = tensor("k_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_39_dilations_0 = const()[name = tensor("k_39_dilations_0"), val = tensor([1, 1])]; + tensor k_39_groups_0 = const()[name = tensor("k_39_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_key_weight_to_fp16 = const()[name = tensor("blocks_19_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490598848)))]; + tensor k_39_cast_fp16 = conv(dilations = k_39_dilations_0, groups = k_39_groups_0, pad = k_39_pad_0, pad_type = k_39_pad_type_0, strides = k_39_strides_0, weight = blocks_19_attn_key_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor var_4660_pad_type_0 = const()[name = tensor("op_4660_pad_type_0"), val = tensor("valid")]; + tensor var_4660_strides_0 = const()[name = tensor("op_4660_strides_0"), val = tensor([1, 1])]; + tensor var_4660_pad_0 = const()[name = tensor("op_4660_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4660_dilations_0 = const()[name = tensor("op_4660_dilations_0"), val = tensor([1, 1])]; + tensor var_4660_groups_0 = const()[name = tensor("op_4660_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_value_weight_to_fp16 = const()[name = tensor("blocks_19_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492696064)))]; + tensor blocks_19_attn_value_bias_to_fp16 = const()[name = tensor("blocks_19_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494793280)))]; + tensor var_4660_cast_fp16 = conv(bias = blocks_19_attn_value_bias_to_fp16, dilations = var_4660_dilations_0, groups = var_4660_groups_0, pad = var_4660_pad_0, pad_type = var_4660_pad_type_0, strides = var_4660_strides_0, weight = blocks_19_attn_value_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor tile_57 = const()[name = tensor("tile_57"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4663_axis_0 = const()[name = tensor("op_4663_axis_0"), val = tensor(1)]; + tensor var_4663_cast_fp16_0, tensor var_4663_cast_fp16_1, tensor var_4663_cast_fp16_2, tensor var_4663_cast_fp16_3, tensor var_4663_cast_fp16_4, tensor var_4663_cast_fp16_5, tensor var_4663_cast_fp16_6, tensor var_4663_cast_fp16_7, tensor var_4663_cast_fp16_8, tensor var_4663_cast_fp16_9, tensor var_4663_cast_fp16_10, tensor var_4663_cast_fp16_11, tensor var_4663_cast_fp16_12, tensor var_4663_cast_fp16_13, tensor var_4663_cast_fp16_14, tensor var_4663_cast_fp16_15 = split(axis = var_4663_axis_0, split_sizes = tile_57, x = var_4662_cast_fp16)[name = tensor("op_4663_cast_fp16")]; + tensor var_4680_perm_0 = const()[name = tensor("op_4680_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_58 = const()[name = tensor("tile_58"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(3)]; + tensor var_4680_cast_fp16 = transpose(perm = var_4680_perm_0, x = k_39_cast_fp16)[name = tensor("transpose_5")]; + tensor var_4681_cast_fp16_0, tensor var_4681_cast_fp16_1, tensor var_4681_cast_fp16_2, tensor var_4681_cast_fp16_3, tensor var_4681_cast_fp16_4, tensor var_4681_cast_fp16_5, tensor var_4681_cast_fp16_6, tensor var_4681_cast_fp16_7, tensor var_4681_cast_fp16_8, tensor var_4681_cast_fp16_9, tensor var_4681_cast_fp16_10, tensor var_4681_cast_fp16_11, tensor var_4681_cast_fp16_12, tensor var_4681_cast_fp16_13, tensor var_4681_cast_fp16_14, tensor var_4681_cast_fp16_15 = split(axis = var_4681_axis_0, split_sizes = tile_58, x = var_4680_cast_fp16)[name = tensor("op_4681_cast_fp16")]; + tensor tile_59 = const()[name = tensor("tile_59"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4698_axis_0 = const()[name = tensor("op_4698_axis_0"), val = tensor(1)]; + tensor var_4698_cast_fp16_0, tensor var_4698_cast_fp16_1, tensor var_4698_cast_fp16_2, tensor var_4698_cast_fp16_3, tensor var_4698_cast_fp16_4, tensor var_4698_cast_fp16_5, tensor var_4698_cast_fp16_6, tensor var_4698_cast_fp16_7, tensor var_4698_cast_fp16_8, tensor var_4698_cast_fp16_9, tensor var_4698_cast_fp16_10, tensor var_4698_cast_fp16_11, tensor var_4698_cast_fp16_12, tensor var_4698_cast_fp16_13, tensor var_4698_cast_fp16_14, tensor var_4698_cast_fp16_15 = split(axis = var_4698_axis_0, split_sizes = tile_59, x = var_4660_cast_fp16)[name = tensor("op_4698_cast_fp16")]; + tensor aw_609_equation_0 = const()[name = tensor("aw_609_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_609_cast_fp16 = einsum(equation = aw_609_equation_0, values = (var_4681_cast_fp16_0, var_4663_cast_fp16_0))[name = tensor("aw_609_cast_fp16")]; + tensor aw_611_equation_0 = const()[name = tensor("aw_611_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_611_cast_fp16 = einsum(equation = aw_611_equation_0, values = (var_4681_cast_fp16_1, var_4663_cast_fp16_1))[name = tensor("aw_611_cast_fp16")]; + tensor aw_613_equation_0 = const()[name = tensor("aw_613_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_613_cast_fp16 = einsum(equation = aw_613_equation_0, values = (var_4681_cast_fp16_2, var_4663_cast_fp16_2))[name = tensor("aw_613_cast_fp16")]; + tensor aw_615_equation_0 = const()[name = tensor("aw_615_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_615_cast_fp16 = einsum(equation = aw_615_equation_0, values = (var_4681_cast_fp16_3, var_4663_cast_fp16_3))[name = tensor("aw_615_cast_fp16")]; + tensor aw_617_equation_0 = const()[name = tensor("aw_617_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_617_cast_fp16 = einsum(equation = aw_617_equation_0, values = (var_4681_cast_fp16_4, var_4663_cast_fp16_4))[name = tensor("aw_617_cast_fp16")]; + tensor aw_619_equation_0 = const()[name = tensor("aw_619_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_619_cast_fp16 = einsum(equation = aw_619_equation_0, values = (var_4681_cast_fp16_5, var_4663_cast_fp16_5))[name = tensor("aw_619_cast_fp16")]; + tensor aw_621_equation_0 = const()[name = tensor("aw_621_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_621_cast_fp16 = einsum(equation = aw_621_equation_0, values = (var_4681_cast_fp16_6, var_4663_cast_fp16_6))[name = tensor("aw_621_cast_fp16")]; + tensor aw_623_equation_0 = const()[name = tensor("aw_623_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_623_cast_fp16 = einsum(equation = aw_623_equation_0, values = (var_4681_cast_fp16_7, var_4663_cast_fp16_7))[name = tensor("aw_623_cast_fp16")]; + tensor aw_625_equation_0 = const()[name = tensor("aw_625_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_625_cast_fp16 = einsum(equation = aw_625_equation_0, values = (var_4681_cast_fp16_8, var_4663_cast_fp16_8))[name = tensor("aw_625_cast_fp16")]; + tensor aw_627_equation_0 = const()[name = tensor("aw_627_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_627_cast_fp16 = einsum(equation = aw_627_equation_0, values = (var_4681_cast_fp16_9, var_4663_cast_fp16_9))[name = tensor("aw_627_cast_fp16")]; + tensor aw_629_equation_0 = const()[name = tensor("aw_629_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_629_cast_fp16 = einsum(equation = aw_629_equation_0, values = (var_4681_cast_fp16_10, var_4663_cast_fp16_10))[name = tensor("aw_629_cast_fp16")]; + tensor aw_631_equation_0 = const()[name = tensor("aw_631_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_631_cast_fp16 = einsum(equation = aw_631_equation_0, values = (var_4681_cast_fp16_11, var_4663_cast_fp16_11))[name = tensor("aw_631_cast_fp16")]; + tensor aw_633_equation_0 = const()[name = tensor("aw_633_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_633_cast_fp16 = einsum(equation = aw_633_equation_0, values = (var_4681_cast_fp16_12, var_4663_cast_fp16_12))[name = tensor("aw_633_cast_fp16")]; + tensor aw_635_equation_0 = const()[name = tensor("aw_635_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_635_cast_fp16 = einsum(equation = aw_635_equation_0, values = (var_4681_cast_fp16_13, var_4663_cast_fp16_13))[name = tensor("aw_635_cast_fp16")]; + tensor aw_637_equation_0 = const()[name = tensor("aw_637_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_637_cast_fp16 = einsum(equation = aw_637_equation_0, values = (var_4681_cast_fp16_14, var_4663_cast_fp16_14))[name = tensor("aw_637_cast_fp16")]; + tensor aw_639_equation_0 = const()[name = tensor("aw_639_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_639_cast_fp16 = einsum(equation = aw_639_equation_0, values = (var_4681_cast_fp16_15, var_4663_cast_fp16_15))[name = tensor("aw_639_cast_fp16")]; + tensor var_4747_cast_fp16 = softmax(axis = var_4611, x = aw_609_cast_fp16)[name = tensor("op_4747_cast_fp16")]; + tensor var_4748_cast_fp16 = softmax(axis = var_4611, x = aw_611_cast_fp16)[name = tensor("op_4748_cast_fp16")]; + tensor var_4749_cast_fp16 = softmax(axis = var_4611, x = aw_613_cast_fp16)[name = tensor("op_4749_cast_fp16")]; + tensor var_4750_cast_fp16 = softmax(axis = var_4611, x = aw_615_cast_fp16)[name = tensor("op_4750_cast_fp16")]; + tensor var_4751_cast_fp16 = softmax(axis = var_4611, x = aw_617_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor var_4752_cast_fp16 = softmax(axis = var_4611, x = aw_619_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4753_cast_fp16 = softmax(axis = var_4611, x = aw_621_cast_fp16)[name = tensor("op_4753_cast_fp16")]; + tensor var_4754_cast_fp16 = softmax(axis = var_4611, x = aw_623_cast_fp16)[name = tensor("op_4754_cast_fp16")]; + tensor var_4755_cast_fp16 = softmax(axis = var_4611, x = aw_625_cast_fp16)[name = tensor("op_4755_cast_fp16")]; + tensor var_4756_cast_fp16 = softmax(axis = var_4611, x = aw_627_cast_fp16)[name = tensor("op_4756_cast_fp16")]; + tensor var_4757_cast_fp16 = softmax(axis = var_4611, x = aw_629_cast_fp16)[name = tensor("op_4757_cast_fp16")]; + tensor var_4758_cast_fp16 = softmax(axis = var_4611, x = aw_631_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = softmax(axis = var_4611, x = aw_633_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor var_4760_cast_fp16 = softmax(axis = var_4611, x = aw_635_cast_fp16)[name = tensor("op_4760_cast_fp16")]; + tensor var_4761_cast_fp16 = softmax(axis = var_4611, x = aw_637_cast_fp16)[name = tensor("op_4761_cast_fp16")]; + tensor var_4762_cast_fp16 = softmax(axis = var_4611, x = aw_639_cast_fp16)[name = tensor("op_4762_cast_fp16")]; + tensor var_4764_equation_0 = const()[name = tensor("op_4764_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4764_cast_fp16 = einsum(equation = var_4764_equation_0, values = (var_4698_cast_fp16_0, var_4747_cast_fp16))[name = tensor("op_4764_cast_fp16")]; + tensor var_4766_equation_0 = const()[name = tensor("op_4766_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4766_cast_fp16 = einsum(equation = var_4766_equation_0, values = (var_4698_cast_fp16_1, var_4748_cast_fp16))[name = tensor("op_4766_cast_fp16")]; + tensor var_4768_equation_0 = const()[name = tensor("op_4768_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4768_cast_fp16 = einsum(equation = var_4768_equation_0, values = (var_4698_cast_fp16_2, var_4749_cast_fp16))[name = tensor("op_4768_cast_fp16")]; + tensor var_4770_equation_0 = const()[name = tensor("op_4770_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4770_cast_fp16 = einsum(equation = var_4770_equation_0, values = (var_4698_cast_fp16_3, var_4750_cast_fp16))[name = tensor("op_4770_cast_fp16")]; + tensor var_4772_equation_0 = const()[name = tensor("op_4772_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4772_cast_fp16 = einsum(equation = var_4772_equation_0, values = (var_4698_cast_fp16_4, var_4751_cast_fp16))[name = tensor("op_4772_cast_fp16")]; + tensor var_4774_equation_0 = const()[name = tensor("op_4774_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4774_cast_fp16 = einsum(equation = var_4774_equation_0, values = (var_4698_cast_fp16_5, var_4752_cast_fp16))[name = tensor("op_4774_cast_fp16")]; + tensor var_4776_equation_0 = const()[name = tensor("op_4776_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4776_cast_fp16 = einsum(equation = var_4776_equation_0, values = (var_4698_cast_fp16_6, var_4753_cast_fp16))[name = tensor("op_4776_cast_fp16")]; + tensor var_4778_equation_0 = const()[name = tensor("op_4778_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4778_cast_fp16 = einsum(equation = var_4778_equation_0, values = (var_4698_cast_fp16_7, var_4754_cast_fp16))[name = tensor("op_4778_cast_fp16")]; + tensor var_4780_equation_0 = const()[name = tensor("op_4780_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4780_cast_fp16 = einsum(equation = var_4780_equation_0, values = (var_4698_cast_fp16_8, var_4755_cast_fp16))[name = tensor("op_4780_cast_fp16")]; + tensor var_4782_equation_0 = const()[name = tensor("op_4782_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4782_cast_fp16 = einsum(equation = var_4782_equation_0, values = (var_4698_cast_fp16_9, var_4756_cast_fp16))[name = tensor("op_4782_cast_fp16")]; + tensor var_4784_equation_0 = const()[name = tensor("op_4784_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4784_cast_fp16 = einsum(equation = var_4784_equation_0, values = (var_4698_cast_fp16_10, var_4757_cast_fp16))[name = tensor("op_4784_cast_fp16")]; + tensor var_4786_equation_0 = const()[name = tensor("op_4786_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4786_cast_fp16 = einsum(equation = var_4786_equation_0, values = (var_4698_cast_fp16_11, var_4758_cast_fp16))[name = tensor("op_4786_cast_fp16")]; + tensor var_4788_equation_0 = const()[name = tensor("op_4788_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4788_cast_fp16 = einsum(equation = var_4788_equation_0, values = (var_4698_cast_fp16_12, var_4759_cast_fp16))[name = tensor("op_4788_cast_fp16")]; + tensor var_4790_equation_0 = const()[name = tensor("op_4790_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4790_cast_fp16 = einsum(equation = var_4790_equation_0, values = (var_4698_cast_fp16_13, var_4760_cast_fp16))[name = tensor("op_4790_cast_fp16")]; + tensor var_4792_equation_0 = const()[name = tensor("op_4792_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4792_cast_fp16 = einsum(equation = var_4792_equation_0, values = (var_4698_cast_fp16_14, var_4761_cast_fp16))[name = tensor("op_4792_cast_fp16")]; + tensor var_4794_equation_0 = const()[name = tensor("op_4794_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4794_cast_fp16 = einsum(equation = var_4794_equation_0, values = (var_4698_cast_fp16_15, var_4762_cast_fp16))[name = tensor("op_4794_cast_fp16")]; + tensor input_195_interleave_0 = const()[name = tensor("input_195_interleave_0"), val = tensor(false)]; + tensor input_195_cast_fp16 = concat(axis = var_4611, interleave = input_195_interleave_0, values = (var_4764_cast_fp16, var_4766_cast_fp16, var_4768_cast_fp16, var_4770_cast_fp16, var_4772_cast_fp16, var_4774_cast_fp16, var_4776_cast_fp16, var_4778_cast_fp16, var_4780_cast_fp16, var_4782_cast_fp16, var_4784_cast_fp16, var_4786_cast_fp16, var_4788_cast_fp16, var_4790_cast_fp16, var_4792_cast_fp16, var_4794_cast_fp16))[name = tensor("input_195_cast_fp16")]; + tensor var_4803_pad_type_0 = const()[name = tensor("op_4803_pad_type_0"), val = tensor("valid")]; + tensor var_4803_strides_0 = const()[name = tensor("op_4803_strides_0"), val = tensor([1, 1])]; + tensor var_4803_pad_0 = const()[name = tensor("op_4803_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4803_dilations_0 = const()[name = tensor("op_4803_dilations_0"), val = tensor([1, 1])]; + tensor var_4803_groups_0 = const()[name = tensor("op_4803_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_out_weight_to_fp16 = const()[name = tensor("blocks_19_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494795392)))]; + tensor blocks_19_attn_out_bias_to_fp16 = const()[name = tensor("blocks_19_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496892608)))]; + tensor var_4803_cast_fp16 = conv(bias = blocks_19_attn_out_bias_to_fp16, dilations = var_4803_dilations_0, groups = var_4803_groups_0, pad = var_4803_pad_0, pad_type = var_4803_pad_type_0, strides = var_4803_strides_0, weight = blocks_19_attn_out_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_4803_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([1])]; + tensor input_197_gamma_0_to_fp16 = const()[name = tensor("input_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496894720)))]; + tensor input_197_beta_0_to_fp16 = const()[name = tensor("input_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496896832)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = input_197_beta_0_to_fp16, epsilon = var_4813_to_fp16, gamma = input_197_gamma_0_to_fp16, x = inputs_79_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("valid")]; + tensor input_199_strides_0 = const()[name = tensor("input_199_strides_0"), val = tensor([1, 1])]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_199_dilations_0 = const()[name = tensor("input_199_dilations_0"), val = tensor([1, 1])]; + tensor input_199_groups_0 = const()[name = tensor("input_199_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496898944)))]; + tensor blocks_19_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505287616)))]; + tensor input_199_cast_fp16 = conv(bias = blocks_19_mlp_0_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = blocks_19_mlp_0_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_4839_pad_type_0 = const()[name = tensor("op_4839_pad_type_0"), val = tensor("valid")]; + tensor var_4839_strides_0 = const()[name = tensor("op_4839_strides_0"), val = tensor([1, 1])]; + tensor var_4839_pad_0 = const()[name = tensor("op_4839_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4839_dilations_0 = const()[name = tensor("op_4839_dilations_0"), val = tensor([1, 1])]; + tensor var_4839_groups_0 = const()[name = tensor("op_4839_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505295872)))]; + tensor blocks_19_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513684544)))]; + tensor var_4839_cast_fp16 = conv(bias = blocks_19_mlp_2_bias_to_fp16, dilations = var_4839_dilations_0, groups = var_4839_groups_0, pad = var_4839_pad_0, pad_type = var_4839_pad_type_0, strides = var_4839_strides_0, weight = blocks_19_mlp_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("op_4839_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = var_4839_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor(1)]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([1])]; + tensor input_203_gamma_0_to_fp16 = const()[name = tensor("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513686656)))]; + tensor input_203_beta_0_to_fp16 = const()[name = tensor("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513688768)))]; + tensor var_4864_to_fp16 = const()[name = tensor("op_4864_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = input_203_beta_0_to_fp16, epsilon = var_4864_to_fp16, gamma = input_203_gamma_0_to_fp16, x = inputs_81_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor q_41_pad_type_0 = const()[name = tensor("q_41_pad_type_0"), val = tensor("valid")]; + tensor q_41_strides_0 = const()[name = tensor("q_41_strides_0"), val = tensor([1, 1])]; + tensor q_41_pad_0 = const()[name = tensor("q_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_41_dilations_0 = const()[name = tensor("q_41_dilations_0"), val = tensor([1, 1])]; + tensor q_41_groups_0 = const()[name = tensor("q_41_groups_0"), val = tensor(1)]; + tensor var_4899_weight_0_to_fp16 = const()[name = tensor("op_4899_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513690880)))]; + tensor var_4899_bias_0_to_fp16 = const()[name = tensor("op_4899_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515788096)))]; + tensor var_4899_cast_fp16 = conv(bias = var_4899_bias_0_to_fp16, dilations = q_41_dilations_0, groups = q_41_groups_0, pad = q_41_pad_0, pad_type = q_41_pad_type_0, strides = q_41_strides_0, weight = var_4899_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("op_4899_cast_fp16")]; + tensor k_41_pad_type_0 = const()[name = tensor("k_41_pad_type_0"), val = tensor("valid")]; + tensor k_41_strides_0 = const()[name = tensor("k_41_strides_0"), val = tensor([1, 1])]; + tensor k_41_pad_0 = const()[name = tensor("k_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_41_dilations_0 = const()[name = tensor("k_41_dilations_0"), val = tensor([1, 1])]; + tensor k_41_groups_0 = const()[name = tensor("k_41_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_key_weight_to_fp16 = const()[name = tensor("blocks_20_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515790208)))]; + tensor k_41_cast_fp16 = conv(dilations = k_41_dilations_0, groups = k_41_groups_0, pad = k_41_pad_0, pad_type = k_41_pad_type_0, strides = k_41_strides_0, weight = blocks_20_attn_key_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor var_4897_pad_type_0 = const()[name = tensor("op_4897_pad_type_0"), val = tensor("valid")]; + tensor var_4897_strides_0 = const()[name = tensor("op_4897_strides_0"), val = tensor([1, 1])]; + tensor var_4897_pad_0 = const()[name = tensor("op_4897_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4897_dilations_0 = const()[name = tensor("op_4897_dilations_0"), val = tensor([1, 1])]; + tensor var_4897_groups_0 = const()[name = tensor("op_4897_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_value_weight_to_fp16 = const()[name = tensor("blocks_20_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517887424)))]; + tensor blocks_20_attn_value_bias_to_fp16 = const()[name = tensor("blocks_20_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519984640)))]; + tensor var_4897_cast_fp16 = conv(bias = blocks_20_attn_value_bias_to_fp16, dilations = var_4897_dilations_0, groups = var_4897_groups_0, pad = var_4897_pad_0, pad_type = var_4897_pad_type_0, strides = var_4897_strides_0, weight = blocks_20_attn_value_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("op_4897_cast_fp16")]; + tensor tile_60 = const()[name = tensor("tile_60"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4900_axis_0 = const()[name = tensor("op_4900_axis_0"), val = tensor(1)]; + tensor var_4900_cast_fp16_0, tensor var_4900_cast_fp16_1, tensor var_4900_cast_fp16_2, tensor var_4900_cast_fp16_3, tensor var_4900_cast_fp16_4, tensor var_4900_cast_fp16_5, tensor var_4900_cast_fp16_6, tensor var_4900_cast_fp16_7, tensor var_4900_cast_fp16_8, tensor var_4900_cast_fp16_9, tensor var_4900_cast_fp16_10, tensor var_4900_cast_fp16_11, tensor var_4900_cast_fp16_12, tensor var_4900_cast_fp16_13, tensor var_4900_cast_fp16_14, tensor var_4900_cast_fp16_15 = split(axis = var_4900_axis_0, split_sizes = tile_60, x = var_4899_cast_fp16)[name = tensor("op_4900_cast_fp16")]; + tensor var_4917_perm_0 = const()[name = tensor("op_4917_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_61 = const()[name = tensor("tile_61"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4918_axis_0 = const()[name = tensor("op_4918_axis_0"), val = tensor(3)]; + tensor var_4917_cast_fp16 = transpose(perm = var_4917_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_4")]; + tensor var_4918_cast_fp16_0, tensor var_4918_cast_fp16_1, tensor var_4918_cast_fp16_2, tensor var_4918_cast_fp16_3, tensor var_4918_cast_fp16_4, tensor var_4918_cast_fp16_5, tensor var_4918_cast_fp16_6, tensor var_4918_cast_fp16_7, tensor var_4918_cast_fp16_8, tensor var_4918_cast_fp16_9, tensor var_4918_cast_fp16_10, tensor var_4918_cast_fp16_11, tensor var_4918_cast_fp16_12, tensor var_4918_cast_fp16_13, tensor var_4918_cast_fp16_14, tensor var_4918_cast_fp16_15 = split(axis = var_4918_axis_0, split_sizes = tile_61, x = var_4917_cast_fp16)[name = tensor("op_4918_cast_fp16")]; + tensor tile_62 = const()[name = tensor("tile_62"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4935_axis_0 = const()[name = tensor("op_4935_axis_0"), val = tensor(1)]; + tensor var_4935_cast_fp16_0, tensor var_4935_cast_fp16_1, tensor var_4935_cast_fp16_2, tensor var_4935_cast_fp16_3, tensor var_4935_cast_fp16_4, tensor var_4935_cast_fp16_5, tensor var_4935_cast_fp16_6, tensor var_4935_cast_fp16_7, tensor var_4935_cast_fp16_8, tensor var_4935_cast_fp16_9, tensor var_4935_cast_fp16_10, tensor var_4935_cast_fp16_11, tensor var_4935_cast_fp16_12, tensor var_4935_cast_fp16_13, tensor var_4935_cast_fp16_14, tensor var_4935_cast_fp16_15 = split(axis = var_4935_axis_0, split_sizes = tile_62, x = var_4897_cast_fp16)[name = tensor("op_4935_cast_fp16")]; + tensor aw_641_equation_0 = const()[name = tensor("aw_641_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_641_cast_fp16 = einsum(equation = aw_641_equation_0, values = (var_4918_cast_fp16_0, var_4900_cast_fp16_0))[name = tensor("aw_641_cast_fp16")]; + tensor aw_643_equation_0 = const()[name = tensor("aw_643_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_643_cast_fp16 = einsum(equation = aw_643_equation_0, values = (var_4918_cast_fp16_1, var_4900_cast_fp16_1))[name = tensor("aw_643_cast_fp16")]; + tensor aw_645_equation_0 = const()[name = tensor("aw_645_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_645_cast_fp16 = einsum(equation = aw_645_equation_0, values = (var_4918_cast_fp16_2, var_4900_cast_fp16_2))[name = tensor("aw_645_cast_fp16")]; + tensor aw_647_equation_0 = const()[name = tensor("aw_647_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_647_cast_fp16 = einsum(equation = aw_647_equation_0, values = (var_4918_cast_fp16_3, var_4900_cast_fp16_3))[name = tensor("aw_647_cast_fp16")]; + tensor aw_649_equation_0 = const()[name = tensor("aw_649_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_649_cast_fp16 = einsum(equation = aw_649_equation_0, values = (var_4918_cast_fp16_4, var_4900_cast_fp16_4))[name = tensor("aw_649_cast_fp16")]; + tensor aw_651_equation_0 = const()[name = tensor("aw_651_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_651_cast_fp16 = einsum(equation = aw_651_equation_0, values = (var_4918_cast_fp16_5, var_4900_cast_fp16_5))[name = tensor("aw_651_cast_fp16")]; + tensor aw_653_equation_0 = const()[name = tensor("aw_653_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_653_cast_fp16 = einsum(equation = aw_653_equation_0, values = (var_4918_cast_fp16_6, var_4900_cast_fp16_6))[name = tensor("aw_653_cast_fp16")]; + tensor aw_655_equation_0 = const()[name = tensor("aw_655_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_655_cast_fp16 = einsum(equation = aw_655_equation_0, values = (var_4918_cast_fp16_7, var_4900_cast_fp16_7))[name = tensor("aw_655_cast_fp16")]; + tensor aw_657_equation_0 = const()[name = tensor("aw_657_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_657_cast_fp16 = einsum(equation = aw_657_equation_0, values = (var_4918_cast_fp16_8, var_4900_cast_fp16_8))[name = tensor("aw_657_cast_fp16")]; + tensor aw_659_equation_0 = const()[name = tensor("aw_659_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_659_cast_fp16 = einsum(equation = aw_659_equation_0, values = (var_4918_cast_fp16_9, var_4900_cast_fp16_9))[name = tensor("aw_659_cast_fp16")]; + tensor aw_661_equation_0 = const()[name = tensor("aw_661_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_661_cast_fp16 = einsum(equation = aw_661_equation_0, values = (var_4918_cast_fp16_10, var_4900_cast_fp16_10))[name = tensor("aw_661_cast_fp16")]; + tensor aw_663_equation_0 = const()[name = tensor("aw_663_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_663_cast_fp16 = einsum(equation = aw_663_equation_0, values = (var_4918_cast_fp16_11, var_4900_cast_fp16_11))[name = tensor("aw_663_cast_fp16")]; + tensor aw_665_equation_0 = const()[name = tensor("aw_665_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_665_cast_fp16 = einsum(equation = aw_665_equation_0, values = (var_4918_cast_fp16_12, var_4900_cast_fp16_12))[name = tensor("aw_665_cast_fp16")]; + tensor aw_667_equation_0 = const()[name = tensor("aw_667_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_667_cast_fp16 = einsum(equation = aw_667_equation_0, values = (var_4918_cast_fp16_13, var_4900_cast_fp16_13))[name = tensor("aw_667_cast_fp16")]; + tensor aw_669_equation_0 = const()[name = tensor("aw_669_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_669_cast_fp16 = einsum(equation = aw_669_equation_0, values = (var_4918_cast_fp16_14, var_4900_cast_fp16_14))[name = tensor("aw_669_cast_fp16")]; + tensor aw_671_equation_0 = const()[name = tensor("aw_671_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_671_cast_fp16 = einsum(equation = aw_671_equation_0, values = (var_4918_cast_fp16_15, var_4900_cast_fp16_15))[name = tensor("aw_671_cast_fp16")]; + tensor var_4984_cast_fp16 = softmax(axis = var_4848, x = aw_641_cast_fp16)[name = tensor("op_4984_cast_fp16")]; + tensor var_4985_cast_fp16 = softmax(axis = var_4848, x = aw_643_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor var_4986_cast_fp16 = softmax(axis = var_4848, x = aw_645_cast_fp16)[name = tensor("op_4986_cast_fp16")]; + tensor var_4987_cast_fp16 = softmax(axis = var_4848, x = aw_647_cast_fp16)[name = tensor("op_4987_cast_fp16")]; + tensor var_4988_cast_fp16 = softmax(axis = var_4848, x = aw_649_cast_fp16)[name = tensor("op_4988_cast_fp16")]; + tensor var_4989_cast_fp16 = softmax(axis = var_4848, x = aw_651_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor var_4990_cast_fp16 = softmax(axis = var_4848, x = aw_653_cast_fp16)[name = tensor("op_4990_cast_fp16")]; + tensor var_4991_cast_fp16 = softmax(axis = var_4848, x = aw_655_cast_fp16)[name = tensor("op_4991_cast_fp16")]; + tensor var_4992_cast_fp16 = softmax(axis = var_4848, x = aw_657_cast_fp16)[name = tensor("op_4992_cast_fp16")]; + tensor var_4993_cast_fp16 = softmax(axis = var_4848, x = aw_659_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor var_4994_cast_fp16 = softmax(axis = var_4848, x = aw_661_cast_fp16)[name = tensor("op_4994_cast_fp16")]; + tensor var_4995_cast_fp16 = softmax(axis = var_4848, x = aw_663_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor var_4996_cast_fp16 = softmax(axis = var_4848, x = aw_665_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_4997_cast_fp16 = softmax(axis = var_4848, x = aw_667_cast_fp16)[name = tensor("op_4997_cast_fp16")]; + tensor var_4998_cast_fp16 = softmax(axis = var_4848, x = aw_669_cast_fp16)[name = tensor("op_4998_cast_fp16")]; + tensor var_4999_cast_fp16 = softmax(axis = var_4848, x = aw_671_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5001_equation_0 = const()[name = tensor("op_5001_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5001_cast_fp16 = einsum(equation = var_5001_equation_0, values = (var_4935_cast_fp16_0, var_4984_cast_fp16))[name = tensor("op_5001_cast_fp16")]; + tensor var_5003_equation_0 = const()[name = tensor("op_5003_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5003_cast_fp16 = einsum(equation = var_5003_equation_0, values = (var_4935_cast_fp16_1, var_4985_cast_fp16))[name = tensor("op_5003_cast_fp16")]; + tensor var_5005_equation_0 = const()[name = tensor("op_5005_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5005_cast_fp16 = einsum(equation = var_5005_equation_0, values = (var_4935_cast_fp16_2, var_4986_cast_fp16))[name = tensor("op_5005_cast_fp16")]; + tensor var_5007_equation_0 = const()[name = tensor("op_5007_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5007_cast_fp16 = einsum(equation = var_5007_equation_0, values = (var_4935_cast_fp16_3, var_4987_cast_fp16))[name = tensor("op_5007_cast_fp16")]; + tensor var_5009_equation_0 = const()[name = tensor("op_5009_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5009_cast_fp16 = einsum(equation = var_5009_equation_0, values = (var_4935_cast_fp16_4, var_4988_cast_fp16))[name = tensor("op_5009_cast_fp16")]; + tensor var_5011_equation_0 = const()[name = tensor("op_5011_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5011_cast_fp16 = einsum(equation = var_5011_equation_0, values = (var_4935_cast_fp16_5, var_4989_cast_fp16))[name = tensor("op_5011_cast_fp16")]; + tensor var_5013_equation_0 = const()[name = tensor("op_5013_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5013_cast_fp16 = einsum(equation = var_5013_equation_0, values = (var_4935_cast_fp16_6, var_4990_cast_fp16))[name = tensor("op_5013_cast_fp16")]; + tensor var_5015_equation_0 = const()[name = tensor("op_5015_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5015_cast_fp16 = einsum(equation = var_5015_equation_0, values = (var_4935_cast_fp16_7, var_4991_cast_fp16))[name = tensor("op_5015_cast_fp16")]; + tensor var_5017_equation_0 = const()[name = tensor("op_5017_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5017_cast_fp16 = einsum(equation = var_5017_equation_0, values = (var_4935_cast_fp16_8, var_4992_cast_fp16))[name = tensor("op_5017_cast_fp16")]; + tensor var_5019_equation_0 = const()[name = tensor("op_5019_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5019_cast_fp16 = einsum(equation = var_5019_equation_0, values = (var_4935_cast_fp16_9, var_4993_cast_fp16))[name = tensor("op_5019_cast_fp16")]; + tensor var_5021_equation_0 = const()[name = tensor("op_5021_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5021_cast_fp16 = einsum(equation = var_5021_equation_0, values = (var_4935_cast_fp16_10, var_4994_cast_fp16))[name = tensor("op_5021_cast_fp16")]; + tensor var_5023_equation_0 = const()[name = tensor("op_5023_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5023_cast_fp16 = einsum(equation = var_5023_equation_0, values = (var_4935_cast_fp16_11, var_4995_cast_fp16))[name = tensor("op_5023_cast_fp16")]; + tensor var_5025_equation_0 = const()[name = tensor("op_5025_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5025_cast_fp16 = einsum(equation = var_5025_equation_0, values = (var_4935_cast_fp16_12, var_4996_cast_fp16))[name = tensor("op_5025_cast_fp16")]; + tensor var_5027_equation_0 = const()[name = tensor("op_5027_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5027_cast_fp16 = einsum(equation = var_5027_equation_0, values = (var_4935_cast_fp16_13, var_4997_cast_fp16))[name = tensor("op_5027_cast_fp16")]; + tensor var_5029_equation_0 = const()[name = tensor("op_5029_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5029_cast_fp16 = einsum(equation = var_5029_equation_0, values = (var_4935_cast_fp16_14, var_4998_cast_fp16))[name = tensor("op_5029_cast_fp16")]; + tensor var_5031_equation_0 = const()[name = tensor("op_5031_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5031_cast_fp16 = einsum(equation = var_5031_equation_0, values = (var_4935_cast_fp16_15, var_4999_cast_fp16))[name = tensor("op_5031_cast_fp16")]; + tensor input_205_interleave_0 = const()[name = tensor("input_205_interleave_0"), val = tensor(false)]; + tensor input_205_cast_fp16 = concat(axis = var_4848, interleave = input_205_interleave_0, values = (var_5001_cast_fp16, var_5003_cast_fp16, var_5005_cast_fp16, var_5007_cast_fp16, var_5009_cast_fp16, var_5011_cast_fp16, var_5013_cast_fp16, var_5015_cast_fp16, var_5017_cast_fp16, var_5019_cast_fp16, var_5021_cast_fp16, var_5023_cast_fp16, var_5025_cast_fp16, var_5027_cast_fp16, var_5029_cast_fp16, var_5031_cast_fp16))[name = tensor("input_205_cast_fp16")]; + tensor var_5040_pad_type_0 = const()[name = tensor("op_5040_pad_type_0"), val = tensor("valid")]; + tensor var_5040_strides_0 = const()[name = tensor("op_5040_strides_0"), val = tensor([1, 1])]; + tensor var_5040_pad_0 = const()[name = tensor("op_5040_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5040_dilations_0 = const()[name = tensor("op_5040_dilations_0"), val = tensor([1, 1])]; + tensor var_5040_groups_0 = const()[name = tensor("op_5040_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_out_weight_to_fp16 = const()[name = tensor("blocks_20_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519986752)))]; + tensor blocks_20_attn_out_bias_to_fp16 = const()[name = tensor("blocks_20_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522083968)))]; + tensor var_5040_cast_fp16 = conv(bias = blocks_20_attn_out_bias_to_fp16, dilations = var_5040_dilations_0, groups = var_5040_groups_0, pad = var_5040_pad_0, pad_type = var_5040_pad_type_0, strides = var_5040_strides_0, weight = blocks_20_attn_out_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_5040_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([1])]; + tensor input_207_gamma_0_to_fp16 = const()[name = tensor("input_207_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522086080)))]; + tensor input_207_beta_0_to_fp16 = const()[name = tensor("input_207_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522088192)))]; + tensor var_5050_to_fp16 = const()[name = tensor("op_5050_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = input_207_beta_0_to_fp16, epsilon = var_5050_to_fp16, gamma = input_207_gamma_0_to_fp16, x = inputs_83_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("valid")]; + tensor input_209_strides_0 = const()[name = tensor("input_209_strides_0"), val = tensor([1, 1])]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_209_dilations_0 = const()[name = tensor("input_209_dilations_0"), val = tensor([1, 1])]; + tensor input_209_groups_0 = const()[name = tensor("input_209_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522090304)))]; + tensor blocks_20_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530478976)))]; + tensor input_209_cast_fp16 = conv(bias = blocks_20_mlp_0_bias_to_fp16, dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = blocks_20_mlp_0_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = input_209_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_5076_pad_type_0 = const()[name = tensor("op_5076_pad_type_0"), val = tensor("valid")]; + tensor var_5076_strides_0 = const()[name = tensor("op_5076_strides_0"), val = tensor([1, 1])]; + tensor var_5076_pad_0 = const()[name = tensor("op_5076_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5076_dilations_0 = const()[name = tensor("op_5076_dilations_0"), val = tensor([1, 1])]; + tensor var_5076_groups_0 = const()[name = tensor("op_5076_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530487232)))]; + tensor blocks_20_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538875904)))]; + tensor var_5076_cast_fp16 = conv(bias = blocks_20_mlp_2_bias_to_fp16, dilations = var_5076_dilations_0, groups = var_5076_groups_0, pad = var_5076_pad_0, pad_type = var_5076_pad_type_0, strides = var_5076_strides_0, weight = blocks_20_mlp_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("op_5076_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = var_5076_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_5085 = const()[name = tensor("op_5085"), val = tensor(1)]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([1])]; + tensor input_213_gamma_0_to_fp16 = const()[name = tensor("input_213_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538878016)))]; + tensor input_213_beta_0_to_fp16 = const()[name = tensor("input_213_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538880128)))]; + tensor var_5101_to_fp16 = const()[name = tensor("op_5101_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = input_213_beta_0_to_fp16, epsilon = var_5101_to_fp16, gamma = input_213_gamma_0_to_fp16, x = inputs_85_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor q_43_pad_type_0 = const()[name = tensor("q_43_pad_type_0"), val = tensor("valid")]; + tensor q_43_strides_0 = const()[name = tensor("q_43_strides_0"), val = tensor([1, 1])]; + tensor q_43_pad_0 = const()[name = tensor("q_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_43_dilations_0 = const()[name = tensor("q_43_dilations_0"), val = tensor([1, 1])]; + tensor q_43_groups_0 = const()[name = tensor("q_43_groups_0"), val = tensor(1)]; + tensor var_5136_weight_0_to_fp16 = const()[name = tensor("op_5136_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538882240)))]; + tensor var_5136_bias_0_to_fp16 = const()[name = tensor("op_5136_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540979456)))]; + tensor var_5136_cast_fp16 = conv(bias = var_5136_bias_0_to_fp16, dilations = q_43_dilations_0, groups = q_43_groups_0, pad = q_43_pad_0, pad_type = q_43_pad_type_0, strides = q_43_strides_0, weight = var_5136_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5136_cast_fp16")]; + tensor k_43_pad_type_0 = const()[name = tensor("k_43_pad_type_0"), val = tensor("valid")]; + tensor k_43_strides_0 = const()[name = tensor("k_43_strides_0"), val = tensor([1, 1])]; + tensor k_43_pad_0 = const()[name = tensor("k_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_43_dilations_0 = const()[name = tensor("k_43_dilations_0"), val = tensor([1, 1])]; + tensor k_43_groups_0 = const()[name = tensor("k_43_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_key_weight_to_fp16 = const()[name = tensor("blocks_21_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540981568)))]; + tensor k_43_cast_fp16 = conv(dilations = k_43_dilations_0, groups = k_43_groups_0, pad = k_43_pad_0, pad_type = k_43_pad_type_0, strides = k_43_strides_0, weight = blocks_21_attn_key_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor var_5134_pad_type_0 = const()[name = tensor("op_5134_pad_type_0"), val = tensor("valid")]; + tensor var_5134_strides_0 = const()[name = tensor("op_5134_strides_0"), val = tensor([1, 1])]; + tensor var_5134_pad_0 = const()[name = tensor("op_5134_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5134_dilations_0 = const()[name = tensor("op_5134_dilations_0"), val = tensor([1, 1])]; + tensor var_5134_groups_0 = const()[name = tensor("op_5134_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_value_weight_to_fp16 = const()[name = tensor("blocks_21_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543078784)))]; + tensor blocks_21_attn_value_bias_to_fp16 = const()[name = tensor("blocks_21_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545176000)))]; + tensor var_5134_cast_fp16 = conv(bias = blocks_21_attn_value_bias_to_fp16, dilations = var_5134_dilations_0, groups = var_5134_groups_0, pad = var_5134_pad_0, pad_type = var_5134_pad_type_0, strides = var_5134_strides_0, weight = blocks_21_attn_value_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5134_cast_fp16")]; + tensor tile_63 = const()[name = tensor("tile_63"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5137_axis_0 = const()[name = tensor("op_5137_axis_0"), val = tensor(1)]; + tensor var_5137_cast_fp16_0, tensor var_5137_cast_fp16_1, tensor var_5137_cast_fp16_2, tensor var_5137_cast_fp16_3, tensor var_5137_cast_fp16_4, tensor var_5137_cast_fp16_5, tensor var_5137_cast_fp16_6, tensor var_5137_cast_fp16_7, tensor var_5137_cast_fp16_8, tensor var_5137_cast_fp16_9, tensor var_5137_cast_fp16_10, tensor var_5137_cast_fp16_11, tensor var_5137_cast_fp16_12, tensor var_5137_cast_fp16_13, tensor var_5137_cast_fp16_14, tensor var_5137_cast_fp16_15 = split(axis = var_5137_axis_0, split_sizes = tile_63, x = var_5136_cast_fp16)[name = tensor("op_5137_cast_fp16")]; + tensor var_5154_perm_0 = const()[name = tensor("op_5154_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_64 = const()[name = tensor("tile_64"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5155_axis_0 = const()[name = tensor("op_5155_axis_0"), val = tensor(3)]; + tensor var_5154_cast_fp16 = transpose(perm = var_5154_perm_0, x = k_43_cast_fp16)[name = tensor("transpose_3")]; + tensor var_5155_cast_fp16_0, tensor var_5155_cast_fp16_1, tensor var_5155_cast_fp16_2, tensor var_5155_cast_fp16_3, tensor var_5155_cast_fp16_4, tensor var_5155_cast_fp16_5, tensor var_5155_cast_fp16_6, tensor var_5155_cast_fp16_7, tensor var_5155_cast_fp16_8, tensor var_5155_cast_fp16_9, tensor var_5155_cast_fp16_10, tensor var_5155_cast_fp16_11, tensor var_5155_cast_fp16_12, tensor var_5155_cast_fp16_13, tensor var_5155_cast_fp16_14, tensor var_5155_cast_fp16_15 = split(axis = var_5155_axis_0, split_sizes = tile_64, x = var_5154_cast_fp16)[name = tensor("op_5155_cast_fp16")]; + tensor tile_65 = const()[name = tensor("tile_65"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5172_axis_0 = const()[name = tensor("op_5172_axis_0"), val = tensor(1)]; + tensor var_5172_cast_fp16_0, tensor var_5172_cast_fp16_1, tensor var_5172_cast_fp16_2, tensor var_5172_cast_fp16_3, tensor var_5172_cast_fp16_4, tensor var_5172_cast_fp16_5, tensor var_5172_cast_fp16_6, tensor var_5172_cast_fp16_7, tensor var_5172_cast_fp16_8, tensor var_5172_cast_fp16_9, tensor var_5172_cast_fp16_10, tensor var_5172_cast_fp16_11, tensor var_5172_cast_fp16_12, tensor var_5172_cast_fp16_13, tensor var_5172_cast_fp16_14, tensor var_5172_cast_fp16_15 = split(axis = var_5172_axis_0, split_sizes = tile_65, x = var_5134_cast_fp16)[name = tensor("op_5172_cast_fp16")]; + tensor aw_673_equation_0 = const()[name = tensor("aw_673_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_673_cast_fp16 = einsum(equation = aw_673_equation_0, values = (var_5155_cast_fp16_0, var_5137_cast_fp16_0))[name = tensor("aw_673_cast_fp16")]; + tensor aw_675_equation_0 = const()[name = tensor("aw_675_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_675_cast_fp16 = einsum(equation = aw_675_equation_0, values = (var_5155_cast_fp16_1, var_5137_cast_fp16_1))[name = tensor("aw_675_cast_fp16")]; + tensor aw_677_equation_0 = const()[name = tensor("aw_677_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_677_cast_fp16 = einsum(equation = aw_677_equation_0, values = (var_5155_cast_fp16_2, var_5137_cast_fp16_2))[name = tensor("aw_677_cast_fp16")]; + tensor aw_679_equation_0 = const()[name = tensor("aw_679_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_679_cast_fp16 = einsum(equation = aw_679_equation_0, values = (var_5155_cast_fp16_3, var_5137_cast_fp16_3))[name = tensor("aw_679_cast_fp16")]; + tensor aw_681_equation_0 = const()[name = tensor("aw_681_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_681_cast_fp16 = einsum(equation = aw_681_equation_0, values = (var_5155_cast_fp16_4, var_5137_cast_fp16_4))[name = tensor("aw_681_cast_fp16")]; + tensor aw_683_equation_0 = const()[name = tensor("aw_683_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_683_cast_fp16 = einsum(equation = aw_683_equation_0, values = (var_5155_cast_fp16_5, var_5137_cast_fp16_5))[name = tensor("aw_683_cast_fp16")]; + tensor aw_685_equation_0 = const()[name = tensor("aw_685_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_685_cast_fp16 = einsum(equation = aw_685_equation_0, values = (var_5155_cast_fp16_6, var_5137_cast_fp16_6))[name = tensor("aw_685_cast_fp16")]; + tensor aw_687_equation_0 = const()[name = tensor("aw_687_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_687_cast_fp16 = einsum(equation = aw_687_equation_0, values = (var_5155_cast_fp16_7, var_5137_cast_fp16_7))[name = tensor("aw_687_cast_fp16")]; + tensor aw_689_equation_0 = const()[name = tensor("aw_689_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_689_cast_fp16 = einsum(equation = aw_689_equation_0, values = (var_5155_cast_fp16_8, var_5137_cast_fp16_8))[name = tensor("aw_689_cast_fp16")]; + tensor aw_691_equation_0 = const()[name = tensor("aw_691_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_691_cast_fp16 = einsum(equation = aw_691_equation_0, values = (var_5155_cast_fp16_9, var_5137_cast_fp16_9))[name = tensor("aw_691_cast_fp16")]; + tensor aw_693_equation_0 = const()[name = tensor("aw_693_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_693_cast_fp16 = einsum(equation = aw_693_equation_0, values = (var_5155_cast_fp16_10, var_5137_cast_fp16_10))[name = tensor("aw_693_cast_fp16")]; + tensor aw_695_equation_0 = const()[name = tensor("aw_695_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_695_cast_fp16 = einsum(equation = aw_695_equation_0, values = (var_5155_cast_fp16_11, var_5137_cast_fp16_11))[name = tensor("aw_695_cast_fp16")]; + tensor aw_697_equation_0 = const()[name = tensor("aw_697_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_697_cast_fp16 = einsum(equation = aw_697_equation_0, values = (var_5155_cast_fp16_12, var_5137_cast_fp16_12))[name = tensor("aw_697_cast_fp16")]; + tensor aw_699_equation_0 = const()[name = tensor("aw_699_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_699_cast_fp16 = einsum(equation = aw_699_equation_0, values = (var_5155_cast_fp16_13, var_5137_cast_fp16_13))[name = tensor("aw_699_cast_fp16")]; + tensor aw_701_equation_0 = const()[name = tensor("aw_701_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_701_cast_fp16 = einsum(equation = aw_701_equation_0, values = (var_5155_cast_fp16_14, var_5137_cast_fp16_14))[name = tensor("aw_701_cast_fp16")]; + tensor aw_703_equation_0 = const()[name = tensor("aw_703_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_703_cast_fp16 = einsum(equation = aw_703_equation_0, values = (var_5155_cast_fp16_15, var_5137_cast_fp16_15))[name = tensor("aw_703_cast_fp16")]; + tensor var_5221_cast_fp16 = softmax(axis = var_5085, x = aw_673_cast_fp16)[name = tensor("op_5221_cast_fp16")]; + tensor var_5222_cast_fp16 = softmax(axis = var_5085, x = aw_675_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5223_cast_fp16 = softmax(axis = var_5085, x = aw_677_cast_fp16)[name = tensor("op_5223_cast_fp16")]; + tensor var_5224_cast_fp16 = softmax(axis = var_5085, x = aw_679_cast_fp16)[name = tensor("op_5224_cast_fp16")]; + tensor var_5225_cast_fp16 = softmax(axis = var_5085, x = aw_681_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5226_cast_fp16 = softmax(axis = var_5085, x = aw_683_cast_fp16)[name = tensor("op_5226_cast_fp16")]; + tensor var_5227_cast_fp16 = softmax(axis = var_5085, x = aw_685_cast_fp16)[name = tensor("op_5227_cast_fp16")]; + tensor var_5228_cast_fp16 = softmax(axis = var_5085, x = aw_687_cast_fp16)[name = tensor("op_5228_cast_fp16")]; + tensor var_5229_cast_fp16 = softmax(axis = var_5085, x = aw_689_cast_fp16)[name = tensor("op_5229_cast_fp16")]; + tensor var_5230_cast_fp16 = softmax(axis = var_5085, x = aw_691_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5231_cast_fp16 = softmax(axis = var_5085, x = aw_693_cast_fp16)[name = tensor("op_5231_cast_fp16")]; + tensor var_5232_cast_fp16 = softmax(axis = var_5085, x = aw_695_cast_fp16)[name = tensor("op_5232_cast_fp16")]; + tensor var_5233_cast_fp16 = softmax(axis = var_5085, x = aw_697_cast_fp16)[name = tensor("op_5233_cast_fp16")]; + tensor var_5234_cast_fp16 = softmax(axis = var_5085, x = aw_699_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = softmax(axis = var_5085, x = aw_701_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor var_5236_cast_fp16 = softmax(axis = var_5085, x = aw_703_cast_fp16)[name = tensor("op_5236_cast_fp16")]; + tensor var_5238_equation_0 = const()[name = tensor("op_5238_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5238_cast_fp16 = einsum(equation = var_5238_equation_0, values = (var_5172_cast_fp16_0, var_5221_cast_fp16))[name = tensor("op_5238_cast_fp16")]; + tensor var_5240_equation_0 = const()[name = tensor("op_5240_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5240_cast_fp16 = einsum(equation = var_5240_equation_0, values = (var_5172_cast_fp16_1, var_5222_cast_fp16))[name = tensor("op_5240_cast_fp16")]; + tensor var_5242_equation_0 = const()[name = tensor("op_5242_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5242_cast_fp16 = einsum(equation = var_5242_equation_0, values = (var_5172_cast_fp16_2, var_5223_cast_fp16))[name = tensor("op_5242_cast_fp16")]; + tensor var_5244_equation_0 = const()[name = tensor("op_5244_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5244_cast_fp16 = einsum(equation = var_5244_equation_0, values = (var_5172_cast_fp16_3, var_5224_cast_fp16))[name = tensor("op_5244_cast_fp16")]; + tensor var_5246_equation_0 = const()[name = tensor("op_5246_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5246_cast_fp16 = einsum(equation = var_5246_equation_0, values = (var_5172_cast_fp16_4, var_5225_cast_fp16))[name = tensor("op_5246_cast_fp16")]; + tensor var_5248_equation_0 = const()[name = tensor("op_5248_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5248_cast_fp16 = einsum(equation = var_5248_equation_0, values = (var_5172_cast_fp16_5, var_5226_cast_fp16))[name = tensor("op_5248_cast_fp16")]; + tensor var_5250_equation_0 = const()[name = tensor("op_5250_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5250_cast_fp16 = einsum(equation = var_5250_equation_0, values = (var_5172_cast_fp16_6, var_5227_cast_fp16))[name = tensor("op_5250_cast_fp16")]; + tensor var_5252_equation_0 = const()[name = tensor("op_5252_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5252_cast_fp16 = einsum(equation = var_5252_equation_0, values = (var_5172_cast_fp16_7, var_5228_cast_fp16))[name = tensor("op_5252_cast_fp16")]; + tensor var_5254_equation_0 = const()[name = tensor("op_5254_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5254_cast_fp16 = einsum(equation = var_5254_equation_0, values = (var_5172_cast_fp16_8, var_5229_cast_fp16))[name = tensor("op_5254_cast_fp16")]; + tensor var_5256_equation_0 = const()[name = tensor("op_5256_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5256_cast_fp16 = einsum(equation = var_5256_equation_0, values = (var_5172_cast_fp16_9, var_5230_cast_fp16))[name = tensor("op_5256_cast_fp16")]; + tensor var_5258_equation_0 = const()[name = tensor("op_5258_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5258_cast_fp16 = einsum(equation = var_5258_equation_0, values = (var_5172_cast_fp16_10, var_5231_cast_fp16))[name = tensor("op_5258_cast_fp16")]; + tensor var_5260_equation_0 = const()[name = tensor("op_5260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5260_cast_fp16 = einsum(equation = var_5260_equation_0, values = (var_5172_cast_fp16_11, var_5232_cast_fp16))[name = tensor("op_5260_cast_fp16")]; + tensor var_5262_equation_0 = const()[name = tensor("op_5262_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5262_cast_fp16 = einsum(equation = var_5262_equation_0, values = (var_5172_cast_fp16_12, var_5233_cast_fp16))[name = tensor("op_5262_cast_fp16")]; + tensor var_5264_equation_0 = const()[name = tensor("op_5264_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5264_cast_fp16 = einsum(equation = var_5264_equation_0, values = (var_5172_cast_fp16_13, var_5234_cast_fp16))[name = tensor("op_5264_cast_fp16")]; + tensor var_5266_equation_0 = const()[name = tensor("op_5266_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5266_cast_fp16 = einsum(equation = var_5266_equation_0, values = (var_5172_cast_fp16_14, var_5235_cast_fp16))[name = tensor("op_5266_cast_fp16")]; + tensor var_5268_equation_0 = const()[name = tensor("op_5268_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5268_cast_fp16 = einsum(equation = var_5268_equation_0, values = (var_5172_cast_fp16_15, var_5236_cast_fp16))[name = tensor("op_5268_cast_fp16")]; + tensor input_215_interleave_0 = const()[name = tensor("input_215_interleave_0"), val = tensor(false)]; + tensor input_215_cast_fp16 = concat(axis = var_5085, interleave = input_215_interleave_0, values = (var_5238_cast_fp16, var_5240_cast_fp16, var_5242_cast_fp16, var_5244_cast_fp16, var_5246_cast_fp16, var_5248_cast_fp16, var_5250_cast_fp16, var_5252_cast_fp16, var_5254_cast_fp16, var_5256_cast_fp16, var_5258_cast_fp16, var_5260_cast_fp16, var_5262_cast_fp16, var_5264_cast_fp16, var_5266_cast_fp16, var_5268_cast_fp16))[name = tensor("input_215_cast_fp16")]; + tensor var_5277_pad_type_0 = const()[name = tensor("op_5277_pad_type_0"), val = tensor("valid")]; + tensor var_5277_strides_0 = const()[name = tensor("op_5277_strides_0"), val = tensor([1, 1])]; + tensor var_5277_pad_0 = const()[name = tensor("op_5277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5277_dilations_0 = const()[name = tensor("op_5277_dilations_0"), val = tensor([1, 1])]; + tensor var_5277_groups_0 = const()[name = tensor("op_5277_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_out_weight_to_fp16 = const()[name = tensor("blocks_21_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545178112)))]; + tensor blocks_21_attn_out_bias_to_fp16 = const()[name = tensor("blocks_21_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547275328)))]; + tensor var_5277_cast_fp16 = conv(bias = blocks_21_attn_out_bias_to_fp16, dilations = var_5277_dilations_0, groups = var_5277_groups_0, pad = var_5277_pad_0, pad_type = var_5277_pad_type_0, strides = var_5277_strides_0, weight = blocks_21_attn_out_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("op_5277_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = var_5277_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor input_217_axes_0 = const()[name = tensor("input_217_axes_0"), val = tensor([1])]; + tensor input_217_gamma_0_to_fp16 = const()[name = tensor("input_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547277440)))]; + tensor input_217_beta_0_to_fp16 = const()[name = tensor("input_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547279552)))]; + tensor var_5287_to_fp16 = const()[name = tensor("op_5287_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_217_cast_fp16 = layer_norm(axes = input_217_axes_0, beta = input_217_beta_0_to_fp16, epsilon = var_5287_to_fp16, gamma = input_217_gamma_0_to_fp16, x = inputs_87_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("valid")]; + tensor input_219_strides_0 = const()[name = tensor("input_219_strides_0"), val = tensor([1, 1])]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_219_dilations_0 = const()[name = tensor("input_219_dilations_0"), val = tensor([1, 1])]; + tensor input_219_groups_0 = const()[name = tensor("input_219_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547281664)))]; + tensor blocks_21_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555670336)))]; + tensor input_219_cast_fp16 = conv(bias = blocks_21_mlp_0_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = blocks_21_mlp_0_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_5313_pad_type_0 = const()[name = tensor("op_5313_pad_type_0"), val = tensor("valid")]; + tensor var_5313_strides_0 = const()[name = tensor("op_5313_strides_0"), val = tensor([1, 1])]; + tensor var_5313_pad_0 = const()[name = tensor("op_5313_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5313_dilations_0 = const()[name = tensor("op_5313_dilations_0"), val = tensor([1, 1])]; + tensor var_5313_groups_0 = const()[name = tensor("op_5313_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555678592)))]; + tensor blocks_21_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564067264)))]; + tensor var_5313_cast_fp16 = conv(bias = blocks_21_mlp_2_bias_to_fp16, dilations = var_5313_dilations_0, groups = var_5313_groups_0, pad = var_5313_pad_0, pad_type = var_5313_pad_type_0, strides = var_5313_strides_0, weight = blocks_21_mlp_2_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("op_5313_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_5313_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_5322 = const()[name = tensor("op_5322"), val = tensor(1)]; + tensor input_223_axes_0 = const()[name = tensor("input_223_axes_0"), val = tensor([1])]; + tensor input_223_gamma_0_to_fp16 = const()[name = tensor("input_223_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564069376)))]; + tensor input_223_beta_0_to_fp16 = const()[name = tensor("input_223_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564071488)))]; + tensor var_5338_to_fp16 = const()[name = tensor("op_5338_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_223_cast_fp16 = layer_norm(axes = input_223_axes_0, beta = input_223_beta_0_to_fp16, epsilon = var_5338_to_fp16, gamma = input_223_gamma_0_to_fp16, x = inputs_89_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor q_45_pad_type_0 = const()[name = tensor("q_45_pad_type_0"), val = tensor("valid")]; + tensor q_45_strides_0 = const()[name = tensor("q_45_strides_0"), val = tensor([1, 1])]; + tensor q_45_pad_0 = const()[name = tensor("q_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_45_dilations_0 = const()[name = tensor("q_45_dilations_0"), val = tensor([1, 1])]; + tensor q_45_groups_0 = const()[name = tensor("q_45_groups_0"), val = tensor(1)]; + tensor var_5373_weight_0_to_fp16 = const()[name = tensor("op_5373_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564073600)))]; + tensor var_5373_bias_0_to_fp16 = const()[name = tensor("op_5373_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566170816)))]; + tensor var_5373_cast_fp16 = conv(bias = var_5373_bias_0_to_fp16, dilations = q_45_dilations_0, groups = q_45_groups_0, pad = q_45_pad_0, pad_type = q_45_pad_type_0, strides = q_45_strides_0, weight = var_5373_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("op_5373_cast_fp16")]; + tensor k_45_pad_type_0 = const()[name = tensor("k_45_pad_type_0"), val = tensor("valid")]; + tensor k_45_strides_0 = const()[name = tensor("k_45_strides_0"), val = tensor([1, 1])]; + tensor k_45_pad_0 = const()[name = tensor("k_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_45_dilations_0 = const()[name = tensor("k_45_dilations_0"), val = tensor([1, 1])]; + tensor k_45_groups_0 = const()[name = tensor("k_45_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_key_weight_to_fp16 = const()[name = tensor("blocks_22_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566172928)))]; + tensor k_45_cast_fp16 = conv(dilations = k_45_dilations_0, groups = k_45_groups_0, pad = k_45_pad_0, pad_type = k_45_pad_type_0, strides = k_45_strides_0, weight = blocks_22_attn_key_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor var_5371_pad_type_0 = const()[name = tensor("op_5371_pad_type_0"), val = tensor("valid")]; + tensor var_5371_strides_0 = const()[name = tensor("op_5371_strides_0"), val = tensor([1, 1])]; + tensor var_5371_pad_0 = const()[name = tensor("op_5371_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5371_dilations_0 = const()[name = tensor("op_5371_dilations_0"), val = tensor([1, 1])]; + tensor var_5371_groups_0 = const()[name = tensor("op_5371_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_value_weight_to_fp16 = const()[name = tensor("blocks_22_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568270144)))]; + tensor blocks_22_attn_value_bias_to_fp16 = const()[name = tensor("blocks_22_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570367360)))]; + tensor var_5371_cast_fp16 = conv(bias = blocks_22_attn_value_bias_to_fp16, dilations = var_5371_dilations_0, groups = var_5371_groups_0, pad = var_5371_pad_0, pad_type = var_5371_pad_type_0, strides = var_5371_strides_0, weight = blocks_22_attn_value_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("op_5371_cast_fp16")]; + tensor tile_66 = const()[name = tensor("tile_66"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_cast_fp16_0, tensor var_5374_cast_fp16_1, tensor var_5374_cast_fp16_2, tensor var_5374_cast_fp16_3, tensor var_5374_cast_fp16_4, tensor var_5374_cast_fp16_5, tensor var_5374_cast_fp16_6, tensor var_5374_cast_fp16_7, tensor var_5374_cast_fp16_8, tensor var_5374_cast_fp16_9, tensor var_5374_cast_fp16_10, tensor var_5374_cast_fp16_11, tensor var_5374_cast_fp16_12, tensor var_5374_cast_fp16_13, tensor var_5374_cast_fp16_14, tensor var_5374_cast_fp16_15 = split(axis = var_5374_axis_0, split_sizes = tile_66, x = var_5373_cast_fp16)[name = tensor("op_5374_cast_fp16")]; + tensor var_5391_perm_0 = const()[name = tensor("op_5391_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_67 = const()[name = tensor("tile_67"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5392_axis_0 = const()[name = tensor("op_5392_axis_0"), val = tensor(3)]; + tensor var_5391_cast_fp16 = transpose(perm = var_5391_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_2")]; + tensor var_5392_cast_fp16_0, tensor var_5392_cast_fp16_1, tensor var_5392_cast_fp16_2, tensor var_5392_cast_fp16_3, tensor var_5392_cast_fp16_4, tensor var_5392_cast_fp16_5, tensor var_5392_cast_fp16_6, tensor var_5392_cast_fp16_7, tensor var_5392_cast_fp16_8, tensor var_5392_cast_fp16_9, tensor var_5392_cast_fp16_10, tensor var_5392_cast_fp16_11, tensor var_5392_cast_fp16_12, tensor var_5392_cast_fp16_13, tensor var_5392_cast_fp16_14, tensor var_5392_cast_fp16_15 = split(axis = var_5392_axis_0, split_sizes = tile_67, x = var_5391_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor tile_68 = const()[name = tensor("tile_68"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5409_axis_0 = const()[name = tensor("op_5409_axis_0"), val = tensor(1)]; + tensor var_5409_cast_fp16_0, tensor var_5409_cast_fp16_1, tensor var_5409_cast_fp16_2, tensor var_5409_cast_fp16_3, tensor var_5409_cast_fp16_4, tensor var_5409_cast_fp16_5, tensor var_5409_cast_fp16_6, tensor var_5409_cast_fp16_7, tensor var_5409_cast_fp16_8, tensor var_5409_cast_fp16_9, tensor var_5409_cast_fp16_10, tensor var_5409_cast_fp16_11, tensor var_5409_cast_fp16_12, tensor var_5409_cast_fp16_13, tensor var_5409_cast_fp16_14, tensor var_5409_cast_fp16_15 = split(axis = var_5409_axis_0, split_sizes = tile_68, x = var_5371_cast_fp16)[name = tensor("op_5409_cast_fp16")]; + tensor aw_705_equation_0 = const()[name = tensor("aw_705_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_705_cast_fp16 = einsum(equation = aw_705_equation_0, values = (var_5392_cast_fp16_0, var_5374_cast_fp16_0))[name = tensor("aw_705_cast_fp16")]; + tensor aw_707_equation_0 = const()[name = tensor("aw_707_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_707_cast_fp16 = einsum(equation = aw_707_equation_0, values = (var_5392_cast_fp16_1, var_5374_cast_fp16_1))[name = tensor("aw_707_cast_fp16")]; + tensor aw_709_equation_0 = const()[name = tensor("aw_709_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_709_cast_fp16 = einsum(equation = aw_709_equation_0, values = (var_5392_cast_fp16_2, var_5374_cast_fp16_2))[name = tensor("aw_709_cast_fp16")]; + tensor aw_711_equation_0 = const()[name = tensor("aw_711_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_711_cast_fp16 = einsum(equation = aw_711_equation_0, values = (var_5392_cast_fp16_3, var_5374_cast_fp16_3))[name = tensor("aw_711_cast_fp16")]; + tensor aw_713_equation_0 = const()[name = tensor("aw_713_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_713_cast_fp16 = einsum(equation = aw_713_equation_0, values = (var_5392_cast_fp16_4, var_5374_cast_fp16_4))[name = tensor("aw_713_cast_fp16")]; + tensor aw_715_equation_0 = const()[name = tensor("aw_715_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_715_cast_fp16 = einsum(equation = aw_715_equation_0, values = (var_5392_cast_fp16_5, var_5374_cast_fp16_5))[name = tensor("aw_715_cast_fp16")]; + tensor aw_717_equation_0 = const()[name = tensor("aw_717_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_717_cast_fp16 = einsum(equation = aw_717_equation_0, values = (var_5392_cast_fp16_6, var_5374_cast_fp16_6))[name = tensor("aw_717_cast_fp16")]; + tensor aw_719_equation_0 = const()[name = tensor("aw_719_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_719_cast_fp16 = einsum(equation = aw_719_equation_0, values = (var_5392_cast_fp16_7, var_5374_cast_fp16_7))[name = tensor("aw_719_cast_fp16")]; + tensor aw_721_equation_0 = const()[name = tensor("aw_721_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_721_cast_fp16 = einsum(equation = aw_721_equation_0, values = (var_5392_cast_fp16_8, var_5374_cast_fp16_8))[name = tensor("aw_721_cast_fp16")]; + tensor aw_723_equation_0 = const()[name = tensor("aw_723_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_723_cast_fp16 = einsum(equation = aw_723_equation_0, values = (var_5392_cast_fp16_9, var_5374_cast_fp16_9))[name = tensor("aw_723_cast_fp16")]; + tensor aw_725_equation_0 = const()[name = tensor("aw_725_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_725_cast_fp16 = einsum(equation = aw_725_equation_0, values = (var_5392_cast_fp16_10, var_5374_cast_fp16_10))[name = tensor("aw_725_cast_fp16")]; + tensor aw_727_equation_0 = const()[name = tensor("aw_727_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_727_cast_fp16 = einsum(equation = aw_727_equation_0, values = (var_5392_cast_fp16_11, var_5374_cast_fp16_11))[name = tensor("aw_727_cast_fp16")]; + tensor aw_729_equation_0 = const()[name = tensor("aw_729_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_729_cast_fp16 = einsum(equation = aw_729_equation_0, values = (var_5392_cast_fp16_12, var_5374_cast_fp16_12))[name = tensor("aw_729_cast_fp16")]; + tensor aw_731_equation_0 = const()[name = tensor("aw_731_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_731_cast_fp16 = einsum(equation = aw_731_equation_0, values = (var_5392_cast_fp16_13, var_5374_cast_fp16_13))[name = tensor("aw_731_cast_fp16")]; + tensor aw_733_equation_0 = const()[name = tensor("aw_733_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_733_cast_fp16 = einsum(equation = aw_733_equation_0, values = (var_5392_cast_fp16_14, var_5374_cast_fp16_14))[name = tensor("aw_733_cast_fp16")]; + tensor aw_735_equation_0 = const()[name = tensor("aw_735_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_735_cast_fp16 = einsum(equation = aw_735_equation_0, values = (var_5392_cast_fp16_15, var_5374_cast_fp16_15))[name = tensor("aw_735_cast_fp16")]; + tensor var_5458_cast_fp16 = softmax(axis = var_5322, x = aw_705_cast_fp16)[name = tensor("op_5458_cast_fp16")]; + tensor var_5459_cast_fp16 = softmax(axis = var_5322, x = aw_707_cast_fp16)[name = tensor("op_5459_cast_fp16")]; + tensor var_5460_cast_fp16 = softmax(axis = var_5322, x = aw_709_cast_fp16)[name = tensor("op_5460_cast_fp16")]; + tensor var_5461_cast_fp16 = softmax(axis = var_5322, x = aw_711_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5462_cast_fp16 = softmax(axis = var_5322, x = aw_713_cast_fp16)[name = tensor("op_5462_cast_fp16")]; + tensor var_5463_cast_fp16 = softmax(axis = var_5322, x = aw_715_cast_fp16)[name = tensor("op_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = softmax(axis = var_5322, x = aw_717_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_cast_fp16 = softmax(axis = var_5322, x = aw_719_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = softmax(axis = var_5322, x = aw_721_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor var_5467_cast_fp16 = softmax(axis = var_5322, x = aw_723_cast_fp16)[name = tensor("op_5467_cast_fp16")]; + tensor var_5468_cast_fp16 = softmax(axis = var_5322, x = aw_725_cast_fp16)[name = tensor("op_5468_cast_fp16")]; + tensor var_5469_cast_fp16 = softmax(axis = var_5322, x = aw_727_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = softmax(axis = var_5322, x = aw_729_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor var_5471_cast_fp16 = softmax(axis = var_5322, x = aw_731_cast_fp16)[name = tensor("op_5471_cast_fp16")]; + tensor var_5472_cast_fp16 = softmax(axis = var_5322, x = aw_733_cast_fp16)[name = tensor("op_5472_cast_fp16")]; + tensor var_5473_cast_fp16 = softmax(axis = var_5322, x = aw_735_cast_fp16)[name = tensor("op_5473_cast_fp16")]; + tensor var_5475_equation_0 = const()[name = tensor("op_5475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5475_cast_fp16 = einsum(equation = var_5475_equation_0, values = (var_5409_cast_fp16_0, var_5458_cast_fp16))[name = tensor("op_5475_cast_fp16")]; + tensor var_5477_equation_0 = const()[name = tensor("op_5477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5477_cast_fp16 = einsum(equation = var_5477_equation_0, values = (var_5409_cast_fp16_1, var_5459_cast_fp16))[name = tensor("op_5477_cast_fp16")]; + tensor var_5479_equation_0 = const()[name = tensor("op_5479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5479_cast_fp16 = einsum(equation = var_5479_equation_0, values = (var_5409_cast_fp16_2, var_5460_cast_fp16))[name = tensor("op_5479_cast_fp16")]; + tensor var_5481_equation_0 = const()[name = tensor("op_5481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5481_cast_fp16 = einsum(equation = var_5481_equation_0, values = (var_5409_cast_fp16_3, var_5461_cast_fp16))[name = tensor("op_5481_cast_fp16")]; + tensor var_5483_equation_0 = const()[name = tensor("op_5483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5483_cast_fp16 = einsum(equation = var_5483_equation_0, values = (var_5409_cast_fp16_4, var_5462_cast_fp16))[name = tensor("op_5483_cast_fp16")]; + tensor var_5485_equation_0 = const()[name = tensor("op_5485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5485_cast_fp16 = einsum(equation = var_5485_equation_0, values = (var_5409_cast_fp16_5, var_5463_cast_fp16))[name = tensor("op_5485_cast_fp16")]; + tensor var_5487_equation_0 = const()[name = tensor("op_5487_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5487_cast_fp16 = einsum(equation = var_5487_equation_0, values = (var_5409_cast_fp16_6, var_5464_cast_fp16))[name = tensor("op_5487_cast_fp16")]; + tensor var_5489_equation_0 = const()[name = tensor("op_5489_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5489_cast_fp16 = einsum(equation = var_5489_equation_0, values = (var_5409_cast_fp16_7, var_5465_cast_fp16))[name = tensor("op_5489_cast_fp16")]; + tensor var_5491_equation_0 = const()[name = tensor("op_5491_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5491_cast_fp16 = einsum(equation = var_5491_equation_0, values = (var_5409_cast_fp16_8, var_5466_cast_fp16))[name = tensor("op_5491_cast_fp16")]; + tensor var_5493_equation_0 = const()[name = tensor("op_5493_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5493_cast_fp16 = einsum(equation = var_5493_equation_0, values = (var_5409_cast_fp16_9, var_5467_cast_fp16))[name = tensor("op_5493_cast_fp16")]; + tensor var_5495_equation_0 = const()[name = tensor("op_5495_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5495_cast_fp16 = einsum(equation = var_5495_equation_0, values = (var_5409_cast_fp16_10, var_5468_cast_fp16))[name = tensor("op_5495_cast_fp16")]; + tensor var_5497_equation_0 = const()[name = tensor("op_5497_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5497_cast_fp16 = einsum(equation = var_5497_equation_0, values = (var_5409_cast_fp16_11, var_5469_cast_fp16))[name = tensor("op_5497_cast_fp16")]; + tensor var_5499_equation_0 = const()[name = tensor("op_5499_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5499_cast_fp16 = einsum(equation = var_5499_equation_0, values = (var_5409_cast_fp16_12, var_5470_cast_fp16))[name = tensor("op_5499_cast_fp16")]; + tensor var_5501_equation_0 = const()[name = tensor("op_5501_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5501_cast_fp16 = einsum(equation = var_5501_equation_0, values = (var_5409_cast_fp16_13, var_5471_cast_fp16))[name = tensor("op_5501_cast_fp16")]; + tensor var_5503_equation_0 = const()[name = tensor("op_5503_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5503_cast_fp16 = einsum(equation = var_5503_equation_0, values = (var_5409_cast_fp16_14, var_5472_cast_fp16))[name = tensor("op_5503_cast_fp16")]; + tensor var_5505_equation_0 = const()[name = tensor("op_5505_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5505_cast_fp16 = einsum(equation = var_5505_equation_0, values = (var_5409_cast_fp16_15, var_5473_cast_fp16))[name = tensor("op_5505_cast_fp16")]; + tensor input_225_interleave_0 = const()[name = tensor("input_225_interleave_0"), val = tensor(false)]; + tensor input_225_cast_fp16 = concat(axis = var_5322, interleave = input_225_interleave_0, values = (var_5475_cast_fp16, var_5477_cast_fp16, var_5479_cast_fp16, var_5481_cast_fp16, var_5483_cast_fp16, var_5485_cast_fp16, var_5487_cast_fp16, var_5489_cast_fp16, var_5491_cast_fp16, var_5493_cast_fp16, var_5495_cast_fp16, var_5497_cast_fp16, var_5499_cast_fp16, var_5501_cast_fp16, var_5503_cast_fp16, var_5505_cast_fp16))[name = tensor("input_225_cast_fp16")]; + tensor var_5514_pad_type_0 = const()[name = tensor("op_5514_pad_type_0"), val = tensor("valid")]; + tensor var_5514_strides_0 = const()[name = tensor("op_5514_strides_0"), val = tensor([1, 1])]; + tensor var_5514_pad_0 = const()[name = tensor("op_5514_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5514_dilations_0 = const()[name = tensor("op_5514_dilations_0"), val = tensor([1, 1])]; + tensor var_5514_groups_0 = const()[name = tensor("op_5514_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_out_weight_to_fp16 = const()[name = tensor("blocks_22_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570369472)))]; + tensor blocks_22_attn_out_bias_to_fp16 = const()[name = tensor("blocks_22_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572466688)))]; + tensor var_5514_cast_fp16 = conv(bias = blocks_22_attn_out_bias_to_fp16, dilations = var_5514_dilations_0, groups = var_5514_groups_0, pad = var_5514_pad_0, pad_type = var_5514_pad_type_0, strides = var_5514_strides_0, weight = blocks_22_attn_out_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = var_5514_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([1])]; + tensor input_227_gamma_0_to_fp16 = const()[name = tensor("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572468800)))]; + tensor input_227_beta_0_to_fp16 = const()[name = tensor("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572470912)))]; + tensor var_5524_to_fp16 = const()[name = tensor("op_5524_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = input_227_beta_0_to_fp16, epsilon = var_5524_to_fp16, gamma = input_227_gamma_0_to_fp16, x = inputs_91_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("valid")]; + tensor input_229_strides_0 = const()[name = tensor("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = tensor("input_229_dilations_0"), val = tensor([1, 1])]; + tensor input_229_groups_0 = const()[name = tensor("input_229_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572473024)))]; + tensor blocks_22_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580861696)))]; + tensor input_229_cast_fp16 = conv(bias = blocks_22_mlp_0_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = blocks_22_mlp_0_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_5550_pad_type_0 = const()[name = tensor("op_5550_pad_type_0"), val = tensor("valid")]; + tensor var_5550_strides_0 = const()[name = tensor("op_5550_strides_0"), val = tensor([1, 1])]; + tensor var_5550_pad_0 = const()[name = tensor("op_5550_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5550_dilations_0 = const()[name = tensor("op_5550_dilations_0"), val = tensor([1, 1])]; + tensor var_5550_groups_0 = const()[name = tensor("op_5550_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580869952)))]; + tensor blocks_22_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589258624)))]; + tensor var_5550_cast_fp16 = conv(bias = blocks_22_mlp_2_bias_to_fp16, dilations = var_5550_dilations_0, groups = var_5550_groups_0, pad = var_5550_pad_0, pad_type = var_5550_pad_type_0, strides = var_5550_strides_0, weight = blocks_22_mlp_2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("op_5550_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_5550_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_5559 = const()[name = tensor("op_5559"), val = tensor(1)]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([1])]; + tensor input_233_gamma_0_to_fp16 = const()[name = tensor("input_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589260736)))]; + tensor input_233_beta_0_to_fp16 = const()[name = tensor("input_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589262848)))]; + tensor var_5575_to_fp16 = const()[name = tensor("op_5575_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = input_233_beta_0_to_fp16, epsilon = var_5575_to_fp16, gamma = input_233_gamma_0_to_fp16, x = inputs_93_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_5610_weight_0_to_fp16 = const()[name = tensor("op_5610_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589264960)))]; + tensor var_5610_bias_0_to_fp16 = const()[name = tensor("op_5610_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591362176)))]; + tensor var_5610_cast_fp16 = conv(bias = var_5610_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_5610_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("op_5610_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_key_weight_to_fp16 = const()[name = tensor("blocks_23_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591364288)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_23_attn_key_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_5608_pad_type_0 = const()[name = tensor("op_5608_pad_type_0"), val = tensor("valid")]; + tensor var_5608_strides_0 = const()[name = tensor("op_5608_strides_0"), val = tensor([1, 1])]; + tensor var_5608_pad_0 = const()[name = tensor("op_5608_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5608_dilations_0 = const()[name = tensor("op_5608_dilations_0"), val = tensor([1, 1])]; + tensor var_5608_groups_0 = const()[name = tensor("op_5608_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_value_weight_to_fp16 = const()[name = tensor("blocks_23_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593461504)))]; + tensor blocks_23_attn_value_bias_to_fp16 = const()[name = tensor("blocks_23_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595558720)))]; + tensor var_5608_cast_fp16 = conv(bias = blocks_23_attn_value_bias_to_fp16, dilations = var_5608_dilations_0, groups = var_5608_groups_0, pad = var_5608_pad_0, pad_type = var_5608_pad_type_0, strides = var_5608_strides_0, weight = blocks_23_attn_value_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor tile_69 = const()[name = tensor("tile_69"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5611_axis_0 = const()[name = tensor("op_5611_axis_0"), val = tensor(1)]; + tensor var_5611_cast_fp16_0, tensor var_5611_cast_fp16_1, tensor var_5611_cast_fp16_2, tensor var_5611_cast_fp16_3, tensor var_5611_cast_fp16_4, tensor var_5611_cast_fp16_5, tensor var_5611_cast_fp16_6, tensor var_5611_cast_fp16_7, tensor var_5611_cast_fp16_8, tensor var_5611_cast_fp16_9, tensor var_5611_cast_fp16_10, tensor var_5611_cast_fp16_11, tensor var_5611_cast_fp16_12, tensor var_5611_cast_fp16_13, tensor var_5611_cast_fp16_14, tensor var_5611_cast_fp16_15 = split(axis = var_5611_axis_0, split_sizes = tile_69, x = var_5610_cast_fp16)[name = tensor("op_5611_cast_fp16")]; + tensor var_5628_perm_0 = const()[name = tensor("op_5628_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_70 = const()[name = tensor("tile_70"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5629_axis_0 = const()[name = tensor("op_5629_axis_0"), val = tensor(3)]; + tensor var_5628_cast_fp16 = transpose(perm = var_5628_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_5629_cast_fp16_0, tensor var_5629_cast_fp16_1, tensor var_5629_cast_fp16_2, tensor var_5629_cast_fp16_3, tensor var_5629_cast_fp16_4, tensor var_5629_cast_fp16_5, tensor var_5629_cast_fp16_6, tensor var_5629_cast_fp16_7, tensor var_5629_cast_fp16_8, tensor var_5629_cast_fp16_9, tensor var_5629_cast_fp16_10, tensor var_5629_cast_fp16_11, tensor var_5629_cast_fp16_12, tensor var_5629_cast_fp16_13, tensor var_5629_cast_fp16_14, tensor var_5629_cast_fp16_15 = split(axis = var_5629_axis_0, split_sizes = tile_70, x = var_5628_cast_fp16)[name = tensor("op_5629_cast_fp16")]; + tensor tile_71 = const()[name = tensor("tile_71"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5646_axis_0 = const()[name = tensor("op_5646_axis_0"), val = tensor(1)]; + tensor var_5646_cast_fp16_0, tensor var_5646_cast_fp16_1, tensor var_5646_cast_fp16_2, tensor var_5646_cast_fp16_3, tensor var_5646_cast_fp16_4, tensor var_5646_cast_fp16_5, tensor var_5646_cast_fp16_6, tensor var_5646_cast_fp16_7, tensor var_5646_cast_fp16_8, tensor var_5646_cast_fp16_9, tensor var_5646_cast_fp16_10, tensor var_5646_cast_fp16_11, tensor var_5646_cast_fp16_12, tensor var_5646_cast_fp16_13, tensor var_5646_cast_fp16_14, tensor var_5646_cast_fp16_15 = split(axis = var_5646_axis_0, split_sizes = tile_71, x = var_5608_cast_fp16)[name = tensor("op_5646_cast_fp16")]; + tensor aw_737_equation_0 = const()[name = tensor("aw_737_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_737_cast_fp16 = einsum(equation = aw_737_equation_0, values = (var_5629_cast_fp16_0, var_5611_cast_fp16_0))[name = tensor("aw_737_cast_fp16")]; + tensor aw_739_equation_0 = const()[name = tensor("aw_739_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_739_cast_fp16 = einsum(equation = aw_739_equation_0, values = (var_5629_cast_fp16_1, var_5611_cast_fp16_1))[name = tensor("aw_739_cast_fp16")]; + tensor aw_741_equation_0 = const()[name = tensor("aw_741_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_741_cast_fp16 = einsum(equation = aw_741_equation_0, values = (var_5629_cast_fp16_2, var_5611_cast_fp16_2))[name = tensor("aw_741_cast_fp16")]; + tensor aw_743_equation_0 = const()[name = tensor("aw_743_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_743_cast_fp16 = einsum(equation = aw_743_equation_0, values = (var_5629_cast_fp16_3, var_5611_cast_fp16_3))[name = tensor("aw_743_cast_fp16")]; + tensor aw_745_equation_0 = const()[name = tensor("aw_745_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_745_cast_fp16 = einsum(equation = aw_745_equation_0, values = (var_5629_cast_fp16_4, var_5611_cast_fp16_4))[name = tensor("aw_745_cast_fp16")]; + tensor aw_747_equation_0 = const()[name = tensor("aw_747_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_747_cast_fp16 = einsum(equation = aw_747_equation_0, values = (var_5629_cast_fp16_5, var_5611_cast_fp16_5))[name = tensor("aw_747_cast_fp16")]; + tensor aw_749_equation_0 = const()[name = tensor("aw_749_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_749_cast_fp16 = einsum(equation = aw_749_equation_0, values = (var_5629_cast_fp16_6, var_5611_cast_fp16_6))[name = tensor("aw_749_cast_fp16")]; + tensor aw_751_equation_0 = const()[name = tensor("aw_751_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_751_cast_fp16 = einsum(equation = aw_751_equation_0, values = (var_5629_cast_fp16_7, var_5611_cast_fp16_7))[name = tensor("aw_751_cast_fp16")]; + tensor aw_753_equation_0 = const()[name = tensor("aw_753_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_753_cast_fp16 = einsum(equation = aw_753_equation_0, values = (var_5629_cast_fp16_8, var_5611_cast_fp16_8))[name = tensor("aw_753_cast_fp16")]; + tensor aw_755_equation_0 = const()[name = tensor("aw_755_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_755_cast_fp16 = einsum(equation = aw_755_equation_0, values = (var_5629_cast_fp16_9, var_5611_cast_fp16_9))[name = tensor("aw_755_cast_fp16")]; + tensor aw_757_equation_0 = const()[name = tensor("aw_757_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_757_cast_fp16 = einsum(equation = aw_757_equation_0, values = (var_5629_cast_fp16_10, var_5611_cast_fp16_10))[name = tensor("aw_757_cast_fp16")]; + tensor aw_759_equation_0 = const()[name = tensor("aw_759_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_759_cast_fp16 = einsum(equation = aw_759_equation_0, values = (var_5629_cast_fp16_11, var_5611_cast_fp16_11))[name = tensor("aw_759_cast_fp16")]; + tensor aw_761_equation_0 = const()[name = tensor("aw_761_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_761_cast_fp16 = einsum(equation = aw_761_equation_0, values = (var_5629_cast_fp16_12, var_5611_cast_fp16_12))[name = tensor("aw_761_cast_fp16")]; + tensor aw_763_equation_0 = const()[name = tensor("aw_763_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_763_cast_fp16 = einsum(equation = aw_763_equation_0, values = (var_5629_cast_fp16_13, var_5611_cast_fp16_13))[name = tensor("aw_763_cast_fp16")]; + tensor aw_765_equation_0 = const()[name = tensor("aw_765_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_765_cast_fp16 = einsum(equation = aw_765_equation_0, values = (var_5629_cast_fp16_14, var_5611_cast_fp16_14))[name = tensor("aw_765_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_5629_cast_fp16_15, var_5611_cast_fp16_15))[name = tensor("aw_cast_fp16")]; + tensor var_5695_cast_fp16 = softmax(axis = var_5559, x = aw_737_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor var_5696_cast_fp16 = softmax(axis = var_5559, x = aw_739_cast_fp16)[name = tensor("op_5696_cast_fp16")]; + tensor var_5697_cast_fp16 = softmax(axis = var_5559, x = aw_741_cast_fp16)[name = tensor("op_5697_cast_fp16")]; + tensor var_5698_cast_fp16 = softmax(axis = var_5559, x = aw_743_cast_fp16)[name = tensor("op_5698_cast_fp16")]; + tensor var_5699_cast_fp16 = softmax(axis = var_5559, x = aw_745_cast_fp16)[name = tensor("op_5699_cast_fp16")]; + tensor var_5700_cast_fp16 = softmax(axis = var_5559, x = aw_747_cast_fp16)[name = tensor("op_5700_cast_fp16")]; + tensor var_5701_cast_fp16 = softmax(axis = var_5559, x = aw_749_cast_fp16)[name = tensor("op_5701_cast_fp16")]; + tensor var_5702_cast_fp16 = softmax(axis = var_5559, x = aw_751_cast_fp16)[name = tensor("op_5702_cast_fp16")]; + tensor var_5703_cast_fp16 = softmax(axis = var_5559, x = aw_753_cast_fp16)[name = tensor("op_5703_cast_fp16")]; + tensor var_5704_cast_fp16 = softmax(axis = var_5559, x = aw_755_cast_fp16)[name = tensor("op_5704_cast_fp16")]; + tensor var_5705_cast_fp16 = softmax(axis = var_5559, x = aw_757_cast_fp16)[name = tensor("op_5705_cast_fp16")]; + tensor var_5706_cast_fp16 = softmax(axis = var_5559, x = aw_759_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5707_cast_fp16 = softmax(axis = var_5559, x = aw_761_cast_fp16)[name = tensor("op_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = softmax(axis = var_5559, x = aw_763_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_cast_fp16 = softmax(axis = var_5559, x = aw_765_cast_fp16)[name = tensor("op_5709_cast_fp16")]; + tensor var_5710_cast_fp16 = softmax(axis = var_5559, x = aw_cast_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor var_5712_equation_0 = const()[name = tensor("op_5712_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5712_cast_fp16 = einsum(equation = var_5712_equation_0, values = (var_5646_cast_fp16_0, var_5695_cast_fp16))[name = tensor("op_5712_cast_fp16")]; + tensor var_5714_equation_0 = const()[name = tensor("op_5714_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5714_cast_fp16 = einsum(equation = var_5714_equation_0, values = (var_5646_cast_fp16_1, var_5696_cast_fp16))[name = tensor("op_5714_cast_fp16")]; + tensor var_5716_equation_0 = const()[name = tensor("op_5716_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5716_cast_fp16 = einsum(equation = var_5716_equation_0, values = (var_5646_cast_fp16_2, var_5697_cast_fp16))[name = tensor("op_5716_cast_fp16")]; + tensor var_5718_equation_0 = const()[name = tensor("op_5718_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5718_cast_fp16 = einsum(equation = var_5718_equation_0, values = (var_5646_cast_fp16_3, var_5698_cast_fp16))[name = tensor("op_5718_cast_fp16")]; + tensor var_5720_equation_0 = const()[name = tensor("op_5720_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5720_cast_fp16 = einsum(equation = var_5720_equation_0, values = (var_5646_cast_fp16_4, var_5699_cast_fp16))[name = tensor("op_5720_cast_fp16")]; + tensor var_5722_equation_0 = const()[name = tensor("op_5722_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5722_cast_fp16 = einsum(equation = var_5722_equation_0, values = (var_5646_cast_fp16_5, var_5700_cast_fp16))[name = tensor("op_5722_cast_fp16")]; + tensor var_5724_equation_0 = const()[name = tensor("op_5724_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5724_cast_fp16 = einsum(equation = var_5724_equation_0, values = (var_5646_cast_fp16_6, var_5701_cast_fp16))[name = tensor("op_5724_cast_fp16")]; + tensor var_5726_equation_0 = const()[name = tensor("op_5726_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5726_cast_fp16 = einsum(equation = var_5726_equation_0, values = (var_5646_cast_fp16_7, var_5702_cast_fp16))[name = tensor("op_5726_cast_fp16")]; + tensor var_5728_equation_0 = const()[name = tensor("op_5728_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5728_cast_fp16 = einsum(equation = var_5728_equation_0, values = (var_5646_cast_fp16_8, var_5703_cast_fp16))[name = tensor("op_5728_cast_fp16")]; + tensor var_5730_equation_0 = const()[name = tensor("op_5730_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5730_cast_fp16 = einsum(equation = var_5730_equation_0, values = (var_5646_cast_fp16_9, var_5704_cast_fp16))[name = tensor("op_5730_cast_fp16")]; + tensor var_5732_equation_0 = const()[name = tensor("op_5732_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5732_cast_fp16 = einsum(equation = var_5732_equation_0, values = (var_5646_cast_fp16_10, var_5705_cast_fp16))[name = tensor("op_5732_cast_fp16")]; + tensor var_5734_equation_0 = const()[name = tensor("op_5734_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5734_cast_fp16 = einsum(equation = var_5734_equation_0, values = (var_5646_cast_fp16_11, var_5706_cast_fp16))[name = tensor("op_5734_cast_fp16")]; + tensor var_5736_equation_0 = const()[name = tensor("op_5736_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5736_cast_fp16 = einsum(equation = var_5736_equation_0, values = (var_5646_cast_fp16_12, var_5707_cast_fp16))[name = tensor("op_5736_cast_fp16")]; + tensor var_5738_equation_0 = const()[name = tensor("op_5738_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5738_cast_fp16 = einsum(equation = var_5738_equation_0, values = (var_5646_cast_fp16_13, var_5708_cast_fp16))[name = tensor("op_5738_cast_fp16")]; + tensor var_5740_equation_0 = const()[name = tensor("op_5740_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5740_cast_fp16 = einsum(equation = var_5740_equation_0, values = (var_5646_cast_fp16_14, var_5709_cast_fp16))[name = tensor("op_5740_cast_fp16")]; + tensor var_5742_equation_0 = const()[name = tensor("op_5742_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5742_cast_fp16 = einsum(equation = var_5742_equation_0, values = (var_5646_cast_fp16_15, var_5710_cast_fp16))[name = tensor("op_5742_cast_fp16")]; + tensor input_235_interleave_0 = const()[name = tensor("input_235_interleave_0"), val = tensor(false)]; + tensor input_235_cast_fp16 = concat(axis = var_5559, interleave = input_235_interleave_0, values = (var_5712_cast_fp16, var_5714_cast_fp16, var_5716_cast_fp16, var_5718_cast_fp16, var_5720_cast_fp16, var_5722_cast_fp16, var_5724_cast_fp16, var_5726_cast_fp16, var_5728_cast_fp16, var_5730_cast_fp16, var_5732_cast_fp16, var_5734_cast_fp16, var_5736_cast_fp16, var_5738_cast_fp16, var_5740_cast_fp16, var_5742_cast_fp16))[name = tensor("input_235_cast_fp16")]; + tensor var_5751_pad_type_0 = const()[name = tensor("op_5751_pad_type_0"), val = tensor("valid")]; + tensor var_5751_strides_0 = const()[name = tensor("op_5751_strides_0"), val = tensor([1, 1])]; + tensor var_5751_pad_0 = const()[name = tensor("op_5751_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5751_dilations_0 = const()[name = tensor("op_5751_dilations_0"), val = tensor([1, 1])]; + tensor var_5751_groups_0 = const()[name = tensor("op_5751_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_out_weight_to_fp16 = const()[name = tensor("blocks_23_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595560832)))]; + tensor blocks_23_attn_out_bias_to_fp16 = const()[name = tensor("blocks_23_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597658048)))]; + tensor var_5751_cast_fp16 = conv(bias = blocks_23_attn_out_bias_to_fp16, dilations = var_5751_dilations_0, groups = var_5751_groups_0, pad = var_5751_pad_0, pad_type = var_5751_pad_type_0, strides = var_5751_strides_0, weight = blocks_23_attn_out_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("op_5751_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = var_5751_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([1])]; + tensor input_237_gamma_0_to_fp16 = const()[name = tensor("input_237_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597660160)))]; + tensor input_237_beta_0_to_fp16 = const()[name = tensor("input_237_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597662272)))]; + tensor var_5761_to_fp16 = const()[name = tensor("op_5761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = input_237_beta_0_to_fp16, epsilon = var_5761_to_fp16, gamma = input_237_gamma_0_to_fp16, x = inputs_95_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("valid")]; + tensor input_239_strides_0 = const()[name = tensor("input_239_strides_0"), val = tensor([1, 1])]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_239_dilations_0 = const()[name = tensor("input_239_dilations_0"), val = tensor([1, 1])]; + tensor input_239_groups_0 = const()[name = tensor("input_239_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597664384)))]; + tensor blocks_23_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606053056)))]; + tensor input_239_cast_fp16 = conv(bias = blocks_23_mlp_0_bias_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = blocks_23_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_239_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_5787_pad_type_0 = const()[name = tensor("op_5787_pad_type_0"), val = tensor("valid")]; + tensor var_5787_strides_0 = const()[name = tensor("op_5787_strides_0"), val = tensor([1, 1])]; + tensor var_5787_pad_0 = const()[name = tensor("op_5787_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5787_dilations_0 = const()[name = tensor("op_5787_dilations_0"), val = tensor([1, 1])]; + tensor var_5787_groups_0 = const()[name = tensor("op_5787_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606061312)))]; + tensor blocks_23_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614449984)))]; + tensor var_5787_cast_fp16 = conv(bias = blocks_23_mlp_2_bias_to_fp16, dilations = var_5787_dilations_0, groups = var_5787_groups_0, pad = var_5787_pad_0, pad_type = var_5787_pad_type_0, strides = var_5787_strides_0, weight = blocks_23_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_95_cast_fp16, y = var_5787_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614452096)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614454208)))]; + tensor var_5801_to_fp16 = const()[name = tensor("op_5801_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_5801_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_5812_axes_0 = const()[name = tensor("op_5812_axes_0"), val = tensor([2])]; + tensor var_5812_cast_fp16 = squeeze(axes = var_5812_axes_0, x = x_cast_fp16)[name = tensor("op_5812_cast_fp16")]; + tensor var_5815_perm_0 = const()[name = tensor("op_5815_perm_0"), val = tensor([0, 2, 1])]; + tensor var_5815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5815_cast_fp16 = transpose(perm = var_5815_perm_0, x = var_5812_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_5815_cast_fp16_to_fp32_dtype_0, x = var_5815_cast_fp16)[name = tensor("cast_99")]; + } -> (output); +} \ No newline at end of file diff --git a/medium.en/ggml-medium.en-encoder.mlmodelc/weights/weight.bin b/medium.en/ggml-medium.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a1ff84b198cc02ec2b63b8ee094dce3f239565c5 --- /dev/null +++ b/medium.en/ggml-medium.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74029c2746f01f4f43ebbcdaa98dfb6e0597eb7e3d4008fdef72c2113a4d2483 +size 614456320 diff --git a/medium.en/ggml-medium.en.bin b/medium.en/ggml-medium.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..f8d7f988b60916d7f7e7feee9897c037a09b2f85 --- /dev/null +++ b/medium.en/ggml-medium.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc37e93478338ec7700281a7ac30a10128929eb8f427dda2e865faa8f6da4356 +size 1533774781 diff --git a/medium/.DS_Store b/medium/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..79ccbf9bfb80488f459329a093e7ac4031bea9cc Binary files /dev/null and b/medium/.DS_Store differ diff --git a/medium/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin b/medium/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..328e5c6707519542579df61365ef4c4391c16353 --- /dev/null +++ b/medium/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:423209035d398f1e6945478cc49cf8e83022a3dc4e00d735c54289b3f62cdf89 +size 243 diff --git a/medium/ggml-medium-encoder.mlmodelc/coremldata.bin b/medium/ggml-medium-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a7e2b54475c4dbc86435a318e0608d48945bae16 --- /dev/null +++ b/medium/ggml-medium-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d20a4fe17a031efc213c5c295df6967c6e87eba9cca3f07fa63c2beb835ca420 +size 320 diff --git a/medium/ggml-medium-encoder.mlmodelc/metadata.json b/medium/ggml-medium-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..a4281c486a43b6865cadcf5135988a026525daaf --- /dev/null +++ b/medium/ggml-medium-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1500, 1024]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 24, + "Gelu" : 26, + "LayerNorm" : 49, + "Transpose" : 25, + "Softmax" : 384, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 49, + "Einsum" : 768, + "ExpandDims" : 1, + "Split" : 72, + "Conv" : 146 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_medium", + "method" : "predict" + } +] \ No newline at end of file diff --git a/medium/ggml-medium-encoder.mlmodelc/model.mil b/medium/ggml-medium-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..e6dfb1231cc22b9b4af2039b332f526ac8800584 --- /dev/null +++ b/medium/ggml-medium-encoder.mlmodelc/model.mil @@ -0,0 +1,3763 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_68_pad_type_0 = const()[name = tensor("op_68_pad_type_0"), val = tensor("custom")]; + tensor var_68_pad_0 = const()[name = tensor("op_68_pad_0"), val = tensor([1, 1])]; + tensor var_68_strides_0 = const()[name = tensor("op_68_strides_0"), val = tensor([1])]; + tensor var_68_dilations_0 = const()[name = tensor("op_68_dilations_0"), val = tensor([1])]; + tensor var_68_groups_0 = const()[name = tensor("op_68_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491648)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_100")]; + tensor var_68_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_68_dilations_0, groups = var_68_groups_0, pad = var_68_pad_0, pad_type = var_68_pad_type_0, strides = var_68_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_68_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_68_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_86_pad_type_0 = const()[name = tensor("op_86_pad_type_0"), val = tensor("custom")]; + tensor var_86_pad_0 = const()[name = tensor("op_86_pad_0"), val = tensor([1, 1])]; + tensor var_86_strides_0 = const()[name = tensor("op_86_strides_0"), val = tensor([2])]; + tensor var_86_dilations_0 = const()[name = tensor("op_86_dilations_0"), val = tensor([1])]; + tensor var_86_groups_0 = const()[name = tensor("op_86_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493760)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6785280)))]; + tensor var_86_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_86_dilations_0, groups = var_86_groups_0, pad = var_86_pad_0, pad_type = var_86_pad_type_0, strides = var_86_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_86_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_86_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_91_to_fp16 = const()[name = tensor("op_91_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6787392)))]; + tensor var_93_cast_fp16 = add(x = x_3_cast_fp16, y = var_91_to_fp16)[name = tensor("op_93_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_93_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_108 = const()[name = tensor("op_108"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9859456)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9861568)))]; + tensor var_124_to_fp16 = const()[name = tensor("op_124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_124_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_159_weight_0_to_fp16 = const()[name = tensor("op_159_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9863680)))]; + tensor var_159_bias_0_to_fp16 = const()[name = tensor("op_159_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11960896)))]; + tensor var_159_cast_fp16 = conv(bias = var_159_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_159_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_159_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11963008)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_157_pad_type_0 = const()[name = tensor("op_157_pad_type_0"), val = tensor("valid")]; + tensor var_157_strides_0 = const()[name = tensor("op_157_strides_0"), val = tensor([1, 1])]; + tensor var_157_pad_0 = const()[name = tensor("op_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_157_dilations_0 = const()[name = tensor("op_157_dilations_0"), val = tensor([1, 1])]; + tensor var_157_groups_0 = const()[name = tensor("op_157_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14060224)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16157440)))]; + tensor var_157_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_157_dilations_0, groups = var_157_groups_0, pad = var_157_pad_0, pad_type = var_157_pad_type_0, strides = var_157_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_157_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_160_axis_0 = const()[name = tensor("op_160_axis_0"), val = tensor(1)]; + tensor var_160_cast_fp16_0, tensor var_160_cast_fp16_1, tensor var_160_cast_fp16_2, tensor var_160_cast_fp16_3, tensor var_160_cast_fp16_4, tensor var_160_cast_fp16_5, tensor var_160_cast_fp16_6, tensor var_160_cast_fp16_7, tensor var_160_cast_fp16_8, tensor var_160_cast_fp16_9, tensor var_160_cast_fp16_10, tensor var_160_cast_fp16_11, tensor var_160_cast_fp16_12, tensor var_160_cast_fp16_13, tensor var_160_cast_fp16_14, tensor var_160_cast_fp16_15 = split(axis = var_160_axis_0, split_sizes = tile_0, x = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_177_perm_0 = const()[name = tensor("op_177_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_178_axis_0 = const()[name = tensor("op_178_axis_0"), val = tensor(3)]; + tensor var_177_cast_fp16 = transpose(perm = var_177_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_24")]; + tensor var_178_cast_fp16_0, tensor var_178_cast_fp16_1, tensor var_178_cast_fp16_2, tensor var_178_cast_fp16_3, tensor var_178_cast_fp16_4, tensor var_178_cast_fp16_5, tensor var_178_cast_fp16_6, tensor var_178_cast_fp16_7, tensor var_178_cast_fp16_8, tensor var_178_cast_fp16_9, tensor var_178_cast_fp16_10, tensor var_178_cast_fp16_11, tensor var_178_cast_fp16_12, tensor var_178_cast_fp16_13, tensor var_178_cast_fp16_14, tensor var_178_cast_fp16_15 = split(axis = var_178_axis_0, split_sizes = tile_1, x = var_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_195_axis_0 = const()[name = tensor("op_195_axis_0"), val = tensor(1)]; + tensor var_195_cast_fp16_0, tensor var_195_cast_fp16_1, tensor var_195_cast_fp16_2, tensor var_195_cast_fp16_3, tensor var_195_cast_fp16_4, tensor var_195_cast_fp16_5, tensor var_195_cast_fp16_6, tensor var_195_cast_fp16_7, tensor var_195_cast_fp16_8, tensor var_195_cast_fp16_9, tensor var_195_cast_fp16_10, tensor var_195_cast_fp16_11, tensor var_195_cast_fp16_12, tensor var_195_cast_fp16_13, tensor var_195_cast_fp16_14, tensor var_195_cast_fp16_15 = split(axis = var_195_axis_0, split_sizes = tile_2, x = var_157_cast_fp16)[name = tensor("op_195_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_178_cast_fp16_0, var_160_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_178_cast_fp16_1, var_160_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_178_cast_fp16_2, var_160_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_178_cast_fp16_3, var_160_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_178_cast_fp16_4, var_160_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_178_cast_fp16_5, var_160_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_178_cast_fp16_6, var_160_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_178_cast_fp16_7, var_160_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_178_cast_fp16_8, var_160_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_178_cast_fp16_9, var_160_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_178_cast_fp16_10, var_160_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_178_cast_fp16_11, var_160_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_178_cast_fp16_12, var_160_cast_fp16_12))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_178_cast_fp16_13, var_160_cast_fp16_13))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_178_cast_fp16_14, var_160_cast_fp16_14))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_178_cast_fp16_15, var_160_cast_fp16_15))[name = tensor("aw_31_cast_fp16")]; + tensor var_244_cast_fp16 = softmax(axis = var_108, x = aw_1_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor var_245_cast_fp16 = softmax(axis = var_108, x = aw_3_cast_fp16)[name = tensor("op_245_cast_fp16")]; + tensor var_246_cast_fp16 = softmax(axis = var_108, x = aw_5_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = softmax(axis = var_108, x = aw_7_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor var_248_cast_fp16 = softmax(axis = var_108, x = aw_9_cast_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_249_cast_fp16 = softmax(axis = var_108, x = aw_11_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor var_250_cast_fp16 = softmax(axis = var_108, x = aw_13_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = softmax(axis = var_108, x = aw_15_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor var_252_cast_fp16 = softmax(axis = var_108, x = aw_17_cast_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_cast_fp16 = softmax(axis = var_108, x = aw_19_cast_fp16)[name = tensor("op_253_cast_fp16")]; + tensor var_254_cast_fp16 = softmax(axis = var_108, x = aw_21_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = softmax(axis = var_108, x = aw_23_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_cast_fp16 = softmax(axis = var_108, x = aw_25_cast_fp16)[name = tensor("op_256_cast_fp16")]; + tensor var_257_cast_fp16 = softmax(axis = var_108, x = aw_27_cast_fp16)[name = tensor("op_257_cast_fp16")]; + tensor var_258_cast_fp16 = softmax(axis = var_108, x = aw_29_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = softmax(axis = var_108, x = aw_31_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor var_261_equation_0 = const()[name = tensor("op_261_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_261_cast_fp16 = einsum(equation = var_261_equation_0, values = (var_195_cast_fp16_0, var_244_cast_fp16))[name = tensor("op_261_cast_fp16")]; + tensor var_263_equation_0 = const()[name = tensor("op_263_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_263_cast_fp16 = einsum(equation = var_263_equation_0, values = (var_195_cast_fp16_1, var_245_cast_fp16))[name = tensor("op_263_cast_fp16")]; + tensor var_265_equation_0 = const()[name = tensor("op_265_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_265_cast_fp16 = einsum(equation = var_265_equation_0, values = (var_195_cast_fp16_2, var_246_cast_fp16))[name = tensor("op_265_cast_fp16")]; + tensor var_267_equation_0 = const()[name = tensor("op_267_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_267_cast_fp16 = einsum(equation = var_267_equation_0, values = (var_195_cast_fp16_3, var_247_cast_fp16))[name = tensor("op_267_cast_fp16")]; + tensor var_269_equation_0 = const()[name = tensor("op_269_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_269_cast_fp16 = einsum(equation = var_269_equation_0, values = (var_195_cast_fp16_4, var_248_cast_fp16))[name = tensor("op_269_cast_fp16")]; + tensor var_271_equation_0 = const()[name = tensor("op_271_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_271_cast_fp16 = einsum(equation = var_271_equation_0, values = (var_195_cast_fp16_5, var_249_cast_fp16))[name = tensor("op_271_cast_fp16")]; + tensor var_273_equation_0 = const()[name = tensor("op_273_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_273_cast_fp16 = einsum(equation = var_273_equation_0, values = (var_195_cast_fp16_6, var_250_cast_fp16))[name = tensor("op_273_cast_fp16")]; + tensor var_275_equation_0 = const()[name = tensor("op_275_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_275_cast_fp16 = einsum(equation = var_275_equation_0, values = (var_195_cast_fp16_7, var_251_cast_fp16))[name = tensor("op_275_cast_fp16")]; + tensor var_277_equation_0 = const()[name = tensor("op_277_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_277_cast_fp16 = einsum(equation = var_277_equation_0, values = (var_195_cast_fp16_8, var_252_cast_fp16))[name = tensor("op_277_cast_fp16")]; + tensor var_279_equation_0 = const()[name = tensor("op_279_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_279_cast_fp16 = einsum(equation = var_279_equation_0, values = (var_195_cast_fp16_9, var_253_cast_fp16))[name = tensor("op_279_cast_fp16")]; + tensor var_281_equation_0 = const()[name = tensor("op_281_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_281_cast_fp16 = einsum(equation = var_281_equation_0, values = (var_195_cast_fp16_10, var_254_cast_fp16))[name = tensor("op_281_cast_fp16")]; + tensor var_283_equation_0 = const()[name = tensor("op_283_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_283_cast_fp16 = einsum(equation = var_283_equation_0, values = (var_195_cast_fp16_11, var_255_cast_fp16))[name = tensor("op_283_cast_fp16")]; + tensor var_285_equation_0 = const()[name = tensor("op_285_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_285_cast_fp16 = einsum(equation = var_285_equation_0, values = (var_195_cast_fp16_12, var_256_cast_fp16))[name = tensor("op_285_cast_fp16")]; + tensor var_287_equation_0 = const()[name = tensor("op_287_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_287_cast_fp16 = einsum(equation = var_287_equation_0, values = (var_195_cast_fp16_13, var_257_cast_fp16))[name = tensor("op_287_cast_fp16")]; + tensor var_289_equation_0 = const()[name = tensor("op_289_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_289_cast_fp16 = einsum(equation = var_289_equation_0, values = (var_195_cast_fp16_14, var_258_cast_fp16))[name = tensor("op_289_cast_fp16")]; + tensor var_291_equation_0 = const()[name = tensor("op_291_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_291_cast_fp16 = einsum(equation = var_291_equation_0, values = (var_195_cast_fp16_15, var_259_cast_fp16))[name = tensor("op_291_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_108, interleave = input_5_interleave_0, values = (var_261_cast_fp16, var_263_cast_fp16, var_265_cast_fp16, var_267_cast_fp16, var_269_cast_fp16, var_271_cast_fp16, var_273_cast_fp16, var_275_cast_fp16, var_277_cast_fp16, var_279_cast_fp16, var_281_cast_fp16, var_283_cast_fp16, var_285_cast_fp16, var_287_cast_fp16, var_289_cast_fp16, var_291_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_300_pad_type_0 = const()[name = tensor("op_300_pad_type_0"), val = tensor("valid")]; + tensor var_300_strides_0 = const()[name = tensor("op_300_strides_0"), val = tensor([1, 1])]; + tensor var_300_pad_0 = const()[name = tensor("op_300_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_300_dilations_0 = const()[name = tensor("op_300_dilations_0"), val = tensor([1, 1])]; + tensor var_300_groups_0 = const()[name = tensor("op_300_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16159552)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18256768)))]; + tensor var_300_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_300_dilations_0, groups = var_300_groups_0, pad = var_300_pad_0, pad_type = var_300_pad_type_0, strides = var_300_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_300_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_300_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18258880)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18260992)))]; + tensor var_310_to_fp16 = const()[name = tensor("op_310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_310_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18263104)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26651776)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_336_pad_type_0 = const()[name = tensor("op_336_pad_type_0"), val = tensor("valid")]; + tensor var_336_strides_0 = const()[name = tensor("op_336_strides_0"), val = tensor([1, 1])]; + tensor var_336_pad_0 = const()[name = tensor("op_336_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_336_dilations_0 = const()[name = tensor("op_336_dilations_0"), val = tensor([1, 1])]; + tensor var_336_groups_0 = const()[name = tensor("op_336_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26660032)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35048704)))]; + tensor var_336_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_336_dilations_0, groups = var_336_groups_0, pad = var_336_pad_0, pad_type = var_336_pad_type_0, strides = var_336_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_336_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35050816)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35052928)))]; + tensor var_361_to_fp16 = const()[name = tensor("op_361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_361_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_396_weight_0_to_fp16 = const()[name = tensor("op_396_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35055040)))]; + tensor var_396_bias_0_to_fp16 = const()[name = tensor("op_396_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37152256)))]; + tensor var_396_cast_fp16 = conv(bias = var_396_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_396_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_396_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37154368)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_394_pad_type_0 = const()[name = tensor("op_394_pad_type_0"), val = tensor("valid")]; + tensor var_394_strides_0 = const()[name = tensor("op_394_strides_0"), val = tensor([1, 1])]; + tensor var_394_pad_0 = const()[name = tensor("op_394_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_394_dilations_0 = const()[name = tensor("op_394_dilations_0"), val = tensor([1, 1])]; + tensor var_394_groups_0 = const()[name = tensor("op_394_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39251584)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41348800)))]; + tensor var_394_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_394_dilations_0, groups = var_394_groups_0, pad = var_394_pad_0, pad_type = var_394_pad_type_0, strides = var_394_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_397_axis_0 = const()[name = tensor("op_397_axis_0"), val = tensor(1)]; + tensor var_397_cast_fp16_0, tensor var_397_cast_fp16_1, tensor var_397_cast_fp16_2, tensor var_397_cast_fp16_3, tensor var_397_cast_fp16_4, tensor var_397_cast_fp16_5, tensor var_397_cast_fp16_6, tensor var_397_cast_fp16_7, tensor var_397_cast_fp16_8, tensor var_397_cast_fp16_9, tensor var_397_cast_fp16_10, tensor var_397_cast_fp16_11, tensor var_397_cast_fp16_12, tensor var_397_cast_fp16_13, tensor var_397_cast_fp16_14, tensor var_397_cast_fp16_15 = split(axis = var_397_axis_0, split_sizes = tile_3, x = var_396_cast_fp16)[name = tensor("op_397_cast_fp16")]; + tensor var_414_perm_0 = const()[name = tensor("op_414_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_415_axis_0 = const()[name = tensor("op_415_axis_0"), val = tensor(3)]; + tensor var_414_cast_fp16 = transpose(perm = var_414_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_23")]; + tensor var_415_cast_fp16_0, tensor var_415_cast_fp16_1, tensor var_415_cast_fp16_2, tensor var_415_cast_fp16_3, tensor var_415_cast_fp16_4, tensor var_415_cast_fp16_5, tensor var_415_cast_fp16_6, tensor var_415_cast_fp16_7, tensor var_415_cast_fp16_8, tensor var_415_cast_fp16_9, tensor var_415_cast_fp16_10, tensor var_415_cast_fp16_11, tensor var_415_cast_fp16_12, tensor var_415_cast_fp16_13, tensor var_415_cast_fp16_14, tensor var_415_cast_fp16_15 = split(axis = var_415_axis_0, split_sizes = tile_4, x = var_414_cast_fp16)[name = tensor("op_415_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_432_axis_0 = const()[name = tensor("op_432_axis_0"), val = tensor(1)]; + tensor var_432_cast_fp16_0, tensor var_432_cast_fp16_1, tensor var_432_cast_fp16_2, tensor var_432_cast_fp16_3, tensor var_432_cast_fp16_4, tensor var_432_cast_fp16_5, tensor var_432_cast_fp16_6, tensor var_432_cast_fp16_7, tensor var_432_cast_fp16_8, tensor var_432_cast_fp16_9, tensor var_432_cast_fp16_10, tensor var_432_cast_fp16_11, tensor var_432_cast_fp16_12, tensor var_432_cast_fp16_13, tensor var_432_cast_fp16_14, tensor var_432_cast_fp16_15 = split(axis = var_432_axis_0, split_sizes = tile_5, x = var_394_cast_fp16)[name = tensor("op_432_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_415_cast_fp16_0, var_397_cast_fp16_0))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_415_cast_fp16_1, var_397_cast_fp16_1))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_415_cast_fp16_2, var_397_cast_fp16_2))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_415_cast_fp16_3, var_397_cast_fp16_3))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_415_cast_fp16_4, var_397_cast_fp16_4))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_415_cast_fp16_5, var_397_cast_fp16_5))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_415_cast_fp16_6, var_397_cast_fp16_6))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_415_cast_fp16_7, var_397_cast_fp16_7))[name = tensor("aw_47_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_415_cast_fp16_8, var_397_cast_fp16_8))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_415_cast_fp16_9, var_397_cast_fp16_9))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_415_cast_fp16_10, var_397_cast_fp16_10))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_415_cast_fp16_11, var_397_cast_fp16_11))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_415_cast_fp16_12, var_397_cast_fp16_12))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_415_cast_fp16_13, var_397_cast_fp16_13))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_415_cast_fp16_14, var_397_cast_fp16_14))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_415_cast_fp16_15, var_397_cast_fp16_15))[name = tensor("aw_63_cast_fp16")]; + tensor var_481_cast_fp16 = softmax(axis = var_345, x = aw_33_cast_fp16)[name = tensor("op_481_cast_fp16")]; + tensor var_482_cast_fp16 = softmax(axis = var_345, x = aw_35_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor var_483_cast_fp16 = softmax(axis = var_345, x = aw_37_cast_fp16)[name = tensor("op_483_cast_fp16")]; + tensor var_484_cast_fp16 = softmax(axis = var_345, x = aw_39_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_485_cast_fp16 = softmax(axis = var_345, x = aw_41_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor var_486_cast_fp16 = softmax(axis = var_345, x = aw_43_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_487_cast_fp16 = softmax(axis = var_345, x = aw_45_cast_fp16)[name = tensor("op_487_cast_fp16")]; + tensor var_488_cast_fp16 = softmax(axis = var_345, x = aw_47_cast_fp16)[name = tensor("op_488_cast_fp16")]; + tensor var_489_cast_fp16 = softmax(axis = var_345, x = aw_49_cast_fp16)[name = tensor("op_489_cast_fp16")]; + tensor var_490_cast_fp16 = softmax(axis = var_345, x = aw_51_cast_fp16)[name = tensor("op_490_cast_fp16")]; + tensor var_491_cast_fp16 = softmax(axis = var_345, x = aw_53_cast_fp16)[name = tensor("op_491_cast_fp16")]; + tensor var_492_cast_fp16 = softmax(axis = var_345, x = aw_55_cast_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_493_cast_fp16 = softmax(axis = var_345, x = aw_57_cast_fp16)[name = tensor("op_493_cast_fp16")]; + tensor var_494_cast_fp16 = softmax(axis = var_345, x = aw_59_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor var_495_cast_fp16 = softmax(axis = var_345, x = aw_61_cast_fp16)[name = tensor("op_495_cast_fp16")]; + tensor var_496_cast_fp16 = softmax(axis = var_345, x = aw_63_cast_fp16)[name = tensor("op_496_cast_fp16")]; + tensor var_498_equation_0 = const()[name = tensor("op_498_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_498_cast_fp16 = einsum(equation = var_498_equation_0, values = (var_432_cast_fp16_0, var_481_cast_fp16))[name = tensor("op_498_cast_fp16")]; + tensor var_500_equation_0 = const()[name = tensor("op_500_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_500_cast_fp16 = einsum(equation = var_500_equation_0, values = (var_432_cast_fp16_1, var_482_cast_fp16))[name = tensor("op_500_cast_fp16")]; + tensor var_502_equation_0 = const()[name = tensor("op_502_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_502_cast_fp16 = einsum(equation = var_502_equation_0, values = (var_432_cast_fp16_2, var_483_cast_fp16))[name = tensor("op_502_cast_fp16")]; + tensor var_504_equation_0 = const()[name = tensor("op_504_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_504_cast_fp16 = einsum(equation = var_504_equation_0, values = (var_432_cast_fp16_3, var_484_cast_fp16))[name = tensor("op_504_cast_fp16")]; + tensor var_506_equation_0 = const()[name = tensor("op_506_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_506_cast_fp16 = einsum(equation = var_506_equation_0, values = (var_432_cast_fp16_4, var_485_cast_fp16))[name = tensor("op_506_cast_fp16")]; + tensor var_508_equation_0 = const()[name = tensor("op_508_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_508_cast_fp16 = einsum(equation = var_508_equation_0, values = (var_432_cast_fp16_5, var_486_cast_fp16))[name = tensor("op_508_cast_fp16")]; + tensor var_510_equation_0 = const()[name = tensor("op_510_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_510_cast_fp16 = einsum(equation = var_510_equation_0, values = (var_432_cast_fp16_6, var_487_cast_fp16))[name = tensor("op_510_cast_fp16")]; + tensor var_512_equation_0 = const()[name = tensor("op_512_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_512_cast_fp16 = einsum(equation = var_512_equation_0, values = (var_432_cast_fp16_7, var_488_cast_fp16))[name = tensor("op_512_cast_fp16")]; + tensor var_514_equation_0 = const()[name = tensor("op_514_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_514_cast_fp16 = einsum(equation = var_514_equation_0, values = (var_432_cast_fp16_8, var_489_cast_fp16))[name = tensor("op_514_cast_fp16")]; + tensor var_516_equation_0 = const()[name = tensor("op_516_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_516_cast_fp16 = einsum(equation = var_516_equation_0, values = (var_432_cast_fp16_9, var_490_cast_fp16))[name = tensor("op_516_cast_fp16")]; + tensor var_518_equation_0 = const()[name = tensor("op_518_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_518_cast_fp16 = einsum(equation = var_518_equation_0, values = (var_432_cast_fp16_10, var_491_cast_fp16))[name = tensor("op_518_cast_fp16")]; + tensor var_520_equation_0 = const()[name = tensor("op_520_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_520_cast_fp16 = einsum(equation = var_520_equation_0, values = (var_432_cast_fp16_11, var_492_cast_fp16))[name = tensor("op_520_cast_fp16")]; + tensor var_522_equation_0 = const()[name = tensor("op_522_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_522_cast_fp16 = einsum(equation = var_522_equation_0, values = (var_432_cast_fp16_12, var_493_cast_fp16))[name = tensor("op_522_cast_fp16")]; + tensor var_524_equation_0 = const()[name = tensor("op_524_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_524_cast_fp16 = einsum(equation = var_524_equation_0, values = (var_432_cast_fp16_13, var_494_cast_fp16))[name = tensor("op_524_cast_fp16")]; + tensor var_526_equation_0 = const()[name = tensor("op_526_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_526_cast_fp16 = einsum(equation = var_526_equation_0, values = (var_432_cast_fp16_14, var_495_cast_fp16))[name = tensor("op_526_cast_fp16")]; + tensor var_528_equation_0 = const()[name = tensor("op_528_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_528_cast_fp16 = einsum(equation = var_528_equation_0, values = (var_432_cast_fp16_15, var_496_cast_fp16))[name = tensor("op_528_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_345, interleave = input_15_interleave_0, values = (var_498_cast_fp16, var_500_cast_fp16, var_502_cast_fp16, var_504_cast_fp16, var_506_cast_fp16, var_508_cast_fp16, var_510_cast_fp16, var_512_cast_fp16, var_514_cast_fp16, var_516_cast_fp16, var_518_cast_fp16, var_520_cast_fp16, var_522_cast_fp16, var_524_cast_fp16, var_526_cast_fp16, var_528_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_537_pad_type_0 = const()[name = tensor("op_537_pad_type_0"), val = tensor("valid")]; + tensor var_537_strides_0 = const()[name = tensor("op_537_strides_0"), val = tensor([1, 1])]; + tensor var_537_pad_0 = const()[name = tensor("op_537_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_537_dilations_0 = const()[name = tensor("op_537_dilations_0"), val = tensor([1, 1])]; + tensor var_537_groups_0 = const()[name = tensor("op_537_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41350912)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43448128)))]; + tensor var_537_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_537_dilations_0, groups = var_537_groups_0, pad = var_537_pad_0, pad_type = var_537_pad_type_0, strides = var_537_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_537_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_537_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43450240)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43452352)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_547_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43454464)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51843136)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_573_pad_type_0 = const()[name = tensor("op_573_pad_type_0"), val = tensor("valid")]; + tensor var_573_strides_0 = const()[name = tensor("op_573_strides_0"), val = tensor([1, 1])]; + tensor var_573_pad_0 = const()[name = tensor("op_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_573_dilations_0 = const()[name = tensor("op_573_dilations_0"), val = tensor([1, 1])]; + tensor var_573_groups_0 = const()[name = tensor("op_573_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51851392)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60240064)))]; + tensor var_573_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_573_dilations_0, groups = var_573_groups_0, pad = var_573_pad_0, pad_type = var_573_pad_type_0, strides = var_573_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_573_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_582 = const()[name = tensor("op_582"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60242176)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60244288)))]; + tensor var_598_to_fp16 = const()[name = tensor("op_598_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_598_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_633_weight_0_to_fp16 = const()[name = tensor("op_633_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60246400)))]; + tensor var_633_bias_0_to_fp16 = const()[name = tensor("op_633_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62343616)))]; + tensor var_633_cast_fp16 = conv(bias = var_633_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_633_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_633_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62345728)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_631_pad_type_0 = const()[name = tensor("op_631_pad_type_0"), val = tensor("valid")]; + tensor var_631_strides_0 = const()[name = tensor("op_631_strides_0"), val = tensor([1, 1])]; + tensor var_631_pad_0 = const()[name = tensor("op_631_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_631_dilations_0 = const()[name = tensor("op_631_dilations_0"), val = tensor([1, 1])]; + tensor var_631_groups_0 = const()[name = tensor("op_631_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64442944)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66540160)))]; + tensor var_631_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_631_dilations_0, groups = var_631_groups_0, pad = var_631_pad_0, pad_type = var_631_pad_type_0, strides = var_631_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_634_axis_0 = const()[name = tensor("op_634_axis_0"), val = tensor(1)]; + tensor var_634_cast_fp16_0, tensor var_634_cast_fp16_1, tensor var_634_cast_fp16_2, tensor var_634_cast_fp16_3, tensor var_634_cast_fp16_4, tensor var_634_cast_fp16_5, tensor var_634_cast_fp16_6, tensor var_634_cast_fp16_7, tensor var_634_cast_fp16_8, tensor var_634_cast_fp16_9, tensor var_634_cast_fp16_10, tensor var_634_cast_fp16_11, tensor var_634_cast_fp16_12, tensor var_634_cast_fp16_13, tensor var_634_cast_fp16_14, tensor var_634_cast_fp16_15 = split(axis = var_634_axis_0, split_sizes = tile_6, x = var_633_cast_fp16)[name = tensor("op_634_cast_fp16")]; + tensor var_651_perm_0 = const()[name = tensor("op_651_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(3)]; + tensor var_651_cast_fp16 = transpose(perm = var_651_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_22")]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2, tensor var_652_cast_fp16_3, tensor var_652_cast_fp16_4, tensor var_652_cast_fp16_5, tensor var_652_cast_fp16_6, tensor var_652_cast_fp16_7, tensor var_652_cast_fp16_8, tensor var_652_cast_fp16_9, tensor var_652_cast_fp16_10, tensor var_652_cast_fp16_11, tensor var_652_cast_fp16_12, tensor var_652_cast_fp16_13, tensor var_652_cast_fp16_14, tensor var_652_cast_fp16_15 = split(axis = var_652_axis_0, split_sizes = tile_7, x = var_651_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_669_axis_0 = const()[name = tensor("op_669_axis_0"), val = tensor(1)]; + tensor var_669_cast_fp16_0, tensor var_669_cast_fp16_1, tensor var_669_cast_fp16_2, tensor var_669_cast_fp16_3, tensor var_669_cast_fp16_4, tensor var_669_cast_fp16_5, tensor var_669_cast_fp16_6, tensor var_669_cast_fp16_7, tensor var_669_cast_fp16_8, tensor var_669_cast_fp16_9, tensor var_669_cast_fp16_10, tensor var_669_cast_fp16_11, tensor var_669_cast_fp16_12, tensor var_669_cast_fp16_13, tensor var_669_cast_fp16_14, tensor var_669_cast_fp16_15 = split(axis = var_669_axis_0, split_sizes = tile_8, x = var_631_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_652_cast_fp16_0, var_634_cast_fp16_0))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_652_cast_fp16_1, var_634_cast_fp16_1))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_652_cast_fp16_2, var_634_cast_fp16_2))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_652_cast_fp16_3, var_634_cast_fp16_3))[name = tensor("aw_71_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_652_cast_fp16_4, var_634_cast_fp16_4))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_652_cast_fp16_5, var_634_cast_fp16_5))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_652_cast_fp16_6, var_634_cast_fp16_6))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_652_cast_fp16_7, var_634_cast_fp16_7))[name = tensor("aw_79_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_652_cast_fp16_8, var_634_cast_fp16_8))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_652_cast_fp16_9, var_634_cast_fp16_9))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_652_cast_fp16_10, var_634_cast_fp16_10))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_652_cast_fp16_11, var_634_cast_fp16_11))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_652_cast_fp16_12, var_634_cast_fp16_12))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_652_cast_fp16_13, var_634_cast_fp16_13))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_652_cast_fp16_14, var_634_cast_fp16_14))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_652_cast_fp16_15, var_634_cast_fp16_15))[name = tensor("aw_95_cast_fp16")]; + tensor var_718_cast_fp16 = softmax(axis = var_582, x = aw_65_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor var_719_cast_fp16 = softmax(axis = var_582, x = aw_67_cast_fp16)[name = tensor("op_719_cast_fp16")]; + tensor var_720_cast_fp16 = softmax(axis = var_582, x = aw_69_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor var_721_cast_fp16 = softmax(axis = var_582, x = aw_71_cast_fp16)[name = tensor("op_721_cast_fp16")]; + tensor var_722_cast_fp16 = softmax(axis = var_582, x = aw_73_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = softmax(axis = var_582, x = aw_75_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor var_724_cast_fp16 = softmax(axis = var_582, x = aw_77_cast_fp16)[name = tensor("op_724_cast_fp16")]; + tensor var_725_cast_fp16 = softmax(axis = var_582, x = aw_79_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor var_726_cast_fp16 = softmax(axis = var_582, x = aw_81_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = softmax(axis = var_582, x = aw_83_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor var_728_cast_fp16 = softmax(axis = var_582, x = aw_85_cast_fp16)[name = tensor("op_728_cast_fp16")]; + tensor var_729_cast_fp16 = softmax(axis = var_582, x = aw_87_cast_fp16)[name = tensor("op_729_cast_fp16")]; + tensor var_730_cast_fp16 = softmax(axis = var_582, x = aw_89_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = softmax(axis = var_582, x = aw_91_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor var_732_cast_fp16 = softmax(axis = var_582, x = aw_93_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_cast_fp16 = softmax(axis = var_582, x = aw_95_cast_fp16)[name = tensor("op_733_cast_fp16")]; + tensor var_735_equation_0 = const()[name = tensor("op_735_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_735_cast_fp16 = einsum(equation = var_735_equation_0, values = (var_669_cast_fp16_0, var_718_cast_fp16))[name = tensor("op_735_cast_fp16")]; + tensor var_737_equation_0 = const()[name = tensor("op_737_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_737_cast_fp16 = einsum(equation = var_737_equation_0, values = (var_669_cast_fp16_1, var_719_cast_fp16))[name = tensor("op_737_cast_fp16")]; + tensor var_739_equation_0 = const()[name = tensor("op_739_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_739_cast_fp16 = einsum(equation = var_739_equation_0, values = (var_669_cast_fp16_2, var_720_cast_fp16))[name = tensor("op_739_cast_fp16")]; + tensor var_741_equation_0 = const()[name = tensor("op_741_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_741_cast_fp16 = einsum(equation = var_741_equation_0, values = (var_669_cast_fp16_3, var_721_cast_fp16))[name = tensor("op_741_cast_fp16")]; + tensor var_743_equation_0 = const()[name = tensor("op_743_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_743_cast_fp16 = einsum(equation = var_743_equation_0, values = (var_669_cast_fp16_4, var_722_cast_fp16))[name = tensor("op_743_cast_fp16")]; + tensor var_745_equation_0 = const()[name = tensor("op_745_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_745_cast_fp16 = einsum(equation = var_745_equation_0, values = (var_669_cast_fp16_5, var_723_cast_fp16))[name = tensor("op_745_cast_fp16")]; + tensor var_747_equation_0 = const()[name = tensor("op_747_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_747_cast_fp16 = einsum(equation = var_747_equation_0, values = (var_669_cast_fp16_6, var_724_cast_fp16))[name = tensor("op_747_cast_fp16")]; + tensor var_749_equation_0 = const()[name = tensor("op_749_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_749_cast_fp16 = einsum(equation = var_749_equation_0, values = (var_669_cast_fp16_7, var_725_cast_fp16))[name = tensor("op_749_cast_fp16")]; + tensor var_751_equation_0 = const()[name = tensor("op_751_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_751_cast_fp16 = einsum(equation = var_751_equation_0, values = (var_669_cast_fp16_8, var_726_cast_fp16))[name = tensor("op_751_cast_fp16")]; + tensor var_753_equation_0 = const()[name = tensor("op_753_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_753_cast_fp16 = einsum(equation = var_753_equation_0, values = (var_669_cast_fp16_9, var_727_cast_fp16))[name = tensor("op_753_cast_fp16")]; + tensor var_755_equation_0 = const()[name = tensor("op_755_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_755_cast_fp16 = einsum(equation = var_755_equation_0, values = (var_669_cast_fp16_10, var_728_cast_fp16))[name = tensor("op_755_cast_fp16")]; + tensor var_757_equation_0 = const()[name = tensor("op_757_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_757_cast_fp16 = einsum(equation = var_757_equation_0, values = (var_669_cast_fp16_11, var_729_cast_fp16))[name = tensor("op_757_cast_fp16")]; + tensor var_759_equation_0 = const()[name = tensor("op_759_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_759_cast_fp16 = einsum(equation = var_759_equation_0, values = (var_669_cast_fp16_12, var_730_cast_fp16))[name = tensor("op_759_cast_fp16")]; + tensor var_761_equation_0 = const()[name = tensor("op_761_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_761_cast_fp16 = einsum(equation = var_761_equation_0, values = (var_669_cast_fp16_13, var_731_cast_fp16))[name = tensor("op_761_cast_fp16")]; + tensor var_763_equation_0 = const()[name = tensor("op_763_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_763_cast_fp16 = einsum(equation = var_763_equation_0, values = (var_669_cast_fp16_14, var_732_cast_fp16))[name = tensor("op_763_cast_fp16")]; + tensor var_765_equation_0 = const()[name = tensor("op_765_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_765_cast_fp16 = einsum(equation = var_765_equation_0, values = (var_669_cast_fp16_15, var_733_cast_fp16))[name = tensor("op_765_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_582, interleave = input_25_interleave_0, values = (var_735_cast_fp16, var_737_cast_fp16, var_739_cast_fp16, var_741_cast_fp16, var_743_cast_fp16, var_745_cast_fp16, var_747_cast_fp16, var_749_cast_fp16, var_751_cast_fp16, var_753_cast_fp16, var_755_cast_fp16, var_757_cast_fp16, var_759_cast_fp16, var_761_cast_fp16, var_763_cast_fp16, var_765_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_774_pad_type_0 = const()[name = tensor("op_774_pad_type_0"), val = tensor("valid")]; + tensor var_774_strides_0 = const()[name = tensor("op_774_strides_0"), val = tensor([1, 1])]; + tensor var_774_pad_0 = const()[name = tensor("op_774_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_774_dilations_0 = const()[name = tensor("op_774_dilations_0"), val = tensor([1, 1])]; + tensor var_774_groups_0 = const()[name = tensor("op_774_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66542272)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68639488)))]; + tensor var_774_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_774_dilations_0, groups = var_774_groups_0, pad = var_774_pad_0, pad_type = var_774_pad_type_0, strides = var_774_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_774_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68641600)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68643712)))]; + tensor var_784_to_fp16 = const()[name = tensor("op_784_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_784_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68645824)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77034496)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_810_pad_type_0 = const()[name = tensor("op_810_pad_type_0"), val = tensor("valid")]; + tensor var_810_strides_0 = const()[name = tensor("op_810_strides_0"), val = tensor([1, 1])]; + tensor var_810_pad_0 = const()[name = tensor("op_810_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_810_dilations_0 = const()[name = tensor("op_810_dilations_0"), val = tensor([1, 1])]; + tensor var_810_groups_0 = const()[name = tensor("op_810_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77042752)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85431424)))]; + tensor var_810_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_810_dilations_0, groups = var_810_groups_0, pad = var_810_pad_0, pad_type = var_810_pad_type_0, strides = var_810_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_810_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_819 = const()[name = tensor("op_819"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85433536)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85435648)))]; + tensor var_835_to_fp16 = const()[name = tensor("op_835_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_835_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_870_weight_0_to_fp16 = const()[name = tensor("op_870_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85437760)))]; + tensor var_870_bias_0_to_fp16 = const()[name = tensor("op_870_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87534976)))]; + tensor var_870_cast_fp16 = conv(bias = var_870_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_870_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_870_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87537088)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_868_pad_type_0 = const()[name = tensor("op_868_pad_type_0"), val = tensor("valid")]; + tensor var_868_strides_0 = const()[name = tensor("op_868_strides_0"), val = tensor([1, 1])]; + tensor var_868_pad_0 = const()[name = tensor("op_868_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_868_dilations_0 = const()[name = tensor("op_868_dilations_0"), val = tensor([1, 1])]; + tensor var_868_groups_0 = const()[name = tensor("op_868_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89634304)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91731520)))]; + tensor var_868_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_868_dilations_0, groups = var_868_groups_0, pad = var_868_pad_0, pad_type = var_868_pad_type_0, strides = var_868_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_871_axis_0 = const()[name = tensor("op_871_axis_0"), val = tensor(1)]; + tensor var_871_cast_fp16_0, tensor var_871_cast_fp16_1, tensor var_871_cast_fp16_2, tensor var_871_cast_fp16_3, tensor var_871_cast_fp16_4, tensor var_871_cast_fp16_5, tensor var_871_cast_fp16_6, tensor var_871_cast_fp16_7, tensor var_871_cast_fp16_8, tensor var_871_cast_fp16_9, tensor var_871_cast_fp16_10, tensor var_871_cast_fp16_11, tensor var_871_cast_fp16_12, tensor var_871_cast_fp16_13, tensor var_871_cast_fp16_14, tensor var_871_cast_fp16_15 = split(axis = var_871_axis_0, split_sizes = tile_9, x = var_870_cast_fp16)[name = tensor("op_871_cast_fp16")]; + tensor var_888_perm_0 = const()[name = tensor("op_888_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_889_axis_0 = const()[name = tensor("op_889_axis_0"), val = tensor(3)]; + tensor var_888_cast_fp16 = transpose(perm = var_888_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_21")]; + tensor var_889_cast_fp16_0, tensor var_889_cast_fp16_1, tensor var_889_cast_fp16_2, tensor var_889_cast_fp16_3, tensor var_889_cast_fp16_4, tensor var_889_cast_fp16_5, tensor var_889_cast_fp16_6, tensor var_889_cast_fp16_7, tensor var_889_cast_fp16_8, tensor var_889_cast_fp16_9, tensor var_889_cast_fp16_10, tensor var_889_cast_fp16_11, tensor var_889_cast_fp16_12, tensor var_889_cast_fp16_13, tensor var_889_cast_fp16_14, tensor var_889_cast_fp16_15 = split(axis = var_889_axis_0, split_sizes = tile_10, x = var_888_cast_fp16)[name = tensor("op_889_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_906_axis_0 = const()[name = tensor("op_906_axis_0"), val = tensor(1)]; + tensor var_906_cast_fp16_0, tensor var_906_cast_fp16_1, tensor var_906_cast_fp16_2, tensor var_906_cast_fp16_3, tensor var_906_cast_fp16_4, tensor var_906_cast_fp16_5, tensor var_906_cast_fp16_6, tensor var_906_cast_fp16_7, tensor var_906_cast_fp16_8, tensor var_906_cast_fp16_9, tensor var_906_cast_fp16_10, tensor var_906_cast_fp16_11, tensor var_906_cast_fp16_12, tensor var_906_cast_fp16_13, tensor var_906_cast_fp16_14, tensor var_906_cast_fp16_15 = split(axis = var_906_axis_0, split_sizes = tile_11, x = var_868_cast_fp16)[name = tensor("op_906_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_889_cast_fp16_0, var_871_cast_fp16_0))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_889_cast_fp16_1, var_871_cast_fp16_1))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_889_cast_fp16_2, var_871_cast_fp16_2))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_889_cast_fp16_3, var_871_cast_fp16_3))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_889_cast_fp16_4, var_871_cast_fp16_4))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_889_cast_fp16_5, var_871_cast_fp16_5))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_889_cast_fp16_6, var_871_cast_fp16_6))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_889_cast_fp16_7, var_871_cast_fp16_7))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_889_cast_fp16_8, var_871_cast_fp16_8))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_889_cast_fp16_9, var_871_cast_fp16_9))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_889_cast_fp16_10, var_871_cast_fp16_10))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_889_cast_fp16_11, var_871_cast_fp16_11))[name = tensor("aw_119_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_889_cast_fp16_12, var_871_cast_fp16_12))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_889_cast_fp16_13, var_871_cast_fp16_13))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_889_cast_fp16_14, var_871_cast_fp16_14))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_889_cast_fp16_15, var_871_cast_fp16_15))[name = tensor("aw_127_cast_fp16")]; + tensor var_955_cast_fp16 = softmax(axis = var_819, x = aw_97_cast_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_956_cast_fp16 = softmax(axis = var_819, x = aw_99_cast_fp16)[name = tensor("op_956_cast_fp16")]; + tensor var_957_cast_fp16 = softmax(axis = var_819, x = aw_101_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor var_958_cast_fp16 = softmax(axis = var_819, x = aw_103_cast_fp16)[name = tensor("op_958_cast_fp16")]; + tensor var_959_cast_fp16 = softmax(axis = var_819, x = aw_105_cast_fp16)[name = tensor("op_959_cast_fp16")]; + tensor var_960_cast_fp16 = softmax(axis = var_819, x = aw_107_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor var_961_cast_fp16 = softmax(axis = var_819, x = aw_109_cast_fp16)[name = tensor("op_961_cast_fp16")]; + tensor var_962_cast_fp16 = softmax(axis = var_819, x = aw_111_cast_fp16)[name = tensor("op_962_cast_fp16")]; + tensor var_963_cast_fp16 = softmax(axis = var_819, x = aw_113_cast_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_964_cast_fp16 = softmax(axis = var_819, x = aw_115_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = softmax(axis = var_819, x = aw_117_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor var_966_cast_fp16 = softmax(axis = var_819, x = aw_119_cast_fp16)[name = tensor("op_966_cast_fp16")]; + tensor var_967_cast_fp16 = softmax(axis = var_819, x = aw_121_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor var_968_cast_fp16 = softmax(axis = var_819, x = aw_123_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = softmax(axis = var_819, x = aw_125_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor var_970_cast_fp16 = softmax(axis = var_819, x = aw_127_cast_fp16)[name = tensor("op_970_cast_fp16")]; + tensor var_972_equation_0 = const()[name = tensor("op_972_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_972_cast_fp16 = einsum(equation = var_972_equation_0, values = (var_906_cast_fp16_0, var_955_cast_fp16))[name = tensor("op_972_cast_fp16")]; + tensor var_974_equation_0 = const()[name = tensor("op_974_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_974_cast_fp16 = einsum(equation = var_974_equation_0, values = (var_906_cast_fp16_1, var_956_cast_fp16))[name = tensor("op_974_cast_fp16")]; + tensor var_976_equation_0 = const()[name = tensor("op_976_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_976_cast_fp16 = einsum(equation = var_976_equation_0, values = (var_906_cast_fp16_2, var_957_cast_fp16))[name = tensor("op_976_cast_fp16")]; + tensor var_978_equation_0 = const()[name = tensor("op_978_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_978_cast_fp16 = einsum(equation = var_978_equation_0, values = (var_906_cast_fp16_3, var_958_cast_fp16))[name = tensor("op_978_cast_fp16")]; + tensor var_980_equation_0 = const()[name = tensor("op_980_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_980_cast_fp16 = einsum(equation = var_980_equation_0, values = (var_906_cast_fp16_4, var_959_cast_fp16))[name = tensor("op_980_cast_fp16")]; + tensor var_982_equation_0 = const()[name = tensor("op_982_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_982_cast_fp16 = einsum(equation = var_982_equation_0, values = (var_906_cast_fp16_5, var_960_cast_fp16))[name = tensor("op_982_cast_fp16")]; + tensor var_984_equation_0 = const()[name = tensor("op_984_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_984_cast_fp16 = einsum(equation = var_984_equation_0, values = (var_906_cast_fp16_6, var_961_cast_fp16))[name = tensor("op_984_cast_fp16")]; + tensor var_986_equation_0 = const()[name = tensor("op_986_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_986_cast_fp16 = einsum(equation = var_986_equation_0, values = (var_906_cast_fp16_7, var_962_cast_fp16))[name = tensor("op_986_cast_fp16")]; + tensor var_988_equation_0 = const()[name = tensor("op_988_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_988_cast_fp16 = einsum(equation = var_988_equation_0, values = (var_906_cast_fp16_8, var_963_cast_fp16))[name = tensor("op_988_cast_fp16")]; + tensor var_990_equation_0 = const()[name = tensor("op_990_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_990_cast_fp16 = einsum(equation = var_990_equation_0, values = (var_906_cast_fp16_9, var_964_cast_fp16))[name = tensor("op_990_cast_fp16")]; + tensor var_992_equation_0 = const()[name = tensor("op_992_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_992_cast_fp16 = einsum(equation = var_992_equation_0, values = (var_906_cast_fp16_10, var_965_cast_fp16))[name = tensor("op_992_cast_fp16")]; + tensor var_994_equation_0 = const()[name = tensor("op_994_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_994_cast_fp16 = einsum(equation = var_994_equation_0, values = (var_906_cast_fp16_11, var_966_cast_fp16))[name = tensor("op_994_cast_fp16")]; + tensor var_996_equation_0 = const()[name = tensor("op_996_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_996_cast_fp16 = einsum(equation = var_996_equation_0, values = (var_906_cast_fp16_12, var_967_cast_fp16))[name = tensor("op_996_cast_fp16")]; + tensor var_998_equation_0 = const()[name = tensor("op_998_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_998_cast_fp16 = einsum(equation = var_998_equation_0, values = (var_906_cast_fp16_13, var_968_cast_fp16))[name = tensor("op_998_cast_fp16")]; + tensor var_1000_equation_0 = const()[name = tensor("op_1000_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1000_cast_fp16 = einsum(equation = var_1000_equation_0, values = (var_906_cast_fp16_14, var_969_cast_fp16))[name = tensor("op_1000_cast_fp16")]; + tensor var_1002_equation_0 = const()[name = tensor("op_1002_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1002_cast_fp16 = einsum(equation = var_1002_equation_0, values = (var_906_cast_fp16_15, var_970_cast_fp16))[name = tensor("op_1002_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_819, interleave = input_35_interleave_0, values = (var_972_cast_fp16, var_974_cast_fp16, var_976_cast_fp16, var_978_cast_fp16, var_980_cast_fp16, var_982_cast_fp16, var_984_cast_fp16, var_986_cast_fp16, var_988_cast_fp16, var_990_cast_fp16, var_992_cast_fp16, var_994_cast_fp16, var_996_cast_fp16, var_998_cast_fp16, var_1000_cast_fp16, var_1002_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_1011_pad_type_0 = const()[name = tensor("op_1011_pad_type_0"), val = tensor("valid")]; + tensor var_1011_strides_0 = const()[name = tensor("op_1011_strides_0"), val = tensor([1, 1])]; + tensor var_1011_pad_0 = const()[name = tensor("op_1011_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1011_dilations_0 = const()[name = tensor("op_1011_dilations_0"), val = tensor([1, 1])]; + tensor var_1011_groups_0 = const()[name = tensor("op_1011_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91733632)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93830848)))]; + tensor var_1011_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_1011_dilations_0, groups = var_1011_groups_0, pad = var_1011_pad_0, pad_type = var_1011_pad_type_0, strides = var_1011_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_1011_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93832960)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93835072)))]; + tensor var_1021_to_fp16 = const()[name = tensor("op_1021_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_1021_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93837184)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102225856)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1047_pad_type_0 = const()[name = tensor("op_1047_pad_type_0"), val = tensor("valid")]; + tensor var_1047_strides_0 = const()[name = tensor("op_1047_strides_0"), val = tensor([1, 1])]; + tensor var_1047_pad_0 = const()[name = tensor("op_1047_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1047_dilations_0 = const()[name = tensor("op_1047_dilations_0"), val = tensor([1, 1])]; + tensor var_1047_groups_0 = const()[name = tensor("op_1047_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102234112)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110622784)))]; + tensor var_1047_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_1047_dilations_0, groups = var_1047_groups_0, pad = var_1047_pad_0, pad_type = var_1047_pad_type_0, strides = var_1047_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_1047_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_1047_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110624896)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110627008)))]; + tensor var_1072_to_fp16 = const()[name = tensor("op_1072_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_1072_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_1107_weight_0_to_fp16 = const()[name = tensor("op_1107_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110629120)))]; + tensor var_1107_bias_0_to_fp16 = const()[name = tensor("op_1107_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112726336)))]; + tensor var_1107_cast_fp16 = conv(bias = var_1107_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_1107_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1107_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112728448)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_1105_pad_type_0 = const()[name = tensor("op_1105_pad_type_0"), val = tensor("valid")]; + tensor var_1105_strides_0 = const()[name = tensor("op_1105_strides_0"), val = tensor([1, 1])]; + tensor var_1105_pad_0 = const()[name = tensor("op_1105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1105_dilations_0 = const()[name = tensor("op_1105_dilations_0"), val = tensor([1, 1])]; + tensor var_1105_groups_0 = const()[name = tensor("op_1105_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114825664)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116922880)))]; + tensor var_1105_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_1105_dilations_0, groups = var_1105_groups_0, pad = var_1105_pad_0, pad_type = var_1105_pad_type_0, strides = var_1105_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1108_axis_0 = const()[name = tensor("op_1108_axis_0"), val = tensor(1)]; + tensor var_1108_cast_fp16_0, tensor var_1108_cast_fp16_1, tensor var_1108_cast_fp16_2, tensor var_1108_cast_fp16_3, tensor var_1108_cast_fp16_4, tensor var_1108_cast_fp16_5, tensor var_1108_cast_fp16_6, tensor var_1108_cast_fp16_7, tensor var_1108_cast_fp16_8, tensor var_1108_cast_fp16_9, tensor var_1108_cast_fp16_10, tensor var_1108_cast_fp16_11, tensor var_1108_cast_fp16_12, tensor var_1108_cast_fp16_13, tensor var_1108_cast_fp16_14, tensor var_1108_cast_fp16_15 = split(axis = var_1108_axis_0, split_sizes = tile_12, x = var_1107_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1126_axis_0 = const()[name = tensor("op_1126_axis_0"), val = tensor(3)]; + tensor var_1125_cast_fp16 = transpose(perm = var_1125_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_20")]; + tensor var_1126_cast_fp16_0, tensor var_1126_cast_fp16_1, tensor var_1126_cast_fp16_2, tensor var_1126_cast_fp16_3, tensor var_1126_cast_fp16_4, tensor var_1126_cast_fp16_5, tensor var_1126_cast_fp16_6, tensor var_1126_cast_fp16_7, tensor var_1126_cast_fp16_8, tensor var_1126_cast_fp16_9, tensor var_1126_cast_fp16_10, tensor var_1126_cast_fp16_11, tensor var_1126_cast_fp16_12, tensor var_1126_cast_fp16_13, tensor var_1126_cast_fp16_14, tensor var_1126_cast_fp16_15 = split(axis = var_1126_axis_0, split_sizes = tile_13, x = var_1125_cast_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1143_axis_0 = const()[name = tensor("op_1143_axis_0"), val = tensor(1)]; + tensor var_1143_cast_fp16_0, tensor var_1143_cast_fp16_1, tensor var_1143_cast_fp16_2, tensor var_1143_cast_fp16_3, tensor var_1143_cast_fp16_4, tensor var_1143_cast_fp16_5, tensor var_1143_cast_fp16_6, tensor var_1143_cast_fp16_7, tensor var_1143_cast_fp16_8, tensor var_1143_cast_fp16_9, tensor var_1143_cast_fp16_10, tensor var_1143_cast_fp16_11, tensor var_1143_cast_fp16_12, tensor var_1143_cast_fp16_13, tensor var_1143_cast_fp16_14, tensor var_1143_cast_fp16_15 = split(axis = var_1143_axis_0, split_sizes = tile_14, x = var_1105_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1126_cast_fp16_0, var_1108_cast_fp16_0))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1126_cast_fp16_1, var_1108_cast_fp16_1))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1126_cast_fp16_2, var_1108_cast_fp16_2))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1126_cast_fp16_3, var_1108_cast_fp16_3))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1126_cast_fp16_4, var_1108_cast_fp16_4))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1126_cast_fp16_5, var_1108_cast_fp16_5))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1126_cast_fp16_6, var_1108_cast_fp16_6))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1126_cast_fp16_7, var_1108_cast_fp16_7))[name = tensor("aw_143_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1126_cast_fp16_8, var_1108_cast_fp16_8))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1126_cast_fp16_9, var_1108_cast_fp16_9))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1126_cast_fp16_10, var_1108_cast_fp16_10))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1126_cast_fp16_11, var_1108_cast_fp16_11))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1126_cast_fp16_12, var_1108_cast_fp16_12))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1126_cast_fp16_13, var_1108_cast_fp16_13))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1126_cast_fp16_14, var_1108_cast_fp16_14))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1126_cast_fp16_15, var_1108_cast_fp16_15))[name = tensor("aw_159_cast_fp16")]; + tensor var_1192_cast_fp16 = softmax(axis = var_1056, x = aw_129_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor var_1193_cast_fp16 = softmax(axis = var_1056, x = aw_131_cast_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194_cast_fp16 = softmax(axis = var_1056, x = aw_133_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1195_cast_fp16 = softmax(axis = var_1056, x = aw_135_cast_fp16)[name = tensor("op_1195_cast_fp16")]; + tensor var_1196_cast_fp16 = softmax(axis = var_1056, x = aw_137_cast_fp16)[name = tensor("op_1196_cast_fp16")]; + tensor var_1197_cast_fp16 = softmax(axis = var_1056, x = aw_139_cast_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1198_cast_fp16 = softmax(axis = var_1056, x = aw_141_cast_fp16)[name = tensor("op_1198_cast_fp16")]; + tensor var_1199_cast_fp16 = softmax(axis = var_1056, x = aw_143_cast_fp16)[name = tensor("op_1199_cast_fp16")]; + tensor var_1200_cast_fp16 = softmax(axis = var_1056, x = aw_145_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor var_1201_cast_fp16 = softmax(axis = var_1056, x = aw_147_cast_fp16)[name = tensor("op_1201_cast_fp16")]; + tensor var_1202_cast_fp16 = softmax(axis = var_1056, x = aw_149_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1203_cast_fp16 = softmax(axis = var_1056, x = aw_151_cast_fp16)[name = tensor("op_1203_cast_fp16")]; + tensor var_1204_cast_fp16 = softmax(axis = var_1056, x = aw_153_cast_fp16)[name = tensor("op_1204_cast_fp16")]; + tensor var_1205_cast_fp16 = softmax(axis = var_1056, x = aw_155_cast_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1206_cast_fp16 = softmax(axis = var_1056, x = aw_157_cast_fp16)[name = tensor("op_1206_cast_fp16")]; + tensor var_1207_cast_fp16 = softmax(axis = var_1056, x = aw_159_cast_fp16)[name = tensor("op_1207_cast_fp16")]; + tensor var_1209_equation_0 = const()[name = tensor("op_1209_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1209_cast_fp16 = einsum(equation = var_1209_equation_0, values = (var_1143_cast_fp16_0, var_1192_cast_fp16))[name = tensor("op_1209_cast_fp16")]; + tensor var_1211_equation_0 = const()[name = tensor("op_1211_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1211_cast_fp16 = einsum(equation = var_1211_equation_0, values = (var_1143_cast_fp16_1, var_1193_cast_fp16))[name = tensor("op_1211_cast_fp16")]; + tensor var_1213_equation_0 = const()[name = tensor("op_1213_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1213_cast_fp16 = einsum(equation = var_1213_equation_0, values = (var_1143_cast_fp16_2, var_1194_cast_fp16))[name = tensor("op_1213_cast_fp16")]; + tensor var_1215_equation_0 = const()[name = tensor("op_1215_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1215_cast_fp16 = einsum(equation = var_1215_equation_0, values = (var_1143_cast_fp16_3, var_1195_cast_fp16))[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_equation_0 = const()[name = tensor("op_1217_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1217_cast_fp16 = einsum(equation = var_1217_equation_0, values = (var_1143_cast_fp16_4, var_1196_cast_fp16))[name = tensor("op_1217_cast_fp16")]; + tensor var_1219_equation_0 = const()[name = tensor("op_1219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1219_cast_fp16 = einsum(equation = var_1219_equation_0, values = (var_1143_cast_fp16_5, var_1197_cast_fp16))[name = tensor("op_1219_cast_fp16")]; + tensor var_1221_equation_0 = const()[name = tensor("op_1221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1221_cast_fp16 = einsum(equation = var_1221_equation_0, values = (var_1143_cast_fp16_6, var_1198_cast_fp16))[name = tensor("op_1221_cast_fp16")]; + tensor var_1223_equation_0 = const()[name = tensor("op_1223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1223_cast_fp16 = einsum(equation = var_1223_equation_0, values = (var_1143_cast_fp16_7, var_1199_cast_fp16))[name = tensor("op_1223_cast_fp16")]; + tensor var_1225_equation_0 = const()[name = tensor("op_1225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1225_cast_fp16 = einsum(equation = var_1225_equation_0, values = (var_1143_cast_fp16_8, var_1200_cast_fp16))[name = tensor("op_1225_cast_fp16")]; + tensor var_1227_equation_0 = const()[name = tensor("op_1227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1227_cast_fp16 = einsum(equation = var_1227_equation_0, values = (var_1143_cast_fp16_9, var_1201_cast_fp16))[name = tensor("op_1227_cast_fp16")]; + tensor var_1229_equation_0 = const()[name = tensor("op_1229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1229_cast_fp16 = einsum(equation = var_1229_equation_0, values = (var_1143_cast_fp16_10, var_1202_cast_fp16))[name = tensor("op_1229_cast_fp16")]; + tensor var_1231_equation_0 = const()[name = tensor("op_1231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1231_cast_fp16 = einsum(equation = var_1231_equation_0, values = (var_1143_cast_fp16_11, var_1203_cast_fp16))[name = tensor("op_1231_cast_fp16")]; + tensor var_1233_equation_0 = const()[name = tensor("op_1233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1233_cast_fp16 = einsum(equation = var_1233_equation_0, values = (var_1143_cast_fp16_12, var_1204_cast_fp16))[name = tensor("op_1233_cast_fp16")]; + tensor var_1235_equation_0 = const()[name = tensor("op_1235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1235_cast_fp16 = einsum(equation = var_1235_equation_0, values = (var_1143_cast_fp16_13, var_1205_cast_fp16))[name = tensor("op_1235_cast_fp16")]; + tensor var_1237_equation_0 = const()[name = tensor("op_1237_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1237_cast_fp16 = einsum(equation = var_1237_equation_0, values = (var_1143_cast_fp16_14, var_1206_cast_fp16))[name = tensor("op_1237_cast_fp16")]; + tensor var_1239_equation_0 = const()[name = tensor("op_1239_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1239_cast_fp16 = einsum(equation = var_1239_equation_0, values = (var_1143_cast_fp16_15, var_1207_cast_fp16))[name = tensor("op_1239_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_1056, interleave = input_45_interleave_0, values = (var_1209_cast_fp16, var_1211_cast_fp16, var_1213_cast_fp16, var_1215_cast_fp16, var_1217_cast_fp16, var_1219_cast_fp16, var_1221_cast_fp16, var_1223_cast_fp16, var_1225_cast_fp16, var_1227_cast_fp16, var_1229_cast_fp16, var_1231_cast_fp16, var_1233_cast_fp16, var_1235_cast_fp16, var_1237_cast_fp16, var_1239_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1248_pad_type_0 = const()[name = tensor("op_1248_pad_type_0"), val = tensor("valid")]; + tensor var_1248_strides_0 = const()[name = tensor("op_1248_strides_0"), val = tensor([1, 1])]; + tensor var_1248_pad_0 = const()[name = tensor("op_1248_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1248_dilations_0 = const()[name = tensor("op_1248_dilations_0"), val = tensor([1, 1])]; + tensor var_1248_groups_0 = const()[name = tensor("op_1248_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116924992)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119022208)))]; + tensor var_1248_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1248_dilations_0, groups = var_1248_groups_0, pad = var_1248_pad_0, pad_type = var_1248_pad_type_0, strides = var_1248_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1248_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1248_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119024320)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119026432)))]; + tensor var_1258_to_fp16 = const()[name = tensor("op_1258_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1258_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119028544)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127417216)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1284_pad_type_0 = const()[name = tensor("op_1284_pad_type_0"), val = tensor("valid")]; + tensor var_1284_strides_0 = const()[name = tensor("op_1284_strides_0"), val = tensor([1, 1])]; + tensor var_1284_pad_0 = const()[name = tensor("op_1284_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1284_dilations_0 = const()[name = tensor("op_1284_dilations_0"), val = tensor([1, 1])]; + tensor var_1284_groups_0 = const()[name = tensor("op_1284_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127425472)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135814144)))]; + tensor var_1284_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1284_dilations_0, groups = var_1284_groups_0, pad = var_1284_pad_0, pad_type = var_1284_pad_type_0, strides = var_1284_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1284_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1284_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135816256)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135818368)))]; + tensor var_1309_to_fp16 = const()[name = tensor("op_1309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1309_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1344_weight_0_to_fp16 = const()[name = tensor("op_1344_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135820480)))]; + tensor var_1344_bias_0_to_fp16 = const()[name = tensor("op_1344_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137917696)))]; + tensor var_1344_cast_fp16 = conv(bias = var_1344_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1344_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137919808)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1342_pad_type_0 = const()[name = tensor("op_1342_pad_type_0"), val = tensor("valid")]; + tensor var_1342_strides_0 = const()[name = tensor("op_1342_strides_0"), val = tensor([1, 1])]; + tensor var_1342_pad_0 = const()[name = tensor("op_1342_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1342_dilations_0 = const()[name = tensor("op_1342_dilations_0"), val = tensor([1, 1])]; + tensor var_1342_groups_0 = const()[name = tensor("op_1342_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140017024)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142114240)))]; + tensor var_1342_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1342_dilations_0, groups = var_1342_groups_0, pad = var_1342_pad_0, pad_type = var_1342_pad_type_0, strides = var_1342_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1345_axis_0 = const()[name = tensor("op_1345_axis_0"), val = tensor(1)]; + tensor var_1345_cast_fp16_0, tensor var_1345_cast_fp16_1, tensor var_1345_cast_fp16_2, tensor var_1345_cast_fp16_3, tensor var_1345_cast_fp16_4, tensor var_1345_cast_fp16_5, tensor var_1345_cast_fp16_6, tensor var_1345_cast_fp16_7, tensor var_1345_cast_fp16_8, tensor var_1345_cast_fp16_9, tensor var_1345_cast_fp16_10, tensor var_1345_cast_fp16_11, tensor var_1345_cast_fp16_12, tensor var_1345_cast_fp16_13, tensor var_1345_cast_fp16_14, tensor var_1345_cast_fp16_15 = split(axis = var_1345_axis_0, split_sizes = tile_15, x = var_1344_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor var_1362_perm_0 = const()[name = tensor("op_1362_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1363_axis_0 = const()[name = tensor("op_1363_axis_0"), val = tensor(3)]; + tensor var_1362_cast_fp16 = transpose(perm = var_1362_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_19")]; + tensor var_1363_cast_fp16_0, tensor var_1363_cast_fp16_1, tensor var_1363_cast_fp16_2, tensor var_1363_cast_fp16_3, tensor var_1363_cast_fp16_4, tensor var_1363_cast_fp16_5, tensor var_1363_cast_fp16_6, tensor var_1363_cast_fp16_7, tensor var_1363_cast_fp16_8, tensor var_1363_cast_fp16_9, tensor var_1363_cast_fp16_10, tensor var_1363_cast_fp16_11, tensor var_1363_cast_fp16_12, tensor var_1363_cast_fp16_13, tensor var_1363_cast_fp16_14, tensor var_1363_cast_fp16_15 = split(axis = var_1363_axis_0, split_sizes = tile_16, x = var_1362_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(1)]; + tensor var_1380_cast_fp16_0, tensor var_1380_cast_fp16_1, tensor var_1380_cast_fp16_2, tensor var_1380_cast_fp16_3, tensor var_1380_cast_fp16_4, tensor var_1380_cast_fp16_5, tensor var_1380_cast_fp16_6, tensor var_1380_cast_fp16_7, tensor var_1380_cast_fp16_8, tensor var_1380_cast_fp16_9, tensor var_1380_cast_fp16_10, tensor var_1380_cast_fp16_11, tensor var_1380_cast_fp16_12, tensor var_1380_cast_fp16_13, tensor var_1380_cast_fp16_14, tensor var_1380_cast_fp16_15 = split(axis = var_1380_axis_0, split_sizes = tile_17, x = var_1342_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1363_cast_fp16_0, var_1345_cast_fp16_0))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1363_cast_fp16_1, var_1345_cast_fp16_1))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1363_cast_fp16_2, var_1345_cast_fp16_2))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1363_cast_fp16_3, var_1345_cast_fp16_3))[name = tensor("aw_167_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1363_cast_fp16_4, var_1345_cast_fp16_4))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1363_cast_fp16_5, var_1345_cast_fp16_5))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1363_cast_fp16_6, var_1345_cast_fp16_6))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1363_cast_fp16_7, var_1345_cast_fp16_7))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1363_cast_fp16_8, var_1345_cast_fp16_8))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1363_cast_fp16_9, var_1345_cast_fp16_9))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1363_cast_fp16_10, var_1345_cast_fp16_10))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1363_cast_fp16_11, var_1345_cast_fp16_11))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1363_cast_fp16_12, var_1345_cast_fp16_12))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1363_cast_fp16_13, var_1345_cast_fp16_13))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1363_cast_fp16_14, var_1345_cast_fp16_14))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1363_cast_fp16_15, var_1345_cast_fp16_15))[name = tensor("aw_191_cast_fp16")]; + tensor var_1429_cast_fp16 = softmax(axis = var_1293, x = aw_161_cast_fp16)[name = tensor("op_1429_cast_fp16")]; + tensor var_1430_cast_fp16 = softmax(axis = var_1293, x = aw_163_cast_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor var_1431_cast_fp16 = softmax(axis = var_1293, x = aw_165_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1432_cast_fp16 = softmax(axis = var_1293, x = aw_167_cast_fp16)[name = tensor("op_1432_cast_fp16")]; + tensor var_1433_cast_fp16 = softmax(axis = var_1293, x = aw_169_cast_fp16)[name = tensor("op_1433_cast_fp16")]; + tensor var_1434_cast_fp16 = softmax(axis = var_1293, x = aw_171_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor var_1435_cast_fp16 = softmax(axis = var_1293, x = aw_173_cast_fp16)[name = tensor("op_1435_cast_fp16")]; + tensor var_1436_cast_fp16 = softmax(axis = var_1293, x = aw_175_cast_fp16)[name = tensor("op_1436_cast_fp16")]; + tensor var_1437_cast_fp16 = softmax(axis = var_1293, x = aw_177_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = softmax(axis = var_1293, x = aw_179_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor var_1439_cast_fp16 = softmax(axis = var_1293, x = aw_181_cast_fp16)[name = tensor("op_1439_cast_fp16")]; + tensor var_1440_cast_fp16 = softmax(axis = var_1293, x = aw_183_cast_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1441_cast_fp16 = softmax(axis = var_1293, x = aw_185_cast_fp16)[name = tensor("op_1441_cast_fp16")]; + tensor var_1442_cast_fp16 = softmax(axis = var_1293, x = aw_187_cast_fp16)[name = tensor("op_1442_cast_fp16")]; + tensor var_1443_cast_fp16 = softmax(axis = var_1293, x = aw_189_cast_fp16)[name = tensor("op_1443_cast_fp16")]; + tensor var_1444_cast_fp16 = softmax(axis = var_1293, x = aw_191_cast_fp16)[name = tensor("op_1444_cast_fp16")]; + tensor var_1446_equation_0 = const()[name = tensor("op_1446_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1446_cast_fp16 = einsum(equation = var_1446_equation_0, values = (var_1380_cast_fp16_0, var_1429_cast_fp16))[name = tensor("op_1446_cast_fp16")]; + tensor var_1448_equation_0 = const()[name = tensor("op_1448_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1448_cast_fp16 = einsum(equation = var_1448_equation_0, values = (var_1380_cast_fp16_1, var_1430_cast_fp16))[name = tensor("op_1448_cast_fp16")]; + tensor var_1450_equation_0 = const()[name = tensor("op_1450_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1450_cast_fp16 = einsum(equation = var_1450_equation_0, values = (var_1380_cast_fp16_2, var_1431_cast_fp16))[name = tensor("op_1450_cast_fp16")]; + tensor var_1452_equation_0 = const()[name = tensor("op_1452_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1452_cast_fp16 = einsum(equation = var_1452_equation_0, values = (var_1380_cast_fp16_3, var_1432_cast_fp16))[name = tensor("op_1452_cast_fp16")]; + tensor var_1454_equation_0 = const()[name = tensor("op_1454_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1454_cast_fp16 = einsum(equation = var_1454_equation_0, values = (var_1380_cast_fp16_4, var_1433_cast_fp16))[name = tensor("op_1454_cast_fp16")]; + tensor var_1456_equation_0 = const()[name = tensor("op_1456_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1456_cast_fp16 = einsum(equation = var_1456_equation_0, values = (var_1380_cast_fp16_5, var_1434_cast_fp16))[name = tensor("op_1456_cast_fp16")]; + tensor var_1458_equation_0 = const()[name = tensor("op_1458_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1458_cast_fp16 = einsum(equation = var_1458_equation_0, values = (var_1380_cast_fp16_6, var_1435_cast_fp16))[name = tensor("op_1458_cast_fp16")]; + tensor var_1460_equation_0 = const()[name = tensor("op_1460_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1460_cast_fp16 = einsum(equation = var_1460_equation_0, values = (var_1380_cast_fp16_7, var_1436_cast_fp16))[name = tensor("op_1460_cast_fp16")]; + tensor var_1462_equation_0 = const()[name = tensor("op_1462_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1462_cast_fp16 = einsum(equation = var_1462_equation_0, values = (var_1380_cast_fp16_8, var_1437_cast_fp16))[name = tensor("op_1462_cast_fp16")]; + tensor var_1464_equation_0 = const()[name = tensor("op_1464_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1464_cast_fp16 = einsum(equation = var_1464_equation_0, values = (var_1380_cast_fp16_9, var_1438_cast_fp16))[name = tensor("op_1464_cast_fp16")]; + tensor var_1466_equation_0 = const()[name = tensor("op_1466_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1466_cast_fp16 = einsum(equation = var_1466_equation_0, values = (var_1380_cast_fp16_10, var_1439_cast_fp16))[name = tensor("op_1466_cast_fp16")]; + tensor var_1468_equation_0 = const()[name = tensor("op_1468_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1468_cast_fp16 = einsum(equation = var_1468_equation_0, values = (var_1380_cast_fp16_11, var_1440_cast_fp16))[name = tensor("op_1468_cast_fp16")]; + tensor var_1470_equation_0 = const()[name = tensor("op_1470_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1470_cast_fp16 = einsum(equation = var_1470_equation_0, values = (var_1380_cast_fp16_12, var_1441_cast_fp16))[name = tensor("op_1470_cast_fp16")]; + tensor var_1472_equation_0 = const()[name = tensor("op_1472_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1472_cast_fp16 = einsum(equation = var_1472_equation_0, values = (var_1380_cast_fp16_13, var_1442_cast_fp16))[name = tensor("op_1472_cast_fp16")]; + tensor var_1474_equation_0 = const()[name = tensor("op_1474_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1474_cast_fp16 = einsum(equation = var_1474_equation_0, values = (var_1380_cast_fp16_14, var_1443_cast_fp16))[name = tensor("op_1474_cast_fp16")]; + tensor var_1476_equation_0 = const()[name = tensor("op_1476_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1476_cast_fp16 = einsum(equation = var_1476_equation_0, values = (var_1380_cast_fp16_15, var_1444_cast_fp16))[name = tensor("op_1476_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1293, interleave = input_55_interleave_0, values = (var_1446_cast_fp16, var_1448_cast_fp16, var_1450_cast_fp16, var_1452_cast_fp16, var_1454_cast_fp16, var_1456_cast_fp16, var_1458_cast_fp16, var_1460_cast_fp16, var_1462_cast_fp16, var_1464_cast_fp16, var_1466_cast_fp16, var_1468_cast_fp16, var_1470_cast_fp16, var_1472_cast_fp16, var_1474_cast_fp16, var_1476_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1485_pad_type_0 = const()[name = tensor("op_1485_pad_type_0"), val = tensor("valid")]; + tensor var_1485_strides_0 = const()[name = tensor("op_1485_strides_0"), val = tensor([1, 1])]; + tensor var_1485_pad_0 = const()[name = tensor("op_1485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1485_dilations_0 = const()[name = tensor("op_1485_dilations_0"), val = tensor([1, 1])]; + tensor var_1485_groups_0 = const()[name = tensor("op_1485_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142116352)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144213568)))]; + tensor var_1485_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1485_dilations_0, groups = var_1485_groups_0, pad = var_1485_pad_0, pad_type = var_1485_pad_type_0, strides = var_1485_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1485_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144215680)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144217792)))]; + tensor var_1495_to_fp16 = const()[name = tensor("op_1495_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1495_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144219904)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152608576)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1521_pad_type_0 = const()[name = tensor("op_1521_pad_type_0"), val = tensor("valid")]; + tensor var_1521_strides_0 = const()[name = tensor("op_1521_strides_0"), val = tensor([1, 1])]; + tensor var_1521_pad_0 = const()[name = tensor("op_1521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1521_dilations_0 = const()[name = tensor("op_1521_dilations_0"), val = tensor([1, 1])]; + tensor var_1521_groups_0 = const()[name = tensor("op_1521_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152616832)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161005504)))]; + tensor var_1521_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1521_dilations_0, groups = var_1521_groups_0, pad = var_1521_pad_0, pad_type = var_1521_pad_type_0, strides = var_1521_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1521_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1521_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161007616)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161009728)))]; + tensor var_1546_to_fp16 = const()[name = tensor("op_1546_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1546_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1581_weight_0_to_fp16 = const()[name = tensor("op_1581_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161011840)))]; + tensor var_1581_bias_0_to_fp16 = const()[name = tensor("op_1581_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163109056)))]; + tensor var_1581_cast_fp16 = conv(bias = var_1581_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1581_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1581_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163111168)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1579_pad_type_0 = const()[name = tensor("op_1579_pad_type_0"), val = tensor("valid")]; + tensor var_1579_strides_0 = const()[name = tensor("op_1579_strides_0"), val = tensor([1, 1])]; + tensor var_1579_pad_0 = const()[name = tensor("op_1579_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1579_dilations_0 = const()[name = tensor("op_1579_dilations_0"), val = tensor([1, 1])]; + tensor var_1579_groups_0 = const()[name = tensor("op_1579_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165208384)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167305600)))]; + tensor var_1579_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1579_dilations_0, groups = var_1579_groups_0, pad = var_1579_pad_0, pad_type = var_1579_pad_type_0, strides = var_1579_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1582_axis_0 = const()[name = tensor("op_1582_axis_0"), val = tensor(1)]; + tensor var_1582_cast_fp16_0, tensor var_1582_cast_fp16_1, tensor var_1582_cast_fp16_2, tensor var_1582_cast_fp16_3, tensor var_1582_cast_fp16_4, tensor var_1582_cast_fp16_5, tensor var_1582_cast_fp16_6, tensor var_1582_cast_fp16_7, tensor var_1582_cast_fp16_8, tensor var_1582_cast_fp16_9, tensor var_1582_cast_fp16_10, tensor var_1582_cast_fp16_11, tensor var_1582_cast_fp16_12, tensor var_1582_cast_fp16_13, tensor var_1582_cast_fp16_14, tensor var_1582_cast_fp16_15 = split(axis = var_1582_axis_0, split_sizes = tile_18, x = var_1581_cast_fp16)[name = tensor("op_1582_cast_fp16")]; + tensor var_1599_perm_0 = const()[name = tensor("op_1599_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1600_axis_0 = const()[name = tensor("op_1600_axis_0"), val = tensor(3)]; + tensor var_1599_cast_fp16 = transpose(perm = var_1599_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_18")]; + tensor var_1600_cast_fp16_0, tensor var_1600_cast_fp16_1, tensor var_1600_cast_fp16_2, tensor var_1600_cast_fp16_3, tensor var_1600_cast_fp16_4, tensor var_1600_cast_fp16_5, tensor var_1600_cast_fp16_6, tensor var_1600_cast_fp16_7, tensor var_1600_cast_fp16_8, tensor var_1600_cast_fp16_9, tensor var_1600_cast_fp16_10, tensor var_1600_cast_fp16_11, tensor var_1600_cast_fp16_12, tensor var_1600_cast_fp16_13, tensor var_1600_cast_fp16_14, tensor var_1600_cast_fp16_15 = split(axis = var_1600_axis_0, split_sizes = tile_19, x = var_1599_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1617_axis_0 = const()[name = tensor("op_1617_axis_0"), val = tensor(1)]; + tensor var_1617_cast_fp16_0, tensor var_1617_cast_fp16_1, tensor var_1617_cast_fp16_2, tensor var_1617_cast_fp16_3, tensor var_1617_cast_fp16_4, tensor var_1617_cast_fp16_5, tensor var_1617_cast_fp16_6, tensor var_1617_cast_fp16_7, tensor var_1617_cast_fp16_8, tensor var_1617_cast_fp16_9, tensor var_1617_cast_fp16_10, tensor var_1617_cast_fp16_11, tensor var_1617_cast_fp16_12, tensor var_1617_cast_fp16_13, tensor var_1617_cast_fp16_14, tensor var_1617_cast_fp16_15 = split(axis = var_1617_axis_0, split_sizes = tile_20, x = var_1579_cast_fp16)[name = tensor("op_1617_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1600_cast_fp16_0, var_1582_cast_fp16_0))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1600_cast_fp16_1, var_1582_cast_fp16_1))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1600_cast_fp16_2, var_1582_cast_fp16_2))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1600_cast_fp16_3, var_1582_cast_fp16_3))[name = tensor("aw_199_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1600_cast_fp16_4, var_1582_cast_fp16_4))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1600_cast_fp16_5, var_1582_cast_fp16_5))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1600_cast_fp16_6, var_1582_cast_fp16_6))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1600_cast_fp16_7, var_1582_cast_fp16_7))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1600_cast_fp16_8, var_1582_cast_fp16_8))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1600_cast_fp16_9, var_1582_cast_fp16_9))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1600_cast_fp16_10, var_1582_cast_fp16_10))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1600_cast_fp16_11, var_1582_cast_fp16_11))[name = tensor("aw_215_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1600_cast_fp16_12, var_1582_cast_fp16_12))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1600_cast_fp16_13, var_1582_cast_fp16_13))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1600_cast_fp16_14, var_1582_cast_fp16_14))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1600_cast_fp16_15, var_1582_cast_fp16_15))[name = tensor("aw_223_cast_fp16")]; + tensor var_1666_cast_fp16 = softmax(axis = var_1530, x = aw_193_cast_fp16)[name = tensor("op_1666_cast_fp16")]; + tensor var_1667_cast_fp16 = softmax(axis = var_1530, x = aw_195_cast_fp16)[name = tensor("op_1667_cast_fp16")]; + tensor var_1668_cast_fp16 = softmax(axis = var_1530, x = aw_197_cast_fp16)[name = tensor("op_1668_cast_fp16")]; + tensor var_1669_cast_fp16 = softmax(axis = var_1530, x = aw_199_cast_fp16)[name = tensor("op_1669_cast_fp16")]; + tensor var_1670_cast_fp16 = softmax(axis = var_1530, x = aw_201_cast_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1671_cast_fp16 = softmax(axis = var_1530, x = aw_203_cast_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1672_cast_fp16 = softmax(axis = var_1530, x = aw_205_cast_fp16)[name = tensor("op_1672_cast_fp16")]; + tensor var_1673_cast_fp16 = softmax(axis = var_1530, x = aw_207_cast_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1674_cast_fp16 = softmax(axis = var_1530, x = aw_209_cast_fp16)[name = tensor("op_1674_cast_fp16")]; + tensor var_1675_cast_fp16 = softmax(axis = var_1530, x = aw_211_cast_fp16)[name = tensor("op_1675_cast_fp16")]; + tensor var_1676_cast_fp16 = softmax(axis = var_1530, x = aw_213_cast_fp16)[name = tensor("op_1676_cast_fp16")]; + tensor var_1677_cast_fp16 = softmax(axis = var_1530, x = aw_215_cast_fp16)[name = tensor("op_1677_cast_fp16")]; + tensor var_1678_cast_fp16 = softmax(axis = var_1530, x = aw_217_cast_fp16)[name = tensor("op_1678_cast_fp16")]; + tensor var_1679_cast_fp16 = softmax(axis = var_1530, x = aw_219_cast_fp16)[name = tensor("op_1679_cast_fp16")]; + tensor var_1680_cast_fp16 = softmax(axis = var_1530, x = aw_221_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor var_1681_cast_fp16 = softmax(axis = var_1530, x = aw_223_cast_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1683_equation_0 = const()[name = tensor("op_1683_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1683_cast_fp16 = einsum(equation = var_1683_equation_0, values = (var_1617_cast_fp16_0, var_1666_cast_fp16))[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_equation_0 = const()[name = tensor("op_1685_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1685_cast_fp16 = einsum(equation = var_1685_equation_0, values = (var_1617_cast_fp16_1, var_1667_cast_fp16))[name = tensor("op_1685_cast_fp16")]; + tensor var_1687_equation_0 = const()[name = tensor("op_1687_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1687_cast_fp16 = einsum(equation = var_1687_equation_0, values = (var_1617_cast_fp16_2, var_1668_cast_fp16))[name = tensor("op_1687_cast_fp16")]; + tensor var_1689_equation_0 = const()[name = tensor("op_1689_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1689_cast_fp16 = einsum(equation = var_1689_equation_0, values = (var_1617_cast_fp16_3, var_1669_cast_fp16))[name = tensor("op_1689_cast_fp16")]; + tensor var_1691_equation_0 = const()[name = tensor("op_1691_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1691_cast_fp16 = einsum(equation = var_1691_equation_0, values = (var_1617_cast_fp16_4, var_1670_cast_fp16))[name = tensor("op_1691_cast_fp16")]; + tensor var_1693_equation_0 = const()[name = tensor("op_1693_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1693_cast_fp16 = einsum(equation = var_1693_equation_0, values = (var_1617_cast_fp16_5, var_1671_cast_fp16))[name = tensor("op_1693_cast_fp16")]; + tensor var_1695_equation_0 = const()[name = tensor("op_1695_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1695_cast_fp16 = einsum(equation = var_1695_equation_0, values = (var_1617_cast_fp16_6, var_1672_cast_fp16))[name = tensor("op_1695_cast_fp16")]; + tensor var_1697_equation_0 = const()[name = tensor("op_1697_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1697_cast_fp16 = einsum(equation = var_1697_equation_0, values = (var_1617_cast_fp16_7, var_1673_cast_fp16))[name = tensor("op_1697_cast_fp16")]; + tensor var_1699_equation_0 = const()[name = tensor("op_1699_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1699_cast_fp16 = einsum(equation = var_1699_equation_0, values = (var_1617_cast_fp16_8, var_1674_cast_fp16))[name = tensor("op_1699_cast_fp16")]; + tensor var_1701_equation_0 = const()[name = tensor("op_1701_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1701_cast_fp16 = einsum(equation = var_1701_equation_0, values = (var_1617_cast_fp16_9, var_1675_cast_fp16))[name = tensor("op_1701_cast_fp16")]; + tensor var_1703_equation_0 = const()[name = tensor("op_1703_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1703_cast_fp16 = einsum(equation = var_1703_equation_0, values = (var_1617_cast_fp16_10, var_1676_cast_fp16))[name = tensor("op_1703_cast_fp16")]; + tensor var_1705_equation_0 = const()[name = tensor("op_1705_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1705_cast_fp16 = einsum(equation = var_1705_equation_0, values = (var_1617_cast_fp16_11, var_1677_cast_fp16))[name = tensor("op_1705_cast_fp16")]; + tensor var_1707_equation_0 = const()[name = tensor("op_1707_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1707_cast_fp16 = einsum(equation = var_1707_equation_0, values = (var_1617_cast_fp16_12, var_1678_cast_fp16))[name = tensor("op_1707_cast_fp16")]; + tensor var_1709_equation_0 = const()[name = tensor("op_1709_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1709_cast_fp16 = einsum(equation = var_1709_equation_0, values = (var_1617_cast_fp16_13, var_1679_cast_fp16))[name = tensor("op_1709_cast_fp16")]; + tensor var_1711_equation_0 = const()[name = tensor("op_1711_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1711_cast_fp16 = einsum(equation = var_1711_equation_0, values = (var_1617_cast_fp16_14, var_1680_cast_fp16))[name = tensor("op_1711_cast_fp16")]; + tensor var_1713_equation_0 = const()[name = tensor("op_1713_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1713_cast_fp16 = einsum(equation = var_1713_equation_0, values = (var_1617_cast_fp16_15, var_1681_cast_fp16))[name = tensor("op_1713_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1530, interleave = input_65_interleave_0, values = (var_1683_cast_fp16, var_1685_cast_fp16, var_1687_cast_fp16, var_1689_cast_fp16, var_1691_cast_fp16, var_1693_cast_fp16, var_1695_cast_fp16, var_1697_cast_fp16, var_1699_cast_fp16, var_1701_cast_fp16, var_1703_cast_fp16, var_1705_cast_fp16, var_1707_cast_fp16, var_1709_cast_fp16, var_1711_cast_fp16, var_1713_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1722_pad_type_0 = const()[name = tensor("op_1722_pad_type_0"), val = tensor("valid")]; + tensor var_1722_strides_0 = const()[name = tensor("op_1722_strides_0"), val = tensor([1, 1])]; + tensor var_1722_pad_0 = const()[name = tensor("op_1722_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1722_dilations_0 = const()[name = tensor("op_1722_dilations_0"), val = tensor([1, 1])]; + tensor var_1722_groups_0 = const()[name = tensor("op_1722_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167307712)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169404928)))]; + tensor var_1722_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1722_dilations_0, groups = var_1722_groups_0, pad = var_1722_pad_0, pad_type = var_1722_pad_type_0, strides = var_1722_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1722_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1722_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169407040)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169409152)))]; + tensor var_1732_to_fp16 = const()[name = tensor("op_1732_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1732_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169411264)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177799936)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1758_pad_type_0 = const()[name = tensor("op_1758_pad_type_0"), val = tensor("valid")]; + tensor var_1758_strides_0 = const()[name = tensor("op_1758_strides_0"), val = tensor([1, 1])]; + tensor var_1758_pad_0 = const()[name = tensor("op_1758_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1758_dilations_0 = const()[name = tensor("op_1758_dilations_0"), val = tensor([1, 1])]; + tensor var_1758_groups_0 = const()[name = tensor("op_1758_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177808192)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186196864)))]; + tensor var_1758_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1758_dilations_0, groups = var_1758_groups_0, pad = var_1758_pad_0, pad_type = var_1758_pad_type_0, strides = var_1758_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1758_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1758_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1767 = const()[name = tensor("op_1767"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186198976)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186201088)))]; + tensor var_1783_to_fp16 = const()[name = tensor("op_1783_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_1783_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_1818_weight_0_to_fp16 = const()[name = tensor("op_1818_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186203200)))]; + tensor var_1818_bias_0_to_fp16 = const()[name = tensor("op_1818_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188300416)))]; + tensor var_1818_cast_fp16 = conv(bias = var_1818_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_1818_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1818_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188302528)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_1816_pad_type_0 = const()[name = tensor("op_1816_pad_type_0"), val = tensor("valid")]; + tensor var_1816_strides_0 = const()[name = tensor("op_1816_strides_0"), val = tensor([1, 1])]; + tensor var_1816_pad_0 = const()[name = tensor("op_1816_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1816_dilations_0 = const()[name = tensor("op_1816_dilations_0"), val = tensor([1, 1])]; + tensor var_1816_groups_0 = const()[name = tensor("op_1816_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190399744)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192496960)))]; + tensor var_1816_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_1816_dilations_0, groups = var_1816_groups_0, pad = var_1816_pad_0, pad_type = var_1816_pad_type_0, strides = var_1816_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1819_axis_0 = const()[name = tensor("op_1819_axis_0"), val = tensor(1)]; + tensor var_1819_cast_fp16_0, tensor var_1819_cast_fp16_1, tensor var_1819_cast_fp16_2, tensor var_1819_cast_fp16_3, tensor var_1819_cast_fp16_4, tensor var_1819_cast_fp16_5, tensor var_1819_cast_fp16_6, tensor var_1819_cast_fp16_7, tensor var_1819_cast_fp16_8, tensor var_1819_cast_fp16_9, tensor var_1819_cast_fp16_10, tensor var_1819_cast_fp16_11, tensor var_1819_cast_fp16_12, tensor var_1819_cast_fp16_13, tensor var_1819_cast_fp16_14, tensor var_1819_cast_fp16_15 = split(axis = var_1819_axis_0, split_sizes = tile_21, x = var_1818_cast_fp16)[name = tensor("op_1819_cast_fp16")]; + tensor var_1836_perm_0 = const()[name = tensor("op_1836_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1837_axis_0 = const()[name = tensor("op_1837_axis_0"), val = tensor(3)]; + tensor var_1836_cast_fp16 = transpose(perm = var_1836_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_17")]; + tensor var_1837_cast_fp16_0, tensor var_1837_cast_fp16_1, tensor var_1837_cast_fp16_2, tensor var_1837_cast_fp16_3, tensor var_1837_cast_fp16_4, tensor var_1837_cast_fp16_5, tensor var_1837_cast_fp16_6, tensor var_1837_cast_fp16_7, tensor var_1837_cast_fp16_8, tensor var_1837_cast_fp16_9, tensor var_1837_cast_fp16_10, tensor var_1837_cast_fp16_11, tensor var_1837_cast_fp16_12, tensor var_1837_cast_fp16_13, tensor var_1837_cast_fp16_14, tensor var_1837_cast_fp16_15 = split(axis = var_1837_axis_0, split_sizes = tile_22, x = var_1836_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1854_axis_0 = const()[name = tensor("op_1854_axis_0"), val = tensor(1)]; + tensor var_1854_cast_fp16_0, tensor var_1854_cast_fp16_1, tensor var_1854_cast_fp16_2, tensor var_1854_cast_fp16_3, tensor var_1854_cast_fp16_4, tensor var_1854_cast_fp16_5, tensor var_1854_cast_fp16_6, tensor var_1854_cast_fp16_7, tensor var_1854_cast_fp16_8, tensor var_1854_cast_fp16_9, tensor var_1854_cast_fp16_10, tensor var_1854_cast_fp16_11, tensor var_1854_cast_fp16_12, tensor var_1854_cast_fp16_13, tensor var_1854_cast_fp16_14, tensor var_1854_cast_fp16_15 = split(axis = var_1854_axis_0, split_sizes = tile_23, x = var_1816_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1837_cast_fp16_0, var_1819_cast_fp16_0))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1837_cast_fp16_1, var_1819_cast_fp16_1))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1837_cast_fp16_2, var_1819_cast_fp16_2))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1837_cast_fp16_3, var_1819_cast_fp16_3))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1837_cast_fp16_4, var_1819_cast_fp16_4))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1837_cast_fp16_5, var_1819_cast_fp16_5))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1837_cast_fp16_6, var_1819_cast_fp16_6))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1837_cast_fp16_7, var_1819_cast_fp16_7))[name = tensor("aw_239_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_1837_cast_fp16_8, var_1819_cast_fp16_8))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_1837_cast_fp16_9, var_1819_cast_fp16_9))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_1837_cast_fp16_10, var_1819_cast_fp16_10))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_1837_cast_fp16_11, var_1819_cast_fp16_11))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_1837_cast_fp16_12, var_1819_cast_fp16_12))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_1837_cast_fp16_13, var_1819_cast_fp16_13))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_1837_cast_fp16_14, var_1819_cast_fp16_14))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_1837_cast_fp16_15, var_1819_cast_fp16_15))[name = tensor("aw_255_cast_fp16")]; + tensor var_1903_cast_fp16 = softmax(axis = var_1767, x = aw_225_cast_fp16)[name = tensor("op_1903_cast_fp16")]; + tensor var_1904_cast_fp16 = softmax(axis = var_1767, x = aw_227_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1905_cast_fp16 = softmax(axis = var_1767, x = aw_229_cast_fp16)[name = tensor("op_1905_cast_fp16")]; + tensor var_1906_cast_fp16 = softmax(axis = var_1767, x = aw_231_cast_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1907_cast_fp16 = softmax(axis = var_1767, x = aw_233_cast_fp16)[name = tensor("op_1907_cast_fp16")]; + tensor var_1908_cast_fp16 = softmax(axis = var_1767, x = aw_235_cast_fp16)[name = tensor("op_1908_cast_fp16")]; + tensor var_1909_cast_fp16 = softmax(axis = var_1767, x = aw_237_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor var_1910_cast_fp16 = softmax(axis = var_1767, x = aw_239_cast_fp16)[name = tensor("op_1910_cast_fp16")]; + tensor var_1911_cast_fp16 = softmax(axis = var_1767, x = aw_241_cast_fp16)[name = tensor("op_1911_cast_fp16")]; + tensor var_1912_cast_fp16 = softmax(axis = var_1767, x = aw_243_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1913_cast_fp16 = softmax(axis = var_1767, x = aw_245_cast_fp16)[name = tensor("op_1913_cast_fp16")]; + tensor var_1914_cast_fp16 = softmax(axis = var_1767, x = aw_247_cast_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1915_cast_fp16 = softmax(axis = var_1767, x = aw_249_cast_fp16)[name = tensor("op_1915_cast_fp16")]; + tensor var_1916_cast_fp16 = softmax(axis = var_1767, x = aw_251_cast_fp16)[name = tensor("op_1916_cast_fp16")]; + tensor var_1917_cast_fp16 = softmax(axis = var_1767, x = aw_253_cast_fp16)[name = tensor("op_1917_cast_fp16")]; + tensor var_1918_cast_fp16 = softmax(axis = var_1767, x = aw_255_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor var_1920_equation_0 = const()[name = tensor("op_1920_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1920_cast_fp16 = einsum(equation = var_1920_equation_0, values = (var_1854_cast_fp16_0, var_1903_cast_fp16))[name = tensor("op_1920_cast_fp16")]; + tensor var_1922_equation_0 = const()[name = tensor("op_1922_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1922_cast_fp16 = einsum(equation = var_1922_equation_0, values = (var_1854_cast_fp16_1, var_1904_cast_fp16))[name = tensor("op_1922_cast_fp16")]; + tensor var_1924_equation_0 = const()[name = tensor("op_1924_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1924_cast_fp16 = einsum(equation = var_1924_equation_0, values = (var_1854_cast_fp16_2, var_1905_cast_fp16))[name = tensor("op_1924_cast_fp16")]; + tensor var_1926_equation_0 = const()[name = tensor("op_1926_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1926_cast_fp16 = einsum(equation = var_1926_equation_0, values = (var_1854_cast_fp16_3, var_1906_cast_fp16))[name = tensor("op_1926_cast_fp16")]; + tensor var_1928_equation_0 = const()[name = tensor("op_1928_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1928_cast_fp16 = einsum(equation = var_1928_equation_0, values = (var_1854_cast_fp16_4, var_1907_cast_fp16))[name = tensor("op_1928_cast_fp16")]; + tensor var_1930_equation_0 = const()[name = tensor("op_1930_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1930_cast_fp16 = einsum(equation = var_1930_equation_0, values = (var_1854_cast_fp16_5, var_1908_cast_fp16))[name = tensor("op_1930_cast_fp16")]; + tensor var_1932_equation_0 = const()[name = tensor("op_1932_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1932_cast_fp16 = einsum(equation = var_1932_equation_0, values = (var_1854_cast_fp16_6, var_1909_cast_fp16))[name = tensor("op_1932_cast_fp16")]; + tensor var_1934_equation_0 = const()[name = tensor("op_1934_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1934_cast_fp16 = einsum(equation = var_1934_equation_0, values = (var_1854_cast_fp16_7, var_1910_cast_fp16))[name = tensor("op_1934_cast_fp16")]; + tensor var_1936_equation_0 = const()[name = tensor("op_1936_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1936_cast_fp16 = einsum(equation = var_1936_equation_0, values = (var_1854_cast_fp16_8, var_1911_cast_fp16))[name = tensor("op_1936_cast_fp16")]; + tensor var_1938_equation_0 = const()[name = tensor("op_1938_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1938_cast_fp16 = einsum(equation = var_1938_equation_0, values = (var_1854_cast_fp16_9, var_1912_cast_fp16))[name = tensor("op_1938_cast_fp16")]; + tensor var_1940_equation_0 = const()[name = tensor("op_1940_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1940_cast_fp16 = einsum(equation = var_1940_equation_0, values = (var_1854_cast_fp16_10, var_1913_cast_fp16))[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_equation_0 = const()[name = tensor("op_1942_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1942_cast_fp16 = einsum(equation = var_1942_equation_0, values = (var_1854_cast_fp16_11, var_1914_cast_fp16))[name = tensor("op_1942_cast_fp16")]; + tensor var_1944_equation_0 = const()[name = tensor("op_1944_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1944_cast_fp16 = einsum(equation = var_1944_equation_0, values = (var_1854_cast_fp16_12, var_1915_cast_fp16))[name = tensor("op_1944_cast_fp16")]; + tensor var_1946_equation_0 = const()[name = tensor("op_1946_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1946_cast_fp16 = einsum(equation = var_1946_equation_0, values = (var_1854_cast_fp16_13, var_1916_cast_fp16))[name = tensor("op_1946_cast_fp16")]; + tensor var_1948_equation_0 = const()[name = tensor("op_1948_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1948_cast_fp16 = einsum(equation = var_1948_equation_0, values = (var_1854_cast_fp16_14, var_1917_cast_fp16))[name = tensor("op_1948_cast_fp16")]; + tensor var_1950_equation_0 = const()[name = tensor("op_1950_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1950_cast_fp16 = einsum(equation = var_1950_equation_0, values = (var_1854_cast_fp16_15, var_1918_cast_fp16))[name = tensor("op_1950_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_1767, interleave = input_75_interleave_0, values = (var_1920_cast_fp16, var_1922_cast_fp16, var_1924_cast_fp16, var_1926_cast_fp16, var_1928_cast_fp16, var_1930_cast_fp16, var_1932_cast_fp16, var_1934_cast_fp16, var_1936_cast_fp16, var_1938_cast_fp16, var_1940_cast_fp16, var_1942_cast_fp16, var_1944_cast_fp16, var_1946_cast_fp16, var_1948_cast_fp16, var_1950_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_1959_pad_type_0 = const()[name = tensor("op_1959_pad_type_0"), val = tensor("valid")]; + tensor var_1959_strides_0 = const()[name = tensor("op_1959_strides_0"), val = tensor([1, 1])]; + tensor var_1959_pad_0 = const()[name = tensor("op_1959_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1959_dilations_0 = const()[name = tensor("op_1959_dilations_0"), val = tensor([1, 1])]; + tensor var_1959_groups_0 = const()[name = tensor("op_1959_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192499072)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194596288)))]; + tensor var_1959_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_1959_dilations_0, groups = var_1959_groups_0, pad = var_1959_pad_0, pad_type = var_1959_pad_type_0, strides = var_1959_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_1959_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194598400)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194600512)))]; + tensor var_1969_to_fp16 = const()[name = tensor("op_1969_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_1969_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194602624)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202991296)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1995_pad_type_0 = const()[name = tensor("op_1995_pad_type_0"), val = tensor("valid")]; + tensor var_1995_strides_0 = const()[name = tensor("op_1995_strides_0"), val = tensor([1, 1])]; + tensor var_1995_pad_0 = const()[name = tensor("op_1995_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1995_dilations_0 = const()[name = tensor("op_1995_dilations_0"), val = tensor([1, 1])]; + tensor var_1995_groups_0 = const()[name = tensor("op_1995_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202999552)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211388224)))]; + tensor var_1995_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_1995_dilations_0, groups = var_1995_groups_0, pad = var_1995_pad_0, pad_type = var_1995_pad_type_0, strides = var_1995_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1995_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211390336)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211392448)))]; + tensor var_2020_to_fp16 = const()[name = tensor("op_2020_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_2020_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_2055_weight_0_to_fp16 = const()[name = tensor("op_2055_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211394560)))]; + tensor var_2055_bias_0_to_fp16 = const()[name = tensor("op_2055_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213491776)))]; + tensor var_2055_cast_fp16 = conv(bias = var_2055_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_2055_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2055_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213493888)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_2053_pad_type_0 = const()[name = tensor("op_2053_pad_type_0"), val = tensor("valid")]; + tensor var_2053_strides_0 = const()[name = tensor("op_2053_strides_0"), val = tensor([1, 1])]; + tensor var_2053_pad_0 = const()[name = tensor("op_2053_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2053_dilations_0 = const()[name = tensor("op_2053_dilations_0"), val = tensor([1, 1])]; + tensor var_2053_groups_0 = const()[name = tensor("op_2053_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215591104)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217688320)))]; + tensor var_2053_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_2053_dilations_0, groups = var_2053_groups_0, pad = var_2053_pad_0, pad_type = var_2053_pad_type_0, strides = var_2053_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2056_axis_0 = const()[name = tensor("op_2056_axis_0"), val = tensor(1)]; + tensor var_2056_cast_fp16_0, tensor var_2056_cast_fp16_1, tensor var_2056_cast_fp16_2, tensor var_2056_cast_fp16_3, tensor var_2056_cast_fp16_4, tensor var_2056_cast_fp16_5, tensor var_2056_cast_fp16_6, tensor var_2056_cast_fp16_7, tensor var_2056_cast_fp16_8, tensor var_2056_cast_fp16_9, tensor var_2056_cast_fp16_10, tensor var_2056_cast_fp16_11, tensor var_2056_cast_fp16_12, tensor var_2056_cast_fp16_13, tensor var_2056_cast_fp16_14, tensor var_2056_cast_fp16_15 = split(axis = var_2056_axis_0, split_sizes = tile_24, x = var_2055_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor var_2073_perm_0 = const()[name = tensor("op_2073_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2074_axis_0 = const()[name = tensor("op_2074_axis_0"), val = tensor(3)]; + tensor var_2073_cast_fp16 = transpose(perm = var_2073_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_16")]; + tensor var_2074_cast_fp16_0, tensor var_2074_cast_fp16_1, tensor var_2074_cast_fp16_2, tensor var_2074_cast_fp16_3, tensor var_2074_cast_fp16_4, tensor var_2074_cast_fp16_5, tensor var_2074_cast_fp16_6, tensor var_2074_cast_fp16_7, tensor var_2074_cast_fp16_8, tensor var_2074_cast_fp16_9, tensor var_2074_cast_fp16_10, tensor var_2074_cast_fp16_11, tensor var_2074_cast_fp16_12, tensor var_2074_cast_fp16_13, tensor var_2074_cast_fp16_14, tensor var_2074_cast_fp16_15 = split(axis = var_2074_axis_0, split_sizes = tile_25, x = var_2073_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2091_axis_0 = const()[name = tensor("op_2091_axis_0"), val = tensor(1)]; + tensor var_2091_cast_fp16_0, tensor var_2091_cast_fp16_1, tensor var_2091_cast_fp16_2, tensor var_2091_cast_fp16_3, tensor var_2091_cast_fp16_4, tensor var_2091_cast_fp16_5, tensor var_2091_cast_fp16_6, tensor var_2091_cast_fp16_7, tensor var_2091_cast_fp16_8, tensor var_2091_cast_fp16_9, tensor var_2091_cast_fp16_10, tensor var_2091_cast_fp16_11, tensor var_2091_cast_fp16_12, tensor var_2091_cast_fp16_13, tensor var_2091_cast_fp16_14, tensor var_2091_cast_fp16_15 = split(axis = var_2091_axis_0, split_sizes = tile_26, x = var_2053_cast_fp16)[name = tensor("op_2091_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_2074_cast_fp16_0, var_2056_cast_fp16_0))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_2074_cast_fp16_1, var_2056_cast_fp16_1))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_2074_cast_fp16_2, var_2056_cast_fp16_2))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_2074_cast_fp16_3, var_2056_cast_fp16_3))[name = tensor("aw_263_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_2074_cast_fp16_4, var_2056_cast_fp16_4))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_2074_cast_fp16_5, var_2056_cast_fp16_5))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_2074_cast_fp16_6, var_2056_cast_fp16_6))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_2074_cast_fp16_7, var_2056_cast_fp16_7))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_2074_cast_fp16_8, var_2056_cast_fp16_8))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_2074_cast_fp16_9, var_2056_cast_fp16_9))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_2074_cast_fp16_10, var_2056_cast_fp16_10))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_2074_cast_fp16_11, var_2056_cast_fp16_11))[name = tensor("aw_279_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2074_cast_fp16_12, var_2056_cast_fp16_12))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2074_cast_fp16_13, var_2056_cast_fp16_13))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2074_cast_fp16_14, var_2056_cast_fp16_14))[name = tensor("aw_285_cast_fp16")]; + tensor aw_287_equation_0 = const()[name = tensor("aw_287_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_287_cast_fp16 = einsum(equation = aw_287_equation_0, values = (var_2074_cast_fp16_15, var_2056_cast_fp16_15))[name = tensor("aw_287_cast_fp16")]; + tensor var_2140_cast_fp16 = softmax(axis = var_2004, x = aw_257_cast_fp16)[name = tensor("op_2140_cast_fp16")]; + tensor var_2141_cast_fp16 = softmax(axis = var_2004, x = aw_259_cast_fp16)[name = tensor("op_2141_cast_fp16")]; + tensor var_2142_cast_fp16 = softmax(axis = var_2004, x = aw_261_cast_fp16)[name = tensor("op_2142_cast_fp16")]; + tensor var_2143_cast_fp16 = softmax(axis = var_2004, x = aw_263_cast_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2144_cast_fp16 = softmax(axis = var_2004, x = aw_265_cast_fp16)[name = tensor("op_2144_cast_fp16")]; + tensor var_2145_cast_fp16 = softmax(axis = var_2004, x = aw_267_cast_fp16)[name = tensor("op_2145_cast_fp16")]; + tensor var_2146_cast_fp16 = softmax(axis = var_2004, x = aw_269_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor var_2147_cast_fp16 = softmax(axis = var_2004, x = aw_271_cast_fp16)[name = tensor("op_2147_cast_fp16")]; + tensor var_2148_cast_fp16 = softmax(axis = var_2004, x = aw_273_cast_fp16)[name = tensor("op_2148_cast_fp16")]; + tensor var_2149_cast_fp16 = softmax(axis = var_2004, x = aw_275_cast_fp16)[name = tensor("op_2149_cast_fp16")]; + tensor var_2150_cast_fp16 = softmax(axis = var_2004, x = aw_277_cast_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2151_cast_fp16 = softmax(axis = var_2004, x = aw_279_cast_fp16)[name = tensor("op_2151_cast_fp16")]; + tensor var_2152_cast_fp16 = softmax(axis = var_2004, x = aw_281_cast_fp16)[name = tensor("op_2152_cast_fp16")]; + tensor var_2153_cast_fp16 = softmax(axis = var_2004, x = aw_283_cast_fp16)[name = tensor("op_2153_cast_fp16")]; + tensor var_2154_cast_fp16 = softmax(axis = var_2004, x = aw_285_cast_fp16)[name = tensor("op_2154_cast_fp16")]; + tensor var_2155_cast_fp16 = softmax(axis = var_2004, x = aw_287_cast_fp16)[name = tensor("op_2155_cast_fp16")]; + tensor var_2157_equation_0 = const()[name = tensor("op_2157_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2157_cast_fp16 = einsum(equation = var_2157_equation_0, values = (var_2091_cast_fp16_0, var_2140_cast_fp16))[name = tensor("op_2157_cast_fp16")]; + tensor var_2159_equation_0 = const()[name = tensor("op_2159_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2159_cast_fp16 = einsum(equation = var_2159_equation_0, values = (var_2091_cast_fp16_1, var_2141_cast_fp16))[name = tensor("op_2159_cast_fp16")]; + tensor var_2161_equation_0 = const()[name = tensor("op_2161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2161_cast_fp16 = einsum(equation = var_2161_equation_0, values = (var_2091_cast_fp16_2, var_2142_cast_fp16))[name = tensor("op_2161_cast_fp16")]; + tensor var_2163_equation_0 = const()[name = tensor("op_2163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2163_cast_fp16 = einsum(equation = var_2163_equation_0, values = (var_2091_cast_fp16_3, var_2143_cast_fp16))[name = tensor("op_2163_cast_fp16")]; + tensor var_2165_equation_0 = const()[name = tensor("op_2165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2165_cast_fp16 = einsum(equation = var_2165_equation_0, values = (var_2091_cast_fp16_4, var_2144_cast_fp16))[name = tensor("op_2165_cast_fp16")]; + tensor var_2167_equation_0 = const()[name = tensor("op_2167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2167_cast_fp16 = einsum(equation = var_2167_equation_0, values = (var_2091_cast_fp16_5, var_2145_cast_fp16))[name = tensor("op_2167_cast_fp16")]; + tensor var_2169_equation_0 = const()[name = tensor("op_2169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2169_cast_fp16 = einsum(equation = var_2169_equation_0, values = (var_2091_cast_fp16_6, var_2146_cast_fp16))[name = tensor("op_2169_cast_fp16")]; + tensor var_2171_equation_0 = const()[name = tensor("op_2171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2171_cast_fp16 = einsum(equation = var_2171_equation_0, values = (var_2091_cast_fp16_7, var_2147_cast_fp16))[name = tensor("op_2171_cast_fp16")]; + tensor var_2173_equation_0 = const()[name = tensor("op_2173_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2173_cast_fp16 = einsum(equation = var_2173_equation_0, values = (var_2091_cast_fp16_8, var_2148_cast_fp16))[name = tensor("op_2173_cast_fp16")]; + tensor var_2175_equation_0 = const()[name = tensor("op_2175_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2175_cast_fp16 = einsum(equation = var_2175_equation_0, values = (var_2091_cast_fp16_9, var_2149_cast_fp16))[name = tensor("op_2175_cast_fp16")]; + tensor var_2177_equation_0 = const()[name = tensor("op_2177_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2177_cast_fp16 = einsum(equation = var_2177_equation_0, values = (var_2091_cast_fp16_10, var_2150_cast_fp16))[name = tensor("op_2177_cast_fp16")]; + tensor var_2179_equation_0 = const()[name = tensor("op_2179_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2179_cast_fp16 = einsum(equation = var_2179_equation_0, values = (var_2091_cast_fp16_11, var_2151_cast_fp16))[name = tensor("op_2179_cast_fp16")]; + tensor var_2181_equation_0 = const()[name = tensor("op_2181_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2181_cast_fp16 = einsum(equation = var_2181_equation_0, values = (var_2091_cast_fp16_12, var_2152_cast_fp16))[name = tensor("op_2181_cast_fp16")]; + tensor var_2183_equation_0 = const()[name = tensor("op_2183_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2183_cast_fp16 = einsum(equation = var_2183_equation_0, values = (var_2091_cast_fp16_13, var_2153_cast_fp16))[name = tensor("op_2183_cast_fp16")]; + tensor var_2185_equation_0 = const()[name = tensor("op_2185_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2185_cast_fp16 = einsum(equation = var_2185_equation_0, values = (var_2091_cast_fp16_14, var_2154_cast_fp16))[name = tensor("op_2185_cast_fp16")]; + tensor var_2187_equation_0 = const()[name = tensor("op_2187_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2187_cast_fp16 = einsum(equation = var_2187_equation_0, values = (var_2091_cast_fp16_15, var_2155_cast_fp16))[name = tensor("op_2187_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_2004, interleave = input_85_interleave_0, values = (var_2157_cast_fp16, var_2159_cast_fp16, var_2161_cast_fp16, var_2163_cast_fp16, var_2165_cast_fp16, var_2167_cast_fp16, var_2169_cast_fp16, var_2171_cast_fp16, var_2173_cast_fp16, var_2175_cast_fp16, var_2177_cast_fp16, var_2179_cast_fp16, var_2181_cast_fp16, var_2183_cast_fp16, var_2185_cast_fp16, var_2187_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_2196_pad_type_0 = const()[name = tensor("op_2196_pad_type_0"), val = tensor("valid")]; + tensor var_2196_strides_0 = const()[name = tensor("op_2196_strides_0"), val = tensor([1, 1])]; + tensor var_2196_pad_0 = const()[name = tensor("op_2196_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2196_dilations_0 = const()[name = tensor("op_2196_dilations_0"), val = tensor([1, 1])]; + tensor var_2196_groups_0 = const()[name = tensor("op_2196_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217690432)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219787648)))]; + tensor var_2196_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_2196_dilations_0, groups = var_2196_groups_0, pad = var_2196_pad_0, pad_type = var_2196_pad_type_0, strides = var_2196_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_2196_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219789760)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219791872)))]; + tensor var_2206_to_fp16 = const()[name = tensor("op_2206_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_2206_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219793984)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228182656)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_2232_pad_type_0 = const()[name = tensor("op_2232_pad_type_0"), val = tensor("valid")]; + tensor var_2232_strides_0 = const()[name = tensor("op_2232_strides_0"), val = tensor([1, 1])]; + tensor var_2232_pad_0 = const()[name = tensor("op_2232_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2232_dilations_0 = const()[name = tensor("op_2232_dilations_0"), val = tensor([1, 1])]; + tensor var_2232_groups_0 = const()[name = tensor("op_2232_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228190912)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236579584)))]; + tensor var_2232_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_2232_dilations_0, groups = var_2232_groups_0, pad = var_2232_pad_0, pad_type = var_2232_pad_type_0, strides = var_2232_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_2232_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2241 = const()[name = tensor("op_2241"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236581696)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236583808)))]; + tensor var_2257_to_fp16 = const()[name = tensor("op_2257_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_2257_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_2292_weight_0_to_fp16 = const()[name = tensor("op_2292_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236585920)))]; + tensor var_2292_bias_0_to_fp16 = const()[name = tensor("op_2292_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238683136)))]; + tensor var_2292_cast_fp16 = conv(bias = var_2292_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_2292_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238685248)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_2290_pad_type_0 = const()[name = tensor("op_2290_pad_type_0"), val = tensor("valid")]; + tensor var_2290_strides_0 = const()[name = tensor("op_2290_strides_0"), val = tensor([1, 1])]; + tensor var_2290_pad_0 = const()[name = tensor("op_2290_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2290_dilations_0 = const()[name = tensor("op_2290_dilations_0"), val = tensor([1, 1])]; + tensor var_2290_groups_0 = const()[name = tensor("op_2290_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240782464)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242879680)))]; + tensor var_2290_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_2290_dilations_0, groups = var_2290_groups_0, pad = var_2290_pad_0, pad_type = var_2290_pad_type_0, strides = var_2290_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_2290_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2293_axis_0 = const()[name = tensor("op_2293_axis_0"), val = tensor(1)]; + tensor var_2293_cast_fp16_0, tensor var_2293_cast_fp16_1, tensor var_2293_cast_fp16_2, tensor var_2293_cast_fp16_3, tensor var_2293_cast_fp16_4, tensor var_2293_cast_fp16_5, tensor var_2293_cast_fp16_6, tensor var_2293_cast_fp16_7, tensor var_2293_cast_fp16_8, tensor var_2293_cast_fp16_9, tensor var_2293_cast_fp16_10, tensor var_2293_cast_fp16_11, tensor var_2293_cast_fp16_12, tensor var_2293_cast_fp16_13, tensor var_2293_cast_fp16_14, tensor var_2293_cast_fp16_15 = split(axis = var_2293_axis_0, split_sizes = tile_27, x = var_2292_cast_fp16)[name = tensor("op_2293_cast_fp16")]; + tensor var_2310_perm_0 = const()[name = tensor("op_2310_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2311_axis_0 = const()[name = tensor("op_2311_axis_0"), val = tensor(3)]; + tensor var_2310_cast_fp16 = transpose(perm = var_2310_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_15")]; + tensor var_2311_cast_fp16_0, tensor var_2311_cast_fp16_1, tensor var_2311_cast_fp16_2, tensor var_2311_cast_fp16_3, tensor var_2311_cast_fp16_4, tensor var_2311_cast_fp16_5, tensor var_2311_cast_fp16_6, tensor var_2311_cast_fp16_7, tensor var_2311_cast_fp16_8, tensor var_2311_cast_fp16_9, tensor var_2311_cast_fp16_10, tensor var_2311_cast_fp16_11, tensor var_2311_cast_fp16_12, tensor var_2311_cast_fp16_13, tensor var_2311_cast_fp16_14, tensor var_2311_cast_fp16_15 = split(axis = var_2311_axis_0, split_sizes = tile_28, x = var_2310_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2328_axis_0 = const()[name = tensor("op_2328_axis_0"), val = tensor(1)]; + tensor var_2328_cast_fp16_0, tensor var_2328_cast_fp16_1, tensor var_2328_cast_fp16_2, tensor var_2328_cast_fp16_3, tensor var_2328_cast_fp16_4, tensor var_2328_cast_fp16_5, tensor var_2328_cast_fp16_6, tensor var_2328_cast_fp16_7, tensor var_2328_cast_fp16_8, tensor var_2328_cast_fp16_9, tensor var_2328_cast_fp16_10, tensor var_2328_cast_fp16_11, tensor var_2328_cast_fp16_12, tensor var_2328_cast_fp16_13, tensor var_2328_cast_fp16_14, tensor var_2328_cast_fp16_15 = split(axis = var_2328_axis_0, split_sizes = tile_29, x = var_2290_cast_fp16)[name = tensor("op_2328_cast_fp16")]; + tensor aw_289_equation_0 = const()[name = tensor("aw_289_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_289_cast_fp16 = einsum(equation = aw_289_equation_0, values = (var_2311_cast_fp16_0, var_2293_cast_fp16_0))[name = tensor("aw_289_cast_fp16")]; + tensor aw_291_equation_0 = const()[name = tensor("aw_291_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_291_cast_fp16 = einsum(equation = aw_291_equation_0, values = (var_2311_cast_fp16_1, var_2293_cast_fp16_1))[name = tensor("aw_291_cast_fp16")]; + tensor aw_293_equation_0 = const()[name = tensor("aw_293_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_293_cast_fp16 = einsum(equation = aw_293_equation_0, values = (var_2311_cast_fp16_2, var_2293_cast_fp16_2))[name = tensor("aw_293_cast_fp16")]; + tensor aw_295_equation_0 = const()[name = tensor("aw_295_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_295_cast_fp16 = einsum(equation = aw_295_equation_0, values = (var_2311_cast_fp16_3, var_2293_cast_fp16_3))[name = tensor("aw_295_cast_fp16")]; + tensor aw_297_equation_0 = const()[name = tensor("aw_297_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_297_cast_fp16 = einsum(equation = aw_297_equation_0, values = (var_2311_cast_fp16_4, var_2293_cast_fp16_4))[name = tensor("aw_297_cast_fp16")]; + tensor aw_299_equation_0 = const()[name = tensor("aw_299_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_299_cast_fp16 = einsum(equation = aw_299_equation_0, values = (var_2311_cast_fp16_5, var_2293_cast_fp16_5))[name = tensor("aw_299_cast_fp16")]; + tensor aw_301_equation_0 = const()[name = tensor("aw_301_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_301_cast_fp16 = einsum(equation = aw_301_equation_0, values = (var_2311_cast_fp16_6, var_2293_cast_fp16_6))[name = tensor("aw_301_cast_fp16")]; + tensor aw_303_equation_0 = const()[name = tensor("aw_303_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_303_cast_fp16 = einsum(equation = aw_303_equation_0, values = (var_2311_cast_fp16_7, var_2293_cast_fp16_7))[name = tensor("aw_303_cast_fp16")]; + tensor aw_305_equation_0 = const()[name = tensor("aw_305_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_305_cast_fp16 = einsum(equation = aw_305_equation_0, values = (var_2311_cast_fp16_8, var_2293_cast_fp16_8))[name = tensor("aw_305_cast_fp16")]; + tensor aw_307_equation_0 = const()[name = tensor("aw_307_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_307_cast_fp16 = einsum(equation = aw_307_equation_0, values = (var_2311_cast_fp16_9, var_2293_cast_fp16_9))[name = tensor("aw_307_cast_fp16")]; + tensor aw_309_equation_0 = const()[name = tensor("aw_309_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_309_cast_fp16 = einsum(equation = aw_309_equation_0, values = (var_2311_cast_fp16_10, var_2293_cast_fp16_10))[name = tensor("aw_309_cast_fp16")]; + tensor aw_311_equation_0 = const()[name = tensor("aw_311_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_311_cast_fp16 = einsum(equation = aw_311_equation_0, values = (var_2311_cast_fp16_11, var_2293_cast_fp16_11))[name = tensor("aw_311_cast_fp16")]; + tensor aw_313_equation_0 = const()[name = tensor("aw_313_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_313_cast_fp16 = einsum(equation = aw_313_equation_0, values = (var_2311_cast_fp16_12, var_2293_cast_fp16_12))[name = tensor("aw_313_cast_fp16")]; + tensor aw_315_equation_0 = const()[name = tensor("aw_315_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_315_cast_fp16 = einsum(equation = aw_315_equation_0, values = (var_2311_cast_fp16_13, var_2293_cast_fp16_13))[name = tensor("aw_315_cast_fp16")]; + tensor aw_317_equation_0 = const()[name = tensor("aw_317_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_317_cast_fp16 = einsum(equation = aw_317_equation_0, values = (var_2311_cast_fp16_14, var_2293_cast_fp16_14))[name = tensor("aw_317_cast_fp16")]; + tensor aw_319_equation_0 = const()[name = tensor("aw_319_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_319_cast_fp16 = einsum(equation = aw_319_equation_0, values = (var_2311_cast_fp16_15, var_2293_cast_fp16_15))[name = tensor("aw_319_cast_fp16")]; + tensor var_2377_cast_fp16 = softmax(axis = var_2241, x = aw_289_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_cast_fp16 = softmax(axis = var_2241, x = aw_291_cast_fp16)[name = tensor("op_2378_cast_fp16")]; + tensor var_2379_cast_fp16 = softmax(axis = var_2241, x = aw_293_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = softmax(axis = var_2241, x = aw_295_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor var_2381_cast_fp16 = softmax(axis = var_2241, x = aw_297_cast_fp16)[name = tensor("op_2381_cast_fp16")]; + tensor var_2382_cast_fp16 = softmax(axis = var_2241, x = aw_299_cast_fp16)[name = tensor("op_2382_cast_fp16")]; + tensor var_2383_cast_fp16 = softmax(axis = var_2241, x = aw_301_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = softmax(axis = var_2241, x = aw_303_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor var_2385_cast_fp16 = softmax(axis = var_2241, x = aw_305_cast_fp16)[name = tensor("op_2385_cast_fp16")]; + tensor var_2386_cast_fp16 = softmax(axis = var_2241, x = aw_307_cast_fp16)[name = tensor("op_2386_cast_fp16")]; + tensor var_2387_cast_fp16 = softmax(axis = var_2241, x = aw_309_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = softmax(axis = var_2241, x = aw_311_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor var_2389_cast_fp16 = softmax(axis = var_2241, x = aw_313_cast_fp16)[name = tensor("op_2389_cast_fp16")]; + tensor var_2390_cast_fp16 = softmax(axis = var_2241, x = aw_315_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor var_2391_cast_fp16 = softmax(axis = var_2241, x = aw_317_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = softmax(axis = var_2241, x = aw_319_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor var_2394_equation_0 = const()[name = tensor("op_2394_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2394_cast_fp16 = einsum(equation = var_2394_equation_0, values = (var_2328_cast_fp16_0, var_2377_cast_fp16))[name = tensor("op_2394_cast_fp16")]; + tensor var_2396_equation_0 = const()[name = tensor("op_2396_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2396_cast_fp16 = einsum(equation = var_2396_equation_0, values = (var_2328_cast_fp16_1, var_2378_cast_fp16))[name = tensor("op_2396_cast_fp16")]; + tensor var_2398_equation_0 = const()[name = tensor("op_2398_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2398_cast_fp16 = einsum(equation = var_2398_equation_0, values = (var_2328_cast_fp16_2, var_2379_cast_fp16))[name = tensor("op_2398_cast_fp16")]; + tensor var_2400_equation_0 = const()[name = tensor("op_2400_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2400_cast_fp16 = einsum(equation = var_2400_equation_0, values = (var_2328_cast_fp16_3, var_2380_cast_fp16))[name = tensor("op_2400_cast_fp16")]; + tensor var_2402_equation_0 = const()[name = tensor("op_2402_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2402_cast_fp16 = einsum(equation = var_2402_equation_0, values = (var_2328_cast_fp16_4, var_2381_cast_fp16))[name = tensor("op_2402_cast_fp16")]; + tensor var_2404_equation_0 = const()[name = tensor("op_2404_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2404_cast_fp16 = einsum(equation = var_2404_equation_0, values = (var_2328_cast_fp16_5, var_2382_cast_fp16))[name = tensor("op_2404_cast_fp16")]; + tensor var_2406_equation_0 = const()[name = tensor("op_2406_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2406_cast_fp16 = einsum(equation = var_2406_equation_0, values = (var_2328_cast_fp16_6, var_2383_cast_fp16))[name = tensor("op_2406_cast_fp16")]; + tensor var_2408_equation_0 = const()[name = tensor("op_2408_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2408_cast_fp16 = einsum(equation = var_2408_equation_0, values = (var_2328_cast_fp16_7, var_2384_cast_fp16))[name = tensor("op_2408_cast_fp16")]; + tensor var_2410_equation_0 = const()[name = tensor("op_2410_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2410_cast_fp16 = einsum(equation = var_2410_equation_0, values = (var_2328_cast_fp16_8, var_2385_cast_fp16))[name = tensor("op_2410_cast_fp16")]; + tensor var_2412_equation_0 = const()[name = tensor("op_2412_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2412_cast_fp16 = einsum(equation = var_2412_equation_0, values = (var_2328_cast_fp16_9, var_2386_cast_fp16))[name = tensor("op_2412_cast_fp16")]; + tensor var_2414_equation_0 = const()[name = tensor("op_2414_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2414_cast_fp16 = einsum(equation = var_2414_equation_0, values = (var_2328_cast_fp16_10, var_2387_cast_fp16))[name = tensor("op_2414_cast_fp16")]; + tensor var_2416_equation_0 = const()[name = tensor("op_2416_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2416_cast_fp16 = einsum(equation = var_2416_equation_0, values = (var_2328_cast_fp16_11, var_2388_cast_fp16))[name = tensor("op_2416_cast_fp16")]; + tensor var_2418_equation_0 = const()[name = tensor("op_2418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2418_cast_fp16 = einsum(equation = var_2418_equation_0, values = (var_2328_cast_fp16_12, var_2389_cast_fp16))[name = tensor("op_2418_cast_fp16")]; + tensor var_2420_equation_0 = const()[name = tensor("op_2420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2420_cast_fp16 = einsum(equation = var_2420_equation_0, values = (var_2328_cast_fp16_13, var_2390_cast_fp16))[name = tensor("op_2420_cast_fp16")]; + tensor var_2422_equation_0 = const()[name = tensor("op_2422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2422_cast_fp16 = einsum(equation = var_2422_equation_0, values = (var_2328_cast_fp16_14, var_2391_cast_fp16))[name = tensor("op_2422_cast_fp16")]; + tensor var_2424_equation_0 = const()[name = tensor("op_2424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2424_cast_fp16 = einsum(equation = var_2424_equation_0, values = (var_2328_cast_fp16_15, var_2392_cast_fp16))[name = tensor("op_2424_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_2241, interleave = input_95_interleave_0, values = (var_2394_cast_fp16, var_2396_cast_fp16, var_2398_cast_fp16, var_2400_cast_fp16, var_2402_cast_fp16, var_2404_cast_fp16, var_2406_cast_fp16, var_2408_cast_fp16, var_2410_cast_fp16, var_2412_cast_fp16, var_2414_cast_fp16, var_2416_cast_fp16, var_2418_cast_fp16, var_2420_cast_fp16, var_2422_cast_fp16, var_2424_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2433_pad_type_0 = const()[name = tensor("op_2433_pad_type_0"), val = tensor("valid")]; + tensor var_2433_strides_0 = const()[name = tensor("op_2433_strides_0"), val = tensor([1, 1])]; + tensor var_2433_pad_0 = const()[name = tensor("op_2433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2433_dilations_0 = const()[name = tensor("op_2433_dilations_0"), val = tensor([1, 1])]; + tensor var_2433_groups_0 = const()[name = tensor("op_2433_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242881792)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244979008)))]; + tensor var_2433_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2433_dilations_0, groups = var_2433_groups_0, pad = var_2433_pad_0, pad_type = var_2433_pad_type_0, strides = var_2433_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2433_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2433_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244981120)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244983232)))]; + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2443_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244985344)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253374016)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2469_pad_type_0 = const()[name = tensor("op_2469_pad_type_0"), val = tensor("valid")]; + tensor var_2469_strides_0 = const()[name = tensor("op_2469_strides_0"), val = tensor([1, 1])]; + tensor var_2469_pad_0 = const()[name = tensor("op_2469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2469_dilations_0 = const()[name = tensor("op_2469_dilations_0"), val = tensor([1, 1])]; + tensor var_2469_groups_0 = const()[name = tensor("op_2469_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253382272)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261770944)))]; + tensor var_2469_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2469_dilations_0, groups = var_2469_groups_0, pad = var_2469_pad_0, pad_type = var_2469_pad_type_0, strides = var_2469_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2469_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2469_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2478 = const()[name = tensor("op_2478"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261773056)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261775168)))]; + tensor var_2494_to_fp16 = const()[name = tensor("op_2494_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2494_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2529_weight_0_to_fp16 = const()[name = tensor("op_2529_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261777280)))]; + tensor var_2529_bias_0_to_fp16 = const()[name = tensor("op_2529_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263874496)))]; + tensor var_2529_cast_fp16 = conv(bias = var_2529_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2529_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263876608)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2527_pad_type_0 = const()[name = tensor("op_2527_pad_type_0"), val = tensor("valid")]; + tensor var_2527_strides_0 = const()[name = tensor("op_2527_strides_0"), val = tensor([1, 1])]; + tensor var_2527_pad_0 = const()[name = tensor("op_2527_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2527_dilations_0 = const()[name = tensor("op_2527_dilations_0"), val = tensor([1, 1])]; + tensor var_2527_groups_0 = const()[name = tensor("op_2527_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265973824)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268071040)))]; + tensor var_2527_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2527_dilations_0, groups = var_2527_groups_0, pad = var_2527_pad_0, pad_type = var_2527_pad_type_0, strides = var_2527_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2530_axis_0 = const()[name = tensor("op_2530_axis_0"), val = tensor(1)]; + tensor var_2530_cast_fp16_0, tensor var_2530_cast_fp16_1, tensor var_2530_cast_fp16_2, tensor var_2530_cast_fp16_3, tensor var_2530_cast_fp16_4, tensor var_2530_cast_fp16_5, tensor var_2530_cast_fp16_6, tensor var_2530_cast_fp16_7, tensor var_2530_cast_fp16_8, tensor var_2530_cast_fp16_9, tensor var_2530_cast_fp16_10, tensor var_2530_cast_fp16_11, tensor var_2530_cast_fp16_12, tensor var_2530_cast_fp16_13, tensor var_2530_cast_fp16_14, tensor var_2530_cast_fp16_15 = split(axis = var_2530_axis_0, split_sizes = tile_30, x = var_2529_cast_fp16)[name = tensor("op_2530_cast_fp16")]; + tensor var_2547_perm_0 = const()[name = tensor("op_2547_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2548_axis_0 = const()[name = tensor("op_2548_axis_0"), val = tensor(3)]; + tensor var_2547_cast_fp16 = transpose(perm = var_2547_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_14")]; + tensor var_2548_cast_fp16_0, tensor var_2548_cast_fp16_1, tensor var_2548_cast_fp16_2, tensor var_2548_cast_fp16_3, tensor var_2548_cast_fp16_4, tensor var_2548_cast_fp16_5, tensor var_2548_cast_fp16_6, tensor var_2548_cast_fp16_7, tensor var_2548_cast_fp16_8, tensor var_2548_cast_fp16_9, tensor var_2548_cast_fp16_10, tensor var_2548_cast_fp16_11, tensor var_2548_cast_fp16_12, tensor var_2548_cast_fp16_13, tensor var_2548_cast_fp16_14, tensor var_2548_cast_fp16_15 = split(axis = var_2548_axis_0, split_sizes = tile_31, x = var_2547_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2565_axis_0 = const()[name = tensor("op_2565_axis_0"), val = tensor(1)]; + tensor var_2565_cast_fp16_0, tensor var_2565_cast_fp16_1, tensor var_2565_cast_fp16_2, tensor var_2565_cast_fp16_3, tensor var_2565_cast_fp16_4, tensor var_2565_cast_fp16_5, tensor var_2565_cast_fp16_6, tensor var_2565_cast_fp16_7, tensor var_2565_cast_fp16_8, tensor var_2565_cast_fp16_9, tensor var_2565_cast_fp16_10, tensor var_2565_cast_fp16_11, tensor var_2565_cast_fp16_12, tensor var_2565_cast_fp16_13, tensor var_2565_cast_fp16_14, tensor var_2565_cast_fp16_15 = split(axis = var_2565_axis_0, split_sizes = tile_32, x = var_2527_cast_fp16)[name = tensor("op_2565_cast_fp16")]; + tensor aw_321_equation_0 = const()[name = tensor("aw_321_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_321_cast_fp16 = einsum(equation = aw_321_equation_0, values = (var_2548_cast_fp16_0, var_2530_cast_fp16_0))[name = tensor("aw_321_cast_fp16")]; + tensor aw_323_equation_0 = const()[name = tensor("aw_323_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_323_cast_fp16 = einsum(equation = aw_323_equation_0, values = (var_2548_cast_fp16_1, var_2530_cast_fp16_1))[name = tensor("aw_323_cast_fp16")]; + tensor aw_325_equation_0 = const()[name = tensor("aw_325_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_325_cast_fp16 = einsum(equation = aw_325_equation_0, values = (var_2548_cast_fp16_2, var_2530_cast_fp16_2))[name = tensor("aw_325_cast_fp16")]; + tensor aw_327_equation_0 = const()[name = tensor("aw_327_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_327_cast_fp16 = einsum(equation = aw_327_equation_0, values = (var_2548_cast_fp16_3, var_2530_cast_fp16_3))[name = tensor("aw_327_cast_fp16")]; + tensor aw_329_equation_0 = const()[name = tensor("aw_329_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_329_cast_fp16 = einsum(equation = aw_329_equation_0, values = (var_2548_cast_fp16_4, var_2530_cast_fp16_4))[name = tensor("aw_329_cast_fp16")]; + tensor aw_331_equation_0 = const()[name = tensor("aw_331_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_331_cast_fp16 = einsum(equation = aw_331_equation_0, values = (var_2548_cast_fp16_5, var_2530_cast_fp16_5))[name = tensor("aw_331_cast_fp16")]; + tensor aw_333_equation_0 = const()[name = tensor("aw_333_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_333_cast_fp16 = einsum(equation = aw_333_equation_0, values = (var_2548_cast_fp16_6, var_2530_cast_fp16_6))[name = tensor("aw_333_cast_fp16")]; + tensor aw_335_equation_0 = const()[name = tensor("aw_335_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_335_cast_fp16 = einsum(equation = aw_335_equation_0, values = (var_2548_cast_fp16_7, var_2530_cast_fp16_7))[name = tensor("aw_335_cast_fp16")]; + tensor aw_337_equation_0 = const()[name = tensor("aw_337_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_337_cast_fp16 = einsum(equation = aw_337_equation_0, values = (var_2548_cast_fp16_8, var_2530_cast_fp16_8))[name = tensor("aw_337_cast_fp16")]; + tensor aw_339_equation_0 = const()[name = tensor("aw_339_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_339_cast_fp16 = einsum(equation = aw_339_equation_0, values = (var_2548_cast_fp16_9, var_2530_cast_fp16_9))[name = tensor("aw_339_cast_fp16")]; + tensor aw_341_equation_0 = const()[name = tensor("aw_341_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_341_cast_fp16 = einsum(equation = aw_341_equation_0, values = (var_2548_cast_fp16_10, var_2530_cast_fp16_10))[name = tensor("aw_341_cast_fp16")]; + tensor aw_343_equation_0 = const()[name = tensor("aw_343_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_343_cast_fp16 = einsum(equation = aw_343_equation_0, values = (var_2548_cast_fp16_11, var_2530_cast_fp16_11))[name = tensor("aw_343_cast_fp16")]; + tensor aw_345_equation_0 = const()[name = tensor("aw_345_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_345_cast_fp16 = einsum(equation = aw_345_equation_0, values = (var_2548_cast_fp16_12, var_2530_cast_fp16_12))[name = tensor("aw_345_cast_fp16")]; + tensor aw_347_equation_0 = const()[name = tensor("aw_347_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_347_cast_fp16 = einsum(equation = aw_347_equation_0, values = (var_2548_cast_fp16_13, var_2530_cast_fp16_13))[name = tensor("aw_347_cast_fp16")]; + tensor aw_349_equation_0 = const()[name = tensor("aw_349_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_349_cast_fp16 = einsum(equation = aw_349_equation_0, values = (var_2548_cast_fp16_14, var_2530_cast_fp16_14))[name = tensor("aw_349_cast_fp16")]; + tensor aw_351_equation_0 = const()[name = tensor("aw_351_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_351_cast_fp16 = einsum(equation = aw_351_equation_0, values = (var_2548_cast_fp16_15, var_2530_cast_fp16_15))[name = tensor("aw_351_cast_fp16")]; + tensor var_2614_cast_fp16 = softmax(axis = var_2478, x = aw_321_cast_fp16)[name = tensor("op_2614_cast_fp16")]; + tensor var_2615_cast_fp16 = softmax(axis = var_2478, x = aw_323_cast_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2616_cast_fp16 = softmax(axis = var_2478, x = aw_325_cast_fp16)[name = tensor("op_2616_cast_fp16")]; + tensor var_2617_cast_fp16 = softmax(axis = var_2478, x = aw_327_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor var_2618_cast_fp16 = softmax(axis = var_2478, x = aw_329_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2619_cast_fp16 = softmax(axis = var_2478, x = aw_331_cast_fp16)[name = tensor("op_2619_cast_fp16")]; + tensor var_2620_cast_fp16 = softmax(axis = var_2478, x = aw_333_cast_fp16)[name = tensor("op_2620_cast_fp16")]; + tensor var_2621_cast_fp16 = softmax(axis = var_2478, x = aw_335_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_cast_fp16 = softmax(axis = var_2478, x = aw_337_cast_fp16)[name = tensor("op_2622_cast_fp16")]; + tensor var_2623_cast_fp16 = softmax(axis = var_2478, x = aw_339_cast_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor var_2624_cast_fp16 = softmax(axis = var_2478, x = aw_341_cast_fp16)[name = tensor("op_2624_cast_fp16")]; + tensor var_2625_cast_fp16 = softmax(axis = var_2478, x = aw_343_cast_fp16)[name = tensor("op_2625_cast_fp16")]; + tensor var_2626_cast_fp16 = softmax(axis = var_2478, x = aw_345_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2627_cast_fp16 = softmax(axis = var_2478, x = aw_347_cast_fp16)[name = tensor("op_2627_cast_fp16")]; + tensor var_2628_cast_fp16 = softmax(axis = var_2478, x = aw_349_cast_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = softmax(axis = var_2478, x = aw_351_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor var_2631_equation_0 = const()[name = tensor("op_2631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2631_cast_fp16 = einsum(equation = var_2631_equation_0, values = (var_2565_cast_fp16_0, var_2614_cast_fp16))[name = tensor("op_2631_cast_fp16")]; + tensor var_2633_equation_0 = const()[name = tensor("op_2633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2633_cast_fp16 = einsum(equation = var_2633_equation_0, values = (var_2565_cast_fp16_1, var_2615_cast_fp16))[name = tensor("op_2633_cast_fp16")]; + tensor var_2635_equation_0 = const()[name = tensor("op_2635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2635_cast_fp16 = einsum(equation = var_2635_equation_0, values = (var_2565_cast_fp16_2, var_2616_cast_fp16))[name = tensor("op_2635_cast_fp16")]; + tensor var_2637_equation_0 = const()[name = tensor("op_2637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2637_cast_fp16 = einsum(equation = var_2637_equation_0, values = (var_2565_cast_fp16_3, var_2617_cast_fp16))[name = tensor("op_2637_cast_fp16")]; + tensor var_2639_equation_0 = const()[name = tensor("op_2639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2639_cast_fp16 = einsum(equation = var_2639_equation_0, values = (var_2565_cast_fp16_4, var_2618_cast_fp16))[name = tensor("op_2639_cast_fp16")]; + tensor var_2641_equation_0 = const()[name = tensor("op_2641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2641_cast_fp16 = einsum(equation = var_2641_equation_0, values = (var_2565_cast_fp16_5, var_2619_cast_fp16))[name = tensor("op_2641_cast_fp16")]; + tensor var_2643_equation_0 = const()[name = tensor("op_2643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2643_cast_fp16 = einsum(equation = var_2643_equation_0, values = (var_2565_cast_fp16_6, var_2620_cast_fp16))[name = tensor("op_2643_cast_fp16")]; + tensor var_2645_equation_0 = const()[name = tensor("op_2645_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2645_cast_fp16 = einsum(equation = var_2645_equation_0, values = (var_2565_cast_fp16_7, var_2621_cast_fp16))[name = tensor("op_2645_cast_fp16")]; + tensor var_2647_equation_0 = const()[name = tensor("op_2647_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2647_cast_fp16 = einsum(equation = var_2647_equation_0, values = (var_2565_cast_fp16_8, var_2622_cast_fp16))[name = tensor("op_2647_cast_fp16")]; + tensor var_2649_equation_0 = const()[name = tensor("op_2649_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2649_cast_fp16 = einsum(equation = var_2649_equation_0, values = (var_2565_cast_fp16_9, var_2623_cast_fp16))[name = tensor("op_2649_cast_fp16")]; + tensor var_2651_equation_0 = const()[name = tensor("op_2651_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2651_cast_fp16 = einsum(equation = var_2651_equation_0, values = (var_2565_cast_fp16_10, var_2624_cast_fp16))[name = tensor("op_2651_cast_fp16")]; + tensor var_2653_equation_0 = const()[name = tensor("op_2653_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2653_cast_fp16 = einsum(equation = var_2653_equation_0, values = (var_2565_cast_fp16_11, var_2625_cast_fp16))[name = tensor("op_2653_cast_fp16")]; + tensor var_2655_equation_0 = const()[name = tensor("op_2655_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2655_cast_fp16 = einsum(equation = var_2655_equation_0, values = (var_2565_cast_fp16_12, var_2626_cast_fp16))[name = tensor("op_2655_cast_fp16")]; + tensor var_2657_equation_0 = const()[name = tensor("op_2657_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2657_cast_fp16 = einsum(equation = var_2657_equation_0, values = (var_2565_cast_fp16_13, var_2627_cast_fp16))[name = tensor("op_2657_cast_fp16")]; + tensor var_2659_equation_0 = const()[name = tensor("op_2659_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2659_cast_fp16 = einsum(equation = var_2659_equation_0, values = (var_2565_cast_fp16_14, var_2628_cast_fp16))[name = tensor("op_2659_cast_fp16")]; + tensor var_2661_equation_0 = const()[name = tensor("op_2661_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2661_cast_fp16 = einsum(equation = var_2661_equation_0, values = (var_2565_cast_fp16_15, var_2629_cast_fp16))[name = tensor("op_2661_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2478, interleave = input_105_interleave_0, values = (var_2631_cast_fp16, var_2633_cast_fp16, var_2635_cast_fp16, var_2637_cast_fp16, var_2639_cast_fp16, var_2641_cast_fp16, var_2643_cast_fp16, var_2645_cast_fp16, var_2647_cast_fp16, var_2649_cast_fp16, var_2651_cast_fp16, var_2653_cast_fp16, var_2655_cast_fp16, var_2657_cast_fp16, var_2659_cast_fp16, var_2661_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_2670_pad_type_0 = const()[name = tensor("op_2670_pad_type_0"), val = tensor("valid")]; + tensor var_2670_strides_0 = const()[name = tensor("op_2670_strides_0"), val = tensor([1, 1])]; + tensor var_2670_pad_0 = const()[name = tensor("op_2670_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2670_dilations_0 = const()[name = tensor("op_2670_dilations_0"), val = tensor([1, 1])]; + tensor var_2670_groups_0 = const()[name = tensor("op_2670_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268073152)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270170368)))]; + tensor var_2670_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_2670_dilations_0, groups = var_2670_groups_0, pad = var_2670_pad_0, pad_type = var_2670_pad_type_0, strides = var_2670_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_2670_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270172480)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270174592)))]; + tensor var_2680_to_fp16 = const()[name = tensor("op_2680_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_2680_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270176704)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278565376)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_2706_pad_type_0 = const()[name = tensor("op_2706_pad_type_0"), val = tensor("valid")]; + tensor var_2706_strides_0 = const()[name = tensor("op_2706_strides_0"), val = tensor([1, 1])]; + tensor var_2706_pad_0 = const()[name = tensor("op_2706_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2706_dilations_0 = const()[name = tensor("op_2706_dilations_0"), val = tensor([1, 1])]; + tensor var_2706_groups_0 = const()[name = tensor("op_2706_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278573632)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286962304)))]; + tensor var_2706_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_2706_dilations_0, groups = var_2706_groups_0, pad = var_2706_pad_0, pad_type = var_2706_pad_type_0, strides = var_2706_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_2706_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_2715 = const()[name = tensor("op_2715"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286964416)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286966528)))]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_2731_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_23_pad_type_0 = const()[name = tensor("q_23_pad_type_0"), val = tensor("valid")]; + tensor q_23_strides_0 = const()[name = tensor("q_23_strides_0"), val = tensor([1, 1])]; + tensor q_23_pad_0 = const()[name = tensor("q_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_23_dilations_0 = const()[name = tensor("q_23_dilations_0"), val = tensor([1, 1])]; + tensor q_23_groups_0 = const()[name = tensor("q_23_groups_0"), val = tensor(1)]; + tensor var_2766_weight_0_to_fp16 = const()[name = tensor("op_2766_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286968640)))]; + tensor var_2766_bias_0_to_fp16 = const()[name = tensor("op_2766_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289065856)))]; + tensor var_2766_cast_fp16 = conv(bias = var_2766_bias_0_to_fp16, dilations = q_23_dilations_0, groups = q_23_groups_0, pad = q_23_pad_0, pad_type = q_23_pad_type_0, strides = q_23_strides_0, weight = var_2766_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor k_23_pad_type_0 = const()[name = tensor("k_23_pad_type_0"), val = tensor("valid")]; + tensor k_23_strides_0 = const()[name = tensor("k_23_strides_0"), val = tensor([1, 1])]; + tensor k_23_pad_0 = const()[name = tensor("k_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_23_dilations_0 = const()[name = tensor("k_23_dilations_0"), val = tensor([1, 1])]; + tensor k_23_groups_0 = const()[name = tensor("k_23_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289067968)))]; + tensor k_23_cast_fp16 = conv(dilations = k_23_dilations_0, groups = k_23_groups_0, pad = k_23_pad_0, pad_type = k_23_pad_type_0, strides = k_23_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_2764_pad_type_0 = const()[name = tensor("op_2764_pad_type_0"), val = tensor("valid")]; + tensor var_2764_strides_0 = const()[name = tensor("op_2764_strides_0"), val = tensor([1, 1])]; + tensor var_2764_pad_0 = const()[name = tensor("op_2764_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2764_dilations_0 = const()[name = tensor("op_2764_dilations_0"), val = tensor([1, 1])]; + tensor var_2764_groups_0 = const()[name = tensor("op_2764_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291165184)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293262400)))]; + tensor var_2764_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_2764_dilations_0, groups = var_2764_groups_0, pad = var_2764_pad_0, pad_type = var_2764_pad_type_0, strides = var_2764_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2767_axis_0 = const()[name = tensor("op_2767_axis_0"), val = tensor(1)]; + tensor var_2767_cast_fp16_0, tensor var_2767_cast_fp16_1, tensor var_2767_cast_fp16_2, tensor var_2767_cast_fp16_3, tensor var_2767_cast_fp16_4, tensor var_2767_cast_fp16_5, tensor var_2767_cast_fp16_6, tensor var_2767_cast_fp16_7, tensor var_2767_cast_fp16_8, tensor var_2767_cast_fp16_9, tensor var_2767_cast_fp16_10, tensor var_2767_cast_fp16_11, tensor var_2767_cast_fp16_12, tensor var_2767_cast_fp16_13, tensor var_2767_cast_fp16_14, tensor var_2767_cast_fp16_15 = split(axis = var_2767_axis_0, split_sizes = tile_33, x = var_2766_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor var_2784_perm_0 = const()[name = tensor("op_2784_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2785_axis_0 = const()[name = tensor("op_2785_axis_0"), val = tensor(3)]; + tensor var_2784_cast_fp16 = transpose(perm = var_2784_perm_0, x = k_23_cast_fp16)[name = tensor("transpose_13")]; + tensor var_2785_cast_fp16_0, tensor var_2785_cast_fp16_1, tensor var_2785_cast_fp16_2, tensor var_2785_cast_fp16_3, tensor var_2785_cast_fp16_4, tensor var_2785_cast_fp16_5, tensor var_2785_cast_fp16_6, tensor var_2785_cast_fp16_7, tensor var_2785_cast_fp16_8, tensor var_2785_cast_fp16_9, tensor var_2785_cast_fp16_10, tensor var_2785_cast_fp16_11, tensor var_2785_cast_fp16_12, tensor var_2785_cast_fp16_13, tensor var_2785_cast_fp16_14, tensor var_2785_cast_fp16_15 = split(axis = var_2785_axis_0, split_sizes = tile_34, x = var_2784_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2802_axis_0 = const()[name = tensor("op_2802_axis_0"), val = tensor(1)]; + tensor var_2802_cast_fp16_0, tensor var_2802_cast_fp16_1, tensor var_2802_cast_fp16_2, tensor var_2802_cast_fp16_3, tensor var_2802_cast_fp16_4, tensor var_2802_cast_fp16_5, tensor var_2802_cast_fp16_6, tensor var_2802_cast_fp16_7, tensor var_2802_cast_fp16_8, tensor var_2802_cast_fp16_9, tensor var_2802_cast_fp16_10, tensor var_2802_cast_fp16_11, tensor var_2802_cast_fp16_12, tensor var_2802_cast_fp16_13, tensor var_2802_cast_fp16_14, tensor var_2802_cast_fp16_15 = split(axis = var_2802_axis_0, split_sizes = tile_35, x = var_2764_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor aw_353_equation_0 = const()[name = tensor("aw_353_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_353_cast_fp16 = einsum(equation = aw_353_equation_0, values = (var_2785_cast_fp16_0, var_2767_cast_fp16_0))[name = tensor("aw_353_cast_fp16")]; + tensor aw_355_equation_0 = const()[name = tensor("aw_355_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_355_cast_fp16 = einsum(equation = aw_355_equation_0, values = (var_2785_cast_fp16_1, var_2767_cast_fp16_1))[name = tensor("aw_355_cast_fp16")]; + tensor aw_357_equation_0 = const()[name = tensor("aw_357_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_357_cast_fp16 = einsum(equation = aw_357_equation_0, values = (var_2785_cast_fp16_2, var_2767_cast_fp16_2))[name = tensor("aw_357_cast_fp16")]; + tensor aw_359_equation_0 = const()[name = tensor("aw_359_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_359_cast_fp16 = einsum(equation = aw_359_equation_0, values = (var_2785_cast_fp16_3, var_2767_cast_fp16_3))[name = tensor("aw_359_cast_fp16")]; + tensor aw_361_equation_0 = const()[name = tensor("aw_361_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_361_cast_fp16 = einsum(equation = aw_361_equation_0, values = (var_2785_cast_fp16_4, var_2767_cast_fp16_4))[name = tensor("aw_361_cast_fp16")]; + tensor aw_363_equation_0 = const()[name = tensor("aw_363_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_363_cast_fp16 = einsum(equation = aw_363_equation_0, values = (var_2785_cast_fp16_5, var_2767_cast_fp16_5))[name = tensor("aw_363_cast_fp16")]; + tensor aw_365_equation_0 = const()[name = tensor("aw_365_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_365_cast_fp16 = einsum(equation = aw_365_equation_0, values = (var_2785_cast_fp16_6, var_2767_cast_fp16_6))[name = tensor("aw_365_cast_fp16")]; + tensor aw_367_equation_0 = const()[name = tensor("aw_367_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_367_cast_fp16 = einsum(equation = aw_367_equation_0, values = (var_2785_cast_fp16_7, var_2767_cast_fp16_7))[name = tensor("aw_367_cast_fp16")]; + tensor aw_369_equation_0 = const()[name = tensor("aw_369_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_369_cast_fp16 = einsum(equation = aw_369_equation_0, values = (var_2785_cast_fp16_8, var_2767_cast_fp16_8))[name = tensor("aw_369_cast_fp16")]; + tensor aw_371_equation_0 = const()[name = tensor("aw_371_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_371_cast_fp16 = einsum(equation = aw_371_equation_0, values = (var_2785_cast_fp16_9, var_2767_cast_fp16_9))[name = tensor("aw_371_cast_fp16")]; + tensor aw_373_equation_0 = const()[name = tensor("aw_373_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_373_cast_fp16 = einsum(equation = aw_373_equation_0, values = (var_2785_cast_fp16_10, var_2767_cast_fp16_10))[name = tensor("aw_373_cast_fp16")]; + tensor aw_375_equation_0 = const()[name = tensor("aw_375_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_375_cast_fp16 = einsum(equation = aw_375_equation_0, values = (var_2785_cast_fp16_11, var_2767_cast_fp16_11))[name = tensor("aw_375_cast_fp16")]; + tensor aw_377_equation_0 = const()[name = tensor("aw_377_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_377_cast_fp16 = einsum(equation = aw_377_equation_0, values = (var_2785_cast_fp16_12, var_2767_cast_fp16_12))[name = tensor("aw_377_cast_fp16")]; + tensor aw_379_equation_0 = const()[name = tensor("aw_379_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_379_cast_fp16 = einsum(equation = aw_379_equation_0, values = (var_2785_cast_fp16_13, var_2767_cast_fp16_13))[name = tensor("aw_379_cast_fp16")]; + tensor aw_381_equation_0 = const()[name = tensor("aw_381_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_381_cast_fp16 = einsum(equation = aw_381_equation_0, values = (var_2785_cast_fp16_14, var_2767_cast_fp16_14))[name = tensor("aw_381_cast_fp16")]; + tensor aw_383_equation_0 = const()[name = tensor("aw_383_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_383_cast_fp16 = einsum(equation = aw_383_equation_0, values = (var_2785_cast_fp16_15, var_2767_cast_fp16_15))[name = tensor("aw_383_cast_fp16")]; + tensor var_2851_cast_fp16 = softmax(axis = var_2715, x = aw_353_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor var_2852_cast_fp16 = softmax(axis = var_2715, x = aw_355_cast_fp16)[name = tensor("op_2852_cast_fp16")]; + tensor var_2853_cast_fp16 = softmax(axis = var_2715, x = aw_357_cast_fp16)[name = tensor("op_2853_cast_fp16")]; + tensor var_2854_cast_fp16 = softmax(axis = var_2715, x = aw_359_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = softmax(axis = var_2715, x = aw_361_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = softmax(axis = var_2715, x = aw_363_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2857_cast_fp16 = softmax(axis = var_2715, x = aw_365_cast_fp16)[name = tensor("op_2857_cast_fp16")]; + tensor var_2858_cast_fp16 = softmax(axis = var_2715, x = aw_367_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor var_2859_cast_fp16 = softmax(axis = var_2715, x = aw_369_cast_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2860_cast_fp16 = softmax(axis = var_2715, x = aw_371_cast_fp16)[name = tensor("op_2860_cast_fp16")]; + tensor var_2861_cast_fp16 = softmax(axis = var_2715, x = aw_373_cast_fp16)[name = tensor("op_2861_cast_fp16")]; + tensor var_2862_cast_fp16 = softmax(axis = var_2715, x = aw_375_cast_fp16)[name = tensor("op_2862_cast_fp16")]; + tensor var_2863_cast_fp16 = softmax(axis = var_2715, x = aw_377_cast_fp16)[name = tensor("op_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = softmax(axis = var_2715, x = aw_379_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_cast_fp16 = softmax(axis = var_2715, x = aw_381_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_cast_fp16 = softmax(axis = var_2715, x = aw_383_cast_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor var_2868_equation_0 = const()[name = tensor("op_2868_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2868_cast_fp16 = einsum(equation = var_2868_equation_0, values = (var_2802_cast_fp16_0, var_2851_cast_fp16))[name = tensor("op_2868_cast_fp16")]; + tensor var_2870_equation_0 = const()[name = tensor("op_2870_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2870_cast_fp16 = einsum(equation = var_2870_equation_0, values = (var_2802_cast_fp16_1, var_2852_cast_fp16))[name = tensor("op_2870_cast_fp16")]; + tensor var_2872_equation_0 = const()[name = tensor("op_2872_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2872_cast_fp16 = einsum(equation = var_2872_equation_0, values = (var_2802_cast_fp16_2, var_2853_cast_fp16))[name = tensor("op_2872_cast_fp16")]; + tensor var_2874_equation_0 = const()[name = tensor("op_2874_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2874_cast_fp16 = einsum(equation = var_2874_equation_0, values = (var_2802_cast_fp16_3, var_2854_cast_fp16))[name = tensor("op_2874_cast_fp16")]; + tensor var_2876_equation_0 = const()[name = tensor("op_2876_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2876_cast_fp16 = einsum(equation = var_2876_equation_0, values = (var_2802_cast_fp16_4, var_2855_cast_fp16))[name = tensor("op_2876_cast_fp16")]; + tensor var_2878_equation_0 = const()[name = tensor("op_2878_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2878_cast_fp16 = einsum(equation = var_2878_equation_0, values = (var_2802_cast_fp16_5, var_2856_cast_fp16))[name = tensor("op_2878_cast_fp16")]; + tensor var_2880_equation_0 = const()[name = tensor("op_2880_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2880_cast_fp16 = einsum(equation = var_2880_equation_0, values = (var_2802_cast_fp16_6, var_2857_cast_fp16))[name = tensor("op_2880_cast_fp16")]; + tensor var_2882_equation_0 = const()[name = tensor("op_2882_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2882_cast_fp16 = einsum(equation = var_2882_equation_0, values = (var_2802_cast_fp16_7, var_2858_cast_fp16))[name = tensor("op_2882_cast_fp16")]; + tensor var_2884_equation_0 = const()[name = tensor("op_2884_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2884_cast_fp16 = einsum(equation = var_2884_equation_0, values = (var_2802_cast_fp16_8, var_2859_cast_fp16))[name = tensor("op_2884_cast_fp16")]; + tensor var_2886_equation_0 = const()[name = tensor("op_2886_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2886_cast_fp16 = einsum(equation = var_2886_equation_0, values = (var_2802_cast_fp16_9, var_2860_cast_fp16))[name = tensor("op_2886_cast_fp16")]; + tensor var_2888_equation_0 = const()[name = tensor("op_2888_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2888_cast_fp16 = einsum(equation = var_2888_equation_0, values = (var_2802_cast_fp16_10, var_2861_cast_fp16))[name = tensor("op_2888_cast_fp16")]; + tensor var_2890_equation_0 = const()[name = tensor("op_2890_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2890_cast_fp16 = einsum(equation = var_2890_equation_0, values = (var_2802_cast_fp16_11, var_2862_cast_fp16))[name = tensor("op_2890_cast_fp16")]; + tensor var_2892_equation_0 = const()[name = tensor("op_2892_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2892_cast_fp16 = einsum(equation = var_2892_equation_0, values = (var_2802_cast_fp16_12, var_2863_cast_fp16))[name = tensor("op_2892_cast_fp16")]; + tensor var_2894_equation_0 = const()[name = tensor("op_2894_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2894_cast_fp16 = einsum(equation = var_2894_equation_0, values = (var_2802_cast_fp16_13, var_2864_cast_fp16))[name = tensor("op_2894_cast_fp16")]; + tensor var_2896_equation_0 = const()[name = tensor("op_2896_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2896_cast_fp16 = einsum(equation = var_2896_equation_0, values = (var_2802_cast_fp16_14, var_2865_cast_fp16))[name = tensor("op_2896_cast_fp16")]; + tensor var_2898_equation_0 = const()[name = tensor("op_2898_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2898_cast_fp16 = einsum(equation = var_2898_equation_0, values = (var_2802_cast_fp16_15, var_2866_cast_fp16))[name = tensor("op_2898_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_2715, interleave = input_115_interleave_0, values = (var_2868_cast_fp16, var_2870_cast_fp16, var_2872_cast_fp16, var_2874_cast_fp16, var_2876_cast_fp16, var_2878_cast_fp16, var_2880_cast_fp16, var_2882_cast_fp16, var_2884_cast_fp16, var_2886_cast_fp16, var_2888_cast_fp16, var_2890_cast_fp16, var_2892_cast_fp16, var_2894_cast_fp16, var_2896_cast_fp16, var_2898_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_2907_pad_type_0 = const()[name = tensor("op_2907_pad_type_0"), val = tensor("valid")]; + tensor var_2907_strides_0 = const()[name = tensor("op_2907_strides_0"), val = tensor([1, 1])]; + tensor var_2907_pad_0 = const()[name = tensor("op_2907_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2907_dilations_0 = const()[name = tensor("op_2907_dilations_0"), val = tensor([1, 1])]; + tensor var_2907_groups_0 = const()[name = tensor("op_2907_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293264512)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295361728)))]; + tensor var_2907_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_2907_dilations_0, groups = var_2907_groups_0, pad = var_2907_pad_0, pad_type = var_2907_pad_type_0, strides = var_2907_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_2907_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_2907_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295363840)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295365952)))]; + tensor var_2917_to_fp16 = const()[name = tensor("op_2917_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_2917_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295368064)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756736)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_2943_pad_type_0 = const()[name = tensor("op_2943_pad_type_0"), val = tensor("valid")]; + tensor var_2943_strides_0 = const()[name = tensor("op_2943_strides_0"), val = tensor([1, 1])]; + tensor var_2943_pad_0 = const()[name = tensor("op_2943_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2943_dilations_0 = const()[name = tensor("op_2943_dilations_0"), val = tensor([1, 1])]; + tensor var_2943_groups_0 = const()[name = tensor("op_2943_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312153664)))]; + tensor var_2943_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_2943_dilations_0, groups = var_2943_groups_0, pad = var_2943_pad_0, pad_type = var_2943_pad_type_0, strides = var_2943_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("op_2943_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_2943_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_2952 = const()[name = tensor("op_2952"), val = tensor(1)]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([1])]; + tensor input_123_gamma_0_to_fp16 = const()[name = tensor("input_123_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312155776)))]; + tensor input_123_beta_0_to_fp16 = const()[name = tensor("input_123_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312157888)))]; + tensor var_2968_to_fp16 = const()[name = tensor("op_2968_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = input_123_beta_0_to_fp16, epsilon = var_2968_to_fp16, gamma = input_123_gamma_0_to_fp16, x = inputs_49_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor q_25_pad_type_0 = const()[name = tensor("q_25_pad_type_0"), val = tensor("valid")]; + tensor q_25_strides_0 = const()[name = tensor("q_25_strides_0"), val = tensor([1, 1])]; + tensor q_25_pad_0 = const()[name = tensor("q_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_25_dilations_0 = const()[name = tensor("q_25_dilations_0"), val = tensor([1, 1])]; + tensor q_25_groups_0 = const()[name = tensor("q_25_groups_0"), val = tensor(1)]; + tensor var_3003_weight_0_to_fp16 = const()[name = tensor("op_3003_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312160000)))]; + tensor var_3003_bias_0_to_fp16 = const()[name = tensor("op_3003_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314257216)))]; + tensor var_3003_cast_fp16 = conv(bias = var_3003_bias_0_to_fp16, dilations = q_25_dilations_0, groups = q_25_groups_0, pad = q_25_pad_0, pad_type = q_25_pad_type_0, strides = q_25_strides_0, weight = var_3003_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3003_cast_fp16")]; + tensor k_25_pad_type_0 = const()[name = tensor("k_25_pad_type_0"), val = tensor("valid")]; + tensor k_25_strides_0 = const()[name = tensor("k_25_strides_0"), val = tensor([1, 1])]; + tensor k_25_pad_0 = const()[name = tensor("k_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_25_dilations_0 = const()[name = tensor("k_25_dilations_0"), val = tensor([1, 1])]; + tensor k_25_groups_0 = const()[name = tensor("k_25_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_key_weight_to_fp16 = const()[name = tensor("blocks_12_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314259328)))]; + tensor k_25_cast_fp16 = conv(dilations = k_25_dilations_0, groups = k_25_groups_0, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = k_25_strides_0, weight = blocks_12_attn_key_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor var_3001_pad_type_0 = const()[name = tensor("op_3001_pad_type_0"), val = tensor("valid")]; + tensor var_3001_strides_0 = const()[name = tensor("op_3001_strides_0"), val = tensor([1, 1])]; + tensor var_3001_pad_0 = const()[name = tensor("op_3001_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3001_dilations_0 = const()[name = tensor("op_3001_dilations_0"), val = tensor([1, 1])]; + tensor var_3001_groups_0 = const()[name = tensor("op_3001_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_value_weight_to_fp16 = const()[name = tensor("blocks_12_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316356544)))]; + tensor blocks_12_attn_value_bias_to_fp16 = const()[name = tensor("blocks_12_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318453760)))]; + tensor var_3001_cast_fp16 = conv(bias = blocks_12_attn_value_bias_to_fp16, dilations = var_3001_dilations_0, groups = var_3001_groups_0, pad = var_3001_pad_0, pad_type = var_3001_pad_type_0, strides = var_3001_strides_0, weight = blocks_12_attn_value_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor tile_36 = const()[name = tensor("tile_36"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3004_axis_0 = const()[name = tensor("op_3004_axis_0"), val = tensor(1)]; + tensor var_3004_cast_fp16_0, tensor var_3004_cast_fp16_1, tensor var_3004_cast_fp16_2, tensor var_3004_cast_fp16_3, tensor var_3004_cast_fp16_4, tensor var_3004_cast_fp16_5, tensor var_3004_cast_fp16_6, tensor var_3004_cast_fp16_7, tensor var_3004_cast_fp16_8, tensor var_3004_cast_fp16_9, tensor var_3004_cast_fp16_10, tensor var_3004_cast_fp16_11, tensor var_3004_cast_fp16_12, tensor var_3004_cast_fp16_13, tensor var_3004_cast_fp16_14, tensor var_3004_cast_fp16_15 = split(axis = var_3004_axis_0, split_sizes = tile_36, x = var_3003_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3021_perm_0 = const()[name = tensor("op_3021_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_37 = const()[name = tensor("tile_37"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3022_axis_0 = const()[name = tensor("op_3022_axis_0"), val = tensor(3)]; + tensor var_3021_cast_fp16 = transpose(perm = var_3021_perm_0, x = k_25_cast_fp16)[name = tensor("transpose_12")]; + tensor var_3022_cast_fp16_0, tensor var_3022_cast_fp16_1, tensor var_3022_cast_fp16_2, tensor var_3022_cast_fp16_3, tensor var_3022_cast_fp16_4, tensor var_3022_cast_fp16_5, tensor var_3022_cast_fp16_6, tensor var_3022_cast_fp16_7, tensor var_3022_cast_fp16_8, tensor var_3022_cast_fp16_9, tensor var_3022_cast_fp16_10, tensor var_3022_cast_fp16_11, tensor var_3022_cast_fp16_12, tensor var_3022_cast_fp16_13, tensor var_3022_cast_fp16_14, tensor var_3022_cast_fp16_15 = split(axis = var_3022_axis_0, split_sizes = tile_37, x = var_3021_cast_fp16)[name = tensor("op_3022_cast_fp16")]; + tensor tile_38 = const()[name = tensor("tile_38"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3039_axis_0 = const()[name = tensor("op_3039_axis_0"), val = tensor(1)]; + tensor var_3039_cast_fp16_0, tensor var_3039_cast_fp16_1, tensor var_3039_cast_fp16_2, tensor var_3039_cast_fp16_3, tensor var_3039_cast_fp16_4, tensor var_3039_cast_fp16_5, tensor var_3039_cast_fp16_6, tensor var_3039_cast_fp16_7, tensor var_3039_cast_fp16_8, tensor var_3039_cast_fp16_9, tensor var_3039_cast_fp16_10, tensor var_3039_cast_fp16_11, tensor var_3039_cast_fp16_12, tensor var_3039_cast_fp16_13, tensor var_3039_cast_fp16_14, tensor var_3039_cast_fp16_15 = split(axis = var_3039_axis_0, split_sizes = tile_38, x = var_3001_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor aw_385_equation_0 = const()[name = tensor("aw_385_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_385_cast_fp16 = einsum(equation = aw_385_equation_0, values = (var_3022_cast_fp16_0, var_3004_cast_fp16_0))[name = tensor("aw_385_cast_fp16")]; + tensor aw_387_equation_0 = const()[name = tensor("aw_387_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_387_cast_fp16 = einsum(equation = aw_387_equation_0, values = (var_3022_cast_fp16_1, var_3004_cast_fp16_1))[name = tensor("aw_387_cast_fp16")]; + tensor aw_389_equation_0 = const()[name = tensor("aw_389_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_389_cast_fp16 = einsum(equation = aw_389_equation_0, values = (var_3022_cast_fp16_2, var_3004_cast_fp16_2))[name = tensor("aw_389_cast_fp16")]; + tensor aw_391_equation_0 = const()[name = tensor("aw_391_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_391_cast_fp16 = einsum(equation = aw_391_equation_0, values = (var_3022_cast_fp16_3, var_3004_cast_fp16_3))[name = tensor("aw_391_cast_fp16")]; + tensor aw_393_equation_0 = const()[name = tensor("aw_393_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_393_cast_fp16 = einsum(equation = aw_393_equation_0, values = (var_3022_cast_fp16_4, var_3004_cast_fp16_4))[name = tensor("aw_393_cast_fp16")]; + tensor aw_395_equation_0 = const()[name = tensor("aw_395_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_395_cast_fp16 = einsum(equation = aw_395_equation_0, values = (var_3022_cast_fp16_5, var_3004_cast_fp16_5))[name = tensor("aw_395_cast_fp16")]; + tensor aw_397_equation_0 = const()[name = tensor("aw_397_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_397_cast_fp16 = einsum(equation = aw_397_equation_0, values = (var_3022_cast_fp16_6, var_3004_cast_fp16_6))[name = tensor("aw_397_cast_fp16")]; + tensor aw_399_equation_0 = const()[name = tensor("aw_399_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_399_cast_fp16 = einsum(equation = aw_399_equation_0, values = (var_3022_cast_fp16_7, var_3004_cast_fp16_7))[name = tensor("aw_399_cast_fp16")]; + tensor aw_401_equation_0 = const()[name = tensor("aw_401_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_401_cast_fp16 = einsum(equation = aw_401_equation_0, values = (var_3022_cast_fp16_8, var_3004_cast_fp16_8))[name = tensor("aw_401_cast_fp16")]; + tensor aw_403_equation_0 = const()[name = tensor("aw_403_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_403_cast_fp16 = einsum(equation = aw_403_equation_0, values = (var_3022_cast_fp16_9, var_3004_cast_fp16_9))[name = tensor("aw_403_cast_fp16")]; + tensor aw_405_equation_0 = const()[name = tensor("aw_405_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_405_cast_fp16 = einsum(equation = aw_405_equation_0, values = (var_3022_cast_fp16_10, var_3004_cast_fp16_10))[name = tensor("aw_405_cast_fp16")]; + tensor aw_407_equation_0 = const()[name = tensor("aw_407_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_407_cast_fp16 = einsum(equation = aw_407_equation_0, values = (var_3022_cast_fp16_11, var_3004_cast_fp16_11))[name = tensor("aw_407_cast_fp16")]; + tensor aw_409_equation_0 = const()[name = tensor("aw_409_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_409_cast_fp16 = einsum(equation = aw_409_equation_0, values = (var_3022_cast_fp16_12, var_3004_cast_fp16_12))[name = tensor("aw_409_cast_fp16")]; + tensor aw_411_equation_0 = const()[name = tensor("aw_411_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_411_cast_fp16 = einsum(equation = aw_411_equation_0, values = (var_3022_cast_fp16_13, var_3004_cast_fp16_13))[name = tensor("aw_411_cast_fp16")]; + tensor aw_413_equation_0 = const()[name = tensor("aw_413_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_413_cast_fp16 = einsum(equation = aw_413_equation_0, values = (var_3022_cast_fp16_14, var_3004_cast_fp16_14))[name = tensor("aw_413_cast_fp16")]; + tensor aw_415_equation_0 = const()[name = tensor("aw_415_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_415_cast_fp16 = einsum(equation = aw_415_equation_0, values = (var_3022_cast_fp16_15, var_3004_cast_fp16_15))[name = tensor("aw_415_cast_fp16")]; + tensor var_3088_cast_fp16 = softmax(axis = var_2952, x = aw_385_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_cast_fp16 = softmax(axis = var_2952, x = aw_387_cast_fp16)[name = tensor("op_3089_cast_fp16")]; + tensor var_3090_cast_fp16 = softmax(axis = var_2952, x = aw_389_cast_fp16)[name = tensor("op_3090_cast_fp16")]; + tensor var_3091_cast_fp16 = softmax(axis = var_2952, x = aw_391_cast_fp16)[name = tensor("op_3091_cast_fp16")]; + tensor var_3092_cast_fp16 = softmax(axis = var_2952, x = aw_393_cast_fp16)[name = tensor("op_3092_cast_fp16")]; + tensor var_3093_cast_fp16 = softmax(axis = var_2952, x = aw_395_cast_fp16)[name = tensor("op_3093_cast_fp16")]; + tensor var_3094_cast_fp16 = softmax(axis = var_2952, x = aw_397_cast_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = softmax(axis = var_2952, x = aw_399_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor var_3096_cast_fp16 = softmax(axis = var_2952, x = aw_401_cast_fp16)[name = tensor("op_3096_cast_fp16")]; + tensor var_3097_cast_fp16 = softmax(axis = var_2952, x = aw_403_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor var_3098_cast_fp16 = softmax(axis = var_2952, x = aw_405_cast_fp16)[name = tensor("op_3098_cast_fp16")]; + tensor var_3099_cast_fp16 = softmax(axis = var_2952, x = aw_407_cast_fp16)[name = tensor("op_3099_cast_fp16")]; + tensor var_3100_cast_fp16 = softmax(axis = var_2952, x = aw_409_cast_fp16)[name = tensor("op_3100_cast_fp16")]; + tensor var_3101_cast_fp16 = softmax(axis = var_2952, x = aw_411_cast_fp16)[name = tensor("op_3101_cast_fp16")]; + tensor var_3102_cast_fp16 = softmax(axis = var_2952, x = aw_413_cast_fp16)[name = tensor("op_3102_cast_fp16")]; + tensor var_3103_cast_fp16 = softmax(axis = var_2952, x = aw_415_cast_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3105_equation_0 = const()[name = tensor("op_3105_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3105_cast_fp16 = einsum(equation = var_3105_equation_0, values = (var_3039_cast_fp16_0, var_3088_cast_fp16))[name = tensor("op_3105_cast_fp16")]; + tensor var_3107_equation_0 = const()[name = tensor("op_3107_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3107_cast_fp16 = einsum(equation = var_3107_equation_0, values = (var_3039_cast_fp16_1, var_3089_cast_fp16))[name = tensor("op_3107_cast_fp16")]; + tensor var_3109_equation_0 = const()[name = tensor("op_3109_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3109_cast_fp16 = einsum(equation = var_3109_equation_0, values = (var_3039_cast_fp16_2, var_3090_cast_fp16))[name = tensor("op_3109_cast_fp16")]; + tensor var_3111_equation_0 = const()[name = tensor("op_3111_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3111_cast_fp16 = einsum(equation = var_3111_equation_0, values = (var_3039_cast_fp16_3, var_3091_cast_fp16))[name = tensor("op_3111_cast_fp16")]; + tensor var_3113_equation_0 = const()[name = tensor("op_3113_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3113_cast_fp16 = einsum(equation = var_3113_equation_0, values = (var_3039_cast_fp16_4, var_3092_cast_fp16))[name = tensor("op_3113_cast_fp16")]; + tensor var_3115_equation_0 = const()[name = tensor("op_3115_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3115_cast_fp16 = einsum(equation = var_3115_equation_0, values = (var_3039_cast_fp16_5, var_3093_cast_fp16))[name = tensor("op_3115_cast_fp16")]; + tensor var_3117_equation_0 = const()[name = tensor("op_3117_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3117_cast_fp16 = einsum(equation = var_3117_equation_0, values = (var_3039_cast_fp16_6, var_3094_cast_fp16))[name = tensor("op_3117_cast_fp16")]; + tensor var_3119_equation_0 = const()[name = tensor("op_3119_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3119_cast_fp16 = einsum(equation = var_3119_equation_0, values = (var_3039_cast_fp16_7, var_3095_cast_fp16))[name = tensor("op_3119_cast_fp16")]; + tensor var_3121_equation_0 = const()[name = tensor("op_3121_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3121_cast_fp16 = einsum(equation = var_3121_equation_0, values = (var_3039_cast_fp16_8, var_3096_cast_fp16))[name = tensor("op_3121_cast_fp16")]; + tensor var_3123_equation_0 = const()[name = tensor("op_3123_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3123_cast_fp16 = einsum(equation = var_3123_equation_0, values = (var_3039_cast_fp16_9, var_3097_cast_fp16))[name = tensor("op_3123_cast_fp16")]; + tensor var_3125_equation_0 = const()[name = tensor("op_3125_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3125_cast_fp16 = einsum(equation = var_3125_equation_0, values = (var_3039_cast_fp16_10, var_3098_cast_fp16))[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_equation_0 = const()[name = tensor("op_3127_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3127_cast_fp16 = einsum(equation = var_3127_equation_0, values = (var_3039_cast_fp16_11, var_3099_cast_fp16))[name = tensor("op_3127_cast_fp16")]; + tensor var_3129_equation_0 = const()[name = tensor("op_3129_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3129_cast_fp16 = einsum(equation = var_3129_equation_0, values = (var_3039_cast_fp16_12, var_3100_cast_fp16))[name = tensor("op_3129_cast_fp16")]; + tensor var_3131_equation_0 = const()[name = tensor("op_3131_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3131_cast_fp16 = einsum(equation = var_3131_equation_0, values = (var_3039_cast_fp16_13, var_3101_cast_fp16))[name = tensor("op_3131_cast_fp16")]; + tensor var_3133_equation_0 = const()[name = tensor("op_3133_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3133_cast_fp16 = einsum(equation = var_3133_equation_0, values = (var_3039_cast_fp16_14, var_3102_cast_fp16))[name = tensor("op_3133_cast_fp16")]; + tensor var_3135_equation_0 = const()[name = tensor("op_3135_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3135_cast_fp16 = einsum(equation = var_3135_equation_0, values = (var_3039_cast_fp16_15, var_3103_cast_fp16))[name = tensor("op_3135_cast_fp16")]; + tensor input_125_interleave_0 = const()[name = tensor("input_125_interleave_0"), val = tensor(false)]; + tensor input_125_cast_fp16 = concat(axis = var_2952, interleave = input_125_interleave_0, values = (var_3105_cast_fp16, var_3107_cast_fp16, var_3109_cast_fp16, var_3111_cast_fp16, var_3113_cast_fp16, var_3115_cast_fp16, var_3117_cast_fp16, var_3119_cast_fp16, var_3121_cast_fp16, var_3123_cast_fp16, var_3125_cast_fp16, var_3127_cast_fp16, var_3129_cast_fp16, var_3131_cast_fp16, var_3133_cast_fp16, var_3135_cast_fp16))[name = tensor("input_125_cast_fp16")]; + tensor var_3144_pad_type_0 = const()[name = tensor("op_3144_pad_type_0"), val = tensor("valid")]; + tensor var_3144_strides_0 = const()[name = tensor("op_3144_strides_0"), val = tensor([1, 1])]; + tensor var_3144_pad_0 = const()[name = tensor("op_3144_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3144_dilations_0 = const()[name = tensor("op_3144_dilations_0"), val = tensor([1, 1])]; + tensor var_3144_groups_0 = const()[name = tensor("op_3144_groups_0"), val = tensor(1)]; + tensor blocks_12_attn_out_weight_to_fp16 = const()[name = tensor("blocks_12_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318455872)))]; + tensor blocks_12_attn_out_bias_to_fp16 = const()[name = tensor("blocks_12_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320553088)))]; + tensor var_3144_cast_fp16 = conv(bias = blocks_12_attn_out_bias_to_fp16, dilations = var_3144_dilations_0, groups = var_3144_groups_0, pad = var_3144_pad_0, pad_type = var_3144_pad_type_0, strides = var_3144_strides_0, weight = blocks_12_attn_out_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("op_3144_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = var_3144_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([1])]; + tensor input_127_gamma_0_to_fp16 = const()[name = tensor("input_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320555200)))]; + tensor input_127_beta_0_to_fp16 = const()[name = tensor("input_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320557312)))]; + tensor var_3154_to_fp16 = const()[name = tensor("op_3154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = input_127_beta_0_to_fp16, epsilon = var_3154_to_fp16, gamma = input_127_gamma_0_to_fp16, x = inputs_51_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("valid")]; + tensor input_129_strides_0 = const()[name = tensor("input_129_strides_0"), val = tensor([1, 1])]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_129_dilations_0 = const()[name = tensor("input_129_dilations_0"), val = tensor([1, 1])]; + tensor input_129_groups_0 = const()[name = tensor("input_129_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320559424)))]; + tensor blocks_12_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328948096)))]; + tensor input_129_cast_fp16 = conv(bias = blocks_12_mlp_0_bias_to_fp16, dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = blocks_12_mlp_0_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_3180_pad_type_0 = const()[name = tensor("op_3180_pad_type_0"), val = tensor("valid")]; + tensor var_3180_strides_0 = const()[name = tensor("op_3180_strides_0"), val = tensor([1, 1])]; + tensor var_3180_pad_0 = const()[name = tensor("op_3180_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3180_dilations_0 = const()[name = tensor("op_3180_dilations_0"), val = tensor([1, 1])]; + tensor var_3180_groups_0 = const()[name = tensor("op_3180_groups_0"), val = tensor(1)]; + tensor blocks_12_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328956352)))]; + tensor blocks_12_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337345024)))]; + tensor var_3180_cast_fp16 = conv(bias = blocks_12_mlp_2_bias_to_fp16, dilations = var_3180_dilations_0, groups = var_3180_groups_0, pad = var_3180_pad_0, pad_type = var_3180_pad_type_0, strides = var_3180_strides_0, weight = blocks_12_mlp_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("op_3180_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_3180_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3189 = const()[name = tensor("op_3189"), val = tensor(1)]; + tensor input_133_axes_0 = const()[name = tensor("input_133_axes_0"), val = tensor([1])]; + tensor input_133_gamma_0_to_fp16 = const()[name = tensor("input_133_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337347136)))]; + tensor input_133_beta_0_to_fp16 = const()[name = tensor("input_133_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337349248)))]; + tensor var_3205_to_fp16 = const()[name = tensor("op_3205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_133_cast_fp16 = layer_norm(axes = input_133_axes_0, beta = input_133_beta_0_to_fp16, epsilon = var_3205_to_fp16, gamma = input_133_gamma_0_to_fp16, x = inputs_53_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor q_27_pad_type_0 = const()[name = tensor("q_27_pad_type_0"), val = tensor("valid")]; + tensor q_27_strides_0 = const()[name = tensor("q_27_strides_0"), val = tensor([1, 1])]; + tensor q_27_pad_0 = const()[name = tensor("q_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_27_dilations_0 = const()[name = tensor("q_27_dilations_0"), val = tensor([1, 1])]; + tensor q_27_groups_0 = const()[name = tensor("q_27_groups_0"), val = tensor(1)]; + tensor var_3240_weight_0_to_fp16 = const()[name = tensor("op_3240_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337351360)))]; + tensor var_3240_bias_0_to_fp16 = const()[name = tensor("op_3240_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339448576)))]; + tensor var_3240_cast_fp16 = conv(bias = var_3240_bias_0_to_fp16, dilations = q_27_dilations_0, groups = q_27_groups_0, pad = q_27_pad_0, pad_type = q_27_pad_type_0, strides = q_27_strides_0, weight = var_3240_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3240_cast_fp16")]; + tensor k_27_pad_type_0 = const()[name = tensor("k_27_pad_type_0"), val = tensor("valid")]; + tensor k_27_strides_0 = const()[name = tensor("k_27_strides_0"), val = tensor([1, 1])]; + tensor k_27_pad_0 = const()[name = tensor("k_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_27_dilations_0 = const()[name = tensor("k_27_dilations_0"), val = tensor([1, 1])]; + tensor k_27_groups_0 = const()[name = tensor("k_27_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_key_weight_to_fp16 = const()[name = tensor("blocks_13_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339450688)))]; + tensor k_27_cast_fp16 = conv(dilations = k_27_dilations_0, groups = k_27_groups_0, pad = k_27_pad_0, pad_type = k_27_pad_type_0, strides = k_27_strides_0, weight = blocks_13_attn_key_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor var_3238_pad_type_0 = const()[name = tensor("op_3238_pad_type_0"), val = tensor("valid")]; + tensor var_3238_strides_0 = const()[name = tensor("op_3238_strides_0"), val = tensor([1, 1])]; + tensor var_3238_pad_0 = const()[name = tensor("op_3238_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3238_dilations_0 = const()[name = tensor("op_3238_dilations_0"), val = tensor([1, 1])]; + tensor var_3238_groups_0 = const()[name = tensor("op_3238_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_value_weight_to_fp16 = const()[name = tensor("blocks_13_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341547904)))]; + tensor blocks_13_attn_value_bias_to_fp16 = const()[name = tensor("blocks_13_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343645120)))]; + tensor var_3238_cast_fp16 = conv(bias = blocks_13_attn_value_bias_to_fp16, dilations = var_3238_dilations_0, groups = var_3238_groups_0, pad = var_3238_pad_0, pad_type = var_3238_pad_type_0, strides = var_3238_strides_0, weight = blocks_13_attn_value_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("op_3238_cast_fp16")]; + tensor tile_39 = const()[name = tensor("tile_39"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3241_axis_0 = const()[name = tensor("op_3241_axis_0"), val = tensor(1)]; + tensor var_3241_cast_fp16_0, tensor var_3241_cast_fp16_1, tensor var_3241_cast_fp16_2, tensor var_3241_cast_fp16_3, tensor var_3241_cast_fp16_4, tensor var_3241_cast_fp16_5, tensor var_3241_cast_fp16_6, tensor var_3241_cast_fp16_7, tensor var_3241_cast_fp16_8, tensor var_3241_cast_fp16_9, tensor var_3241_cast_fp16_10, tensor var_3241_cast_fp16_11, tensor var_3241_cast_fp16_12, tensor var_3241_cast_fp16_13, tensor var_3241_cast_fp16_14, tensor var_3241_cast_fp16_15 = split(axis = var_3241_axis_0, split_sizes = tile_39, x = var_3240_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor var_3258_perm_0 = const()[name = tensor("op_3258_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_40 = const()[name = tensor("tile_40"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3259_axis_0 = const()[name = tensor("op_3259_axis_0"), val = tensor(3)]; + tensor var_3258_cast_fp16 = transpose(perm = var_3258_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_11")]; + tensor var_3259_cast_fp16_0, tensor var_3259_cast_fp16_1, tensor var_3259_cast_fp16_2, tensor var_3259_cast_fp16_3, tensor var_3259_cast_fp16_4, tensor var_3259_cast_fp16_5, tensor var_3259_cast_fp16_6, tensor var_3259_cast_fp16_7, tensor var_3259_cast_fp16_8, tensor var_3259_cast_fp16_9, tensor var_3259_cast_fp16_10, tensor var_3259_cast_fp16_11, tensor var_3259_cast_fp16_12, tensor var_3259_cast_fp16_13, tensor var_3259_cast_fp16_14, tensor var_3259_cast_fp16_15 = split(axis = var_3259_axis_0, split_sizes = tile_40, x = var_3258_cast_fp16)[name = tensor("op_3259_cast_fp16")]; + tensor tile_41 = const()[name = tensor("tile_41"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3276_axis_0 = const()[name = tensor("op_3276_axis_0"), val = tensor(1)]; + tensor var_3276_cast_fp16_0, tensor var_3276_cast_fp16_1, tensor var_3276_cast_fp16_2, tensor var_3276_cast_fp16_3, tensor var_3276_cast_fp16_4, tensor var_3276_cast_fp16_5, tensor var_3276_cast_fp16_6, tensor var_3276_cast_fp16_7, tensor var_3276_cast_fp16_8, tensor var_3276_cast_fp16_9, tensor var_3276_cast_fp16_10, tensor var_3276_cast_fp16_11, tensor var_3276_cast_fp16_12, tensor var_3276_cast_fp16_13, tensor var_3276_cast_fp16_14, tensor var_3276_cast_fp16_15 = split(axis = var_3276_axis_0, split_sizes = tile_41, x = var_3238_cast_fp16)[name = tensor("op_3276_cast_fp16")]; + tensor aw_417_equation_0 = const()[name = tensor("aw_417_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_417_cast_fp16 = einsum(equation = aw_417_equation_0, values = (var_3259_cast_fp16_0, var_3241_cast_fp16_0))[name = tensor("aw_417_cast_fp16")]; + tensor aw_419_equation_0 = const()[name = tensor("aw_419_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_419_cast_fp16 = einsum(equation = aw_419_equation_0, values = (var_3259_cast_fp16_1, var_3241_cast_fp16_1))[name = tensor("aw_419_cast_fp16")]; + tensor aw_421_equation_0 = const()[name = tensor("aw_421_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_421_cast_fp16 = einsum(equation = aw_421_equation_0, values = (var_3259_cast_fp16_2, var_3241_cast_fp16_2))[name = tensor("aw_421_cast_fp16")]; + tensor aw_423_equation_0 = const()[name = tensor("aw_423_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_423_cast_fp16 = einsum(equation = aw_423_equation_0, values = (var_3259_cast_fp16_3, var_3241_cast_fp16_3))[name = tensor("aw_423_cast_fp16")]; + tensor aw_425_equation_0 = const()[name = tensor("aw_425_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_425_cast_fp16 = einsum(equation = aw_425_equation_0, values = (var_3259_cast_fp16_4, var_3241_cast_fp16_4))[name = tensor("aw_425_cast_fp16")]; + tensor aw_427_equation_0 = const()[name = tensor("aw_427_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_427_cast_fp16 = einsum(equation = aw_427_equation_0, values = (var_3259_cast_fp16_5, var_3241_cast_fp16_5))[name = tensor("aw_427_cast_fp16")]; + tensor aw_429_equation_0 = const()[name = tensor("aw_429_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_429_cast_fp16 = einsum(equation = aw_429_equation_0, values = (var_3259_cast_fp16_6, var_3241_cast_fp16_6))[name = tensor("aw_429_cast_fp16")]; + tensor aw_431_equation_0 = const()[name = tensor("aw_431_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_431_cast_fp16 = einsum(equation = aw_431_equation_0, values = (var_3259_cast_fp16_7, var_3241_cast_fp16_7))[name = tensor("aw_431_cast_fp16")]; + tensor aw_433_equation_0 = const()[name = tensor("aw_433_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_433_cast_fp16 = einsum(equation = aw_433_equation_0, values = (var_3259_cast_fp16_8, var_3241_cast_fp16_8))[name = tensor("aw_433_cast_fp16")]; + tensor aw_435_equation_0 = const()[name = tensor("aw_435_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_435_cast_fp16 = einsum(equation = aw_435_equation_0, values = (var_3259_cast_fp16_9, var_3241_cast_fp16_9))[name = tensor("aw_435_cast_fp16")]; + tensor aw_437_equation_0 = const()[name = tensor("aw_437_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_437_cast_fp16 = einsum(equation = aw_437_equation_0, values = (var_3259_cast_fp16_10, var_3241_cast_fp16_10))[name = tensor("aw_437_cast_fp16")]; + tensor aw_439_equation_0 = const()[name = tensor("aw_439_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_439_cast_fp16 = einsum(equation = aw_439_equation_0, values = (var_3259_cast_fp16_11, var_3241_cast_fp16_11))[name = tensor("aw_439_cast_fp16")]; + tensor aw_441_equation_0 = const()[name = tensor("aw_441_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_441_cast_fp16 = einsum(equation = aw_441_equation_0, values = (var_3259_cast_fp16_12, var_3241_cast_fp16_12))[name = tensor("aw_441_cast_fp16")]; + tensor aw_443_equation_0 = const()[name = tensor("aw_443_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_443_cast_fp16 = einsum(equation = aw_443_equation_0, values = (var_3259_cast_fp16_13, var_3241_cast_fp16_13))[name = tensor("aw_443_cast_fp16")]; + tensor aw_445_equation_0 = const()[name = tensor("aw_445_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_445_cast_fp16 = einsum(equation = aw_445_equation_0, values = (var_3259_cast_fp16_14, var_3241_cast_fp16_14))[name = tensor("aw_445_cast_fp16")]; + tensor aw_447_equation_0 = const()[name = tensor("aw_447_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_447_cast_fp16 = einsum(equation = aw_447_equation_0, values = (var_3259_cast_fp16_15, var_3241_cast_fp16_15))[name = tensor("aw_447_cast_fp16")]; + tensor var_3325_cast_fp16 = softmax(axis = var_3189, x = aw_417_cast_fp16)[name = tensor("op_3325_cast_fp16")]; + tensor var_3326_cast_fp16 = softmax(axis = var_3189, x = aw_419_cast_fp16)[name = tensor("op_3326_cast_fp16")]; + tensor var_3327_cast_fp16 = softmax(axis = var_3189, x = aw_421_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = softmax(axis = var_3189, x = aw_423_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor var_3329_cast_fp16 = softmax(axis = var_3189, x = aw_425_cast_fp16)[name = tensor("op_3329_cast_fp16")]; + tensor var_3330_cast_fp16 = softmax(axis = var_3189, x = aw_427_cast_fp16)[name = tensor("op_3330_cast_fp16")]; + tensor var_3331_cast_fp16 = softmax(axis = var_3189, x = aw_429_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = softmax(axis = var_3189, x = aw_431_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_cast_fp16 = softmax(axis = var_3189, x = aw_433_cast_fp16)[name = tensor("op_3333_cast_fp16")]; + tensor var_3334_cast_fp16 = softmax(axis = var_3189, x = aw_435_cast_fp16)[name = tensor("op_3334_cast_fp16")]; + tensor var_3335_cast_fp16 = softmax(axis = var_3189, x = aw_437_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = softmax(axis = var_3189, x = aw_439_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor var_3337_cast_fp16 = softmax(axis = var_3189, x = aw_441_cast_fp16)[name = tensor("op_3337_cast_fp16")]; + tensor var_3338_cast_fp16 = softmax(axis = var_3189, x = aw_443_cast_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = softmax(axis = var_3189, x = aw_445_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor var_3340_cast_fp16 = softmax(axis = var_3189, x = aw_447_cast_fp16)[name = tensor("op_3340_cast_fp16")]; + tensor var_3342_equation_0 = const()[name = tensor("op_3342_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3342_cast_fp16 = einsum(equation = var_3342_equation_0, values = (var_3276_cast_fp16_0, var_3325_cast_fp16))[name = tensor("op_3342_cast_fp16")]; + tensor var_3344_equation_0 = const()[name = tensor("op_3344_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3344_cast_fp16 = einsum(equation = var_3344_equation_0, values = (var_3276_cast_fp16_1, var_3326_cast_fp16))[name = tensor("op_3344_cast_fp16")]; + tensor var_3346_equation_0 = const()[name = tensor("op_3346_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3346_cast_fp16 = einsum(equation = var_3346_equation_0, values = (var_3276_cast_fp16_2, var_3327_cast_fp16))[name = tensor("op_3346_cast_fp16")]; + tensor var_3348_equation_0 = const()[name = tensor("op_3348_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3348_cast_fp16 = einsum(equation = var_3348_equation_0, values = (var_3276_cast_fp16_3, var_3328_cast_fp16))[name = tensor("op_3348_cast_fp16")]; + tensor var_3350_equation_0 = const()[name = tensor("op_3350_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3350_cast_fp16 = einsum(equation = var_3350_equation_0, values = (var_3276_cast_fp16_4, var_3329_cast_fp16))[name = tensor("op_3350_cast_fp16")]; + tensor var_3352_equation_0 = const()[name = tensor("op_3352_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3352_cast_fp16 = einsum(equation = var_3352_equation_0, values = (var_3276_cast_fp16_5, var_3330_cast_fp16))[name = tensor("op_3352_cast_fp16")]; + tensor var_3354_equation_0 = const()[name = tensor("op_3354_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3354_cast_fp16 = einsum(equation = var_3354_equation_0, values = (var_3276_cast_fp16_6, var_3331_cast_fp16))[name = tensor("op_3354_cast_fp16")]; + tensor var_3356_equation_0 = const()[name = tensor("op_3356_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3356_cast_fp16 = einsum(equation = var_3356_equation_0, values = (var_3276_cast_fp16_7, var_3332_cast_fp16))[name = tensor("op_3356_cast_fp16")]; + tensor var_3358_equation_0 = const()[name = tensor("op_3358_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3358_cast_fp16 = einsum(equation = var_3358_equation_0, values = (var_3276_cast_fp16_8, var_3333_cast_fp16))[name = tensor("op_3358_cast_fp16")]; + tensor var_3360_equation_0 = const()[name = tensor("op_3360_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3360_cast_fp16 = einsum(equation = var_3360_equation_0, values = (var_3276_cast_fp16_9, var_3334_cast_fp16))[name = tensor("op_3360_cast_fp16")]; + tensor var_3362_equation_0 = const()[name = tensor("op_3362_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3362_cast_fp16 = einsum(equation = var_3362_equation_0, values = (var_3276_cast_fp16_10, var_3335_cast_fp16))[name = tensor("op_3362_cast_fp16")]; + tensor var_3364_equation_0 = const()[name = tensor("op_3364_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3364_cast_fp16 = einsum(equation = var_3364_equation_0, values = (var_3276_cast_fp16_11, var_3336_cast_fp16))[name = tensor("op_3364_cast_fp16")]; + tensor var_3366_equation_0 = const()[name = tensor("op_3366_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3366_cast_fp16 = einsum(equation = var_3366_equation_0, values = (var_3276_cast_fp16_12, var_3337_cast_fp16))[name = tensor("op_3366_cast_fp16")]; + tensor var_3368_equation_0 = const()[name = tensor("op_3368_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3368_cast_fp16 = einsum(equation = var_3368_equation_0, values = (var_3276_cast_fp16_13, var_3338_cast_fp16))[name = tensor("op_3368_cast_fp16")]; + tensor var_3370_equation_0 = const()[name = tensor("op_3370_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3370_cast_fp16 = einsum(equation = var_3370_equation_0, values = (var_3276_cast_fp16_14, var_3339_cast_fp16))[name = tensor("op_3370_cast_fp16")]; + tensor var_3372_equation_0 = const()[name = tensor("op_3372_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3372_cast_fp16 = einsum(equation = var_3372_equation_0, values = (var_3276_cast_fp16_15, var_3340_cast_fp16))[name = tensor("op_3372_cast_fp16")]; + tensor input_135_interleave_0 = const()[name = tensor("input_135_interleave_0"), val = tensor(false)]; + tensor input_135_cast_fp16 = concat(axis = var_3189, interleave = input_135_interleave_0, values = (var_3342_cast_fp16, var_3344_cast_fp16, var_3346_cast_fp16, var_3348_cast_fp16, var_3350_cast_fp16, var_3352_cast_fp16, var_3354_cast_fp16, var_3356_cast_fp16, var_3358_cast_fp16, var_3360_cast_fp16, var_3362_cast_fp16, var_3364_cast_fp16, var_3366_cast_fp16, var_3368_cast_fp16, var_3370_cast_fp16, var_3372_cast_fp16))[name = tensor("input_135_cast_fp16")]; + tensor var_3381_pad_type_0 = const()[name = tensor("op_3381_pad_type_0"), val = tensor("valid")]; + tensor var_3381_strides_0 = const()[name = tensor("op_3381_strides_0"), val = tensor([1, 1])]; + tensor var_3381_pad_0 = const()[name = tensor("op_3381_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3381_dilations_0 = const()[name = tensor("op_3381_dilations_0"), val = tensor([1, 1])]; + tensor var_3381_groups_0 = const()[name = tensor("op_3381_groups_0"), val = tensor(1)]; + tensor blocks_13_attn_out_weight_to_fp16 = const()[name = tensor("blocks_13_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343647232)))]; + tensor blocks_13_attn_out_bias_to_fp16 = const()[name = tensor("blocks_13_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345744448)))]; + tensor var_3381_cast_fp16 = conv(bias = blocks_13_attn_out_bias_to_fp16, dilations = var_3381_dilations_0, groups = var_3381_groups_0, pad = var_3381_pad_0, pad_type = var_3381_pad_type_0, strides = var_3381_strides_0, weight = blocks_13_attn_out_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("op_3381_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = var_3381_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([1])]; + tensor input_137_gamma_0_to_fp16 = const()[name = tensor("input_137_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345746560)))]; + tensor input_137_beta_0_to_fp16 = const()[name = tensor("input_137_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345748672)))]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = input_137_beta_0_to_fp16, epsilon = var_3391_to_fp16, gamma = input_137_gamma_0_to_fp16, x = inputs_55_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("valid")]; + tensor input_139_strides_0 = const()[name = tensor("input_139_strides_0"), val = tensor([1, 1])]; + tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_139_dilations_0 = const()[name = tensor("input_139_dilations_0"), val = tensor([1, 1])]; + tensor input_139_groups_0 = const()[name = tensor("input_139_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345750784)))]; + tensor blocks_13_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354139456)))]; + tensor input_139_cast_fp16 = conv(bias = blocks_13_mlp_0_bias_to_fp16, dilations = input_139_dilations_0, groups = input_139_groups_0, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = input_139_strides_0, weight = blocks_13_mlp_0_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_3417_pad_type_0 = const()[name = tensor("op_3417_pad_type_0"), val = tensor("valid")]; + tensor var_3417_strides_0 = const()[name = tensor("op_3417_strides_0"), val = tensor([1, 1])]; + tensor var_3417_pad_0 = const()[name = tensor("op_3417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3417_dilations_0 = const()[name = tensor("op_3417_dilations_0"), val = tensor([1, 1])]; + tensor var_3417_groups_0 = const()[name = tensor("op_3417_groups_0"), val = tensor(1)]; + tensor blocks_13_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354147712)))]; + tensor blocks_13_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362536384)))]; + tensor var_3417_cast_fp16 = conv(bias = blocks_13_mlp_2_bias_to_fp16, dilations = var_3417_dilations_0, groups = var_3417_groups_0, pad = var_3417_pad_0, pad_type = var_3417_pad_type_0, strides = var_3417_strides_0, weight = blocks_13_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = var_3417_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3426 = const()[name = tensor("op_3426"), val = tensor(1)]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([1])]; + tensor input_143_gamma_0_to_fp16 = const()[name = tensor("input_143_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362538496)))]; + tensor input_143_beta_0_to_fp16 = const()[name = tensor("input_143_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362540608)))]; + tensor var_3442_to_fp16 = const()[name = tensor("op_3442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = input_143_beta_0_to_fp16, epsilon = var_3442_to_fp16, gamma = input_143_gamma_0_to_fp16, x = inputs_57_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor q_29_pad_type_0 = const()[name = tensor("q_29_pad_type_0"), val = tensor("valid")]; + tensor q_29_strides_0 = const()[name = tensor("q_29_strides_0"), val = tensor([1, 1])]; + tensor q_29_pad_0 = const()[name = tensor("q_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_29_dilations_0 = const()[name = tensor("q_29_dilations_0"), val = tensor([1, 1])]; + tensor q_29_groups_0 = const()[name = tensor("q_29_groups_0"), val = tensor(1)]; + tensor var_3477_weight_0_to_fp16 = const()[name = tensor("op_3477_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362542720)))]; + tensor var_3477_bias_0_to_fp16 = const()[name = tensor("op_3477_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364639936)))]; + tensor var_3477_cast_fp16 = conv(bias = var_3477_bias_0_to_fp16, dilations = q_29_dilations_0, groups = q_29_groups_0, pad = q_29_pad_0, pad_type = q_29_pad_type_0, strides = q_29_strides_0, weight = var_3477_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor k_29_pad_type_0 = const()[name = tensor("k_29_pad_type_0"), val = tensor("valid")]; + tensor k_29_strides_0 = const()[name = tensor("k_29_strides_0"), val = tensor([1, 1])]; + tensor k_29_pad_0 = const()[name = tensor("k_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_29_dilations_0 = const()[name = tensor("k_29_dilations_0"), val = tensor([1, 1])]; + tensor k_29_groups_0 = const()[name = tensor("k_29_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_key_weight_to_fp16 = const()[name = tensor("blocks_14_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364642048)))]; + tensor k_29_cast_fp16 = conv(dilations = k_29_dilations_0, groups = k_29_groups_0, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = k_29_strides_0, weight = blocks_14_attn_key_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor var_3475_pad_type_0 = const()[name = tensor("op_3475_pad_type_0"), val = tensor("valid")]; + tensor var_3475_strides_0 = const()[name = tensor("op_3475_strides_0"), val = tensor([1, 1])]; + tensor var_3475_pad_0 = const()[name = tensor("op_3475_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3475_dilations_0 = const()[name = tensor("op_3475_dilations_0"), val = tensor([1, 1])]; + tensor var_3475_groups_0 = const()[name = tensor("op_3475_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_value_weight_to_fp16 = const()[name = tensor("blocks_14_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366739264)))]; + tensor blocks_14_attn_value_bias_to_fp16 = const()[name = tensor("blocks_14_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368836480)))]; + tensor var_3475_cast_fp16 = conv(bias = blocks_14_attn_value_bias_to_fp16, dilations = var_3475_dilations_0, groups = var_3475_groups_0, pad = var_3475_pad_0, pad_type = var_3475_pad_type_0, strides = var_3475_strides_0, weight = blocks_14_attn_value_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("op_3475_cast_fp16")]; + tensor tile_42 = const()[name = tensor("tile_42"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3478_axis_0 = const()[name = tensor("op_3478_axis_0"), val = tensor(1)]; + tensor var_3478_cast_fp16_0, tensor var_3478_cast_fp16_1, tensor var_3478_cast_fp16_2, tensor var_3478_cast_fp16_3, tensor var_3478_cast_fp16_4, tensor var_3478_cast_fp16_5, tensor var_3478_cast_fp16_6, tensor var_3478_cast_fp16_7, tensor var_3478_cast_fp16_8, tensor var_3478_cast_fp16_9, tensor var_3478_cast_fp16_10, tensor var_3478_cast_fp16_11, tensor var_3478_cast_fp16_12, tensor var_3478_cast_fp16_13, tensor var_3478_cast_fp16_14, tensor var_3478_cast_fp16_15 = split(axis = var_3478_axis_0, split_sizes = tile_42, x = var_3477_cast_fp16)[name = tensor("op_3478_cast_fp16")]; + tensor var_3495_perm_0 = const()[name = tensor("op_3495_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_43 = const()[name = tensor("tile_43"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3496_axis_0 = const()[name = tensor("op_3496_axis_0"), val = tensor(3)]; + tensor var_3495_cast_fp16 = transpose(perm = var_3495_perm_0, x = k_29_cast_fp16)[name = tensor("transpose_10")]; + tensor var_3496_cast_fp16_0, tensor var_3496_cast_fp16_1, tensor var_3496_cast_fp16_2, tensor var_3496_cast_fp16_3, tensor var_3496_cast_fp16_4, tensor var_3496_cast_fp16_5, tensor var_3496_cast_fp16_6, tensor var_3496_cast_fp16_7, tensor var_3496_cast_fp16_8, tensor var_3496_cast_fp16_9, tensor var_3496_cast_fp16_10, tensor var_3496_cast_fp16_11, tensor var_3496_cast_fp16_12, tensor var_3496_cast_fp16_13, tensor var_3496_cast_fp16_14, tensor var_3496_cast_fp16_15 = split(axis = var_3496_axis_0, split_sizes = tile_43, x = var_3495_cast_fp16)[name = tensor("op_3496_cast_fp16")]; + tensor tile_44 = const()[name = tensor("tile_44"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3513_axis_0 = const()[name = tensor("op_3513_axis_0"), val = tensor(1)]; + tensor var_3513_cast_fp16_0, tensor var_3513_cast_fp16_1, tensor var_3513_cast_fp16_2, tensor var_3513_cast_fp16_3, tensor var_3513_cast_fp16_4, tensor var_3513_cast_fp16_5, tensor var_3513_cast_fp16_6, tensor var_3513_cast_fp16_7, tensor var_3513_cast_fp16_8, tensor var_3513_cast_fp16_9, tensor var_3513_cast_fp16_10, tensor var_3513_cast_fp16_11, tensor var_3513_cast_fp16_12, tensor var_3513_cast_fp16_13, tensor var_3513_cast_fp16_14, tensor var_3513_cast_fp16_15 = split(axis = var_3513_axis_0, split_sizes = tile_44, x = var_3475_cast_fp16)[name = tensor("op_3513_cast_fp16")]; + tensor aw_449_equation_0 = const()[name = tensor("aw_449_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_449_cast_fp16 = einsum(equation = aw_449_equation_0, values = (var_3496_cast_fp16_0, var_3478_cast_fp16_0))[name = tensor("aw_449_cast_fp16")]; + tensor aw_451_equation_0 = const()[name = tensor("aw_451_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_451_cast_fp16 = einsum(equation = aw_451_equation_0, values = (var_3496_cast_fp16_1, var_3478_cast_fp16_1))[name = tensor("aw_451_cast_fp16")]; + tensor aw_453_equation_0 = const()[name = tensor("aw_453_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_453_cast_fp16 = einsum(equation = aw_453_equation_0, values = (var_3496_cast_fp16_2, var_3478_cast_fp16_2))[name = tensor("aw_453_cast_fp16")]; + tensor aw_455_equation_0 = const()[name = tensor("aw_455_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_455_cast_fp16 = einsum(equation = aw_455_equation_0, values = (var_3496_cast_fp16_3, var_3478_cast_fp16_3))[name = tensor("aw_455_cast_fp16")]; + tensor aw_457_equation_0 = const()[name = tensor("aw_457_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_457_cast_fp16 = einsum(equation = aw_457_equation_0, values = (var_3496_cast_fp16_4, var_3478_cast_fp16_4))[name = tensor("aw_457_cast_fp16")]; + tensor aw_459_equation_0 = const()[name = tensor("aw_459_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_459_cast_fp16 = einsum(equation = aw_459_equation_0, values = (var_3496_cast_fp16_5, var_3478_cast_fp16_5))[name = tensor("aw_459_cast_fp16")]; + tensor aw_461_equation_0 = const()[name = tensor("aw_461_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_461_cast_fp16 = einsum(equation = aw_461_equation_0, values = (var_3496_cast_fp16_6, var_3478_cast_fp16_6))[name = tensor("aw_461_cast_fp16")]; + tensor aw_463_equation_0 = const()[name = tensor("aw_463_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_463_cast_fp16 = einsum(equation = aw_463_equation_0, values = (var_3496_cast_fp16_7, var_3478_cast_fp16_7))[name = tensor("aw_463_cast_fp16")]; + tensor aw_465_equation_0 = const()[name = tensor("aw_465_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_465_cast_fp16 = einsum(equation = aw_465_equation_0, values = (var_3496_cast_fp16_8, var_3478_cast_fp16_8))[name = tensor("aw_465_cast_fp16")]; + tensor aw_467_equation_0 = const()[name = tensor("aw_467_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_467_cast_fp16 = einsum(equation = aw_467_equation_0, values = (var_3496_cast_fp16_9, var_3478_cast_fp16_9))[name = tensor("aw_467_cast_fp16")]; + tensor aw_469_equation_0 = const()[name = tensor("aw_469_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_469_cast_fp16 = einsum(equation = aw_469_equation_0, values = (var_3496_cast_fp16_10, var_3478_cast_fp16_10))[name = tensor("aw_469_cast_fp16")]; + tensor aw_471_equation_0 = const()[name = tensor("aw_471_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_471_cast_fp16 = einsum(equation = aw_471_equation_0, values = (var_3496_cast_fp16_11, var_3478_cast_fp16_11))[name = tensor("aw_471_cast_fp16")]; + tensor aw_473_equation_0 = const()[name = tensor("aw_473_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_473_cast_fp16 = einsum(equation = aw_473_equation_0, values = (var_3496_cast_fp16_12, var_3478_cast_fp16_12))[name = tensor("aw_473_cast_fp16")]; + tensor aw_475_equation_0 = const()[name = tensor("aw_475_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_475_cast_fp16 = einsum(equation = aw_475_equation_0, values = (var_3496_cast_fp16_13, var_3478_cast_fp16_13))[name = tensor("aw_475_cast_fp16")]; + tensor aw_477_equation_0 = const()[name = tensor("aw_477_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_477_cast_fp16 = einsum(equation = aw_477_equation_0, values = (var_3496_cast_fp16_14, var_3478_cast_fp16_14))[name = tensor("aw_477_cast_fp16")]; + tensor aw_479_equation_0 = const()[name = tensor("aw_479_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_479_cast_fp16 = einsum(equation = aw_479_equation_0, values = (var_3496_cast_fp16_15, var_3478_cast_fp16_15))[name = tensor("aw_479_cast_fp16")]; + tensor var_3562_cast_fp16 = softmax(axis = var_3426, x = aw_449_cast_fp16)[name = tensor("op_3562_cast_fp16")]; + tensor var_3563_cast_fp16 = softmax(axis = var_3426, x = aw_451_cast_fp16)[name = tensor("op_3563_cast_fp16")]; + tensor var_3564_cast_fp16 = softmax(axis = var_3426, x = aw_453_cast_fp16)[name = tensor("op_3564_cast_fp16")]; + tensor var_3565_cast_fp16 = softmax(axis = var_3426, x = aw_455_cast_fp16)[name = tensor("op_3565_cast_fp16")]; + tensor var_3566_cast_fp16 = softmax(axis = var_3426, x = aw_457_cast_fp16)[name = tensor("op_3566_cast_fp16")]; + tensor var_3567_cast_fp16 = softmax(axis = var_3426, x = aw_459_cast_fp16)[name = tensor("op_3567_cast_fp16")]; + tensor var_3568_cast_fp16 = softmax(axis = var_3426, x = aw_461_cast_fp16)[name = tensor("op_3568_cast_fp16")]; + tensor var_3569_cast_fp16 = softmax(axis = var_3426, x = aw_463_cast_fp16)[name = tensor("op_3569_cast_fp16")]; + tensor var_3570_cast_fp16 = softmax(axis = var_3426, x = aw_465_cast_fp16)[name = tensor("op_3570_cast_fp16")]; + tensor var_3571_cast_fp16 = softmax(axis = var_3426, x = aw_467_cast_fp16)[name = tensor("op_3571_cast_fp16")]; + tensor var_3572_cast_fp16 = softmax(axis = var_3426, x = aw_469_cast_fp16)[name = tensor("op_3572_cast_fp16")]; + tensor var_3573_cast_fp16 = softmax(axis = var_3426, x = aw_471_cast_fp16)[name = tensor("op_3573_cast_fp16")]; + tensor var_3574_cast_fp16 = softmax(axis = var_3426, x = aw_473_cast_fp16)[name = tensor("op_3574_cast_fp16")]; + tensor var_3575_cast_fp16 = softmax(axis = var_3426, x = aw_475_cast_fp16)[name = tensor("op_3575_cast_fp16")]; + tensor var_3576_cast_fp16 = softmax(axis = var_3426, x = aw_477_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_cast_fp16 = softmax(axis = var_3426, x = aw_479_cast_fp16)[name = tensor("op_3577_cast_fp16")]; + tensor var_3579_equation_0 = const()[name = tensor("op_3579_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3579_cast_fp16 = einsum(equation = var_3579_equation_0, values = (var_3513_cast_fp16_0, var_3562_cast_fp16))[name = tensor("op_3579_cast_fp16")]; + tensor var_3581_equation_0 = const()[name = tensor("op_3581_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3581_cast_fp16 = einsum(equation = var_3581_equation_0, values = (var_3513_cast_fp16_1, var_3563_cast_fp16))[name = tensor("op_3581_cast_fp16")]; + tensor var_3583_equation_0 = const()[name = tensor("op_3583_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3583_cast_fp16 = einsum(equation = var_3583_equation_0, values = (var_3513_cast_fp16_2, var_3564_cast_fp16))[name = tensor("op_3583_cast_fp16")]; + tensor var_3585_equation_0 = const()[name = tensor("op_3585_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3585_cast_fp16 = einsum(equation = var_3585_equation_0, values = (var_3513_cast_fp16_3, var_3565_cast_fp16))[name = tensor("op_3585_cast_fp16")]; + tensor var_3587_equation_0 = const()[name = tensor("op_3587_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3587_cast_fp16 = einsum(equation = var_3587_equation_0, values = (var_3513_cast_fp16_4, var_3566_cast_fp16))[name = tensor("op_3587_cast_fp16")]; + tensor var_3589_equation_0 = const()[name = tensor("op_3589_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3589_cast_fp16 = einsum(equation = var_3589_equation_0, values = (var_3513_cast_fp16_5, var_3567_cast_fp16))[name = tensor("op_3589_cast_fp16")]; + tensor var_3591_equation_0 = const()[name = tensor("op_3591_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3591_cast_fp16 = einsum(equation = var_3591_equation_0, values = (var_3513_cast_fp16_6, var_3568_cast_fp16))[name = tensor("op_3591_cast_fp16")]; + tensor var_3593_equation_0 = const()[name = tensor("op_3593_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3593_cast_fp16 = einsum(equation = var_3593_equation_0, values = (var_3513_cast_fp16_7, var_3569_cast_fp16))[name = tensor("op_3593_cast_fp16")]; + tensor var_3595_equation_0 = const()[name = tensor("op_3595_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3595_cast_fp16 = einsum(equation = var_3595_equation_0, values = (var_3513_cast_fp16_8, var_3570_cast_fp16))[name = tensor("op_3595_cast_fp16")]; + tensor var_3597_equation_0 = const()[name = tensor("op_3597_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3597_cast_fp16 = einsum(equation = var_3597_equation_0, values = (var_3513_cast_fp16_9, var_3571_cast_fp16))[name = tensor("op_3597_cast_fp16")]; + tensor var_3599_equation_0 = const()[name = tensor("op_3599_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3599_cast_fp16 = einsum(equation = var_3599_equation_0, values = (var_3513_cast_fp16_10, var_3572_cast_fp16))[name = tensor("op_3599_cast_fp16")]; + tensor var_3601_equation_0 = const()[name = tensor("op_3601_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3601_cast_fp16 = einsum(equation = var_3601_equation_0, values = (var_3513_cast_fp16_11, var_3573_cast_fp16))[name = tensor("op_3601_cast_fp16")]; + tensor var_3603_equation_0 = const()[name = tensor("op_3603_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3603_cast_fp16 = einsum(equation = var_3603_equation_0, values = (var_3513_cast_fp16_12, var_3574_cast_fp16))[name = tensor("op_3603_cast_fp16")]; + tensor var_3605_equation_0 = const()[name = tensor("op_3605_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3605_cast_fp16 = einsum(equation = var_3605_equation_0, values = (var_3513_cast_fp16_13, var_3575_cast_fp16))[name = tensor("op_3605_cast_fp16")]; + tensor var_3607_equation_0 = const()[name = tensor("op_3607_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3607_cast_fp16 = einsum(equation = var_3607_equation_0, values = (var_3513_cast_fp16_14, var_3576_cast_fp16))[name = tensor("op_3607_cast_fp16")]; + tensor var_3609_equation_0 = const()[name = tensor("op_3609_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3609_cast_fp16 = einsum(equation = var_3609_equation_0, values = (var_3513_cast_fp16_15, var_3577_cast_fp16))[name = tensor("op_3609_cast_fp16")]; + tensor input_145_interleave_0 = const()[name = tensor("input_145_interleave_0"), val = tensor(false)]; + tensor input_145_cast_fp16 = concat(axis = var_3426, interleave = input_145_interleave_0, values = (var_3579_cast_fp16, var_3581_cast_fp16, var_3583_cast_fp16, var_3585_cast_fp16, var_3587_cast_fp16, var_3589_cast_fp16, var_3591_cast_fp16, var_3593_cast_fp16, var_3595_cast_fp16, var_3597_cast_fp16, var_3599_cast_fp16, var_3601_cast_fp16, var_3603_cast_fp16, var_3605_cast_fp16, var_3607_cast_fp16, var_3609_cast_fp16))[name = tensor("input_145_cast_fp16")]; + tensor var_3618_pad_type_0 = const()[name = tensor("op_3618_pad_type_0"), val = tensor("valid")]; + tensor var_3618_strides_0 = const()[name = tensor("op_3618_strides_0"), val = tensor([1, 1])]; + tensor var_3618_pad_0 = const()[name = tensor("op_3618_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3618_dilations_0 = const()[name = tensor("op_3618_dilations_0"), val = tensor([1, 1])]; + tensor var_3618_groups_0 = const()[name = tensor("op_3618_groups_0"), val = tensor(1)]; + tensor blocks_14_attn_out_weight_to_fp16 = const()[name = tensor("blocks_14_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368838592)))]; + tensor blocks_14_attn_out_bias_to_fp16 = const()[name = tensor("blocks_14_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370935808)))]; + tensor var_3618_cast_fp16 = conv(bias = blocks_14_attn_out_bias_to_fp16, dilations = var_3618_dilations_0, groups = var_3618_groups_0, pad = var_3618_pad_0, pad_type = var_3618_pad_type_0, strides = var_3618_strides_0, weight = blocks_14_attn_out_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_3618_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([1])]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370937920)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370940032)))]; + tensor var_3628_to_fp16 = const()[name = tensor("op_3628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = input_147_beta_0_to_fp16, epsilon = var_3628_to_fp16, gamma = input_147_gamma_0_to_fp16, x = inputs_59_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370942144)))]; + tensor blocks_14_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379330816)))]; + tensor input_149_cast_fp16 = conv(bias = blocks_14_mlp_0_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = blocks_14_mlp_0_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_3654_pad_type_0 = const()[name = tensor("op_3654_pad_type_0"), val = tensor("valid")]; + tensor var_3654_strides_0 = const()[name = tensor("op_3654_strides_0"), val = tensor([1, 1])]; + tensor var_3654_pad_0 = const()[name = tensor("op_3654_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3654_dilations_0 = const()[name = tensor("op_3654_dilations_0"), val = tensor([1, 1])]; + tensor var_3654_groups_0 = const()[name = tensor("op_3654_groups_0"), val = tensor(1)]; + tensor blocks_14_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379339072)))]; + tensor blocks_14_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387727744)))]; + tensor var_3654_cast_fp16 = conv(bias = blocks_14_mlp_2_bias_to_fp16, dilations = var_3654_dilations_0, groups = var_3654_groups_0, pad = var_3654_pad_0, pad_type = var_3654_pad_type_0, strides = var_3654_strides_0, weight = blocks_14_mlp_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("op_3654_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = var_3654_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_3663 = const()[name = tensor("op_3663"), val = tensor(1)]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([1])]; + tensor input_153_gamma_0_to_fp16 = const()[name = tensor("input_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387729856)))]; + tensor input_153_beta_0_to_fp16 = const()[name = tensor("input_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387731968)))]; + tensor var_3679_to_fp16 = const()[name = tensor("op_3679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = input_153_beta_0_to_fp16, epsilon = var_3679_to_fp16, gamma = input_153_gamma_0_to_fp16, x = inputs_61_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor q_31_pad_type_0 = const()[name = tensor("q_31_pad_type_0"), val = tensor("valid")]; + tensor q_31_strides_0 = const()[name = tensor("q_31_strides_0"), val = tensor([1, 1])]; + tensor q_31_pad_0 = const()[name = tensor("q_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_31_dilations_0 = const()[name = tensor("q_31_dilations_0"), val = tensor([1, 1])]; + tensor q_31_groups_0 = const()[name = tensor("q_31_groups_0"), val = tensor(1)]; + tensor var_3714_weight_0_to_fp16 = const()[name = tensor("op_3714_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387734080)))]; + tensor var_3714_bias_0_to_fp16 = const()[name = tensor("op_3714_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389831296)))]; + tensor var_3714_cast_fp16 = conv(bias = var_3714_bias_0_to_fp16, dilations = q_31_dilations_0, groups = q_31_groups_0, pad = q_31_pad_0, pad_type = q_31_pad_type_0, strides = q_31_strides_0, weight = var_3714_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("op_3714_cast_fp16")]; + tensor k_31_pad_type_0 = const()[name = tensor("k_31_pad_type_0"), val = tensor("valid")]; + tensor k_31_strides_0 = const()[name = tensor("k_31_strides_0"), val = tensor([1, 1])]; + tensor k_31_pad_0 = const()[name = tensor("k_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_31_dilations_0 = const()[name = tensor("k_31_dilations_0"), val = tensor([1, 1])]; + tensor k_31_groups_0 = const()[name = tensor("k_31_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_key_weight_to_fp16 = const()[name = tensor("blocks_15_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389833408)))]; + tensor k_31_cast_fp16 = conv(dilations = k_31_dilations_0, groups = k_31_groups_0, pad = k_31_pad_0, pad_type = k_31_pad_type_0, strides = k_31_strides_0, weight = blocks_15_attn_key_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor var_3712_pad_type_0 = const()[name = tensor("op_3712_pad_type_0"), val = tensor("valid")]; + tensor var_3712_strides_0 = const()[name = tensor("op_3712_strides_0"), val = tensor([1, 1])]; + tensor var_3712_pad_0 = const()[name = tensor("op_3712_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3712_dilations_0 = const()[name = tensor("op_3712_dilations_0"), val = tensor([1, 1])]; + tensor var_3712_groups_0 = const()[name = tensor("op_3712_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_value_weight_to_fp16 = const()[name = tensor("blocks_15_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391930624)))]; + tensor blocks_15_attn_value_bias_to_fp16 = const()[name = tensor("blocks_15_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394027840)))]; + tensor var_3712_cast_fp16 = conv(bias = blocks_15_attn_value_bias_to_fp16, dilations = var_3712_dilations_0, groups = var_3712_groups_0, pad = var_3712_pad_0, pad_type = var_3712_pad_type_0, strides = var_3712_strides_0, weight = blocks_15_attn_value_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("op_3712_cast_fp16")]; + tensor tile_45 = const()[name = tensor("tile_45"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3715_axis_0 = const()[name = tensor("op_3715_axis_0"), val = tensor(1)]; + tensor var_3715_cast_fp16_0, tensor var_3715_cast_fp16_1, tensor var_3715_cast_fp16_2, tensor var_3715_cast_fp16_3, tensor var_3715_cast_fp16_4, tensor var_3715_cast_fp16_5, tensor var_3715_cast_fp16_6, tensor var_3715_cast_fp16_7, tensor var_3715_cast_fp16_8, tensor var_3715_cast_fp16_9, tensor var_3715_cast_fp16_10, tensor var_3715_cast_fp16_11, tensor var_3715_cast_fp16_12, tensor var_3715_cast_fp16_13, tensor var_3715_cast_fp16_14, tensor var_3715_cast_fp16_15 = split(axis = var_3715_axis_0, split_sizes = tile_45, x = var_3714_cast_fp16)[name = tensor("op_3715_cast_fp16")]; + tensor var_3732_perm_0 = const()[name = tensor("op_3732_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_46 = const()[name = tensor("tile_46"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3733_axis_0 = const()[name = tensor("op_3733_axis_0"), val = tensor(3)]; + tensor var_3732_cast_fp16 = transpose(perm = var_3732_perm_0, x = k_31_cast_fp16)[name = tensor("transpose_9")]; + tensor var_3733_cast_fp16_0, tensor var_3733_cast_fp16_1, tensor var_3733_cast_fp16_2, tensor var_3733_cast_fp16_3, tensor var_3733_cast_fp16_4, tensor var_3733_cast_fp16_5, tensor var_3733_cast_fp16_6, tensor var_3733_cast_fp16_7, tensor var_3733_cast_fp16_8, tensor var_3733_cast_fp16_9, tensor var_3733_cast_fp16_10, tensor var_3733_cast_fp16_11, tensor var_3733_cast_fp16_12, tensor var_3733_cast_fp16_13, tensor var_3733_cast_fp16_14, tensor var_3733_cast_fp16_15 = split(axis = var_3733_axis_0, split_sizes = tile_46, x = var_3732_cast_fp16)[name = tensor("op_3733_cast_fp16")]; + tensor tile_47 = const()[name = tensor("tile_47"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3750_axis_0 = const()[name = tensor("op_3750_axis_0"), val = tensor(1)]; + tensor var_3750_cast_fp16_0, tensor var_3750_cast_fp16_1, tensor var_3750_cast_fp16_2, tensor var_3750_cast_fp16_3, tensor var_3750_cast_fp16_4, tensor var_3750_cast_fp16_5, tensor var_3750_cast_fp16_6, tensor var_3750_cast_fp16_7, tensor var_3750_cast_fp16_8, tensor var_3750_cast_fp16_9, tensor var_3750_cast_fp16_10, tensor var_3750_cast_fp16_11, tensor var_3750_cast_fp16_12, tensor var_3750_cast_fp16_13, tensor var_3750_cast_fp16_14, tensor var_3750_cast_fp16_15 = split(axis = var_3750_axis_0, split_sizes = tile_47, x = var_3712_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor aw_481_equation_0 = const()[name = tensor("aw_481_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_481_cast_fp16 = einsum(equation = aw_481_equation_0, values = (var_3733_cast_fp16_0, var_3715_cast_fp16_0))[name = tensor("aw_481_cast_fp16")]; + tensor aw_483_equation_0 = const()[name = tensor("aw_483_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_483_cast_fp16 = einsum(equation = aw_483_equation_0, values = (var_3733_cast_fp16_1, var_3715_cast_fp16_1))[name = tensor("aw_483_cast_fp16")]; + tensor aw_485_equation_0 = const()[name = tensor("aw_485_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_485_cast_fp16 = einsum(equation = aw_485_equation_0, values = (var_3733_cast_fp16_2, var_3715_cast_fp16_2))[name = tensor("aw_485_cast_fp16")]; + tensor aw_487_equation_0 = const()[name = tensor("aw_487_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_487_cast_fp16 = einsum(equation = aw_487_equation_0, values = (var_3733_cast_fp16_3, var_3715_cast_fp16_3))[name = tensor("aw_487_cast_fp16")]; + tensor aw_489_equation_0 = const()[name = tensor("aw_489_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_489_cast_fp16 = einsum(equation = aw_489_equation_0, values = (var_3733_cast_fp16_4, var_3715_cast_fp16_4))[name = tensor("aw_489_cast_fp16")]; + tensor aw_491_equation_0 = const()[name = tensor("aw_491_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_491_cast_fp16 = einsum(equation = aw_491_equation_0, values = (var_3733_cast_fp16_5, var_3715_cast_fp16_5))[name = tensor("aw_491_cast_fp16")]; + tensor aw_493_equation_0 = const()[name = tensor("aw_493_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_493_cast_fp16 = einsum(equation = aw_493_equation_0, values = (var_3733_cast_fp16_6, var_3715_cast_fp16_6))[name = tensor("aw_493_cast_fp16")]; + tensor aw_495_equation_0 = const()[name = tensor("aw_495_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_495_cast_fp16 = einsum(equation = aw_495_equation_0, values = (var_3733_cast_fp16_7, var_3715_cast_fp16_7))[name = tensor("aw_495_cast_fp16")]; + tensor aw_497_equation_0 = const()[name = tensor("aw_497_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_497_cast_fp16 = einsum(equation = aw_497_equation_0, values = (var_3733_cast_fp16_8, var_3715_cast_fp16_8))[name = tensor("aw_497_cast_fp16")]; + tensor aw_499_equation_0 = const()[name = tensor("aw_499_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_499_cast_fp16 = einsum(equation = aw_499_equation_0, values = (var_3733_cast_fp16_9, var_3715_cast_fp16_9))[name = tensor("aw_499_cast_fp16")]; + tensor aw_501_equation_0 = const()[name = tensor("aw_501_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_501_cast_fp16 = einsum(equation = aw_501_equation_0, values = (var_3733_cast_fp16_10, var_3715_cast_fp16_10))[name = tensor("aw_501_cast_fp16")]; + tensor aw_503_equation_0 = const()[name = tensor("aw_503_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_503_cast_fp16 = einsum(equation = aw_503_equation_0, values = (var_3733_cast_fp16_11, var_3715_cast_fp16_11))[name = tensor("aw_503_cast_fp16")]; + tensor aw_505_equation_0 = const()[name = tensor("aw_505_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_505_cast_fp16 = einsum(equation = aw_505_equation_0, values = (var_3733_cast_fp16_12, var_3715_cast_fp16_12))[name = tensor("aw_505_cast_fp16")]; + tensor aw_507_equation_0 = const()[name = tensor("aw_507_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_507_cast_fp16 = einsum(equation = aw_507_equation_0, values = (var_3733_cast_fp16_13, var_3715_cast_fp16_13))[name = tensor("aw_507_cast_fp16")]; + tensor aw_509_equation_0 = const()[name = tensor("aw_509_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_509_cast_fp16 = einsum(equation = aw_509_equation_0, values = (var_3733_cast_fp16_14, var_3715_cast_fp16_14))[name = tensor("aw_509_cast_fp16")]; + tensor aw_511_equation_0 = const()[name = tensor("aw_511_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_511_cast_fp16 = einsum(equation = aw_511_equation_0, values = (var_3733_cast_fp16_15, var_3715_cast_fp16_15))[name = tensor("aw_511_cast_fp16")]; + tensor var_3799_cast_fp16 = softmax(axis = var_3663, x = aw_481_cast_fp16)[name = tensor("op_3799_cast_fp16")]; + tensor var_3800_cast_fp16 = softmax(axis = var_3663, x = aw_483_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor var_3801_cast_fp16 = softmax(axis = var_3663, x = aw_485_cast_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor var_3802_cast_fp16 = softmax(axis = var_3663, x = aw_487_cast_fp16)[name = tensor("op_3802_cast_fp16")]; + tensor var_3803_cast_fp16 = softmax(axis = var_3663, x = aw_489_cast_fp16)[name = tensor("op_3803_cast_fp16")]; + tensor var_3804_cast_fp16 = softmax(axis = var_3663, x = aw_491_cast_fp16)[name = tensor("op_3804_cast_fp16")]; + tensor var_3805_cast_fp16 = softmax(axis = var_3663, x = aw_493_cast_fp16)[name = tensor("op_3805_cast_fp16")]; + tensor var_3806_cast_fp16 = softmax(axis = var_3663, x = aw_495_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor var_3807_cast_fp16 = softmax(axis = var_3663, x = aw_497_cast_fp16)[name = tensor("op_3807_cast_fp16")]; + tensor var_3808_cast_fp16 = softmax(axis = var_3663, x = aw_499_cast_fp16)[name = tensor("op_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = softmax(axis = var_3663, x = aw_501_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_cast_fp16 = softmax(axis = var_3663, x = aw_503_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = softmax(axis = var_3663, x = aw_505_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor var_3812_cast_fp16 = softmax(axis = var_3663, x = aw_507_cast_fp16)[name = tensor("op_3812_cast_fp16")]; + tensor var_3813_cast_fp16 = softmax(axis = var_3663, x = aw_509_cast_fp16)[name = tensor("op_3813_cast_fp16")]; + tensor var_3814_cast_fp16 = softmax(axis = var_3663, x = aw_511_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3816_equation_0 = const()[name = tensor("op_3816_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3816_cast_fp16 = einsum(equation = var_3816_equation_0, values = (var_3750_cast_fp16_0, var_3799_cast_fp16))[name = tensor("op_3816_cast_fp16")]; + tensor var_3818_equation_0 = const()[name = tensor("op_3818_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3818_cast_fp16 = einsum(equation = var_3818_equation_0, values = (var_3750_cast_fp16_1, var_3800_cast_fp16))[name = tensor("op_3818_cast_fp16")]; + tensor var_3820_equation_0 = const()[name = tensor("op_3820_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3820_cast_fp16 = einsum(equation = var_3820_equation_0, values = (var_3750_cast_fp16_2, var_3801_cast_fp16))[name = tensor("op_3820_cast_fp16")]; + tensor var_3822_equation_0 = const()[name = tensor("op_3822_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3822_cast_fp16 = einsum(equation = var_3822_equation_0, values = (var_3750_cast_fp16_3, var_3802_cast_fp16))[name = tensor("op_3822_cast_fp16")]; + tensor var_3824_equation_0 = const()[name = tensor("op_3824_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3824_cast_fp16 = einsum(equation = var_3824_equation_0, values = (var_3750_cast_fp16_4, var_3803_cast_fp16))[name = tensor("op_3824_cast_fp16")]; + tensor var_3826_equation_0 = const()[name = tensor("op_3826_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3826_cast_fp16 = einsum(equation = var_3826_equation_0, values = (var_3750_cast_fp16_5, var_3804_cast_fp16))[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_equation_0 = const()[name = tensor("op_3828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3828_cast_fp16 = einsum(equation = var_3828_equation_0, values = (var_3750_cast_fp16_6, var_3805_cast_fp16))[name = tensor("op_3828_cast_fp16")]; + tensor var_3830_equation_0 = const()[name = tensor("op_3830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3830_cast_fp16 = einsum(equation = var_3830_equation_0, values = (var_3750_cast_fp16_7, var_3806_cast_fp16))[name = tensor("op_3830_cast_fp16")]; + tensor var_3832_equation_0 = const()[name = tensor("op_3832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3832_cast_fp16 = einsum(equation = var_3832_equation_0, values = (var_3750_cast_fp16_8, var_3807_cast_fp16))[name = tensor("op_3832_cast_fp16")]; + tensor var_3834_equation_0 = const()[name = tensor("op_3834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3834_cast_fp16 = einsum(equation = var_3834_equation_0, values = (var_3750_cast_fp16_9, var_3808_cast_fp16))[name = tensor("op_3834_cast_fp16")]; + tensor var_3836_equation_0 = const()[name = tensor("op_3836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3836_cast_fp16 = einsum(equation = var_3836_equation_0, values = (var_3750_cast_fp16_10, var_3809_cast_fp16))[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_equation_0 = const()[name = tensor("op_3838_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3838_cast_fp16 = einsum(equation = var_3838_equation_0, values = (var_3750_cast_fp16_11, var_3810_cast_fp16))[name = tensor("op_3838_cast_fp16")]; + tensor var_3840_equation_0 = const()[name = tensor("op_3840_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3840_cast_fp16 = einsum(equation = var_3840_equation_0, values = (var_3750_cast_fp16_12, var_3811_cast_fp16))[name = tensor("op_3840_cast_fp16")]; + tensor var_3842_equation_0 = const()[name = tensor("op_3842_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3842_cast_fp16 = einsum(equation = var_3842_equation_0, values = (var_3750_cast_fp16_13, var_3812_cast_fp16))[name = tensor("op_3842_cast_fp16")]; + tensor var_3844_equation_0 = const()[name = tensor("op_3844_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3844_cast_fp16 = einsum(equation = var_3844_equation_0, values = (var_3750_cast_fp16_14, var_3813_cast_fp16))[name = tensor("op_3844_cast_fp16")]; + tensor var_3846_equation_0 = const()[name = tensor("op_3846_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_3846_cast_fp16 = einsum(equation = var_3846_equation_0, values = (var_3750_cast_fp16_15, var_3814_cast_fp16))[name = tensor("op_3846_cast_fp16")]; + tensor input_155_interleave_0 = const()[name = tensor("input_155_interleave_0"), val = tensor(false)]; + tensor input_155_cast_fp16 = concat(axis = var_3663, interleave = input_155_interleave_0, values = (var_3816_cast_fp16, var_3818_cast_fp16, var_3820_cast_fp16, var_3822_cast_fp16, var_3824_cast_fp16, var_3826_cast_fp16, var_3828_cast_fp16, var_3830_cast_fp16, var_3832_cast_fp16, var_3834_cast_fp16, var_3836_cast_fp16, var_3838_cast_fp16, var_3840_cast_fp16, var_3842_cast_fp16, var_3844_cast_fp16, var_3846_cast_fp16))[name = tensor("input_155_cast_fp16")]; + tensor var_3855_pad_type_0 = const()[name = tensor("op_3855_pad_type_0"), val = tensor("valid")]; + tensor var_3855_strides_0 = const()[name = tensor("op_3855_strides_0"), val = tensor([1, 1])]; + tensor var_3855_pad_0 = const()[name = tensor("op_3855_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3855_dilations_0 = const()[name = tensor("op_3855_dilations_0"), val = tensor([1, 1])]; + tensor var_3855_groups_0 = const()[name = tensor("op_3855_groups_0"), val = tensor(1)]; + tensor blocks_15_attn_out_weight_to_fp16 = const()[name = tensor("blocks_15_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394029952)))]; + tensor blocks_15_attn_out_bias_to_fp16 = const()[name = tensor("blocks_15_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396127168)))]; + tensor var_3855_cast_fp16 = conv(bias = blocks_15_attn_out_bias_to_fp16, dilations = var_3855_dilations_0, groups = var_3855_groups_0, pad = var_3855_pad_0, pad_type = var_3855_pad_type_0, strides = var_3855_strides_0, weight = blocks_15_attn_out_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("op_3855_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_3855_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor input_157_axes_0 = const()[name = tensor("input_157_axes_0"), val = tensor([1])]; + tensor input_157_gamma_0_to_fp16 = const()[name = tensor("input_157_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396129280)))]; + tensor input_157_beta_0_to_fp16 = const()[name = tensor("input_157_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396131392)))]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = input_157_beta_0_to_fp16, epsilon = var_3865_to_fp16, gamma = input_157_gamma_0_to_fp16, x = inputs_63_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("valid")]; + tensor input_159_strides_0 = const()[name = tensor("input_159_strides_0"), val = tensor([1, 1])]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_159_dilations_0 = const()[name = tensor("input_159_dilations_0"), val = tensor([1, 1])]; + tensor input_159_groups_0 = const()[name = tensor("input_159_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396133504)))]; + tensor blocks_15_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404522176)))]; + tensor input_159_cast_fp16 = conv(bias = blocks_15_mlp_0_bias_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = blocks_15_mlp_0_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_3891_pad_type_0 = const()[name = tensor("op_3891_pad_type_0"), val = tensor("valid")]; + tensor var_3891_strides_0 = const()[name = tensor("op_3891_strides_0"), val = tensor([1, 1])]; + tensor var_3891_pad_0 = const()[name = tensor("op_3891_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3891_dilations_0 = const()[name = tensor("op_3891_dilations_0"), val = tensor([1, 1])]; + tensor var_3891_groups_0 = const()[name = tensor("op_3891_groups_0"), val = tensor(1)]; + tensor blocks_15_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404530432)))]; + tensor blocks_15_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412919104)))]; + tensor var_3891_cast_fp16 = conv(bias = blocks_15_mlp_2_bias_to_fp16, dilations = var_3891_dilations_0, groups = var_3891_groups_0, pad = var_3891_pad_0, pad_type = var_3891_pad_type_0, strides = var_3891_strides_0, weight = blocks_15_mlp_2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("op_3891_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = var_3891_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor(1)]; + tensor input_163_axes_0 = const()[name = tensor("input_163_axes_0"), val = tensor([1])]; + tensor input_163_gamma_0_to_fp16 = const()[name = tensor("input_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412921216)))]; + tensor input_163_beta_0_to_fp16 = const()[name = tensor("input_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412923328)))]; + tensor var_3916_to_fp16 = const()[name = tensor("op_3916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = input_163_beta_0_to_fp16, epsilon = var_3916_to_fp16, gamma = input_163_gamma_0_to_fp16, x = inputs_65_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor q_33_pad_type_0 = const()[name = tensor("q_33_pad_type_0"), val = tensor("valid")]; + tensor q_33_strides_0 = const()[name = tensor("q_33_strides_0"), val = tensor([1, 1])]; + tensor q_33_pad_0 = const()[name = tensor("q_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_33_dilations_0 = const()[name = tensor("q_33_dilations_0"), val = tensor([1, 1])]; + tensor q_33_groups_0 = const()[name = tensor("q_33_groups_0"), val = tensor(1)]; + tensor var_3951_weight_0_to_fp16 = const()[name = tensor("op_3951_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412925440)))]; + tensor var_3951_bias_0_to_fp16 = const()[name = tensor("op_3951_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415022656)))]; + tensor var_3951_cast_fp16 = conv(bias = var_3951_bias_0_to_fp16, dilations = q_33_dilations_0, groups = q_33_groups_0, pad = q_33_pad_0, pad_type = q_33_pad_type_0, strides = q_33_strides_0, weight = var_3951_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor k_33_pad_type_0 = const()[name = tensor("k_33_pad_type_0"), val = tensor("valid")]; + tensor k_33_strides_0 = const()[name = tensor("k_33_strides_0"), val = tensor([1, 1])]; + tensor k_33_pad_0 = const()[name = tensor("k_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_33_dilations_0 = const()[name = tensor("k_33_dilations_0"), val = tensor([1, 1])]; + tensor k_33_groups_0 = const()[name = tensor("k_33_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_key_weight_to_fp16 = const()[name = tensor("blocks_16_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415024768)))]; + tensor k_33_cast_fp16 = conv(dilations = k_33_dilations_0, groups = k_33_groups_0, pad = k_33_pad_0, pad_type = k_33_pad_type_0, strides = k_33_strides_0, weight = blocks_16_attn_key_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor var_3949_pad_type_0 = const()[name = tensor("op_3949_pad_type_0"), val = tensor("valid")]; + tensor var_3949_strides_0 = const()[name = tensor("op_3949_strides_0"), val = tensor([1, 1])]; + tensor var_3949_pad_0 = const()[name = tensor("op_3949_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3949_dilations_0 = const()[name = tensor("op_3949_dilations_0"), val = tensor([1, 1])]; + tensor var_3949_groups_0 = const()[name = tensor("op_3949_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_value_weight_to_fp16 = const()[name = tensor("blocks_16_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417121984)))]; + tensor blocks_16_attn_value_bias_to_fp16 = const()[name = tensor("blocks_16_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419219200)))]; + tensor var_3949_cast_fp16 = conv(bias = blocks_16_attn_value_bias_to_fp16, dilations = var_3949_dilations_0, groups = var_3949_groups_0, pad = var_3949_pad_0, pad_type = var_3949_pad_type_0, strides = var_3949_strides_0, weight = blocks_16_attn_value_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("op_3949_cast_fp16")]; + tensor tile_48 = const()[name = tensor("tile_48"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3952_axis_0 = const()[name = tensor("op_3952_axis_0"), val = tensor(1)]; + tensor var_3952_cast_fp16_0, tensor var_3952_cast_fp16_1, tensor var_3952_cast_fp16_2, tensor var_3952_cast_fp16_3, tensor var_3952_cast_fp16_4, tensor var_3952_cast_fp16_5, tensor var_3952_cast_fp16_6, tensor var_3952_cast_fp16_7, tensor var_3952_cast_fp16_8, tensor var_3952_cast_fp16_9, tensor var_3952_cast_fp16_10, tensor var_3952_cast_fp16_11, tensor var_3952_cast_fp16_12, tensor var_3952_cast_fp16_13, tensor var_3952_cast_fp16_14, tensor var_3952_cast_fp16_15 = split(axis = var_3952_axis_0, split_sizes = tile_48, x = var_3951_cast_fp16)[name = tensor("op_3952_cast_fp16")]; + tensor var_3969_perm_0 = const()[name = tensor("op_3969_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_49 = const()[name = tensor("tile_49"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3970_axis_0 = const()[name = tensor("op_3970_axis_0"), val = tensor(3)]; + tensor var_3969_cast_fp16 = transpose(perm = var_3969_perm_0, x = k_33_cast_fp16)[name = tensor("transpose_8")]; + tensor var_3970_cast_fp16_0, tensor var_3970_cast_fp16_1, tensor var_3970_cast_fp16_2, tensor var_3970_cast_fp16_3, tensor var_3970_cast_fp16_4, tensor var_3970_cast_fp16_5, tensor var_3970_cast_fp16_6, tensor var_3970_cast_fp16_7, tensor var_3970_cast_fp16_8, tensor var_3970_cast_fp16_9, tensor var_3970_cast_fp16_10, tensor var_3970_cast_fp16_11, tensor var_3970_cast_fp16_12, tensor var_3970_cast_fp16_13, tensor var_3970_cast_fp16_14, tensor var_3970_cast_fp16_15 = split(axis = var_3970_axis_0, split_sizes = tile_49, x = var_3969_cast_fp16)[name = tensor("op_3970_cast_fp16")]; + tensor tile_50 = const()[name = tensor("tile_50"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_3987_axis_0 = const()[name = tensor("op_3987_axis_0"), val = tensor(1)]; + tensor var_3987_cast_fp16_0, tensor var_3987_cast_fp16_1, tensor var_3987_cast_fp16_2, tensor var_3987_cast_fp16_3, tensor var_3987_cast_fp16_4, tensor var_3987_cast_fp16_5, tensor var_3987_cast_fp16_6, tensor var_3987_cast_fp16_7, tensor var_3987_cast_fp16_8, tensor var_3987_cast_fp16_9, tensor var_3987_cast_fp16_10, tensor var_3987_cast_fp16_11, tensor var_3987_cast_fp16_12, tensor var_3987_cast_fp16_13, tensor var_3987_cast_fp16_14, tensor var_3987_cast_fp16_15 = split(axis = var_3987_axis_0, split_sizes = tile_50, x = var_3949_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor aw_513_equation_0 = const()[name = tensor("aw_513_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_513_cast_fp16 = einsum(equation = aw_513_equation_0, values = (var_3970_cast_fp16_0, var_3952_cast_fp16_0))[name = tensor("aw_513_cast_fp16")]; + tensor aw_515_equation_0 = const()[name = tensor("aw_515_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_515_cast_fp16 = einsum(equation = aw_515_equation_0, values = (var_3970_cast_fp16_1, var_3952_cast_fp16_1))[name = tensor("aw_515_cast_fp16")]; + tensor aw_517_equation_0 = const()[name = tensor("aw_517_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_517_cast_fp16 = einsum(equation = aw_517_equation_0, values = (var_3970_cast_fp16_2, var_3952_cast_fp16_2))[name = tensor("aw_517_cast_fp16")]; + tensor aw_519_equation_0 = const()[name = tensor("aw_519_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_519_cast_fp16 = einsum(equation = aw_519_equation_0, values = (var_3970_cast_fp16_3, var_3952_cast_fp16_3))[name = tensor("aw_519_cast_fp16")]; + tensor aw_521_equation_0 = const()[name = tensor("aw_521_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_521_cast_fp16 = einsum(equation = aw_521_equation_0, values = (var_3970_cast_fp16_4, var_3952_cast_fp16_4))[name = tensor("aw_521_cast_fp16")]; + tensor aw_523_equation_0 = const()[name = tensor("aw_523_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_523_cast_fp16 = einsum(equation = aw_523_equation_0, values = (var_3970_cast_fp16_5, var_3952_cast_fp16_5))[name = tensor("aw_523_cast_fp16")]; + tensor aw_525_equation_0 = const()[name = tensor("aw_525_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_525_cast_fp16 = einsum(equation = aw_525_equation_0, values = (var_3970_cast_fp16_6, var_3952_cast_fp16_6))[name = tensor("aw_525_cast_fp16")]; + tensor aw_527_equation_0 = const()[name = tensor("aw_527_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_527_cast_fp16 = einsum(equation = aw_527_equation_0, values = (var_3970_cast_fp16_7, var_3952_cast_fp16_7))[name = tensor("aw_527_cast_fp16")]; + tensor aw_529_equation_0 = const()[name = tensor("aw_529_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_529_cast_fp16 = einsum(equation = aw_529_equation_0, values = (var_3970_cast_fp16_8, var_3952_cast_fp16_8))[name = tensor("aw_529_cast_fp16")]; + tensor aw_531_equation_0 = const()[name = tensor("aw_531_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_531_cast_fp16 = einsum(equation = aw_531_equation_0, values = (var_3970_cast_fp16_9, var_3952_cast_fp16_9))[name = tensor("aw_531_cast_fp16")]; + tensor aw_533_equation_0 = const()[name = tensor("aw_533_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_533_cast_fp16 = einsum(equation = aw_533_equation_0, values = (var_3970_cast_fp16_10, var_3952_cast_fp16_10))[name = tensor("aw_533_cast_fp16")]; + tensor aw_535_equation_0 = const()[name = tensor("aw_535_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_535_cast_fp16 = einsum(equation = aw_535_equation_0, values = (var_3970_cast_fp16_11, var_3952_cast_fp16_11))[name = tensor("aw_535_cast_fp16")]; + tensor aw_537_equation_0 = const()[name = tensor("aw_537_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_537_cast_fp16 = einsum(equation = aw_537_equation_0, values = (var_3970_cast_fp16_12, var_3952_cast_fp16_12))[name = tensor("aw_537_cast_fp16")]; + tensor aw_539_equation_0 = const()[name = tensor("aw_539_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_539_cast_fp16 = einsum(equation = aw_539_equation_0, values = (var_3970_cast_fp16_13, var_3952_cast_fp16_13))[name = tensor("aw_539_cast_fp16")]; + tensor aw_541_equation_0 = const()[name = tensor("aw_541_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_541_cast_fp16 = einsum(equation = aw_541_equation_0, values = (var_3970_cast_fp16_14, var_3952_cast_fp16_14))[name = tensor("aw_541_cast_fp16")]; + tensor aw_543_equation_0 = const()[name = tensor("aw_543_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_543_cast_fp16 = einsum(equation = aw_543_equation_0, values = (var_3970_cast_fp16_15, var_3952_cast_fp16_15))[name = tensor("aw_543_cast_fp16")]; + tensor var_4036_cast_fp16 = softmax(axis = var_3900, x = aw_513_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor var_4037_cast_fp16 = softmax(axis = var_3900, x = aw_515_cast_fp16)[name = tensor("op_4037_cast_fp16")]; + tensor var_4038_cast_fp16 = softmax(axis = var_3900, x = aw_517_cast_fp16)[name = tensor("op_4038_cast_fp16")]; + tensor var_4039_cast_fp16 = softmax(axis = var_3900, x = aw_519_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = softmax(axis = var_3900, x = aw_521_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor var_4041_cast_fp16 = softmax(axis = var_3900, x = aw_523_cast_fp16)[name = tensor("op_4041_cast_fp16")]; + tensor var_4042_cast_fp16 = softmax(axis = var_3900, x = aw_525_cast_fp16)[name = tensor("op_4042_cast_fp16")]; + tensor var_4043_cast_fp16 = softmax(axis = var_3900, x = aw_527_cast_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4044_cast_fp16 = softmax(axis = var_3900, x = aw_529_cast_fp16)[name = tensor("op_4044_cast_fp16")]; + tensor var_4045_cast_fp16 = softmax(axis = var_3900, x = aw_531_cast_fp16)[name = tensor("op_4045_cast_fp16")]; + tensor var_4046_cast_fp16 = softmax(axis = var_3900, x = aw_533_cast_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor var_4047_cast_fp16 = softmax(axis = var_3900, x = aw_535_cast_fp16)[name = tensor("op_4047_cast_fp16")]; + tensor var_4048_cast_fp16 = softmax(axis = var_3900, x = aw_537_cast_fp16)[name = tensor("op_4048_cast_fp16")]; + tensor var_4049_cast_fp16 = softmax(axis = var_3900, x = aw_539_cast_fp16)[name = tensor("op_4049_cast_fp16")]; + tensor var_4050_cast_fp16 = softmax(axis = var_3900, x = aw_541_cast_fp16)[name = tensor("op_4050_cast_fp16")]; + tensor var_4051_cast_fp16 = softmax(axis = var_3900, x = aw_543_cast_fp16)[name = tensor("op_4051_cast_fp16")]; + tensor var_4053_equation_0 = const()[name = tensor("op_4053_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4053_cast_fp16 = einsum(equation = var_4053_equation_0, values = (var_3987_cast_fp16_0, var_4036_cast_fp16))[name = tensor("op_4053_cast_fp16")]; + tensor var_4055_equation_0 = const()[name = tensor("op_4055_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4055_cast_fp16 = einsum(equation = var_4055_equation_0, values = (var_3987_cast_fp16_1, var_4037_cast_fp16))[name = tensor("op_4055_cast_fp16")]; + tensor var_4057_equation_0 = const()[name = tensor("op_4057_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4057_cast_fp16 = einsum(equation = var_4057_equation_0, values = (var_3987_cast_fp16_2, var_4038_cast_fp16))[name = tensor("op_4057_cast_fp16")]; + tensor var_4059_equation_0 = const()[name = tensor("op_4059_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4059_cast_fp16 = einsum(equation = var_4059_equation_0, values = (var_3987_cast_fp16_3, var_4039_cast_fp16))[name = tensor("op_4059_cast_fp16")]; + tensor var_4061_equation_0 = const()[name = tensor("op_4061_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4061_cast_fp16 = einsum(equation = var_4061_equation_0, values = (var_3987_cast_fp16_4, var_4040_cast_fp16))[name = tensor("op_4061_cast_fp16")]; + tensor var_4063_equation_0 = const()[name = tensor("op_4063_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4063_cast_fp16 = einsum(equation = var_4063_equation_0, values = (var_3987_cast_fp16_5, var_4041_cast_fp16))[name = tensor("op_4063_cast_fp16")]; + tensor var_4065_equation_0 = const()[name = tensor("op_4065_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4065_cast_fp16 = einsum(equation = var_4065_equation_0, values = (var_3987_cast_fp16_6, var_4042_cast_fp16))[name = tensor("op_4065_cast_fp16")]; + tensor var_4067_equation_0 = const()[name = tensor("op_4067_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4067_cast_fp16 = einsum(equation = var_4067_equation_0, values = (var_3987_cast_fp16_7, var_4043_cast_fp16))[name = tensor("op_4067_cast_fp16")]; + tensor var_4069_equation_0 = const()[name = tensor("op_4069_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4069_cast_fp16 = einsum(equation = var_4069_equation_0, values = (var_3987_cast_fp16_8, var_4044_cast_fp16))[name = tensor("op_4069_cast_fp16")]; + tensor var_4071_equation_0 = const()[name = tensor("op_4071_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4071_cast_fp16 = einsum(equation = var_4071_equation_0, values = (var_3987_cast_fp16_9, var_4045_cast_fp16))[name = tensor("op_4071_cast_fp16")]; + tensor var_4073_equation_0 = const()[name = tensor("op_4073_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4073_cast_fp16 = einsum(equation = var_4073_equation_0, values = (var_3987_cast_fp16_10, var_4046_cast_fp16))[name = tensor("op_4073_cast_fp16")]; + tensor var_4075_equation_0 = const()[name = tensor("op_4075_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4075_cast_fp16 = einsum(equation = var_4075_equation_0, values = (var_3987_cast_fp16_11, var_4047_cast_fp16))[name = tensor("op_4075_cast_fp16")]; + tensor var_4077_equation_0 = const()[name = tensor("op_4077_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4077_cast_fp16 = einsum(equation = var_4077_equation_0, values = (var_3987_cast_fp16_12, var_4048_cast_fp16))[name = tensor("op_4077_cast_fp16")]; + tensor var_4079_equation_0 = const()[name = tensor("op_4079_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4079_cast_fp16 = einsum(equation = var_4079_equation_0, values = (var_3987_cast_fp16_13, var_4049_cast_fp16))[name = tensor("op_4079_cast_fp16")]; + tensor var_4081_equation_0 = const()[name = tensor("op_4081_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4081_cast_fp16 = einsum(equation = var_4081_equation_0, values = (var_3987_cast_fp16_14, var_4050_cast_fp16))[name = tensor("op_4081_cast_fp16")]; + tensor var_4083_equation_0 = const()[name = tensor("op_4083_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4083_cast_fp16 = einsum(equation = var_4083_equation_0, values = (var_3987_cast_fp16_15, var_4051_cast_fp16))[name = tensor("op_4083_cast_fp16")]; + tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165_cast_fp16 = concat(axis = var_3900, interleave = input_165_interleave_0, values = (var_4053_cast_fp16, var_4055_cast_fp16, var_4057_cast_fp16, var_4059_cast_fp16, var_4061_cast_fp16, var_4063_cast_fp16, var_4065_cast_fp16, var_4067_cast_fp16, var_4069_cast_fp16, var_4071_cast_fp16, var_4073_cast_fp16, var_4075_cast_fp16, var_4077_cast_fp16, var_4079_cast_fp16, var_4081_cast_fp16, var_4083_cast_fp16))[name = tensor("input_165_cast_fp16")]; + tensor var_4092_pad_type_0 = const()[name = tensor("op_4092_pad_type_0"), val = tensor("valid")]; + tensor var_4092_strides_0 = const()[name = tensor("op_4092_strides_0"), val = tensor([1, 1])]; + tensor var_4092_pad_0 = const()[name = tensor("op_4092_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4092_dilations_0 = const()[name = tensor("op_4092_dilations_0"), val = tensor([1, 1])]; + tensor var_4092_groups_0 = const()[name = tensor("op_4092_groups_0"), val = tensor(1)]; + tensor blocks_16_attn_out_weight_to_fp16 = const()[name = tensor("blocks_16_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419221312)))]; + tensor blocks_16_attn_out_bias_to_fp16 = const()[name = tensor("blocks_16_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421318528)))]; + tensor var_4092_cast_fp16 = conv(bias = blocks_16_attn_out_bias_to_fp16, dilations = var_4092_dilations_0, groups = var_4092_groups_0, pad = var_4092_pad_0, pad_type = var_4092_pad_type_0, strides = var_4092_strides_0, weight = blocks_16_attn_out_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("op_4092_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = var_4092_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([1])]; + tensor input_167_gamma_0_to_fp16 = const()[name = tensor("input_167_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421320640)))]; + tensor input_167_beta_0_to_fp16 = const()[name = tensor("input_167_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421322752)))]; + tensor var_4102_to_fp16 = const()[name = tensor("op_4102_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = input_167_beta_0_to_fp16, epsilon = var_4102_to_fp16, gamma = input_167_gamma_0_to_fp16, x = inputs_67_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("valid")]; + tensor input_169_strides_0 = const()[name = tensor("input_169_strides_0"), val = tensor([1, 1])]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_169_dilations_0 = const()[name = tensor("input_169_dilations_0"), val = tensor([1, 1])]; + tensor input_169_groups_0 = const()[name = tensor("input_169_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421324864)))]; + tensor blocks_16_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429713536)))]; + tensor input_169_cast_fp16 = conv(bias = blocks_16_mlp_0_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = blocks_16_mlp_0_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_4128_pad_type_0 = const()[name = tensor("op_4128_pad_type_0"), val = tensor("valid")]; + tensor var_4128_strides_0 = const()[name = tensor("op_4128_strides_0"), val = tensor([1, 1])]; + tensor var_4128_pad_0 = const()[name = tensor("op_4128_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4128_dilations_0 = const()[name = tensor("op_4128_dilations_0"), val = tensor([1, 1])]; + tensor var_4128_groups_0 = const()[name = tensor("op_4128_groups_0"), val = tensor(1)]; + tensor blocks_16_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429721792)))]; + tensor blocks_16_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438110464)))]; + tensor var_4128_cast_fp16 = conv(bias = blocks_16_mlp_2_bias_to_fp16, dilations = var_4128_dilations_0, groups = var_4128_groups_0, pad = var_4128_pad_0, pad_type = var_4128_pad_type_0, strides = var_4128_strides_0, weight = blocks_16_mlp_2_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("op_4128_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_4128_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4137 = const()[name = tensor("op_4137"), val = tensor(1)]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([1])]; + tensor input_173_gamma_0_to_fp16 = const()[name = tensor("input_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438112576)))]; + tensor input_173_beta_0_to_fp16 = const()[name = tensor("input_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438114688)))]; + tensor var_4153_to_fp16 = const()[name = tensor("op_4153_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = input_173_beta_0_to_fp16, epsilon = var_4153_to_fp16, gamma = input_173_gamma_0_to_fp16, x = inputs_69_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor q_35_pad_type_0 = const()[name = tensor("q_35_pad_type_0"), val = tensor("valid")]; + tensor q_35_strides_0 = const()[name = tensor("q_35_strides_0"), val = tensor([1, 1])]; + tensor q_35_pad_0 = const()[name = tensor("q_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_35_dilations_0 = const()[name = tensor("q_35_dilations_0"), val = tensor([1, 1])]; + tensor q_35_groups_0 = const()[name = tensor("q_35_groups_0"), val = tensor(1)]; + tensor var_4188_weight_0_to_fp16 = const()[name = tensor("op_4188_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438116800)))]; + tensor var_4188_bias_0_to_fp16 = const()[name = tensor("op_4188_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440214016)))]; + tensor var_4188_cast_fp16 = conv(bias = var_4188_bias_0_to_fp16, dilations = q_35_dilations_0, groups = q_35_groups_0, pad = q_35_pad_0, pad_type = q_35_pad_type_0, strides = q_35_strides_0, weight = var_4188_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4188_cast_fp16")]; + tensor k_35_pad_type_0 = const()[name = tensor("k_35_pad_type_0"), val = tensor("valid")]; + tensor k_35_strides_0 = const()[name = tensor("k_35_strides_0"), val = tensor([1, 1])]; + tensor k_35_pad_0 = const()[name = tensor("k_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_35_dilations_0 = const()[name = tensor("k_35_dilations_0"), val = tensor([1, 1])]; + tensor k_35_groups_0 = const()[name = tensor("k_35_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_key_weight_to_fp16 = const()[name = tensor("blocks_17_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440216128)))]; + tensor k_35_cast_fp16 = conv(dilations = k_35_dilations_0, groups = k_35_groups_0, pad = k_35_pad_0, pad_type = k_35_pad_type_0, strides = k_35_strides_0, weight = blocks_17_attn_key_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_4186_pad_type_0 = const()[name = tensor("op_4186_pad_type_0"), val = tensor("valid")]; + tensor var_4186_strides_0 = const()[name = tensor("op_4186_strides_0"), val = tensor([1, 1])]; + tensor var_4186_pad_0 = const()[name = tensor("op_4186_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4186_dilations_0 = const()[name = tensor("op_4186_dilations_0"), val = tensor([1, 1])]; + tensor var_4186_groups_0 = const()[name = tensor("op_4186_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_value_weight_to_fp16 = const()[name = tensor("blocks_17_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442313344)))]; + tensor blocks_17_attn_value_bias_to_fp16 = const()[name = tensor("blocks_17_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444410560)))]; + tensor var_4186_cast_fp16 = conv(bias = blocks_17_attn_value_bias_to_fp16, dilations = var_4186_dilations_0, groups = var_4186_groups_0, pad = var_4186_pad_0, pad_type = var_4186_pad_type_0, strides = var_4186_strides_0, weight = blocks_17_attn_value_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor tile_51 = const()[name = tensor("tile_51"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4189_axis_0 = const()[name = tensor("op_4189_axis_0"), val = tensor(1)]; + tensor var_4189_cast_fp16_0, tensor var_4189_cast_fp16_1, tensor var_4189_cast_fp16_2, tensor var_4189_cast_fp16_3, tensor var_4189_cast_fp16_4, tensor var_4189_cast_fp16_5, tensor var_4189_cast_fp16_6, tensor var_4189_cast_fp16_7, tensor var_4189_cast_fp16_8, tensor var_4189_cast_fp16_9, tensor var_4189_cast_fp16_10, tensor var_4189_cast_fp16_11, tensor var_4189_cast_fp16_12, tensor var_4189_cast_fp16_13, tensor var_4189_cast_fp16_14, tensor var_4189_cast_fp16_15 = split(axis = var_4189_axis_0, split_sizes = tile_51, x = var_4188_cast_fp16)[name = tensor("op_4189_cast_fp16")]; + tensor var_4206_perm_0 = const()[name = tensor("op_4206_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_52 = const()[name = tensor("tile_52"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4207_axis_0 = const()[name = tensor("op_4207_axis_0"), val = tensor(3)]; + tensor var_4206_cast_fp16 = transpose(perm = var_4206_perm_0, x = k_35_cast_fp16)[name = tensor("transpose_7")]; + tensor var_4207_cast_fp16_0, tensor var_4207_cast_fp16_1, tensor var_4207_cast_fp16_2, tensor var_4207_cast_fp16_3, tensor var_4207_cast_fp16_4, tensor var_4207_cast_fp16_5, tensor var_4207_cast_fp16_6, tensor var_4207_cast_fp16_7, tensor var_4207_cast_fp16_8, tensor var_4207_cast_fp16_9, tensor var_4207_cast_fp16_10, tensor var_4207_cast_fp16_11, tensor var_4207_cast_fp16_12, tensor var_4207_cast_fp16_13, tensor var_4207_cast_fp16_14, tensor var_4207_cast_fp16_15 = split(axis = var_4207_axis_0, split_sizes = tile_52, x = var_4206_cast_fp16)[name = tensor("op_4207_cast_fp16")]; + tensor tile_53 = const()[name = tensor("tile_53"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4224_axis_0 = const()[name = tensor("op_4224_axis_0"), val = tensor(1)]; + tensor var_4224_cast_fp16_0, tensor var_4224_cast_fp16_1, tensor var_4224_cast_fp16_2, tensor var_4224_cast_fp16_3, tensor var_4224_cast_fp16_4, tensor var_4224_cast_fp16_5, tensor var_4224_cast_fp16_6, tensor var_4224_cast_fp16_7, tensor var_4224_cast_fp16_8, tensor var_4224_cast_fp16_9, tensor var_4224_cast_fp16_10, tensor var_4224_cast_fp16_11, tensor var_4224_cast_fp16_12, tensor var_4224_cast_fp16_13, tensor var_4224_cast_fp16_14, tensor var_4224_cast_fp16_15 = split(axis = var_4224_axis_0, split_sizes = tile_53, x = var_4186_cast_fp16)[name = tensor("op_4224_cast_fp16")]; + tensor aw_545_equation_0 = const()[name = tensor("aw_545_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_545_cast_fp16 = einsum(equation = aw_545_equation_0, values = (var_4207_cast_fp16_0, var_4189_cast_fp16_0))[name = tensor("aw_545_cast_fp16")]; + tensor aw_547_equation_0 = const()[name = tensor("aw_547_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_547_cast_fp16 = einsum(equation = aw_547_equation_0, values = (var_4207_cast_fp16_1, var_4189_cast_fp16_1))[name = tensor("aw_547_cast_fp16")]; + tensor aw_549_equation_0 = const()[name = tensor("aw_549_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_549_cast_fp16 = einsum(equation = aw_549_equation_0, values = (var_4207_cast_fp16_2, var_4189_cast_fp16_2))[name = tensor("aw_549_cast_fp16")]; + tensor aw_551_equation_0 = const()[name = tensor("aw_551_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_551_cast_fp16 = einsum(equation = aw_551_equation_0, values = (var_4207_cast_fp16_3, var_4189_cast_fp16_3))[name = tensor("aw_551_cast_fp16")]; + tensor aw_553_equation_0 = const()[name = tensor("aw_553_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_553_cast_fp16 = einsum(equation = aw_553_equation_0, values = (var_4207_cast_fp16_4, var_4189_cast_fp16_4))[name = tensor("aw_553_cast_fp16")]; + tensor aw_555_equation_0 = const()[name = tensor("aw_555_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_555_cast_fp16 = einsum(equation = aw_555_equation_0, values = (var_4207_cast_fp16_5, var_4189_cast_fp16_5))[name = tensor("aw_555_cast_fp16")]; + tensor aw_557_equation_0 = const()[name = tensor("aw_557_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_557_cast_fp16 = einsum(equation = aw_557_equation_0, values = (var_4207_cast_fp16_6, var_4189_cast_fp16_6))[name = tensor("aw_557_cast_fp16")]; + tensor aw_559_equation_0 = const()[name = tensor("aw_559_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_559_cast_fp16 = einsum(equation = aw_559_equation_0, values = (var_4207_cast_fp16_7, var_4189_cast_fp16_7))[name = tensor("aw_559_cast_fp16")]; + tensor aw_561_equation_0 = const()[name = tensor("aw_561_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_561_cast_fp16 = einsum(equation = aw_561_equation_0, values = (var_4207_cast_fp16_8, var_4189_cast_fp16_8))[name = tensor("aw_561_cast_fp16")]; + tensor aw_563_equation_0 = const()[name = tensor("aw_563_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_563_cast_fp16 = einsum(equation = aw_563_equation_0, values = (var_4207_cast_fp16_9, var_4189_cast_fp16_9))[name = tensor("aw_563_cast_fp16")]; + tensor aw_565_equation_0 = const()[name = tensor("aw_565_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_565_cast_fp16 = einsum(equation = aw_565_equation_0, values = (var_4207_cast_fp16_10, var_4189_cast_fp16_10))[name = tensor("aw_565_cast_fp16")]; + tensor aw_567_equation_0 = const()[name = tensor("aw_567_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_567_cast_fp16 = einsum(equation = aw_567_equation_0, values = (var_4207_cast_fp16_11, var_4189_cast_fp16_11))[name = tensor("aw_567_cast_fp16")]; + tensor aw_569_equation_0 = const()[name = tensor("aw_569_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_569_cast_fp16 = einsum(equation = aw_569_equation_0, values = (var_4207_cast_fp16_12, var_4189_cast_fp16_12))[name = tensor("aw_569_cast_fp16")]; + tensor aw_571_equation_0 = const()[name = tensor("aw_571_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_571_cast_fp16 = einsum(equation = aw_571_equation_0, values = (var_4207_cast_fp16_13, var_4189_cast_fp16_13))[name = tensor("aw_571_cast_fp16")]; + tensor aw_573_equation_0 = const()[name = tensor("aw_573_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_573_cast_fp16 = einsum(equation = aw_573_equation_0, values = (var_4207_cast_fp16_14, var_4189_cast_fp16_14))[name = tensor("aw_573_cast_fp16")]; + tensor aw_575_equation_0 = const()[name = tensor("aw_575_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_575_cast_fp16 = einsum(equation = aw_575_equation_0, values = (var_4207_cast_fp16_15, var_4189_cast_fp16_15))[name = tensor("aw_575_cast_fp16")]; + tensor var_4273_cast_fp16 = softmax(axis = var_4137, x = aw_545_cast_fp16)[name = tensor("op_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = softmax(axis = var_4137, x = aw_547_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_cast_fp16 = softmax(axis = var_4137, x = aw_549_cast_fp16)[name = tensor("op_4275_cast_fp16")]; + tensor var_4276_cast_fp16 = softmax(axis = var_4137, x = aw_551_cast_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor var_4277_cast_fp16 = softmax(axis = var_4137, x = aw_553_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4278_cast_fp16 = softmax(axis = var_4137, x = aw_555_cast_fp16)[name = tensor("op_4278_cast_fp16")]; + tensor var_4279_cast_fp16 = softmax(axis = var_4137, x = aw_557_cast_fp16)[name = tensor("op_4279_cast_fp16")]; + tensor var_4280_cast_fp16 = softmax(axis = var_4137, x = aw_559_cast_fp16)[name = tensor("op_4280_cast_fp16")]; + tensor var_4281_cast_fp16 = softmax(axis = var_4137, x = aw_561_cast_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4282_cast_fp16 = softmax(axis = var_4137, x = aw_563_cast_fp16)[name = tensor("op_4282_cast_fp16")]; + tensor var_4283_cast_fp16 = softmax(axis = var_4137, x = aw_565_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = softmax(axis = var_4137, x = aw_567_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor var_4285_cast_fp16 = softmax(axis = var_4137, x = aw_569_cast_fp16)[name = tensor("op_4285_cast_fp16")]; + tensor var_4286_cast_fp16 = softmax(axis = var_4137, x = aw_571_cast_fp16)[name = tensor("op_4286_cast_fp16")]; + tensor var_4287_cast_fp16 = softmax(axis = var_4137, x = aw_573_cast_fp16)[name = tensor("op_4287_cast_fp16")]; + tensor var_4288_cast_fp16 = softmax(axis = var_4137, x = aw_575_cast_fp16)[name = tensor("op_4288_cast_fp16")]; + tensor var_4290_equation_0 = const()[name = tensor("op_4290_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4290_cast_fp16 = einsum(equation = var_4290_equation_0, values = (var_4224_cast_fp16_0, var_4273_cast_fp16))[name = tensor("op_4290_cast_fp16")]; + tensor var_4292_equation_0 = const()[name = tensor("op_4292_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4292_cast_fp16 = einsum(equation = var_4292_equation_0, values = (var_4224_cast_fp16_1, var_4274_cast_fp16))[name = tensor("op_4292_cast_fp16")]; + tensor var_4294_equation_0 = const()[name = tensor("op_4294_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4294_cast_fp16 = einsum(equation = var_4294_equation_0, values = (var_4224_cast_fp16_2, var_4275_cast_fp16))[name = tensor("op_4294_cast_fp16")]; + tensor var_4296_equation_0 = const()[name = tensor("op_4296_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4296_cast_fp16 = einsum(equation = var_4296_equation_0, values = (var_4224_cast_fp16_3, var_4276_cast_fp16))[name = tensor("op_4296_cast_fp16")]; + tensor var_4298_equation_0 = const()[name = tensor("op_4298_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4298_cast_fp16 = einsum(equation = var_4298_equation_0, values = (var_4224_cast_fp16_4, var_4277_cast_fp16))[name = tensor("op_4298_cast_fp16")]; + tensor var_4300_equation_0 = const()[name = tensor("op_4300_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4300_cast_fp16 = einsum(equation = var_4300_equation_0, values = (var_4224_cast_fp16_5, var_4278_cast_fp16))[name = tensor("op_4300_cast_fp16")]; + tensor var_4302_equation_0 = const()[name = tensor("op_4302_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4302_cast_fp16 = einsum(equation = var_4302_equation_0, values = (var_4224_cast_fp16_6, var_4279_cast_fp16))[name = tensor("op_4302_cast_fp16")]; + tensor var_4304_equation_0 = const()[name = tensor("op_4304_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4304_cast_fp16 = einsum(equation = var_4304_equation_0, values = (var_4224_cast_fp16_7, var_4280_cast_fp16))[name = tensor("op_4304_cast_fp16")]; + tensor var_4306_equation_0 = const()[name = tensor("op_4306_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4306_cast_fp16 = einsum(equation = var_4306_equation_0, values = (var_4224_cast_fp16_8, var_4281_cast_fp16))[name = tensor("op_4306_cast_fp16")]; + tensor var_4308_equation_0 = const()[name = tensor("op_4308_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4308_cast_fp16 = einsum(equation = var_4308_equation_0, values = (var_4224_cast_fp16_9, var_4282_cast_fp16))[name = tensor("op_4308_cast_fp16")]; + tensor var_4310_equation_0 = const()[name = tensor("op_4310_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4310_cast_fp16 = einsum(equation = var_4310_equation_0, values = (var_4224_cast_fp16_10, var_4283_cast_fp16))[name = tensor("op_4310_cast_fp16")]; + tensor var_4312_equation_0 = const()[name = tensor("op_4312_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4312_cast_fp16 = einsum(equation = var_4312_equation_0, values = (var_4224_cast_fp16_11, var_4284_cast_fp16))[name = tensor("op_4312_cast_fp16")]; + tensor var_4314_equation_0 = const()[name = tensor("op_4314_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4314_cast_fp16 = einsum(equation = var_4314_equation_0, values = (var_4224_cast_fp16_12, var_4285_cast_fp16))[name = tensor("op_4314_cast_fp16")]; + tensor var_4316_equation_0 = const()[name = tensor("op_4316_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4316_cast_fp16 = einsum(equation = var_4316_equation_0, values = (var_4224_cast_fp16_13, var_4286_cast_fp16))[name = tensor("op_4316_cast_fp16")]; + tensor var_4318_equation_0 = const()[name = tensor("op_4318_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4318_cast_fp16 = einsum(equation = var_4318_equation_0, values = (var_4224_cast_fp16_14, var_4287_cast_fp16))[name = tensor("op_4318_cast_fp16")]; + tensor var_4320_equation_0 = const()[name = tensor("op_4320_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4320_cast_fp16 = einsum(equation = var_4320_equation_0, values = (var_4224_cast_fp16_15, var_4288_cast_fp16))[name = tensor("op_4320_cast_fp16")]; + tensor input_175_interleave_0 = const()[name = tensor("input_175_interleave_0"), val = tensor(false)]; + tensor input_175_cast_fp16 = concat(axis = var_4137, interleave = input_175_interleave_0, values = (var_4290_cast_fp16, var_4292_cast_fp16, var_4294_cast_fp16, var_4296_cast_fp16, var_4298_cast_fp16, var_4300_cast_fp16, var_4302_cast_fp16, var_4304_cast_fp16, var_4306_cast_fp16, var_4308_cast_fp16, var_4310_cast_fp16, var_4312_cast_fp16, var_4314_cast_fp16, var_4316_cast_fp16, var_4318_cast_fp16, var_4320_cast_fp16))[name = tensor("input_175_cast_fp16")]; + tensor var_4329_pad_type_0 = const()[name = tensor("op_4329_pad_type_0"), val = tensor("valid")]; + tensor var_4329_strides_0 = const()[name = tensor("op_4329_strides_0"), val = tensor([1, 1])]; + tensor var_4329_pad_0 = const()[name = tensor("op_4329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4329_dilations_0 = const()[name = tensor("op_4329_dilations_0"), val = tensor([1, 1])]; + tensor var_4329_groups_0 = const()[name = tensor("op_4329_groups_0"), val = tensor(1)]; + tensor blocks_17_attn_out_weight_to_fp16 = const()[name = tensor("blocks_17_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444412672)))]; + tensor blocks_17_attn_out_bias_to_fp16 = const()[name = tensor("blocks_17_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446509888)))]; + tensor var_4329_cast_fp16 = conv(bias = blocks_17_attn_out_bias_to_fp16, dilations = var_4329_dilations_0, groups = var_4329_groups_0, pad = var_4329_pad_0, pad_type = var_4329_pad_type_0, strides = var_4329_strides_0, weight = blocks_17_attn_out_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = var_4329_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([1])]; + tensor input_177_gamma_0_to_fp16 = const()[name = tensor("input_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446512000)))]; + tensor input_177_beta_0_to_fp16 = const()[name = tensor("input_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446514112)))]; + tensor var_4339_to_fp16 = const()[name = tensor("op_4339_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = input_177_beta_0_to_fp16, epsilon = var_4339_to_fp16, gamma = input_177_gamma_0_to_fp16, x = inputs_71_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("valid")]; + tensor input_179_strides_0 = const()[name = tensor("input_179_strides_0"), val = tensor([1, 1])]; + tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_179_dilations_0 = const()[name = tensor("input_179_dilations_0"), val = tensor([1, 1])]; + tensor input_179_groups_0 = const()[name = tensor("input_179_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446516224)))]; + tensor blocks_17_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454904896)))]; + tensor input_179_cast_fp16 = conv(bias = blocks_17_mlp_0_bias_to_fp16, dilations = input_179_dilations_0, groups = input_179_groups_0, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = input_179_strides_0, weight = blocks_17_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_4365_pad_type_0 = const()[name = tensor("op_4365_pad_type_0"), val = tensor("valid")]; + tensor var_4365_strides_0 = const()[name = tensor("op_4365_strides_0"), val = tensor([1, 1])]; + tensor var_4365_pad_0 = const()[name = tensor("op_4365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4365_dilations_0 = const()[name = tensor("op_4365_dilations_0"), val = tensor([1, 1])]; + tensor var_4365_groups_0 = const()[name = tensor("op_4365_groups_0"), val = tensor(1)]; + tensor blocks_17_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454913152)))]; + tensor blocks_17_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463301824)))]; + tensor var_4365_cast_fp16 = conv(bias = blocks_17_mlp_2_bias_to_fp16, dilations = var_4365_dilations_0, groups = var_4365_groups_0, pad = var_4365_pad_0, pad_type = var_4365_pad_type_0, strides = var_4365_strides_0, weight = blocks_17_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("op_4365_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_4365_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor(1)]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([1])]; + tensor input_183_gamma_0_to_fp16 = const()[name = tensor("input_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463303936)))]; + tensor input_183_beta_0_to_fp16 = const()[name = tensor("input_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463306048)))]; + tensor var_4390_to_fp16 = const()[name = tensor("op_4390_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = input_183_beta_0_to_fp16, epsilon = var_4390_to_fp16, gamma = input_183_gamma_0_to_fp16, x = inputs_73_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor q_37_pad_type_0 = const()[name = tensor("q_37_pad_type_0"), val = tensor("valid")]; + tensor q_37_strides_0 = const()[name = tensor("q_37_strides_0"), val = tensor([1, 1])]; + tensor q_37_pad_0 = const()[name = tensor("q_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_37_dilations_0 = const()[name = tensor("q_37_dilations_0"), val = tensor([1, 1])]; + tensor q_37_groups_0 = const()[name = tensor("q_37_groups_0"), val = tensor(1)]; + tensor var_4425_weight_0_to_fp16 = const()[name = tensor("op_4425_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463308160)))]; + tensor var_4425_bias_0_to_fp16 = const()[name = tensor("op_4425_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465405376)))]; + tensor var_4425_cast_fp16 = conv(bias = var_4425_bias_0_to_fp16, dilations = q_37_dilations_0, groups = q_37_groups_0, pad = q_37_pad_0, pad_type = q_37_pad_type_0, strides = q_37_strides_0, weight = var_4425_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("op_4425_cast_fp16")]; + tensor k_37_pad_type_0 = const()[name = tensor("k_37_pad_type_0"), val = tensor("valid")]; + tensor k_37_strides_0 = const()[name = tensor("k_37_strides_0"), val = tensor([1, 1])]; + tensor k_37_pad_0 = const()[name = tensor("k_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_37_dilations_0 = const()[name = tensor("k_37_dilations_0"), val = tensor([1, 1])]; + tensor k_37_groups_0 = const()[name = tensor("k_37_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_key_weight_to_fp16 = const()[name = tensor("blocks_18_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465407488)))]; + tensor k_37_cast_fp16 = conv(dilations = k_37_dilations_0, groups = k_37_groups_0, pad = k_37_pad_0, pad_type = k_37_pad_type_0, strides = k_37_strides_0, weight = blocks_18_attn_key_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor var_4423_pad_type_0 = const()[name = tensor("op_4423_pad_type_0"), val = tensor("valid")]; + tensor var_4423_strides_0 = const()[name = tensor("op_4423_strides_0"), val = tensor([1, 1])]; + tensor var_4423_pad_0 = const()[name = tensor("op_4423_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4423_dilations_0 = const()[name = tensor("op_4423_dilations_0"), val = tensor([1, 1])]; + tensor var_4423_groups_0 = const()[name = tensor("op_4423_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_value_weight_to_fp16 = const()[name = tensor("blocks_18_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467504704)))]; + tensor blocks_18_attn_value_bias_to_fp16 = const()[name = tensor("blocks_18_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469601920)))]; + tensor var_4423_cast_fp16 = conv(bias = blocks_18_attn_value_bias_to_fp16, dilations = var_4423_dilations_0, groups = var_4423_groups_0, pad = var_4423_pad_0, pad_type = var_4423_pad_type_0, strides = var_4423_strides_0, weight = blocks_18_attn_value_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("op_4423_cast_fp16")]; + tensor tile_54 = const()[name = tensor("tile_54"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4426_axis_0 = const()[name = tensor("op_4426_axis_0"), val = tensor(1)]; + tensor var_4426_cast_fp16_0, tensor var_4426_cast_fp16_1, tensor var_4426_cast_fp16_2, tensor var_4426_cast_fp16_3, tensor var_4426_cast_fp16_4, tensor var_4426_cast_fp16_5, tensor var_4426_cast_fp16_6, tensor var_4426_cast_fp16_7, tensor var_4426_cast_fp16_8, tensor var_4426_cast_fp16_9, tensor var_4426_cast_fp16_10, tensor var_4426_cast_fp16_11, tensor var_4426_cast_fp16_12, tensor var_4426_cast_fp16_13, tensor var_4426_cast_fp16_14, tensor var_4426_cast_fp16_15 = split(axis = var_4426_axis_0, split_sizes = tile_54, x = var_4425_cast_fp16)[name = tensor("op_4426_cast_fp16")]; + tensor var_4443_perm_0 = const()[name = tensor("op_4443_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_55 = const()[name = tensor("tile_55"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4444_axis_0 = const()[name = tensor("op_4444_axis_0"), val = tensor(3)]; + tensor var_4443_cast_fp16 = transpose(perm = var_4443_perm_0, x = k_37_cast_fp16)[name = tensor("transpose_6")]; + tensor var_4444_cast_fp16_0, tensor var_4444_cast_fp16_1, tensor var_4444_cast_fp16_2, tensor var_4444_cast_fp16_3, tensor var_4444_cast_fp16_4, tensor var_4444_cast_fp16_5, tensor var_4444_cast_fp16_6, tensor var_4444_cast_fp16_7, tensor var_4444_cast_fp16_8, tensor var_4444_cast_fp16_9, tensor var_4444_cast_fp16_10, tensor var_4444_cast_fp16_11, tensor var_4444_cast_fp16_12, tensor var_4444_cast_fp16_13, tensor var_4444_cast_fp16_14, tensor var_4444_cast_fp16_15 = split(axis = var_4444_axis_0, split_sizes = tile_55, x = var_4443_cast_fp16)[name = tensor("op_4444_cast_fp16")]; + tensor tile_56 = const()[name = tensor("tile_56"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4461_axis_0 = const()[name = tensor("op_4461_axis_0"), val = tensor(1)]; + tensor var_4461_cast_fp16_0, tensor var_4461_cast_fp16_1, tensor var_4461_cast_fp16_2, tensor var_4461_cast_fp16_3, tensor var_4461_cast_fp16_4, tensor var_4461_cast_fp16_5, tensor var_4461_cast_fp16_6, tensor var_4461_cast_fp16_7, tensor var_4461_cast_fp16_8, tensor var_4461_cast_fp16_9, tensor var_4461_cast_fp16_10, tensor var_4461_cast_fp16_11, tensor var_4461_cast_fp16_12, tensor var_4461_cast_fp16_13, tensor var_4461_cast_fp16_14, tensor var_4461_cast_fp16_15 = split(axis = var_4461_axis_0, split_sizes = tile_56, x = var_4423_cast_fp16)[name = tensor("op_4461_cast_fp16")]; + tensor aw_577_equation_0 = const()[name = tensor("aw_577_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_577_cast_fp16 = einsum(equation = aw_577_equation_0, values = (var_4444_cast_fp16_0, var_4426_cast_fp16_0))[name = tensor("aw_577_cast_fp16")]; + tensor aw_579_equation_0 = const()[name = tensor("aw_579_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_579_cast_fp16 = einsum(equation = aw_579_equation_0, values = (var_4444_cast_fp16_1, var_4426_cast_fp16_1))[name = tensor("aw_579_cast_fp16")]; + tensor aw_581_equation_0 = const()[name = tensor("aw_581_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_581_cast_fp16 = einsum(equation = aw_581_equation_0, values = (var_4444_cast_fp16_2, var_4426_cast_fp16_2))[name = tensor("aw_581_cast_fp16")]; + tensor aw_583_equation_0 = const()[name = tensor("aw_583_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_583_cast_fp16 = einsum(equation = aw_583_equation_0, values = (var_4444_cast_fp16_3, var_4426_cast_fp16_3))[name = tensor("aw_583_cast_fp16")]; + tensor aw_585_equation_0 = const()[name = tensor("aw_585_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_585_cast_fp16 = einsum(equation = aw_585_equation_0, values = (var_4444_cast_fp16_4, var_4426_cast_fp16_4))[name = tensor("aw_585_cast_fp16")]; + tensor aw_587_equation_0 = const()[name = tensor("aw_587_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_587_cast_fp16 = einsum(equation = aw_587_equation_0, values = (var_4444_cast_fp16_5, var_4426_cast_fp16_5))[name = tensor("aw_587_cast_fp16")]; + tensor aw_589_equation_0 = const()[name = tensor("aw_589_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_589_cast_fp16 = einsum(equation = aw_589_equation_0, values = (var_4444_cast_fp16_6, var_4426_cast_fp16_6))[name = tensor("aw_589_cast_fp16")]; + tensor aw_591_equation_0 = const()[name = tensor("aw_591_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_591_cast_fp16 = einsum(equation = aw_591_equation_0, values = (var_4444_cast_fp16_7, var_4426_cast_fp16_7))[name = tensor("aw_591_cast_fp16")]; + tensor aw_593_equation_0 = const()[name = tensor("aw_593_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_593_cast_fp16 = einsum(equation = aw_593_equation_0, values = (var_4444_cast_fp16_8, var_4426_cast_fp16_8))[name = tensor("aw_593_cast_fp16")]; + tensor aw_595_equation_0 = const()[name = tensor("aw_595_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_595_cast_fp16 = einsum(equation = aw_595_equation_0, values = (var_4444_cast_fp16_9, var_4426_cast_fp16_9))[name = tensor("aw_595_cast_fp16")]; + tensor aw_597_equation_0 = const()[name = tensor("aw_597_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_597_cast_fp16 = einsum(equation = aw_597_equation_0, values = (var_4444_cast_fp16_10, var_4426_cast_fp16_10))[name = tensor("aw_597_cast_fp16")]; + tensor aw_599_equation_0 = const()[name = tensor("aw_599_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_599_cast_fp16 = einsum(equation = aw_599_equation_0, values = (var_4444_cast_fp16_11, var_4426_cast_fp16_11))[name = tensor("aw_599_cast_fp16")]; + tensor aw_601_equation_0 = const()[name = tensor("aw_601_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_601_cast_fp16 = einsum(equation = aw_601_equation_0, values = (var_4444_cast_fp16_12, var_4426_cast_fp16_12))[name = tensor("aw_601_cast_fp16")]; + tensor aw_603_equation_0 = const()[name = tensor("aw_603_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_603_cast_fp16 = einsum(equation = aw_603_equation_0, values = (var_4444_cast_fp16_13, var_4426_cast_fp16_13))[name = tensor("aw_603_cast_fp16")]; + tensor aw_605_equation_0 = const()[name = tensor("aw_605_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_605_cast_fp16 = einsum(equation = aw_605_equation_0, values = (var_4444_cast_fp16_14, var_4426_cast_fp16_14))[name = tensor("aw_605_cast_fp16")]; + tensor aw_607_equation_0 = const()[name = tensor("aw_607_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_607_cast_fp16 = einsum(equation = aw_607_equation_0, values = (var_4444_cast_fp16_15, var_4426_cast_fp16_15))[name = tensor("aw_607_cast_fp16")]; + tensor var_4510_cast_fp16 = softmax(axis = var_4374, x = aw_577_cast_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor var_4511_cast_fp16 = softmax(axis = var_4374, x = aw_579_cast_fp16)[name = tensor("op_4511_cast_fp16")]; + tensor var_4512_cast_fp16 = softmax(axis = var_4374, x = aw_581_cast_fp16)[name = tensor("op_4512_cast_fp16")]; + tensor var_4513_cast_fp16 = softmax(axis = var_4374, x = aw_583_cast_fp16)[name = tensor("op_4513_cast_fp16")]; + tensor var_4514_cast_fp16 = softmax(axis = var_4374, x = aw_585_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = softmax(axis = var_4374, x = aw_587_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor var_4516_cast_fp16 = softmax(axis = var_4374, x = aw_589_cast_fp16)[name = tensor("op_4516_cast_fp16")]; + tensor var_4517_cast_fp16 = softmax(axis = var_4374, x = aw_591_cast_fp16)[name = tensor("op_4517_cast_fp16")]; + tensor var_4518_cast_fp16 = softmax(axis = var_4374, x = aw_593_cast_fp16)[name = tensor("op_4518_cast_fp16")]; + tensor var_4519_cast_fp16 = softmax(axis = var_4374, x = aw_595_cast_fp16)[name = tensor("op_4519_cast_fp16")]; + tensor var_4520_cast_fp16 = softmax(axis = var_4374, x = aw_597_cast_fp16)[name = tensor("op_4520_cast_fp16")]; + tensor var_4521_cast_fp16 = softmax(axis = var_4374, x = aw_599_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4522_cast_fp16 = softmax(axis = var_4374, x = aw_601_cast_fp16)[name = tensor("op_4522_cast_fp16")]; + tensor var_4523_cast_fp16 = softmax(axis = var_4374, x = aw_603_cast_fp16)[name = tensor("op_4523_cast_fp16")]; + tensor var_4524_cast_fp16 = softmax(axis = var_4374, x = aw_605_cast_fp16)[name = tensor("op_4524_cast_fp16")]; + tensor var_4525_cast_fp16 = softmax(axis = var_4374, x = aw_607_cast_fp16)[name = tensor("op_4525_cast_fp16")]; + tensor var_4527_equation_0 = const()[name = tensor("op_4527_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4527_cast_fp16 = einsum(equation = var_4527_equation_0, values = (var_4461_cast_fp16_0, var_4510_cast_fp16))[name = tensor("op_4527_cast_fp16")]; + tensor var_4529_equation_0 = const()[name = tensor("op_4529_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4529_cast_fp16 = einsum(equation = var_4529_equation_0, values = (var_4461_cast_fp16_1, var_4511_cast_fp16))[name = tensor("op_4529_cast_fp16")]; + tensor var_4531_equation_0 = const()[name = tensor("op_4531_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4531_cast_fp16 = einsum(equation = var_4531_equation_0, values = (var_4461_cast_fp16_2, var_4512_cast_fp16))[name = tensor("op_4531_cast_fp16")]; + tensor var_4533_equation_0 = const()[name = tensor("op_4533_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4533_cast_fp16 = einsum(equation = var_4533_equation_0, values = (var_4461_cast_fp16_3, var_4513_cast_fp16))[name = tensor("op_4533_cast_fp16")]; + tensor var_4535_equation_0 = const()[name = tensor("op_4535_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4535_cast_fp16 = einsum(equation = var_4535_equation_0, values = (var_4461_cast_fp16_4, var_4514_cast_fp16))[name = tensor("op_4535_cast_fp16")]; + tensor var_4537_equation_0 = const()[name = tensor("op_4537_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4537_cast_fp16 = einsum(equation = var_4537_equation_0, values = (var_4461_cast_fp16_5, var_4515_cast_fp16))[name = tensor("op_4537_cast_fp16")]; + tensor var_4539_equation_0 = const()[name = tensor("op_4539_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4539_cast_fp16 = einsum(equation = var_4539_equation_0, values = (var_4461_cast_fp16_6, var_4516_cast_fp16))[name = tensor("op_4539_cast_fp16")]; + tensor var_4541_equation_0 = const()[name = tensor("op_4541_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4541_cast_fp16 = einsum(equation = var_4541_equation_0, values = (var_4461_cast_fp16_7, var_4517_cast_fp16))[name = tensor("op_4541_cast_fp16")]; + tensor var_4543_equation_0 = const()[name = tensor("op_4543_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4543_cast_fp16 = einsum(equation = var_4543_equation_0, values = (var_4461_cast_fp16_8, var_4518_cast_fp16))[name = tensor("op_4543_cast_fp16")]; + tensor var_4545_equation_0 = const()[name = tensor("op_4545_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4545_cast_fp16 = einsum(equation = var_4545_equation_0, values = (var_4461_cast_fp16_9, var_4519_cast_fp16))[name = tensor("op_4545_cast_fp16")]; + tensor var_4547_equation_0 = const()[name = tensor("op_4547_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4547_cast_fp16 = einsum(equation = var_4547_equation_0, values = (var_4461_cast_fp16_10, var_4520_cast_fp16))[name = tensor("op_4547_cast_fp16")]; + tensor var_4549_equation_0 = const()[name = tensor("op_4549_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4549_cast_fp16 = einsum(equation = var_4549_equation_0, values = (var_4461_cast_fp16_11, var_4521_cast_fp16))[name = tensor("op_4549_cast_fp16")]; + tensor var_4551_equation_0 = const()[name = tensor("op_4551_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4551_cast_fp16 = einsum(equation = var_4551_equation_0, values = (var_4461_cast_fp16_12, var_4522_cast_fp16))[name = tensor("op_4551_cast_fp16")]; + tensor var_4553_equation_0 = const()[name = tensor("op_4553_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4553_cast_fp16 = einsum(equation = var_4553_equation_0, values = (var_4461_cast_fp16_13, var_4523_cast_fp16))[name = tensor("op_4553_cast_fp16")]; + tensor var_4555_equation_0 = const()[name = tensor("op_4555_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4555_cast_fp16 = einsum(equation = var_4555_equation_0, values = (var_4461_cast_fp16_14, var_4524_cast_fp16))[name = tensor("op_4555_cast_fp16")]; + tensor var_4557_equation_0 = const()[name = tensor("op_4557_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4557_cast_fp16 = einsum(equation = var_4557_equation_0, values = (var_4461_cast_fp16_15, var_4525_cast_fp16))[name = tensor("op_4557_cast_fp16")]; + tensor input_185_interleave_0 = const()[name = tensor("input_185_interleave_0"), val = tensor(false)]; + tensor input_185_cast_fp16 = concat(axis = var_4374, interleave = input_185_interleave_0, values = (var_4527_cast_fp16, var_4529_cast_fp16, var_4531_cast_fp16, var_4533_cast_fp16, var_4535_cast_fp16, var_4537_cast_fp16, var_4539_cast_fp16, var_4541_cast_fp16, var_4543_cast_fp16, var_4545_cast_fp16, var_4547_cast_fp16, var_4549_cast_fp16, var_4551_cast_fp16, var_4553_cast_fp16, var_4555_cast_fp16, var_4557_cast_fp16))[name = tensor("input_185_cast_fp16")]; + tensor var_4566_pad_type_0 = const()[name = tensor("op_4566_pad_type_0"), val = tensor("valid")]; + tensor var_4566_strides_0 = const()[name = tensor("op_4566_strides_0"), val = tensor([1, 1])]; + tensor var_4566_pad_0 = const()[name = tensor("op_4566_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4566_dilations_0 = const()[name = tensor("op_4566_dilations_0"), val = tensor([1, 1])]; + tensor var_4566_groups_0 = const()[name = tensor("op_4566_groups_0"), val = tensor(1)]; + tensor blocks_18_attn_out_weight_to_fp16 = const()[name = tensor("blocks_18_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469604032)))]; + tensor blocks_18_attn_out_bias_to_fp16 = const()[name = tensor("blocks_18_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471701248)))]; + tensor var_4566_cast_fp16 = conv(bias = blocks_18_attn_out_bias_to_fp16, dilations = var_4566_dilations_0, groups = var_4566_groups_0, pad = var_4566_pad_0, pad_type = var_4566_pad_type_0, strides = var_4566_strides_0, weight = blocks_18_attn_out_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("op_4566_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = var_4566_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor input_187_axes_0 = const()[name = tensor("input_187_axes_0"), val = tensor([1])]; + tensor input_187_gamma_0_to_fp16 = const()[name = tensor("input_187_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471703360)))]; + tensor input_187_beta_0_to_fp16 = const()[name = tensor("input_187_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471705472)))]; + tensor var_4576_to_fp16 = const()[name = tensor("op_4576_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = input_187_beta_0_to_fp16, epsilon = var_4576_to_fp16, gamma = input_187_gamma_0_to_fp16, x = inputs_75_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("valid")]; + tensor input_189_strides_0 = const()[name = tensor("input_189_strides_0"), val = tensor([1, 1])]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_189_dilations_0 = const()[name = tensor("input_189_dilations_0"), val = tensor([1, 1])]; + tensor input_189_groups_0 = const()[name = tensor("input_189_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471707584)))]; + tensor blocks_18_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480096256)))]; + tensor input_189_cast_fp16 = conv(bias = blocks_18_mlp_0_bias_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = blocks_18_mlp_0_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_4602_pad_type_0 = const()[name = tensor("op_4602_pad_type_0"), val = tensor("valid")]; + tensor var_4602_strides_0 = const()[name = tensor("op_4602_strides_0"), val = tensor([1, 1])]; + tensor var_4602_pad_0 = const()[name = tensor("op_4602_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4602_dilations_0 = const()[name = tensor("op_4602_dilations_0"), val = tensor([1, 1])]; + tensor var_4602_groups_0 = const()[name = tensor("op_4602_groups_0"), val = tensor(1)]; + tensor blocks_18_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480104512)))]; + tensor blocks_18_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488493184)))]; + tensor var_4602_cast_fp16 = conv(bias = blocks_18_mlp_2_bias_to_fp16, dilations = var_4602_dilations_0, groups = var_4602_groups_0, pad = var_4602_pad_0, pad_type = var_4602_pad_type_0, strides = var_4602_strides_0, weight = blocks_18_mlp_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("op_4602_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = var_4602_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_4611 = const()[name = tensor("op_4611"), val = tensor(1)]; + tensor input_193_axes_0 = const()[name = tensor("input_193_axes_0"), val = tensor([1])]; + tensor input_193_gamma_0_to_fp16 = const()[name = tensor("input_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488495296)))]; + tensor input_193_beta_0_to_fp16 = const()[name = tensor("input_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488497408)))]; + tensor var_4627_to_fp16 = const()[name = tensor("op_4627_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = input_193_beta_0_to_fp16, epsilon = var_4627_to_fp16, gamma = input_193_gamma_0_to_fp16, x = inputs_77_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor q_39_pad_type_0 = const()[name = tensor("q_39_pad_type_0"), val = tensor("valid")]; + tensor q_39_strides_0 = const()[name = tensor("q_39_strides_0"), val = tensor([1, 1])]; + tensor q_39_pad_0 = const()[name = tensor("q_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_39_dilations_0 = const()[name = tensor("q_39_dilations_0"), val = tensor([1, 1])]; + tensor q_39_groups_0 = const()[name = tensor("q_39_groups_0"), val = tensor(1)]; + tensor var_4662_weight_0_to_fp16 = const()[name = tensor("op_4662_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488499520)))]; + tensor var_4662_bias_0_to_fp16 = const()[name = tensor("op_4662_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490596736)))]; + tensor var_4662_cast_fp16 = conv(bias = var_4662_bias_0_to_fp16, dilations = q_39_dilations_0, groups = q_39_groups_0, pad = q_39_pad_0, pad_type = q_39_pad_type_0, strides = q_39_strides_0, weight = var_4662_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor k_39_pad_type_0 = const()[name = tensor("k_39_pad_type_0"), val = tensor("valid")]; + tensor k_39_strides_0 = const()[name = tensor("k_39_strides_0"), val = tensor([1, 1])]; + tensor k_39_pad_0 = const()[name = tensor("k_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_39_dilations_0 = const()[name = tensor("k_39_dilations_0"), val = tensor([1, 1])]; + tensor k_39_groups_0 = const()[name = tensor("k_39_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_key_weight_to_fp16 = const()[name = tensor("blocks_19_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490598848)))]; + tensor k_39_cast_fp16 = conv(dilations = k_39_dilations_0, groups = k_39_groups_0, pad = k_39_pad_0, pad_type = k_39_pad_type_0, strides = k_39_strides_0, weight = blocks_19_attn_key_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor var_4660_pad_type_0 = const()[name = tensor("op_4660_pad_type_0"), val = tensor("valid")]; + tensor var_4660_strides_0 = const()[name = tensor("op_4660_strides_0"), val = tensor([1, 1])]; + tensor var_4660_pad_0 = const()[name = tensor("op_4660_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4660_dilations_0 = const()[name = tensor("op_4660_dilations_0"), val = tensor([1, 1])]; + tensor var_4660_groups_0 = const()[name = tensor("op_4660_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_value_weight_to_fp16 = const()[name = tensor("blocks_19_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492696064)))]; + tensor blocks_19_attn_value_bias_to_fp16 = const()[name = tensor("blocks_19_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494793280)))]; + tensor var_4660_cast_fp16 = conv(bias = blocks_19_attn_value_bias_to_fp16, dilations = var_4660_dilations_0, groups = var_4660_groups_0, pad = var_4660_pad_0, pad_type = var_4660_pad_type_0, strides = var_4660_strides_0, weight = blocks_19_attn_value_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor tile_57 = const()[name = tensor("tile_57"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4663_axis_0 = const()[name = tensor("op_4663_axis_0"), val = tensor(1)]; + tensor var_4663_cast_fp16_0, tensor var_4663_cast_fp16_1, tensor var_4663_cast_fp16_2, tensor var_4663_cast_fp16_3, tensor var_4663_cast_fp16_4, tensor var_4663_cast_fp16_5, tensor var_4663_cast_fp16_6, tensor var_4663_cast_fp16_7, tensor var_4663_cast_fp16_8, tensor var_4663_cast_fp16_9, tensor var_4663_cast_fp16_10, tensor var_4663_cast_fp16_11, tensor var_4663_cast_fp16_12, tensor var_4663_cast_fp16_13, tensor var_4663_cast_fp16_14, tensor var_4663_cast_fp16_15 = split(axis = var_4663_axis_0, split_sizes = tile_57, x = var_4662_cast_fp16)[name = tensor("op_4663_cast_fp16")]; + tensor var_4680_perm_0 = const()[name = tensor("op_4680_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_58 = const()[name = tensor("tile_58"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(3)]; + tensor var_4680_cast_fp16 = transpose(perm = var_4680_perm_0, x = k_39_cast_fp16)[name = tensor("transpose_5")]; + tensor var_4681_cast_fp16_0, tensor var_4681_cast_fp16_1, tensor var_4681_cast_fp16_2, tensor var_4681_cast_fp16_3, tensor var_4681_cast_fp16_4, tensor var_4681_cast_fp16_5, tensor var_4681_cast_fp16_6, tensor var_4681_cast_fp16_7, tensor var_4681_cast_fp16_8, tensor var_4681_cast_fp16_9, tensor var_4681_cast_fp16_10, tensor var_4681_cast_fp16_11, tensor var_4681_cast_fp16_12, tensor var_4681_cast_fp16_13, tensor var_4681_cast_fp16_14, tensor var_4681_cast_fp16_15 = split(axis = var_4681_axis_0, split_sizes = tile_58, x = var_4680_cast_fp16)[name = tensor("op_4681_cast_fp16")]; + tensor tile_59 = const()[name = tensor("tile_59"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4698_axis_0 = const()[name = tensor("op_4698_axis_0"), val = tensor(1)]; + tensor var_4698_cast_fp16_0, tensor var_4698_cast_fp16_1, tensor var_4698_cast_fp16_2, tensor var_4698_cast_fp16_3, tensor var_4698_cast_fp16_4, tensor var_4698_cast_fp16_5, tensor var_4698_cast_fp16_6, tensor var_4698_cast_fp16_7, tensor var_4698_cast_fp16_8, tensor var_4698_cast_fp16_9, tensor var_4698_cast_fp16_10, tensor var_4698_cast_fp16_11, tensor var_4698_cast_fp16_12, tensor var_4698_cast_fp16_13, tensor var_4698_cast_fp16_14, tensor var_4698_cast_fp16_15 = split(axis = var_4698_axis_0, split_sizes = tile_59, x = var_4660_cast_fp16)[name = tensor("op_4698_cast_fp16")]; + tensor aw_609_equation_0 = const()[name = tensor("aw_609_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_609_cast_fp16 = einsum(equation = aw_609_equation_0, values = (var_4681_cast_fp16_0, var_4663_cast_fp16_0))[name = tensor("aw_609_cast_fp16")]; + tensor aw_611_equation_0 = const()[name = tensor("aw_611_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_611_cast_fp16 = einsum(equation = aw_611_equation_0, values = (var_4681_cast_fp16_1, var_4663_cast_fp16_1))[name = tensor("aw_611_cast_fp16")]; + tensor aw_613_equation_0 = const()[name = tensor("aw_613_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_613_cast_fp16 = einsum(equation = aw_613_equation_0, values = (var_4681_cast_fp16_2, var_4663_cast_fp16_2))[name = tensor("aw_613_cast_fp16")]; + tensor aw_615_equation_0 = const()[name = tensor("aw_615_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_615_cast_fp16 = einsum(equation = aw_615_equation_0, values = (var_4681_cast_fp16_3, var_4663_cast_fp16_3))[name = tensor("aw_615_cast_fp16")]; + tensor aw_617_equation_0 = const()[name = tensor("aw_617_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_617_cast_fp16 = einsum(equation = aw_617_equation_0, values = (var_4681_cast_fp16_4, var_4663_cast_fp16_4))[name = tensor("aw_617_cast_fp16")]; + tensor aw_619_equation_0 = const()[name = tensor("aw_619_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_619_cast_fp16 = einsum(equation = aw_619_equation_0, values = (var_4681_cast_fp16_5, var_4663_cast_fp16_5))[name = tensor("aw_619_cast_fp16")]; + tensor aw_621_equation_0 = const()[name = tensor("aw_621_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_621_cast_fp16 = einsum(equation = aw_621_equation_0, values = (var_4681_cast_fp16_6, var_4663_cast_fp16_6))[name = tensor("aw_621_cast_fp16")]; + tensor aw_623_equation_0 = const()[name = tensor("aw_623_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_623_cast_fp16 = einsum(equation = aw_623_equation_0, values = (var_4681_cast_fp16_7, var_4663_cast_fp16_7))[name = tensor("aw_623_cast_fp16")]; + tensor aw_625_equation_0 = const()[name = tensor("aw_625_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_625_cast_fp16 = einsum(equation = aw_625_equation_0, values = (var_4681_cast_fp16_8, var_4663_cast_fp16_8))[name = tensor("aw_625_cast_fp16")]; + tensor aw_627_equation_0 = const()[name = tensor("aw_627_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_627_cast_fp16 = einsum(equation = aw_627_equation_0, values = (var_4681_cast_fp16_9, var_4663_cast_fp16_9))[name = tensor("aw_627_cast_fp16")]; + tensor aw_629_equation_0 = const()[name = tensor("aw_629_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_629_cast_fp16 = einsum(equation = aw_629_equation_0, values = (var_4681_cast_fp16_10, var_4663_cast_fp16_10))[name = tensor("aw_629_cast_fp16")]; + tensor aw_631_equation_0 = const()[name = tensor("aw_631_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_631_cast_fp16 = einsum(equation = aw_631_equation_0, values = (var_4681_cast_fp16_11, var_4663_cast_fp16_11))[name = tensor("aw_631_cast_fp16")]; + tensor aw_633_equation_0 = const()[name = tensor("aw_633_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_633_cast_fp16 = einsum(equation = aw_633_equation_0, values = (var_4681_cast_fp16_12, var_4663_cast_fp16_12))[name = tensor("aw_633_cast_fp16")]; + tensor aw_635_equation_0 = const()[name = tensor("aw_635_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_635_cast_fp16 = einsum(equation = aw_635_equation_0, values = (var_4681_cast_fp16_13, var_4663_cast_fp16_13))[name = tensor("aw_635_cast_fp16")]; + tensor aw_637_equation_0 = const()[name = tensor("aw_637_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_637_cast_fp16 = einsum(equation = aw_637_equation_0, values = (var_4681_cast_fp16_14, var_4663_cast_fp16_14))[name = tensor("aw_637_cast_fp16")]; + tensor aw_639_equation_0 = const()[name = tensor("aw_639_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_639_cast_fp16 = einsum(equation = aw_639_equation_0, values = (var_4681_cast_fp16_15, var_4663_cast_fp16_15))[name = tensor("aw_639_cast_fp16")]; + tensor var_4747_cast_fp16 = softmax(axis = var_4611, x = aw_609_cast_fp16)[name = tensor("op_4747_cast_fp16")]; + tensor var_4748_cast_fp16 = softmax(axis = var_4611, x = aw_611_cast_fp16)[name = tensor("op_4748_cast_fp16")]; + tensor var_4749_cast_fp16 = softmax(axis = var_4611, x = aw_613_cast_fp16)[name = tensor("op_4749_cast_fp16")]; + tensor var_4750_cast_fp16 = softmax(axis = var_4611, x = aw_615_cast_fp16)[name = tensor("op_4750_cast_fp16")]; + tensor var_4751_cast_fp16 = softmax(axis = var_4611, x = aw_617_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor var_4752_cast_fp16 = softmax(axis = var_4611, x = aw_619_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4753_cast_fp16 = softmax(axis = var_4611, x = aw_621_cast_fp16)[name = tensor("op_4753_cast_fp16")]; + tensor var_4754_cast_fp16 = softmax(axis = var_4611, x = aw_623_cast_fp16)[name = tensor("op_4754_cast_fp16")]; + tensor var_4755_cast_fp16 = softmax(axis = var_4611, x = aw_625_cast_fp16)[name = tensor("op_4755_cast_fp16")]; + tensor var_4756_cast_fp16 = softmax(axis = var_4611, x = aw_627_cast_fp16)[name = tensor("op_4756_cast_fp16")]; + tensor var_4757_cast_fp16 = softmax(axis = var_4611, x = aw_629_cast_fp16)[name = tensor("op_4757_cast_fp16")]; + tensor var_4758_cast_fp16 = softmax(axis = var_4611, x = aw_631_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = softmax(axis = var_4611, x = aw_633_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor var_4760_cast_fp16 = softmax(axis = var_4611, x = aw_635_cast_fp16)[name = tensor("op_4760_cast_fp16")]; + tensor var_4761_cast_fp16 = softmax(axis = var_4611, x = aw_637_cast_fp16)[name = tensor("op_4761_cast_fp16")]; + tensor var_4762_cast_fp16 = softmax(axis = var_4611, x = aw_639_cast_fp16)[name = tensor("op_4762_cast_fp16")]; + tensor var_4764_equation_0 = const()[name = tensor("op_4764_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4764_cast_fp16 = einsum(equation = var_4764_equation_0, values = (var_4698_cast_fp16_0, var_4747_cast_fp16))[name = tensor("op_4764_cast_fp16")]; + tensor var_4766_equation_0 = const()[name = tensor("op_4766_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4766_cast_fp16 = einsum(equation = var_4766_equation_0, values = (var_4698_cast_fp16_1, var_4748_cast_fp16))[name = tensor("op_4766_cast_fp16")]; + tensor var_4768_equation_0 = const()[name = tensor("op_4768_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4768_cast_fp16 = einsum(equation = var_4768_equation_0, values = (var_4698_cast_fp16_2, var_4749_cast_fp16))[name = tensor("op_4768_cast_fp16")]; + tensor var_4770_equation_0 = const()[name = tensor("op_4770_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4770_cast_fp16 = einsum(equation = var_4770_equation_0, values = (var_4698_cast_fp16_3, var_4750_cast_fp16))[name = tensor("op_4770_cast_fp16")]; + tensor var_4772_equation_0 = const()[name = tensor("op_4772_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4772_cast_fp16 = einsum(equation = var_4772_equation_0, values = (var_4698_cast_fp16_4, var_4751_cast_fp16))[name = tensor("op_4772_cast_fp16")]; + tensor var_4774_equation_0 = const()[name = tensor("op_4774_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4774_cast_fp16 = einsum(equation = var_4774_equation_0, values = (var_4698_cast_fp16_5, var_4752_cast_fp16))[name = tensor("op_4774_cast_fp16")]; + tensor var_4776_equation_0 = const()[name = tensor("op_4776_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4776_cast_fp16 = einsum(equation = var_4776_equation_0, values = (var_4698_cast_fp16_6, var_4753_cast_fp16))[name = tensor("op_4776_cast_fp16")]; + tensor var_4778_equation_0 = const()[name = tensor("op_4778_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4778_cast_fp16 = einsum(equation = var_4778_equation_0, values = (var_4698_cast_fp16_7, var_4754_cast_fp16))[name = tensor("op_4778_cast_fp16")]; + tensor var_4780_equation_0 = const()[name = tensor("op_4780_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4780_cast_fp16 = einsum(equation = var_4780_equation_0, values = (var_4698_cast_fp16_8, var_4755_cast_fp16))[name = tensor("op_4780_cast_fp16")]; + tensor var_4782_equation_0 = const()[name = tensor("op_4782_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4782_cast_fp16 = einsum(equation = var_4782_equation_0, values = (var_4698_cast_fp16_9, var_4756_cast_fp16))[name = tensor("op_4782_cast_fp16")]; + tensor var_4784_equation_0 = const()[name = tensor("op_4784_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4784_cast_fp16 = einsum(equation = var_4784_equation_0, values = (var_4698_cast_fp16_10, var_4757_cast_fp16))[name = tensor("op_4784_cast_fp16")]; + tensor var_4786_equation_0 = const()[name = tensor("op_4786_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4786_cast_fp16 = einsum(equation = var_4786_equation_0, values = (var_4698_cast_fp16_11, var_4758_cast_fp16))[name = tensor("op_4786_cast_fp16")]; + tensor var_4788_equation_0 = const()[name = tensor("op_4788_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4788_cast_fp16 = einsum(equation = var_4788_equation_0, values = (var_4698_cast_fp16_12, var_4759_cast_fp16))[name = tensor("op_4788_cast_fp16")]; + tensor var_4790_equation_0 = const()[name = tensor("op_4790_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4790_cast_fp16 = einsum(equation = var_4790_equation_0, values = (var_4698_cast_fp16_13, var_4760_cast_fp16))[name = tensor("op_4790_cast_fp16")]; + tensor var_4792_equation_0 = const()[name = tensor("op_4792_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4792_cast_fp16 = einsum(equation = var_4792_equation_0, values = (var_4698_cast_fp16_14, var_4761_cast_fp16))[name = tensor("op_4792_cast_fp16")]; + tensor var_4794_equation_0 = const()[name = tensor("op_4794_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_4794_cast_fp16 = einsum(equation = var_4794_equation_0, values = (var_4698_cast_fp16_15, var_4762_cast_fp16))[name = tensor("op_4794_cast_fp16")]; + tensor input_195_interleave_0 = const()[name = tensor("input_195_interleave_0"), val = tensor(false)]; + tensor input_195_cast_fp16 = concat(axis = var_4611, interleave = input_195_interleave_0, values = (var_4764_cast_fp16, var_4766_cast_fp16, var_4768_cast_fp16, var_4770_cast_fp16, var_4772_cast_fp16, var_4774_cast_fp16, var_4776_cast_fp16, var_4778_cast_fp16, var_4780_cast_fp16, var_4782_cast_fp16, var_4784_cast_fp16, var_4786_cast_fp16, var_4788_cast_fp16, var_4790_cast_fp16, var_4792_cast_fp16, var_4794_cast_fp16))[name = tensor("input_195_cast_fp16")]; + tensor var_4803_pad_type_0 = const()[name = tensor("op_4803_pad_type_0"), val = tensor("valid")]; + tensor var_4803_strides_0 = const()[name = tensor("op_4803_strides_0"), val = tensor([1, 1])]; + tensor var_4803_pad_0 = const()[name = tensor("op_4803_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4803_dilations_0 = const()[name = tensor("op_4803_dilations_0"), val = tensor([1, 1])]; + tensor var_4803_groups_0 = const()[name = tensor("op_4803_groups_0"), val = tensor(1)]; + tensor blocks_19_attn_out_weight_to_fp16 = const()[name = tensor("blocks_19_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494795392)))]; + tensor blocks_19_attn_out_bias_to_fp16 = const()[name = tensor("blocks_19_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496892608)))]; + tensor var_4803_cast_fp16 = conv(bias = blocks_19_attn_out_bias_to_fp16, dilations = var_4803_dilations_0, groups = var_4803_groups_0, pad = var_4803_pad_0, pad_type = var_4803_pad_type_0, strides = var_4803_strides_0, weight = blocks_19_attn_out_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_4803_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([1])]; + tensor input_197_gamma_0_to_fp16 = const()[name = tensor("input_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496894720)))]; + tensor input_197_beta_0_to_fp16 = const()[name = tensor("input_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496896832)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = input_197_beta_0_to_fp16, epsilon = var_4813_to_fp16, gamma = input_197_gamma_0_to_fp16, x = inputs_79_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("valid")]; + tensor input_199_strides_0 = const()[name = tensor("input_199_strides_0"), val = tensor([1, 1])]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_199_dilations_0 = const()[name = tensor("input_199_dilations_0"), val = tensor([1, 1])]; + tensor input_199_groups_0 = const()[name = tensor("input_199_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496898944)))]; + tensor blocks_19_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505287616)))]; + tensor input_199_cast_fp16 = conv(bias = blocks_19_mlp_0_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = blocks_19_mlp_0_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_4839_pad_type_0 = const()[name = tensor("op_4839_pad_type_0"), val = tensor("valid")]; + tensor var_4839_strides_0 = const()[name = tensor("op_4839_strides_0"), val = tensor([1, 1])]; + tensor var_4839_pad_0 = const()[name = tensor("op_4839_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4839_dilations_0 = const()[name = tensor("op_4839_dilations_0"), val = tensor([1, 1])]; + tensor var_4839_groups_0 = const()[name = tensor("op_4839_groups_0"), val = tensor(1)]; + tensor blocks_19_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505295872)))]; + tensor blocks_19_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513684544)))]; + tensor var_4839_cast_fp16 = conv(bias = blocks_19_mlp_2_bias_to_fp16, dilations = var_4839_dilations_0, groups = var_4839_groups_0, pad = var_4839_pad_0, pad_type = var_4839_pad_type_0, strides = var_4839_strides_0, weight = blocks_19_mlp_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("op_4839_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = var_4839_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor(1)]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([1])]; + tensor input_203_gamma_0_to_fp16 = const()[name = tensor("input_203_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513686656)))]; + tensor input_203_beta_0_to_fp16 = const()[name = tensor("input_203_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513688768)))]; + tensor var_4864_to_fp16 = const()[name = tensor("op_4864_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = input_203_beta_0_to_fp16, epsilon = var_4864_to_fp16, gamma = input_203_gamma_0_to_fp16, x = inputs_81_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor q_41_pad_type_0 = const()[name = tensor("q_41_pad_type_0"), val = tensor("valid")]; + tensor q_41_strides_0 = const()[name = tensor("q_41_strides_0"), val = tensor([1, 1])]; + tensor q_41_pad_0 = const()[name = tensor("q_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_41_dilations_0 = const()[name = tensor("q_41_dilations_0"), val = tensor([1, 1])]; + tensor q_41_groups_0 = const()[name = tensor("q_41_groups_0"), val = tensor(1)]; + tensor var_4899_weight_0_to_fp16 = const()[name = tensor("op_4899_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513690880)))]; + tensor var_4899_bias_0_to_fp16 = const()[name = tensor("op_4899_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515788096)))]; + tensor var_4899_cast_fp16 = conv(bias = var_4899_bias_0_to_fp16, dilations = q_41_dilations_0, groups = q_41_groups_0, pad = q_41_pad_0, pad_type = q_41_pad_type_0, strides = q_41_strides_0, weight = var_4899_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("op_4899_cast_fp16")]; + tensor k_41_pad_type_0 = const()[name = tensor("k_41_pad_type_0"), val = tensor("valid")]; + tensor k_41_strides_0 = const()[name = tensor("k_41_strides_0"), val = tensor([1, 1])]; + tensor k_41_pad_0 = const()[name = tensor("k_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_41_dilations_0 = const()[name = tensor("k_41_dilations_0"), val = tensor([1, 1])]; + tensor k_41_groups_0 = const()[name = tensor("k_41_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_key_weight_to_fp16 = const()[name = tensor("blocks_20_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515790208)))]; + tensor k_41_cast_fp16 = conv(dilations = k_41_dilations_0, groups = k_41_groups_0, pad = k_41_pad_0, pad_type = k_41_pad_type_0, strides = k_41_strides_0, weight = blocks_20_attn_key_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor var_4897_pad_type_0 = const()[name = tensor("op_4897_pad_type_0"), val = tensor("valid")]; + tensor var_4897_strides_0 = const()[name = tensor("op_4897_strides_0"), val = tensor([1, 1])]; + tensor var_4897_pad_0 = const()[name = tensor("op_4897_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4897_dilations_0 = const()[name = tensor("op_4897_dilations_0"), val = tensor([1, 1])]; + tensor var_4897_groups_0 = const()[name = tensor("op_4897_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_value_weight_to_fp16 = const()[name = tensor("blocks_20_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517887424)))]; + tensor blocks_20_attn_value_bias_to_fp16 = const()[name = tensor("blocks_20_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519984640)))]; + tensor var_4897_cast_fp16 = conv(bias = blocks_20_attn_value_bias_to_fp16, dilations = var_4897_dilations_0, groups = var_4897_groups_0, pad = var_4897_pad_0, pad_type = var_4897_pad_type_0, strides = var_4897_strides_0, weight = blocks_20_attn_value_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("op_4897_cast_fp16")]; + tensor tile_60 = const()[name = tensor("tile_60"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4900_axis_0 = const()[name = tensor("op_4900_axis_0"), val = tensor(1)]; + tensor var_4900_cast_fp16_0, tensor var_4900_cast_fp16_1, tensor var_4900_cast_fp16_2, tensor var_4900_cast_fp16_3, tensor var_4900_cast_fp16_4, tensor var_4900_cast_fp16_5, tensor var_4900_cast_fp16_6, tensor var_4900_cast_fp16_7, tensor var_4900_cast_fp16_8, tensor var_4900_cast_fp16_9, tensor var_4900_cast_fp16_10, tensor var_4900_cast_fp16_11, tensor var_4900_cast_fp16_12, tensor var_4900_cast_fp16_13, tensor var_4900_cast_fp16_14, tensor var_4900_cast_fp16_15 = split(axis = var_4900_axis_0, split_sizes = tile_60, x = var_4899_cast_fp16)[name = tensor("op_4900_cast_fp16")]; + tensor var_4917_perm_0 = const()[name = tensor("op_4917_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_61 = const()[name = tensor("tile_61"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4918_axis_0 = const()[name = tensor("op_4918_axis_0"), val = tensor(3)]; + tensor var_4917_cast_fp16 = transpose(perm = var_4917_perm_0, x = k_41_cast_fp16)[name = tensor("transpose_4")]; + tensor var_4918_cast_fp16_0, tensor var_4918_cast_fp16_1, tensor var_4918_cast_fp16_2, tensor var_4918_cast_fp16_3, tensor var_4918_cast_fp16_4, tensor var_4918_cast_fp16_5, tensor var_4918_cast_fp16_6, tensor var_4918_cast_fp16_7, tensor var_4918_cast_fp16_8, tensor var_4918_cast_fp16_9, tensor var_4918_cast_fp16_10, tensor var_4918_cast_fp16_11, tensor var_4918_cast_fp16_12, tensor var_4918_cast_fp16_13, tensor var_4918_cast_fp16_14, tensor var_4918_cast_fp16_15 = split(axis = var_4918_axis_0, split_sizes = tile_61, x = var_4917_cast_fp16)[name = tensor("op_4918_cast_fp16")]; + tensor tile_62 = const()[name = tensor("tile_62"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_4935_axis_0 = const()[name = tensor("op_4935_axis_0"), val = tensor(1)]; + tensor var_4935_cast_fp16_0, tensor var_4935_cast_fp16_1, tensor var_4935_cast_fp16_2, tensor var_4935_cast_fp16_3, tensor var_4935_cast_fp16_4, tensor var_4935_cast_fp16_5, tensor var_4935_cast_fp16_6, tensor var_4935_cast_fp16_7, tensor var_4935_cast_fp16_8, tensor var_4935_cast_fp16_9, tensor var_4935_cast_fp16_10, tensor var_4935_cast_fp16_11, tensor var_4935_cast_fp16_12, tensor var_4935_cast_fp16_13, tensor var_4935_cast_fp16_14, tensor var_4935_cast_fp16_15 = split(axis = var_4935_axis_0, split_sizes = tile_62, x = var_4897_cast_fp16)[name = tensor("op_4935_cast_fp16")]; + tensor aw_641_equation_0 = const()[name = tensor("aw_641_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_641_cast_fp16 = einsum(equation = aw_641_equation_0, values = (var_4918_cast_fp16_0, var_4900_cast_fp16_0))[name = tensor("aw_641_cast_fp16")]; + tensor aw_643_equation_0 = const()[name = tensor("aw_643_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_643_cast_fp16 = einsum(equation = aw_643_equation_0, values = (var_4918_cast_fp16_1, var_4900_cast_fp16_1))[name = tensor("aw_643_cast_fp16")]; + tensor aw_645_equation_0 = const()[name = tensor("aw_645_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_645_cast_fp16 = einsum(equation = aw_645_equation_0, values = (var_4918_cast_fp16_2, var_4900_cast_fp16_2))[name = tensor("aw_645_cast_fp16")]; + tensor aw_647_equation_0 = const()[name = tensor("aw_647_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_647_cast_fp16 = einsum(equation = aw_647_equation_0, values = (var_4918_cast_fp16_3, var_4900_cast_fp16_3))[name = tensor("aw_647_cast_fp16")]; + tensor aw_649_equation_0 = const()[name = tensor("aw_649_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_649_cast_fp16 = einsum(equation = aw_649_equation_0, values = (var_4918_cast_fp16_4, var_4900_cast_fp16_4))[name = tensor("aw_649_cast_fp16")]; + tensor aw_651_equation_0 = const()[name = tensor("aw_651_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_651_cast_fp16 = einsum(equation = aw_651_equation_0, values = (var_4918_cast_fp16_5, var_4900_cast_fp16_5))[name = tensor("aw_651_cast_fp16")]; + tensor aw_653_equation_0 = const()[name = tensor("aw_653_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_653_cast_fp16 = einsum(equation = aw_653_equation_0, values = (var_4918_cast_fp16_6, var_4900_cast_fp16_6))[name = tensor("aw_653_cast_fp16")]; + tensor aw_655_equation_0 = const()[name = tensor("aw_655_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_655_cast_fp16 = einsum(equation = aw_655_equation_0, values = (var_4918_cast_fp16_7, var_4900_cast_fp16_7))[name = tensor("aw_655_cast_fp16")]; + tensor aw_657_equation_0 = const()[name = tensor("aw_657_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_657_cast_fp16 = einsum(equation = aw_657_equation_0, values = (var_4918_cast_fp16_8, var_4900_cast_fp16_8))[name = tensor("aw_657_cast_fp16")]; + tensor aw_659_equation_0 = const()[name = tensor("aw_659_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_659_cast_fp16 = einsum(equation = aw_659_equation_0, values = (var_4918_cast_fp16_9, var_4900_cast_fp16_9))[name = tensor("aw_659_cast_fp16")]; + tensor aw_661_equation_0 = const()[name = tensor("aw_661_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_661_cast_fp16 = einsum(equation = aw_661_equation_0, values = (var_4918_cast_fp16_10, var_4900_cast_fp16_10))[name = tensor("aw_661_cast_fp16")]; + tensor aw_663_equation_0 = const()[name = tensor("aw_663_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_663_cast_fp16 = einsum(equation = aw_663_equation_0, values = (var_4918_cast_fp16_11, var_4900_cast_fp16_11))[name = tensor("aw_663_cast_fp16")]; + tensor aw_665_equation_0 = const()[name = tensor("aw_665_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_665_cast_fp16 = einsum(equation = aw_665_equation_0, values = (var_4918_cast_fp16_12, var_4900_cast_fp16_12))[name = tensor("aw_665_cast_fp16")]; + tensor aw_667_equation_0 = const()[name = tensor("aw_667_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_667_cast_fp16 = einsum(equation = aw_667_equation_0, values = (var_4918_cast_fp16_13, var_4900_cast_fp16_13))[name = tensor("aw_667_cast_fp16")]; + tensor aw_669_equation_0 = const()[name = tensor("aw_669_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_669_cast_fp16 = einsum(equation = aw_669_equation_0, values = (var_4918_cast_fp16_14, var_4900_cast_fp16_14))[name = tensor("aw_669_cast_fp16")]; + tensor aw_671_equation_0 = const()[name = tensor("aw_671_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_671_cast_fp16 = einsum(equation = aw_671_equation_0, values = (var_4918_cast_fp16_15, var_4900_cast_fp16_15))[name = tensor("aw_671_cast_fp16")]; + tensor var_4984_cast_fp16 = softmax(axis = var_4848, x = aw_641_cast_fp16)[name = tensor("op_4984_cast_fp16")]; + tensor var_4985_cast_fp16 = softmax(axis = var_4848, x = aw_643_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor var_4986_cast_fp16 = softmax(axis = var_4848, x = aw_645_cast_fp16)[name = tensor("op_4986_cast_fp16")]; + tensor var_4987_cast_fp16 = softmax(axis = var_4848, x = aw_647_cast_fp16)[name = tensor("op_4987_cast_fp16")]; + tensor var_4988_cast_fp16 = softmax(axis = var_4848, x = aw_649_cast_fp16)[name = tensor("op_4988_cast_fp16")]; + tensor var_4989_cast_fp16 = softmax(axis = var_4848, x = aw_651_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor var_4990_cast_fp16 = softmax(axis = var_4848, x = aw_653_cast_fp16)[name = tensor("op_4990_cast_fp16")]; + tensor var_4991_cast_fp16 = softmax(axis = var_4848, x = aw_655_cast_fp16)[name = tensor("op_4991_cast_fp16")]; + tensor var_4992_cast_fp16 = softmax(axis = var_4848, x = aw_657_cast_fp16)[name = tensor("op_4992_cast_fp16")]; + tensor var_4993_cast_fp16 = softmax(axis = var_4848, x = aw_659_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor var_4994_cast_fp16 = softmax(axis = var_4848, x = aw_661_cast_fp16)[name = tensor("op_4994_cast_fp16")]; + tensor var_4995_cast_fp16 = softmax(axis = var_4848, x = aw_663_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor var_4996_cast_fp16 = softmax(axis = var_4848, x = aw_665_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_4997_cast_fp16 = softmax(axis = var_4848, x = aw_667_cast_fp16)[name = tensor("op_4997_cast_fp16")]; + tensor var_4998_cast_fp16 = softmax(axis = var_4848, x = aw_669_cast_fp16)[name = tensor("op_4998_cast_fp16")]; + tensor var_4999_cast_fp16 = softmax(axis = var_4848, x = aw_671_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5001_equation_0 = const()[name = tensor("op_5001_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5001_cast_fp16 = einsum(equation = var_5001_equation_0, values = (var_4935_cast_fp16_0, var_4984_cast_fp16))[name = tensor("op_5001_cast_fp16")]; + tensor var_5003_equation_0 = const()[name = tensor("op_5003_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5003_cast_fp16 = einsum(equation = var_5003_equation_0, values = (var_4935_cast_fp16_1, var_4985_cast_fp16))[name = tensor("op_5003_cast_fp16")]; + tensor var_5005_equation_0 = const()[name = tensor("op_5005_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5005_cast_fp16 = einsum(equation = var_5005_equation_0, values = (var_4935_cast_fp16_2, var_4986_cast_fp16))[name = tensor("op_5005_cast_fp16")]; + tensor var_5007_equation_0 = const()[name = tensor("op_5007_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5007_cast_fp16 = einsum(equation = var_5007_equation_0, values = (var_4935_cast_fp16_3, var_4987_cast_fp16))[name = tensor("op_5007_cast_fp16")]; + tensor var_5009_equation_0 = const()[name = tensor("op_5009_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5009_cast_fp16 = einsum(equation = var_5009_equation_0, values = (var_4935_cast_fp16_4, var_4988_cast_fp16))[name = tensor("op_5009_cast_fp16")]; + tensor var_5011_equation_0 = const()[name = tensor("op_5011_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5011_cast_fp16 = einsum(equation = var_5011_equation_0, values = (var_4935_cast_fp16_5, var_4989_cast_fp16))[name = tensor("op_5011_cast_fp16")]; + tensor var_5013_equation_0 = const()[name = tensor("op_5013_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5013_cast_fp16 = einsum(equation = var_5013_equation_0, values = (var_4935_cast_fp16_6, var_4990_cast_fp16))[name = tensor("op_5013_cast_fp16")]; + tensor var_5015_equation_0 = const()[name = tensor("op_5015_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5015_cast_fp16 = einsum(equation = var_5015_equation_0, values = (var_4935_cast_fp16_7, var_4991_cast_fp16))[name = tensor("op_5015_cast_fp16")]; + tensor var_5017_equation_0 = const()[name = tensor("op_5017_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5017_cast_fp16 = einsum(equation = var_5017_equation_0, values = (var_4935_cast_fp16_8, var_4992_cast_fp16))[name = tensor("op_5017_cast_fp16")]; + tensor var_5019_equation_0 = const()[name = tensor("op_5019_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5019_cast_fp16 = einsum(equation = var_5019_equation_0, values = (var_4935_cast_fp16_9, var_4993_cast_fp16))[name = tensor("op_5019_cast_fp16")]; + tensor var_5021_equation_0 = const()[name = tensor("op_5021_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5021_cast_fp16 = einsum(equation = var_5021_equation_0, values = (var_4935_cast_fp16_10, var_4994_cast_fp16))[name = tensor("op_5021_cast_fp16")]; + tensor var_5023_equation_0 = const()[name = tensor("op_5023_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5023_cast_fp16 = einsum(equation = var_5023_equation_0, values = (var_4935_cast_fp16_11, var_4995_cast_fp16))[name = tensor("op_5023_cast_fp16")]; + tensor var_5025_equation_0 = const()[name = tensor("op_5025_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5025_cast_fp16 = einsum(equation = var_5025_equation_0, values = (var_4935_cast_fp16_12, var_4996_cast_fp16))[name = tensor("op_5025_cast_fp16")]; + tensor var_5027_equation_0 = const()[name = tensor("op_5027_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5027_cast_fp16 = einsum(equation = var_5027_equation_0, values = (var_4935_cast_fp16_13, var_4997_cast_fp16))[name = tensor("op_5027_cast_fp16")]; + tensor var_5029_equation_0 = const()[name = tensor("op_5029_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5029_cast_fp16 = einsum(equation = var_5029_equation_0, values = (var_4935_cast_fp16_14, var_4998_cast_fp16))[name = tensor("op_5029_cast_fp16")]; + tensor var_5031_equation_0 = const()[name = tensor("op_5031_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5031_cast_fp16 = einsum(equation = var_5031_equation_0, values = (var_4935_cast_fp16_15, var_4999_cast_fp16))[name = tensor("op_5031_cast_fp16")]; + tensor input_205_interleave_0 = const()[name = tensor("input_205_interleave_0"), val = tensor(false)]; + tensor input_205_cast_fp16 = concat(axis = var_4848, interleave = input_205_interleave_0, values = (var_5001_cast_fp16, var_5003_cast_fp16, var_5005_cast_fp16, var_5007_cast_fp16, var_5009_cast_fp16, var_5011_cast_fp16, var_5013_cast_fp16, var_5015_cast_fp16, var_5017_cast_fp16, var_5019_cast_fp16, var_5021_cast_fp16, var_5023_cast_fp16, var_5025_cast_fp16, var_5027_cast_fp16, var_5029_cast_fp16, var_5031_cast_fp16))[name = tensor("input_205_cast_fp16")]; + tensor var_5040_pad_type_0 = const()[name = tensor("op_5040_pad_type_0"), val = tensor("valid")]; + tensor var_5040_strides_0 = const()[name = tensor("op_5040_strides_0"), val = tensor([1, 1])]; + tensor var_5040_pad_0 = const()[name = tensor("op_5040_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5040_dilations_0 = const()[name = tensor("op_5040_dilations_0"), val = tensor([1, 1])]; + tensor var_5040_groups_0 = const()[name = tensor("op_5040_groups_0"), val = tensor(1)]; + tensor blocks_20_attn_out_weight_to_fp16 = const()[name = tensor("blocks_20_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(519986752)))]; + tensor blocks_20_attn_out_bias_to_fp16 = const()[name = tensor("blocks_20_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522083968)))]; + tensor var_5040_cast_fp16 = conv(bias = blocks_20_attn_out_bias_to_fp16, dilations = var_5040_dilations_0, groups = var_5040_groups_0, pad = var_5040_pad_0, pad_type = var_5040_pad_type_0, strides = var_5040_strides_0, weight = blocks_20_attn_out_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_5040_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([1])]; + tensor input_207_gamma_0_to_fp16 = const()[name = tensor("input_207_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522086080)))]; + tensor input_207_beta_0_to_fp16 = const()[name = tensor("input_207_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522088192)))]; + tensor var_5050_to_fp16 = const()[name = tensor("op_5050_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = input_207_beta_0_to_fp16, epsilon = var_5050_to_fp16, gamma = input_207_gamma_0_to_fp16, x = inputs_83_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("valid")]; + tensor input_209_strides_0 = const()[name = tensor("input_209_strides_0"), val = tensor([1, 1])]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_209_dilations_0 = const()[name = tensor("input_209_dilations_0"), val = tensor([1, 1])]; + tensor input_209_groups_0 = const()[name = tensor("input_209_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522090304)))]; + tensor blocks_20_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530478976)))]; + tensor input_209_cast_fp16 = conv(bias = blocks_20_mlp_0_bias_to_fp16, dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = blocks_20_mlp_0_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = input_209_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_5076_pad_type_0 = const()[name = tensor("op_5076_pad_type_0"), val = tensor("valid")]; + tensor var_5076_strides_0 = const()[name = tensor("op_5076_strides_0"), val = tensor([1, 1])]; + tensor var_5076_pad_0 = const()[name = tensor("op_5076_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5076_dilations_0 = const()[name = tensor("op_5076_dilations_0"), val = tensor([1, 1])]; + tensor var_5076_groups_0 = const()[name = tensor("op_5076_groups_0"), val = tensor(1)]; + tensor blocks_20_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530487232)))]; + tensor blocks_20_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538875904)))]; + tensor var_5076_cast_fp16 = conv(bias = blocks_20_mlp_2_bias_to_fp16, dilations = var_5076_dilations_0, groups = var_5076_groups_0, pad = var_5076_pad_0, pad_type = var_5076_pad_type_0, strides = var_5076_strides_0, weight = blocks_20_mlp_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("op_5076_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = var_5076_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_5085 = const()[name = tensor("op_5085"), val = tensor(1)]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([1])]; + tensor input_213_gamma_0_to_fp16 = const()[name = tensor("input_213_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538878016)))]; + tensor input_213_beta_0_to_fp16 = const()[name = tensor("input_213_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538880128)))]; + tensor var_5101_to_fp16 = const()[name = tensor("op_5101_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = input_213_beta_0_to_fp16, epsilon = var_5101_to_fp16, gamma = input_213_gamma_0_to_fp16, x = inputs_85_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor q_43_pad_type_0 = const()[name = tensor("q_43_pad_type_0"), val = tensor("valid")]; + tensor q_43_strides_0 = const()[name = tensor("q_43_strides_0"), val = tensor([1, 1])]; + tensor q_43_pad_0 = const()[name = tensor("q_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_43_dilations_0 = const()[name = tensor("q_43_dilations_0"), val = tensor([1, 1])]; + tensor q_43_groups_0 = const()[name = tensor("q_43_groups_0"), val = tensor(1)]; + tensor var_5136_weight_0_to_fp16 = const()[name = tensor("op_5136_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538882240)))]; + tensor var_5136_bias_0_to_fp16 = const()[name = tensor("op_5136_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540979456)))]; + tensor var_5136_cast_fp16 = conv(bias = var_5136_bias_0_to_fp16, dilations = q_43_dilations_0, groups = q_43_groups_0, pad = q_43_pad_0, pad_type = q_43_pad_type_0, strides = q_43_strides_0, weight = var_5136_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5136_cast_fp16")]; + tensor k_43_pad_type_0 = const()[name = tensor("k_43_pad_type_0"), val = tensor("valid")]; + tensor k_43_strides_0 = const()[name = tensor("k_43_strides_0"), val = tensor([1, 1])]; + tensor k_43_pad_0 = const()[name = tensor("k_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_43_dilations_0 = const()[name = tensor("k_43_dilations_0"), val = tensor([1, 1])]; + tensor k_43_groups_0 = const()[name = tensor("k_43_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_key_weight_to_fp16 = const()[name = tensor("blocks_21_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540981568)))]; + tensor k_43_cast_fp16 = conv(dilations = k_43_dilations_0, groups = k_43_groups_0, pad = k_43_pad_0, pad_type = k_43_pad_type_0, strides = k_43_strides_0, weight = blocks_21_attn_key_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor var_5134_pad_type_0 = const()[name = tensor("op_5134_pad_type_0"), val = tensor("valid")]; + tensor var_5134_strides_0 = const()[name = tensor("op_5134_strides_0"), val = tensor([1, 1])]; + tensor var_5134_pad_0 = const()[name = tensor("op_5134_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5134_dilations_0 = const()[name = tensor("op_5134_dilations_0"), val = tensor([1, 1])]; + tensor var_5134_groups_0 = const()[name = tensor("op_5134_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_value_weight_to_fp16 = const()[name = tensor("blocks_21_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543078784)))]; + tensor blocks_21_attn_value_bias_to_fp16 = const()[name = tensor("blocks_21_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545176000)))]; + tensor var_5134_cast_fp16 = conv(bias = blocks_21_attn_value_bias_to_fp16, dilations = var_5134_dilations_0, groups = var_5134_groups_0, pad = var_5134_pad_0, pad_type = var_5134_pad_type_0, strides = var_5134_strides_0, weight = blocks_21_attn_value_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("op_5134_cast_fp16")]; + tensor tile_63 = const()[name = tensor("tile_63"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5137_axis_0 = const()[name = tensor("op_5137_axis_0"), val = tensor(1)]; + tensor var_5137_cast_fp16_0, tensor var_5137_cast_fp16_1, tensor var_5137_cast_fp16_2, tensor var_5137_cast_fp16_3, tensor var_5137_cast_fp16_4, tensor var_5137_cast_fp16_5, tensor var_5137_cast_fp16_6, tensor var_5137_cast_fp16_7, tensor var_5137_cast_fp16_8, tensor var_5137_cast_fp16_9, tensor var_5137_cast_fp16_10, tensor var_5137_cast_fp16_11, tensor var_5137_cast_fp16_12, tensor var_5137_cast_fp16_13, tensor var_5137_cast_fp16_14, tensor var_5137_cast_fp16_15 = split(axis = var_5137_axis_0, split_sizes = tile_63, x = var_5136_cast_fp16)[name = tensor("op_5137_cast_fp16")]; + tensor var_5154_perm_0 = const()[name = tensor("op_5154_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_64 = const()[name = tensor("tile_64"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5155_axis_0 = const()[name = tensor("op_5155_axis_0"), val = tensor(3)]; + tensor var_5154_cast_fp16 = transpose(perm = var_5154_perm_0, x = k_43_cast_fp16)[name = tensor("transpose_3")]; + tensor var_5155_cast_fp16_0, tensor var_5155_cast_fp16_1, tensor var_5155_cast_fp16_2, tensor var_5155_cast_fp16_3, tensor var_5155_cast_fp16_4, tensor var_5155_cast_fp16_5, tensor var_5155_cast_fp16_6, tensor var_5155_cast_fp16_7, tensor var_5155_cast_fp16_8, tensor var_5155_cast_fp16_9, tensor var_5155_cast_fp16_10, tensor var_5155_cast_fp16_11, tensor var_5155_cast_fp16_12, tensor var_5155_cast_fp16_13, tensor var_5155_cast_fp16_14, tensor var_5155_cast_fp16_15 = split(axis = var_5155_axis_0, split_sizes = tile_64, x = var_5154_cast_fp16)[name = tensor("op_5155_cast_fp16")]; + tensor tile_65 = const()[name = tensor("tile_65"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5172_axis_0 = const()[name = tensor("op_5172_axis_0"), val = tensor(1)]; + tensor var_5172_cast_fp16_0, tensor var_5172_cast_fp16_1, tensor var_5172_cast_fp16_2, tensor var_5172_cast_fp16_3, tensor var_5172_cast_fp16_4, tensor var_5172_cast_fp16_5, tensor var_5172_cast_fp16_6, tensor var_5172_cast_fp16_7, tensor var_5172_cast_fp16_8, tensor var_5172_cast_fp16_9, tensor var_5172_cast_fp16_10, tensor var_5172_cast_fp16_11, tensor var_5172_cast_fp16_12, tensor var_5172_cast_fp16_13, tensor var_5172_cast_fp16_14, tensor var_5172_cast_fp16_15 = split(axis = var_5172_axis_0, split_sizes = tile_65, x = var_5134_cast_fp16)[name = tensor("op_5172_cast_fp16")]; + tensor aw_673_equation_0 = const()[name = tensor("aw_673_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_673_cast_fp16 = einsum(equation = aw_673_equation_0, values = (var_5155_cast_fp16_0, var_5137_cast_fp16_0))[name = tensor("aw_673_cast_fp16")]; + tensor aw_675_equation_0 = const()[name = tensor("aw_675_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_675_cast_fp16 = einsum(equation = aw_675_equation_0, values = (var_5155_cast_fp16_1, var_5137_cast_fp16_1))[name = tensor("aw_675_cast_fp16")]; + tensor aw_677_equation_0 = const()[name = tensor("aw_677_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_677_cast_fp16 = einsum(equation = aw_677_equation_0, values = (var_5155_cast_fp16_2, var_5137_cast_fp16_2))[name = tensor("aw_677_cast_fp16")]; + tensor aw_679_equation_0 = const()[name = tensor("aw_679_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_679_cast_fp16 = einsum(equation = aw_679_equation_0, values = (var_5155_cast_fp16_3, var_5137_cast_fp16_3))[name = tensor("aw_679_cast_fp16")]; + tensor aw_681_equation_0 = const()[name = tensor("aw_681_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_681_cast_fp16 = einsum(equation = aw_681_equation_0, values = (var_5155_cast_fp16_4, var_5137_cast_fp16_4))[name = tensor("aw_681_cast_fp16")]; + tensor aw_683_equation_0 = const()[name = tensor("aw_683_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_683_cast_fp16 = einsum(equation = aw_683_equation_0, values = (var_5155_cast_fp16_5, var_5137_cast_fp16_5))[name = tensor("aw_683_cast_fp16")]; + tensor aw_685_equation_0 = const()[name = tensor("aw_685_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_685_cast_fp16 = einsum(equation = aw_685_equation_0, values = (var_5155_cast_fp16_6, var_5137_cast_fp16_6))[name = tensor("aw_685_cast_fp16")]; + tensor aw_687_equation_0 = const()[name = tensor("aw_687_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_687_cast_fp16 = einsum(equation = aw_687_equation_0, values = (var_5155_cast_fp16_7, var_5137_cast_fp16_7))[name = tensor("aw_687_cast_fp16")]; + tensor aw_689_equation_0 = const()[name = tensor("aw_689_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_689_cast_fp16 = einsum(equation = aw_689_equation_0, values = (var_5155_cast_fp16_8, var_5137_cast_fp16_8))[name = tensor("aw_689_cast_fp16")]; + tensor aw_691_equation_0 = const()[name = tensor("aw_691_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_691_cast_fp16 = einsum(equation = aw_691_equation_0, values = (var_5155_cast_fp16_9, var_5137_cast_fp16_9))[name = tensor("aw_691_cast_fp16")]; + tensor aw_693_equation_0 = const()[name = tensor("aw_693_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_693_cast_fp16 = einsum(equation = aw_693_equation_0, values = (var_5155_cast_fp16_10, var_5137_cast_fp16_10))[name = tensor("aw_693_cast_fp16")]; + tensor aw_695_equation_0 = const()[name = tensor("aw_695_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_695_cast_fp16 = einsum(equation = aw_695_equation_0, values = (var_5155_cast_fp16_11, var_5137_cast_fp16_11))[name = tensor("aw_695_cast_fp16")]; + tensor aw_697_equation_0 = const()[name = tensor("aw_697_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_697_cast_fp16 = einsum(equation = aw_697_equation_0, values = (var_5155_cast_fp16_12, var_5137_cast_fp16_12))[name = tensor("aw_697_cast_fp16")]; + tensor aw_699_equation_0 = const()[name = tensor("aw_699_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_699_cast_fp16 = einsum(equation = aw_699_equation_0, values = (var_5155_cast_fp16_13, var_5137_cast_fp16_13))[name = tensor("aw_699_cast_fp16")]; + tensor aw_701_equation_0 = const()[name = tensor("aw_701_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_701_cast_fp16 = einsum(equation = aw_701_equation_0, values = (var_5155_cast_fp16_14, var_5137_cast_fp16_14))[name = tensor("aw_701_cast_fp16")]; + tensor aw_703_equation_0 = const()[name = tensor("aw_703_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_703_cast_fp16 = einsum(equation = aw_703_equation_0, values = (var_5155_cast_fp16_15, var_5137_cast_fp16_15))[name = tensor("aw_703_cast_fp16")]; + tensor var_5221_cast_fp16 = softmax(axis = var_5085, x = aw_673_cast_fp16)[name = tensor("op_5221_cast_fp16")]; + tensor var_5222_cast_fp16 = softmax(axis = var_5085, x = aw_675_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5223_cast_fp16 = softmax(axis = var_5085, x = aw_677_cast_fp16)[name = tensor("op_5223_cast_fp16")]; + tensor var_5224_cast_fp16 = softmax(axis = var_5085, x = aw_679_cast_fp16)[name = tensor("op_5224_cast_fp16")]; + tensor var_5225_cast_fp16 = softmax(axis = var_5085, x = aw_681_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5226_cast_fp16 = softmax(axis = var_5085, x = aw_683_cast_fp16)[name = tensor("op_5226_cast_fp16")]; + tensor var_5227_cast_fp16 = softmax(axis = var_5085, x = aw_685_cast_fp16)[name = tensor("op_5227_cast_fp16")]; + tensor var_5228_cast_fp16 = softmax(axis = var_5085, x = aw_687_cast_fp16)[name = tensor("op_5228_cast_fp16")]; + tensor var_5229_cast_fp16 = softmax(axis = var_5085, x = aw_689_cast_fp16)[name = tensor("op_5229_cast_fp16")]; + tensor var_5230_cast_fp16 = softmax(axis = var_5085, x = aw_691_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5231_cast_fp16 = softmax(axis = var_5085, x = aw_693_cast_fp16)[name = tensor("op_5231_cast_fp16")]; + tensor var_5232_cast_fp16 = softmax(axis = var_5085, x = aw_695_cast_fp16)[name = tensor("op_5232_cast_fp16")]; + tensor var_5233_cast_fp16 = softmax(axis = var_5085, x = aw_697_cast_fp16)[name = tensor("op_5233_cast_fp16")]; + tensor var_5234_cast_fp16 = softmax(axis = var_5085, x = aw_699_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = softmax(axis = var_5085, x = aw_701_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor var_5236_cast_fp16 = softmax(axis = var_5085, x = aw_703_cast_fp16)[name = tensor("op_5236_cast_fp16")]; + tensor var_5238_equation_0 = const()[name = tensor("op_5238_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5238_cast_fp16 = einsum(equation = var_5238_equation_0, values = (var_5172_cast_fp16_0, var_5221_cast_fp16))[name = tensor("op_5238_cast_fp16")]; + tensor var_5240_equation_0 = const()[name = tensor("op_5240_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5240_cast_fp16 = einsum(equation = var_5240_equation_0, values = (var_5172_cast_fp16_1, var_5222_cast_fp16))[name = tensor("op_5240_cast_fp16")]; + tensor var_5242_equation_0 = const()[name = tensor("op_5242_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5242_cast_fp16 = einsum(equation = var_5242_equation_0, values = (var_5172_cast_fp16_2, var_5223_cast_fp16))[name = tensor("op_5242_cast_fp16")]; + tensor var_5244_equation_0 = const()[name = tensor("op_5244_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5244_cast_fp16 = einsum(equation = var_5244_equation_0, values = (var_5172_cast_fp16_3, var_5224_cast_fp16))[name = tensor("op_5244_cast_fp16")]; + tensor var_5246_equation_0 = const()[name = tensor("op_5246_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5246_cast_fp16 = einsum(equation = var_5246_equation_0, values = (var_5172_cast_fp16_4, var_5225_cast_fp16))[name = tensor("op_5246_cast_fp16")]; + tensor var_5248_equation_0 = const()[name = tensor("op_5248_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5248_cast_fp16 = einsum(equation = var_5248_equation_0, values = (var_5172_cast_fp16_5, var_5226_cast_fp16))[name = tensor("op_5248_cast_fp16")]; + tensor var_5250_equation_0 = const()[name = tensor("op_5250_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5250_cast_fp16 = einsum(equation = var_5250_equation_0, values = (var_5172_cast_fp16_6, var_5227_cast_fp16))[name = tensor("op_5250_cast_fp16")]; + tensor var_5252_equation_0 = const()[name = tensor("op_5252_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5252_cast_fp16 = einsum(equation = var_5252_equation_0, values = (var_5172_cast_fp16_7, var_5228_cast_fp16))[name = tensor("op_5252_cast_fp16")]; + tensor var_5254_equation_0 = const()[name = tensor("op_5254_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5254_cast_fp16 = einsum(equation = var_5254_equation_0, values = (var_5172_cast_fp16_8, var_5229_cast_fp16))[name = tensor("op_5254_cast_fp16")]; + tensor var_5256_equation_0 = const()[name = tensor("op_5256_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5256_cast_fp16 = einsum(equation = var_5256_equation_0, values = (var_5172_cast_fp16_9, var_5230_cast_fp16))[name = tensor("op_5256_cast_fp16")]; + tensor var_5258_equation_0 = const()[name = tensor("op_5258_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5258_cast_fp16 = einsum(equation = var_5258_equation_0, values = (var_5172_cast_fp16_10, var_5231_cast_fp16))[name = tensor("op_5258_cast_fp16")]; + tensor var_5260_equation_0 = const()[name = tensor("op_5260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5260_cast_fp16 = einsum(equation = var_5260_equation_0, values = (var_5172_cast_fp16_11, var_5232_cast_fp16))[name = tensor("op_5260_cast_fp16")]; + tensor var_5262_equation_0 = const()[name = tensor("op_5262_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5262_cast_fp16 = einsum(equation = var_5262_equation_0, values = (var_5172_cast_fp16_12, var_5233_cast_fp16))[name = tensor("op_5262_cast_fp16")]; + tensor var_5264_equation_0 = const()[name = tensor("op_5264_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5264_cast_fp16 = einsum(equation = var_5264_equation_0, values = (var_5172_cast_fp16_13, var_5234_cast_fp16))[name = tensor("op_5264_cast_fp16")]; + tensor var_5266_equation_0 = const()[name = tensor("op_5266_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5266_cast_fp16 = einsum(equation = var_5266_equation_0, values = (var_5172_cast_fp16_14, var_5235_cast_fp16))[name = tensor("op_5266_cast_fp16")]; + tensor var_5268_equation_0 = const()[name = tensor("op_5268_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5268_cast_fp16 = einsum(equation = var_5268_equation_0, values = (var_5172_cast_fp16_15, var_5236_cast_fp16))[name = tensor("op_5268_cast_fp16")]; + tensor input_215_interleave_0 = const()[name = tensor("input_215_interleave_0"), val = tensor(false)]; + tensor input_215_cast_fp16 = concat(axis = var_5085, interleave = input_215_interleave_0, values = (var_5238_cast_fp16, var_5240_cast_fp16, var_5242_cast_fp16, var_5244_cast_fp16, var_5246_cast_fp16, var_5248_cast_fp16, var_5250_cast_fp16, var_5252_cast_fp16, var_5254_cast_fp16, var_5256_cast_fp16, var_5258_cast_fp16, var_5260_cast_fp16, var_5262_cast_fp16, var_5264_cast_fp16, var_5266_cast_fp16, var_5268_cast_fp16))[name = tensor("input_215_cast_fp16")]; + tensor var_5277_pad_type_0 = const()[name = tensor("op_5277_pad_type_0"), val = tensor("valid")]; + tensor var_5277_strides_0 = const()[name = tensor("op_5277_strides_0"), val = tensor([1, 1])]; + tensor var_5277_pad_0 = const()[name = tensor("op_5277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5277_dilations_0 = const()[name = tensor("op_5277_dilations_0"), val = tensor([1, 1])]; + tensor var_5277_groups_0 = const()[name = tensor("op_5277_groups_0"), val = tensor(1)]; + tensor blocks_21_attn_out_weight_to_fp16 = const()[name = tensor("blocks_21_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545178112)))]; + tensor blocks_21_attn_out_bias_to_fp16 = const()[name = tensor("blocks_21_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547275328)))]; + tensor var_5277_cast_fp16 = conv(bias = blocks_21_attn_out_bias_to_fp16, dilations = var_5277_dilations_0, groups = var_5277_groups_0, pad = var_5277_pad_0, pad_type = var_5277_pad_type_0, strides = var_5277_strides_0, weight = blocks_21_attn_out_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("op_5277_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = var_5277_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor input_217_axes_0 = const()[name = tensor("input_217_axes_0"), val = tensor([1])]; + tensor input_217_gamma_0_to_fp16 = const()[name = tensor("input_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547277440)))]; + tensor input_217_beta_0_to_fp16 = const()[name = tensor("input_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547279552)))]; + tensor var_5287_to_fp16 = const()[name = tensor("op_5287_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_217_cast_fp16 = layer_norm(axes = input_217_axes_0, beta = input_217_beta_0_to_fp16, epsilon = var_5287_to_fp16, gamma = input_217_gamma_0_to_fp16, x = inputs_87_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("valid")]; + tensor input_219_strides_0 = const()[name = tensor("input_219_strides_0"), val = tensor([1, 1])]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_219_dilations_0 = const()[name = tensor("input_219_dilations_0"), val = tensor([1, 1])]; + tensor input_219_groups_0 = const()[name = tensor("input_219_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547281664)))]; + tensor blocks_21_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555670336)))]; + tensor input_219_cast_fp16 = conv(bias = blocks_21_mlp_0_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = blocks_21_mlp_0_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_5313_pad_type_0 = const()[name = tensor("op_5313_pad_type_0"), val = tensor("valid")]; + tensor var_5313_strides_0 = const()[name = tensor("op_5313_strides_0"), val = tensor([1, 1])]; + tensor var_5313_pad_0 = const()[name = tensor("op_5313_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5313_dilations_0 = const()[name = tensor("op_5313_dilations_0"), val = tensor([1, 1])]; + tensor var_5313_groups_0 = const()[name = tensor("op_5313_groups_0"), val = tensor(1)]; + tensor blocks_21_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555678592)))]; + tensor blocks_21_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564067264)))]; + tensor var_5313_cast_fp16 = conv(bias = blocks_21_mlp_2_bias_to_fp16, dilations = var_5313_dilations_0, groups = var_5313_groups_0, pad = var_5313_pad_0, pad_type = var_5313_pad_type_0, strides = var_5313_strides_0, weight = blocks_21_mlp_2_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("op_5313_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_5313_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_5322 = const()[name = tensor("op_5322"), val = tensor(1)]; + tensor input_223_axes_0 = const()[name = tensor("input_223_axes_0"), val = tensor([1])]; + tensor input_223_gamma_0_to_fp16 = const()[name = tensor("input_223_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564069376)))]; + tensor input_223_beta_0_to_fp16 = const()[name = tensor("input_223_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564071488)))]; + tensor var_5338_to_fp16 = const()[name = tensor("op_5338_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_223_cast_fp16 = layer_norm(axes = input_223_axes_0, beta = input_223_beta_0_to_fp16, epsilon = var_5338_to_fp16, gamma = input_223_gamma_0_to_fp16, x = inputs_89_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor q_45_pad_type_0 = const()[name = tensor("q_45_pad_type_0"), val = tensor("valid")]; + tensor q_45_strides_0 = const()[name = tensor("q_45_strides_0"), val = tensor([1, 1])]; + tensor q_45_pad_0 = const()[name = tensor("q_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_45_dilations_0 = const()[name = tensor("q_45_dilations_0"), val = tensor([1, 1])]; + tensor q_45_groups_0 = const()[name = tensor("q_45_groups_0"), val = tensor(1)]; + tensor var_5373_weight_0_to_fp16 = const()[name = tensor("op_5373_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564073600)))]; + tensor var_5373_bias_0_to_fp16 = const()[name = tensor("op_5373_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566170816)))]; + tensor var_5373_cast_fp16 = conv(bias = var_5373_bias_0_to_fp16, dilations = q_45_dilations_0, groups = q_45_groups_0, pad = q_45_pad_0, pad_type = q_45_pad_type_0, strides = q_45_strides_0, weight = var_5373_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("op_5373_cast_fp16")]; + tensor k_45_pad_type_0 = const()[name = tensor("k_45_pad_type_0"), val = tensor("valid")]; + tensor k_45_strides_0 = const()[name = tensor("k_45_strides_0"), val = tensor([1, 1])]; + tensor k_45_pad_0 = const()[name = tensor("k_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_45_dilations_0 = const()[name = tensor("k_45_dilations_0"), val = tensor([1, 1])]; + tensor k_45_groups_0 = const()[name = tensor("k_45_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_key_weight_to_fp16 = const()[name = tensor("blocks_22_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566172928)))]; + tensor k_45_cast_fp16 = conv(dilations = k_45_dilations_0, groups = k_45_groups_0, pad = k_45_pad_0, pad_type = k_45_pad_type_0, strides = k_45_strides_0, weight = blocks_22_attn_key_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor var_5371_pad_type_0 = const()[name = tensor("op_5371_pad_type_0"), val = tensor("valid")]; + tensor var_5371_strides_0 = const()[name = tensor("op_5371_strides_0"), val = tensor([1, 1])]; + tensor var_5371_pad_0 = const()[name = tensor("op_5371_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5371_dilations_0 = const()[name = tensor("op_5371_dilations_0"), val = tensor([1, 1])]; + tensor var_5371_groups_0 = const()[name = tensor("op_5371_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_value_weight_to_fp16 = const()[name = tensor("blocks_22_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568270144)))]; + tensor blocks_22_attn_value_bias_to_fp16 = const()[name = tensor("blocks_22_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570367360)))]; + tensor var_5371_cast_fp16 = conv(bias = blocks_22_attn_value_bias_to_fp16, dilations = var_5371_dilations_0, groups = var_5371_groups_0, pad = var_5371_pad_0, pad_type = var_5371_pad_type_0, strides = var_5371_strides_0, weight = blocks_22_attn_value_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("op_5371_cast_fp16")]; + tensor tile_66 = const()[name = tensor("tile_66"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_cast_fp16_0, tensor var_5374_cast_fp16_1, tensor var_5374_cast_fp16_2, tensor var_5374_cast_fp16_3, tensor var_5374_cast_fp16_4, tensor var_5374_cast_fp16_5, tensor var_5374_cast_fp16_6, tensor var_5374_cast_fp16_7, tensor var_5374_cast_fp16_8, tensor var_5374_cast_fp16_9, tensor var_5374_cast_fp16_10, tensor var_5374_cast_fp16_11, tensor var_5374_cast_fp16_12, tensor var_5374_cast_fp16_13, tensor var_5374_cast_fp16_14, tensor var_5374_cast_fp16_15 = split(axis = var_5374_axis_0, split_sizes = tile_66, x = var_5373_cast_fp16)[name = tensor("op_5374_cast_fp16")]; + tensor var_5391_perm_0 = const()[name = tensor("op_5391_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_67 = const()[name = tensor("tile_67"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5392_axis_0 = const()[name = tensor("op_5392_axis_0"), val = tensor(3)]; + tensor var_5391_cast_fp16 = transpose(perm = var_5391_perm_0, x = k_45_cast_fp16)[name = tensor("transpose_2")]; + tensor var_5392_cast_fp16_0, tensor var_5392_cast_fp16_1, tensor var_5392_cast_fp16_2, tensor var_5392_cast_fp16_3, tensor var_5392_cast_fp16_4, tensor var_5392_cast_fp16_5, tensor var_5392_cast_fp16_6, tensor var_5392_cast_fp16_7, tensor var_5392_cast_fp16_8, tensor var_5392_cast_fp16_9, tensor var_5392_cast_fp16_10, tensor var_5392_cast_fp16_11, tensor var_5392_cast_fp16_12, tensor var_5392_cast_fp16_13, tensor var_5392_cast_fp16_14, tensor var_5392_cast_fp16_15 = split(axis = var_5392_axis_0, split_sizes = tile_67, x = var_5391_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor tile_68 = const()[name = tensor("tile_68"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5409_axis_0 = const()[name = tensor("op_5409_axis_0"), val = tensor(1)]; + tensor var_5409_cast_fp16_0, tensor var_5409_cast_fp16_1, tensor var_5409_cast_fp16_2, tensor var_5409_cast_fp16_3, tensor var_5409_cast_fp16_4, tensor var_5409_cast_fp16_5, tensor var_5409_cast_fp16_6, tensor var_5409_cast_fp16_7, tensor var_5409_cast_fp16_8, tensor var_5409_cast_fp16_9, tensor var_5409_cast_fp16_10, tensor var_5409_cast_fp16_11, tensor var_5409_cast_fp16_12, tensor var_5409_cast_fp16_13, tensor var_5409_cast_fp16_14, tensor var_5409_cast_fp16_15 = split(axis = var_5409_axis_0, split_sizes = tile_68, x = var_5371_cast_fp16)[name = tensor("op_5409_cast_fp16")]; + tensor aw_705_equation_0 = const()[name = tensor("aw_705_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_705_cast_fp16 = einsum(equation = aw_705_equation_0, values = (var_5392_cast_fp16_0, var_5374_cast_fp16_0))[name = tensor("aw_705_cast_fp16")]; + tensor aw_707_equation_0 = const()[name = tensor("aw_707_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_707_cast_fp16 = einsum(equation = aw_707_equation_0, values = (var_5392_cast_fp16_1, var_5374_cast_fp16_1))[name = tensor("aw_707_cast_fp16")]; + tensor aw_709_equation_0 = const()[name = tensor("aw_709_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_709_cast_fp16 = einsum(equation = aw_709_equation_0, values = (var_5392_cast_fp16_2, var_5374_cast_fp16_2))[name = tensor("aw_709_cast_fp16")]; + tensor aw_711_equation_0 = const()[name = tensor("aw_711_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_711_cast_fp16 = einsum(equation = aw_711_equation_0, values = (var_5392_cast_fp16_3, var_5374_cast_fp16_3))[name = tensor("aw_711_cast_fp16")]; + tensor aw_713_equation_0 = const()[name = tensor("aw_713_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_713_cast_fp16 = einsum(equation = aw_713_equation_0, values = (var_5392_cast_fp16_4, var_5374_cast_fp16_4))[name = tensor("aw_713_cast_fp16")]; + tensor aw_715_equation_0 = const()[name = tensor("aw_715_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_715_cast_fp16 = einsum(equation = aw_715_equation_0, values = (var_5392_cast_fp16_5, var_5374_cast_fp16_5))[name = tensor("aw_715_cast_fp16")]; + tensor aw_717_equation_0 = const()[name = tensor("aw_717_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_717_cast_fp16 = einsum(equation = aw_717_equation_0, values = (var_5392_cast_fp16_6, var_5374_cast_fp16_6))[name = tensor("aw_717_cast_fp16")]; + tensor aw_719_equation_0 = const()[name = tensor("aw_719_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_719_cast_fp16 = einsum(equation = aw_719_equation_0, values = (var_5392_cast_fp16_7, var_5374_cast_fp16_7))[name = tensor("aw_719_cast_fp16")]; + tensor aw_721_equation_0 = const()[name = tensor("aw_721_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_721_cast_fp16 = einsum(equation = aw_721_equation_0, values = (var_5392_cast_fp16_8, var_5374_cast_fp16_8))[name = tensor("aw_721_cast_fp16")]; + tensor aw_723_equation_0 = const()[name = tensor("aw_723_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_723_cast_fp16 = einsum(equation = aw_723_equation_0, values = (var_5392_cast_fp16_9, var_5374_cast_fp16_9))[name = tensor("aw_723_cast_fp16")]; + tensor aw_725_equation_0 = const()[name = tensor("aw_725_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_725_cast_fp16 = einsum(equation = aw_725_equation_0, values = (var_5392_cast_fp16_10, var_5374_cast_fp16_10))[name = tensor("aw_725_cast_fp16")]; + tensor aw_727_equation_0 = const()[name = tensor("aw_727_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_727_cast_fp16 = einsum(equation = aw_727_equation_0, values = (var_5392_cast_fp16_11, var_5374_cast_fp16_11))[name = tensor("aw_727_cast_fp16")]; + tensor aw_729_equation_0 = const()[name = tensor("aw_729_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_729_cast_fp16 = einsum(equation = aw_729_equation_0, values = (var_5392_cast_fp16_12, var_5374_cast_fp16_12))[name = tensor("aw_729_cast_fp16")]; + tensor aw_731_equation_0 = const()[name = tensor("aw_731_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_731_cast_fp16 = einsum(equation = aw_731_equation_0, values = (var_5392_cast_fp16_13, var_5374_cast_fp16_13))[name = tensor("aw_731_cast_fp16")]; + tensor aw_733_equation_0 = const()[name = tensor("aw_733_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_733_cast_fp16 = einsum(equation = aw_733_equation_0, values = (var_5392_cast_fp16_14, var_5374_cast_fp16_14))[name = tensor("aw_733_cast_fp16")]; + tensor aw_735_equation_0 = const()[name = tensor("aw_735_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_735_cast_fp16 = einsum(equation = aw_735_equation_0, values = (var_5392_cast_fp16_15, var_5374_cast_fp16_15))[name = tensor("aw_735_cast_fp16")]; + tensor var_5458_cast_fp16 = softmax(axis = var_5322, x = aw_705_cast_fp16)[name = tensor("op_5458_cast_fp16")]; + tensor var_5459_cast_fp16 = softmax(axis = var_5322, x = aw_707_cast_fp16)[name = tensor("op_5459_cast_fp16")]; + tensor var_5460_cast_fp16 = softmax(axis = var_5322, x = aw_709_cast_fp16)[name = tensor("op_5460_cast_fp16")]; + tensor var_5461_cast_fp16 = softmax(axis = var_5322, x = aw_711_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5462_cast_fp16 = softmax(axis = var_5322, x = aw_713_cast_fp16)[name = tensor("op_5462_cast_fp16")]; + tensor var_5463_cast_fp16 = softmax(axis = var_5322, x = aw_715_cast_fp16)[name = tensor("op_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = softmax(axis = var_5322, x = aw_717_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_cast_fp16 = softmax(axis = var_5322, x = aw_719_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = softmax(axis = var_5322, x = aw_721_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor var_5467_cast_fp16 = softmax(axis = var_5322, x = aw_723_cast_fp16)[name = tensor("op_5467_cast_fp16")]; + tensor var_5468_cast_fp16 = softmax(axis = var_5322, x = aw_725_cast_fp16)[name = tensor("op_5468_cast_fp16")]; + tensor var_5469_cast_fp16 = softmax(axis = var_5322, x = aw_727_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = softmax(axis = var_5322, x = aw_729_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor var_5471_cast_fp16 = softmax(axis = var_5322, x = aw_731_cast_fp16)[name = tensor("op_5471_cast_fp16")]; + tensor var_5472_cast_fp16 = softmax(axis = var_5322, x = aw_733_cast_fp16)[name = tensor("op_5472_cast_fp16")]; + tensor var_5473_cast_fp16 = softmax(axis = var_5322, x = aw_735_cast_fp16)[name = tensor("op_5473_cast_fp16")]; + tensor var_5475_equation_0 = const()[name = tensor("op_5475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5475_cast_fp16 = einsum(equation = var_5475_equation_0, values = (var_5409_cast_fp16_0, var_5458_cast_fp16))[name = tensor("op_5475_cast_fp16")]; + tensor var_5477_equation_0 = const()[name = tensor("op_5477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5477_cast_fp16 = einsum(equation = var_5477_equation_0, values = (var_5409_cast_fp16_1, var_5459_cast_fp16))[name = tensor("op_5477_cast_fp16")]; + tensor var_5479_equation_0 = const()[name = tensor("op_5479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5479_cast_fp16 = einsum(equation = var_5479_equation_0, values = (var_5409_cast_fp16_2, var_5460_cast_fp16))[name = tensor("op_5479_cast_fp16")]; + tensor var_5481_equation_0 = const()[name = tensor("op_5481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5481_cast_fp16 = einsum(equation = var_5481_equation_0, values = (var_5409_cast_fp16_3, var_5461_cast_fp16))[name = tensor("op_5481_cast_fp16")]; + tensor var_5483_equation_0 = const()[name = tensor("op_5483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5483_cast_fp16 = einsum(equation = var_5483_equation_0, values = (var_5409_cast_fp16_4, var_5462_cast_fp16))[name = tensor("op_5483_cast_fp16")]; + tensor var_5485_equation_0 = const()[name = tensor("op_5485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5485_cast_fp16 = einsum(equation = var_5485_equation_0, values = (var_5409_cast_fp16_5, var_5463_cast_fp16))[name = tensor("op_5485_cast_fp16")]; + tensor var_5487_equation_0 = const()[name = tensor("op_5487_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5487_cast_fp16 = einsum(equation = var_5487_equation_0, values = (var_5409_cast_fp16_6, var_5464_cast_fp16))[name = tensor("op_5487_cast_fp16")]; + tensor var_5489_equation_0 = const()[name = tensor("op_5489_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5489_cast_fp16 = einsum(equation = var_5489_equation_0, values = (var_5409_cast_fp16_7, var_5465_cast_fp16))[name = tensor("op_5489_cast_fp16")]; + tensor var_5491_equation_0 = const()[name = tensor("op_5491_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5491_cast_fp16 = einsum(equation = var_5491_equation_0, values = (var_5409_cast_fp16_8, var_5466_cast_fp16))[name = tensor("op_5491_cast_fp16")]; + tensor var_5493_equation_0 = const()[name = tensor("op_5493_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5493_cast_fp16 = einsum(equation = var_5493_equation_0, values = (var_5409_cast_fp16_9, var_5467_cast_fp16))[name = tensor("op_5493_cast_fp16")]; + tensor var_5495_equation_0 = const()[name = tensor("op_5495_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5495_cast_fp16 = einsum(equation = var_5495_equation_0, values = (var_5409_cast_fp16_10, var_5468_cast_fp16))[name = tensor("op_5495_cast_fp16")]; + tensor var_5497_equation_0 = const()[name = tensor("op_5497_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5497_cast_fp16 = einsum(equation = var_5497_equation_0, values = (var_5409_cast_fp16_11, var_5469_cast_fp16))[name = tensor("op_5497_cast_fp16")]; + tensor var_5499_equation_0 = const()[name = tensor("op_5499_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5499_cast_fp16 = einsum(equation = var_5499_equation_0, values = (var_5409_cast_fp16_12, var_5470_cast_fp16))[name = tensor("op_5499_cast_fp16")]; + tensor var_5501_equation_0 = const()[name = tensor("op_5501_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5501_cast_fp16 = einsum(equation = var_5501_equation_0, values = (var_5409_cast_fp16_13, var_5471_cast_fp16))[name = tensor("op_5501_cast_fp16")]; + tensor var_5503_equation_0 = const()[name = tensor("op_5503_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5503_cast_fp16 = einsum(equation = var_5503_equation_0, values = (var_5409_cast_fp16_14, var_5472_cast_fp16))[name = tensor("op_5503_cast_fp16")]; + tensor var_5505_equation_0 = const()[name = tensor("op_5505_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5505_cast_fp16 = einsum(equation = var_5505_equation_0, values = (var_5409_cast_fp16_15, var_5473_cast_fp16))[name = tensor("op_5505_cast_fp16")]; + tensor input_225_interleave_0 = const()[name = tensor("input_225_interleave_0"), val = tensor(false)]; + tensor input_225_cast_fp16 = concat(axis = var_5322, interleave = input_225_interleave_0, values = (var_5475_cast_fp16, var_5477_cast_fp16, var_5479_cast_fp16, var_5481_cast_fp16, var_5483_cast_fp16, var_5485_cast_fp16, var_5487_cast_fp16, var_5489_cast_fp16, var_5491_cast_fp16, var_5493_cast_fp16, var_5495_cast_fp16, var_5497_cast_fp16, var_5499_cast_fp16, var_5501_cast_fp16, var_5503_cast_fp16, var_5505_cast_fp16))[name = tensor("input_225_cast_fp16")]; + tensor var_5514_pad_type_0 = const()[name = tensor("op_5514_pad_type_0"), val = tensor("valid")]; + tensor var_5514_strides_0 = const()[name = tensor("op_5514_strides_0"), val = tensor([1, 1])]; + tensor var_5514_pad_0 = const()[name = tensor("op_5514_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5514_dilations_0 = const()[name = tensor("op_5514_dilations_0"), val = tensor([1, 1])]; + tensor var_5514_groups_0 = const()[name = tensor("op_5514_groups_0"), val = tensor(1)]; + tensor blocks_22_attn_out_weight_to_fp16 = const()[name = tensor("blocks_22_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570369472)))]; + tensor blocks_22_attn_out_bias_to_fp16 = const()[name = tensor("blocks_22_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572466688)))]; + tensor var_5514_cast_fp16 = conv(bias = blocks_22_attn_out_bias_to_fp16, dilations = var_5514_dilations_0, groups = var_5514_groups_0, pad = var_5514_pad_0, pad_type = var_5514_pad_type_0, strides = var_5514_strides_0, weight = blocks_22_attn_out_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = var_5514_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([1])]; + tensor input_227_gamma_0_to_fp16 = const()[name = tensor("input_227_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572468800)))]; + tensor input_227_beta_0_to_fp16 = const()[name = tensor("input_227_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572470912)))]; + tensor var_5524_to_fp16 = const()[name = tensor("op_5524_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = input_227_beta_0_to_fp16, epsilon = var_5524_to_fp16, gamma = input_227_gamma_0_to_fp16, x = inputs_91_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("valid")]; + tensor input_229_strides_0 = const()[name = tensor("input_229_strides_0"), val = tensor([1, 1])]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_229_dilations_0 = const()[name = tensor("input_229_dilations_0"), val = tensor([1, 1])]; + tensor input_229_groups_0 = const()[name = tensor("input_229_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572473024)))]; + tensor blocks_22_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580861696)))]; + tensor input_229_cast_fp16 = conv(bias = blocks_22_mlp_0_bias_to_fp16, dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = blocks_22_mlp_0_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_5550_pad_type_0 = const()[name = tensor("op_5550_pad_type_0"), val = tensor("valid")]; + tensor var_5550_strides_0 = const()[name = tensor("op_5550_strides_0"), val = tensor([1, 1])]; + tensor var_5550_pad_0 = const()[name = tensor("op_5550_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5550_dilations_0 = const()[name = tensor("op_5550_dilations_0"), val = tensor([1, 1])]; + tensor var_5550_groups_0 = const()[name = tensor("op_5550_groups_0"), val = tensor(1)]; + tensor blocks_22_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580869952)))]; + tensor blocks_22_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589258624)))]; + tensor var_5550_cast_fp16 = conv(bias = blocks_22_mlp_2_bias_to_fp16, dilations = var_5550_dilations_0, groups = var_5550_groups_0, pad = var_5550_pad_0, pad_type = var_5550_pad_type_0, strides = var_5550_strides_0, weight = blocks_22_mlp_2_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("op_5550_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_5550_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_5559 = const()[name = tensor("op_5559"), val = tensor(1)]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([1])]; + tensor input_233_gamma_0_to_fp16 = const()[name = tensor("input_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589260736)))]; + tensor input_233_beta_0_to_fp16 = const()[name = tensor("input_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589262848)))]; + tensor var_5575_to_fp16 = const()[name = tensor("op_5575_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = input_233_beta_0_to_fp16, epsilon = var_5575_to_fp16, gamma = input_233_gamma_0_to_fp16, x = inputs_93_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_5610_weight_0_to_fp16 = const()[name = tensor("op_5610_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589264960)))]; + tensor var_5610_bias_0_to_fp16 = const()[name = tensor("op_5610_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591362176)))]; + tensor var_5610_cast_fp16 = conv(bias = var_5610_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_5610_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("op_5610_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_key_weight_to_fp16 = const()[name = tensor("blocks_23_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591364288)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_23_attn_key_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_5608_pad_type_0 = const()[name = tensor("op_5608_pad_type_0"), val = tensor("valid")]; + tensor var_5608_strides_0 = const()[name = tensor("op_5608_strides_0"), val = tensor([1, 1])]; + tensor var_5608_pad_0 = const()[name = tensor("op_5608_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5608_dilations_0 = const()[name = tensor("op_5608_dilations_0"), val = tensor([1, 1])]; + tensor var_5608_groups_0 = const()[name = tensor("op_5608_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_value_weight_to_fp16 = const()[name = tensor("blocks_23_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593461504)))]; + tensor blocks_23_attn_value_bias_to_fp16 = const()[name = tensor("blocks_23_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595558720)))]; + tensor var_5608_cast_fp16 = conv(bias = blocks_23_attn_value_bias_to_fp16, dilations = var_5608_dilations_0, groups = var_5608_groups_0, pad = var_5608_pad_0, pad_type = var_5608_pad_type_0, strides = var_5608_strides_0, weight = blocks_23_attn_value_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor tile_69 = const()[name = tensor("tile_69"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5611_axis_0 = const()[name = tensor("op_5611_axis_0"), val = tensor(1)]; + tensor var_5611_cast_fp16_0, tensor var_5611_cast_fp16_1, tensor var_5611_cast_fp16_2, tensor var_5611_cast_fp16_3, tensor var_5611_cast_fp16_4, tensor var_5611_cast_fp16_5, tensor var_5611_cast_fp16_6, tensor var_5611_cast_fp16_7, tensor var_5611_cast_fp16_8, tensor var_5611_cast_fp16_9, tensor var_5611_cast_fp16_10, tensor var_5611_cast_fp16_11, tensor var_5611_cast_fp16_12, tensor var_5611_cast_fp16_13, tensor var_5611_cast_fp16_14, tensor var_5611_cast_fp16_15 = split(axis = var_5611_axis_0, split_sizes = tile_69, x = var_5610_cast_fp16)[name = tensor("op_5611_cast_fp16")]; + tensor var_5628_perm_0 = const()[name = tensor("op_5628_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_70 = const()[name = tensor("tile_70"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5629_axis_0 = const()[name = tensor("op_5629_axis_0"), val = tensor(3)]; + tensor var_5628_cast_fp16 = transpose(perm = var_5628_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_5629_cast_fp16_0, tensor var_5629_cast_fp16_1, tensor var_5629_cast_fp16_2, tensor var_5629_cast_fp16_3, tensor var_5629_cast_fp16_4, tensor var_5629_cast_fp16_5, tensor var_5629_cast_fp16_6, tensor var_5629_cast_fp16_7, tensor var_5629_cast_fp16_8, tensor var_5629_cast_fp16_9, tensor var_5629_cast_fp16_10, tensor var_5629_cast_fp16_11, tensor var_5629_cast_fp16_12, tensor var_5629_cast_fp16_13, tensor var_5629_cast_fp16_14, tensor var_5629_cast_fp16_15 = split(axis = var_5629_axis_0, split_sizes = tile_70, x = var_5628_cast_fp16)[name = tensor("op_5629_cast_fp16")]; + tensor tile_71 = const()[name = tensor("tile_71"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_5646_axis_0 = const()[name = tensor("op_5646_axis_0"), val = tensor(1)]; + tensor var_5646_cast_fp16_0, tensor var_5646_cast_fp16_1, tensor var_5646_cast_fp16_2, tensor var_5646_cast_fp16_3, tensor var_5646_cast_fp16_4, tensor var_5646_cast_fp16_5, tensor var_5646_cast_fp16_6, tensor var_5646_cast_fp16_7, tensor var_5646_cast_fp16_8, tensor var_5646_cast_fp16_9, tensor var_5646_cast_fp16_10, tensor var_5646_cast_fp16_11, tensor var_5646_cast_fp16_12, tensor var_5646_cast_fp16_13, tensor var_5646_cast_fp16_14, tensor var_5646_cast_fp16_15 = split(axis = var_5646_axis_0, split_sizes = tile_71, x = var_5608_cast_fp16)[name = tensor("op_5646_cast_fp16")]; + tensor aw_737_equation_0 = const()[name = tensor("aw_737_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_737_cast_fp16 = einsum(equation = aw_737_equation_0, values = (var_5629_cast_fp16_0, var_5611_cast_fp16_0))[name = tensor("aw_737_cast_fp16")]; + tensor aw_739_equation_0 = const()[name = tensor("aw_739_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_739_cast_fp16 = einsum(equation = aw_739_equation_0, values = (var_5629_cast_fp16_1, var_5611_cast_fp16_1))[name = tensor("aw_739_cast_fp16")]; + tensor aw_741_equation_0 = const()[name = tensor("aw_741_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_741_cast_fp16 = einsum(equation = aw_741_equation_0, values = (var_5629_cast_fp16_2, var_5611_cast_fp16_2))[name = tensor("aw_741_cast_fp16")]; + tensor aw_743_equation_0 = const()[name = tensor("aw_743_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_743_cast_fp16 = einsum(equation = aw_743_equation_0, values = (var_5629_cast_fp16_3, var_5611_cast_fp16_3))[name = tensor("aw_743_cast_fp16")]; + tensor aw_745_equation_0 = const()[name = tensor("aw_745_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_745_cast_fp16 = einsum(equation = aw_745_equation_0, values = (var_5629_cast_fp16_4, var_5611_cast_fp16_4))[name = tensor("aw_745_cast_fp16")]; + tensor aw_747_equation_0 = const()[name = tensor("aw_747_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_747_cast_fp16 = einsum(equation = aw_747_equation_0, values = (var_5629_cast_fp16_5, var_5611_cast_fp16_5))[name = tensor("aw_747_cast_fp16")]; + tensor aw_749_equation_0 = const()[name = tensor("aw_749_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_749_cast_fp16 = einsum(equation = aw_749_equation_0, values = (var_5629_cast_fp16_6, var_5611_cast_fp16_6))[name = tensor("aw_749_cast_fp16")]; + tensor aw_751_equation_0 = const()[name = tensor("aw_751_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_751_cast_fp16 = einsum(equation = aw_751_equation_0, values = (var_5629_cast_fp16_7, var_5611_cast_fp16_7))[name = tensor("aw_751_cast_fp16")]; + tensor aw_753_equation_0 = const()[name = tensor("aw_753_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_753_cast_fp16 = einsum(equation = aw_753_equation_0, values = (var_5629_cast_fp16_8, var_5611_cast_fp16_8))[name = tensor("aw_753_cast_fp16")]; + tensor aw_755_equation_0 = const()[name = tensor("aw_755_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_755_cast_fp16 = einsum(equation = aw_755_equation_0, values = (var_5629_cast_fp16_9, var_5611_cast_fp16_9))[name = tensor("aw_755_cast_fp16")]; + tensor aw_757_equation_0 = const()[name = tensor("aw_757_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_757_cast_fp16 = einsum(equation = aw_757_equation_0, values = (var_5629_cast_fp16_10, var_5611_cast_fp16_10))[name = tensor("aw_757_cast_fp16")]; + tensor aw_759_equation_0 = const()[name = tensor("aw_759_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_759_cast_fp16 = einsum(equation = aw_759_equation_0, values = (var_5629_cast_fp16_11, var_5611_cast_fp16_11))[name = tensor("aw_759_cast_fp16")]; + tensor aw_761_equation_0 = const()[name = tensor("aw_761_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_761_cast_fp16 = einsum(equation = aw_761_equation_0, values = (var_5629_cast_fp16_12, var_5611_cast_fp16_12))[name = tensor("aw_761_cast_fp16")]; + tensor aw_763_equation_0 = const()[name = tensor("aw_763_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_763_cast_fp16 = einsum(equation = aw_763_equation_0, values = (var_5629_cast_fp16_13, var_5611_cast_fp16_13))[name = tensor("aw_763_cast_fp16")]; + tensor aw_765_equation_0 = const()[name = tensor("aw_765_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_765_cast_fp16 = einsum(equation = aw_765_equation_0, values = (var_5629_cast_fp16_14, var_5611_cast_fp16_14))[name = tensor("aw_765_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_5629_cast_fp16_15, var_5611_cast_fp16_15))[name = tensor("aw_cast_fp16")]; + tensor var_5695_cast_fp16 = softmax(axis = var_5559, x = aw_737_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor var_5696_cast_fp16 = softmax(axis = var_5559, x = aw_739_cast_fp16)[name = tensor("op_5696_cast_fp16")]; + tensor var_5697_cast_fp16 = softmax(axis = var_5559, x = aw_741_cast_fp16)[name = tensor("op_5697_cast_fp16")]; + tensor var_5698_cast_fp16 = softmax(axis = var_5559, x = aw_743_cast_fp16)[name = tensor("op_5698_cast_fp16")]; + tensor var_5699_cast_fp16 = softmax(axis = var_5559, x = aw_745_cast_fp16)[name = tensor("op_5699_cast_fp16")]; + tensor var_5700_cast_fp16 = softmax(axis = var_5559, x = aw_747_cast_fp16)[name = tensor("op_5700_cast_fp16")]; + tensor var_5701_cast_fp16 = softmax(axis = var_5559, x = aw_749_cast_fp16)[name = tensor("op_5701_cast_fp16")]; + tensor var_5702_cast_fp16 = softmax(axis = var_5559, x = aw_751_cast_fp16)[name = tensor("op_5702_cast_fp16")]; + tensor var_5703_cast_fp16 = softmax(axis = var_5559, x = aw_753_cast_fp16)[name = tensor("op_5703_cast_fp16")]; + tensor var_5704_cast_fp16 = softmax(axis = var_5559, x = aw_755_cast_fp16)[name = tensor("op_5704_cast_fp16")]; + tensor var_5705_cast_fp16 = softmax(axis = var_5559, x = aw_757_cast_fp16)[name = tensor("op_5705_cast_fp16")]; + tensor var_5706_cast_fp16 = softmax(axis = var_5559, x = aw_759_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5707_cast_fp16 = softmax(axis = var_5559, x = aw_761_cast_fp16)[name = tensor("op_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = softmax(axis = var_5559, x = aw_763_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_cast_fp16 = softmax(axis = var_5559, x = aw_765_cast_fp16)[name = tensor("op_5709_cast_fp16")]; + tensor var_5710_cast_fp16 = softmax(axis = var_5559, x = aw_cast_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor var_5712_equation_0 = const()[name = tensor("op_5712_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5712_cast_fp16 = einsum(equation = var_5712_equation_0, values = (var_5646_cast_fp16_0, var_5695_cast_fp16))[name = tensor("op_5712_cast_fp16")]; + tensor var_5714_equation_0 = const()[name = tensor("op_5714_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5714_cast_fp16 = einsum(equation = var_5714_equation_0, values = (var_5646_cast_fp16_1, var_5696_cast_fp16))[name = tensor("op_5714_cast_fp16")]; + tensor var_5716_equation_0 = const()[name = tensor("op_5716_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5716_cast_fp16 = einsum(equation = var_5716_equation_0, values = (var_5646_cast_fp16_2, var_5697_cast_fp16))[name = tensor("op_5716_cast_fp16")]; + tensor var_5718_equation_0 = const()[name = tensor("op_5718_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5718_cast_fp16 = einsum(equation = var_5718_equation_0, values = (var_5646_cast_fp16_3, var_5698_cast_fp16))[name = tensor("op_5718_cast_fp16")]; + tensor var_5720_equation_0 = const()[name = tensor("op_5720_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5720_cast_fp16 = einsum(equation = var_5720_equation_0, values = (var_5646_cast_fp16_4, var_5699_cast_fp16))[name = tensor("op_5720_cast_fp16")]; + tensor var_5722_equation_0 = const()[name = tensor("op_5722_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5722_cast_fp16 = einsum(equation = var_5722_equation_0, values = (var_5646_cast_fp16_5, var_5700_cast_fp16))[name = tensor("op_5722_cast_fp16")]; + tensor var_5724_equation_0 = const()[name = tensor("op_5724_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5724_cast_fp16 = einsum(equation = var_5724_equation_0, values = (var_5646_cast_fp16_6, var_5701_cast_fp16))[name = tensor("op_5724_cast_fp16")]; + tensor var_5726_equation_0 = const()[name = tensor("op_5726_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5726_cast_fp16 = einsum(equation = var_5726_equation_0, values = (var_5646_cast_fp16_7, var_5702_cast_fp16))[name = tensor("op_5726_cast_fp16")]; + tensor var_5728_equation_0 = const()[name = tensor("op_5728_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5728_cast_fp16 = einsum(equation = var_5728_equation_0, values = (var_5646_cast_fp16_8, var_5703_cast_fp16))[name = tensor("op_5728_cast_fp16")]; + tensor var_5730_equation_0 = const()[name = tensor("op_5730_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5730_cast_fp16 = einsum(equation = var_5730_equation_0, values = (var_5646_cast_fp16_9, var_5704_cast_fp16))[name = tensor("op_5730_cast_fp16")]; + tensor var_5732_equation_0 = const()[name = tensor("op_5732_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5732_cast_fp16 = einsum(equation = var_5732_equation_0, values = (var_5646_cast_fp16_10, var_5705_cast_fp16))[name = tensor("op_5732_cast_fp16")]; + tensor var_5734_equation_0 = const()[name = tensor("op_5734_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5734_cast_fp16 = einsum(equation = var_5734_equation_0, values = (var_5646_cast_fp16_11, var_5706_cast_fp16))[name = tensor("op_5734_cast_fp16")]; + tensor var_5736_equation_0 = const()[name = tensor("op_5736_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5736_cast_fp16 = einsum(equation = var_5736_equation_0, values = (var_5646_cast_fp16_12, var_5707_cast_fp16))[name = tensor("op_5736_cast_fp16")]; + tensor var_5738_equation_0 = const()[name = tensor("op_5738_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5738_cast_fp16 = einsum(equation = var_5738_equation_0, values = (var_5646_cast_fp16_13, var_5708_cast_fp16))[name = tensor("op_5738_cast_fp16")]; + tensor var_5740_equation_0 = const()[name = tensor("op_5740_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5740_cast_fp16 = einsum(equation = var_5740_equation_0, values = (var_5646_cast_fp16_14, var_5709_cast_fp16))[name = tensor("op_5740_cast_fp16")]; + tensor var_5742_equation_0 = const()[name = tensor("op_5742_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_5742_cast_fp16 = einsum(equation = var_5742_equation_0, values = (var_5646_cast_fp16_15, var_5710_cast_fp16))[name = tensor("op_5742_cast_fp16")]; + tensor input_235_interleave_0 = const()[name = tensor("input_235_interleave_0"), val = tensor(false)]; + tensor input_235_cast_fp16 = concat(axis = var_5559, interleave = input_235_interleave_0, values = (var_5712_cast_fp16, var_5714_cast_fp16, var_5716_cast_fp16, var_5718_cast_fp16, var_5720_cast_fp16, var_5722_cast_fp16, var_5724_cast_fp16, var_5726_cast_fp16, var_5728_cast_fp16, var_5730_cast_fp16, var_5732_cast_fp16, var_5734_cast_fp16, var_5736_cast_fp16, var_5738_cast_fp16, var_5740_cast_fp16, var_5742_cast_fp16))[name = tensor("input_235_cast_fp16")]; + tensor var_5751_pad_type_0 = const()[name = tensor("op_5751_pad_type_0"), val = tensor("valid")]; + tensor var_5751_strides_0 = const()[name = tensor("op_5751_strides_0"), val = tensor([1, 1])]; + tensor var_5751_pad_0 = const()[name = tensor("op_5751_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5751_dilations_0 = const()[name = tensor("op_5751_dilations_0"), val = tensor([1, 1])]; + tensor var_5751_groups_0 = const()[name = tensor("op_5751_groups_0"), val = tensor(1)]; + tensor blocks_23_attn_out_weight_to_fp16 = const()[name = tensor("blocks_23_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595560832)))]; + tensor blocks_23_attn_out_bias_to_fp16 = const()[name = tensor("blocks_23_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597658048)))]; + tensor var_5751_cast_fp16 = conv(bias = blocks_23_attn_out_bias_to_fp16, dilations = var_5751_dilations_0, groups = var_5751_groups_0, pad = var_5751_pad_0, pad_type = var_5751_pad_type_0, strides = var_5751_strides_0, weight = blocks_23_attn_out_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("op_5751_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = var_5751_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([1])]; + tensor input_237_gamma_0_to_fp16 = const()[name = tensor("input_237_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597660160)))]; + tensor input_237_beta_0_to_fp16 = const()[name = tensor("input_237_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597662272)))]; + tensor var_5761_to_fp16 = const()[name = tensor("op_5761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = input_237_beta_0_to_fp16, epsilon = var_5761_to_fp16, gamma = input_237_gamma_0_to_fp16, x = inputs_95_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("valid")]; + tensor input_239_strides_0 = const()[name = tensor("input_239_strides_0"), val = tensor([1, 1])]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_239_dilations_0 = const()[name = tensor("input_239_dilations_0"), val = tensor([1, 1])]; + tensor input_239_groups_0 = const()[name = tensor("input_239_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597664384)))]; + tensor blocks_23_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606053056)))]; + tensor input_239_cast_fp16 = conv(bias = blocks_23_mlp_0_bias_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = blocks_23_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_239_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_5787_pad_type_0 = const()[name = tensor("op_5787_pad_type_0"), val = tensor("valid")]; + tensor var_5787_strides_0 = const()[name = tensor("op_5787_strides_0"), val = tensor([1, 1])]; + tensor var_5787_pad_0 = const()[name = tensor("op_5787_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5787_dilations_0 = const()[name = tensor("op_5787_dilations_0"), val = tensor([1, 1])]; + tensor var_5787_groups_0 = const()[name = tensor("op_5787_groups_0"), val = tensor(1)]; + tensor blocks_23_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606061312)))]; + tensor blocks_23_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614449984)))]; + tensor var_5787_cast_fp16 = conv(bias = blocks_23_mlp_2_bias_to_fp16, dilations = var_5787_dilations_0, groups = var_5787_groups_0, pad = var_5787_pad_0, pad_type = var_5787_pad_type_0, strides = var_5787_strides_0, weight = blocks_23_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_95_cast_fp16, y = var_5787_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614452096)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614454208)))]; + tensor var_5801_to_fp16 = const()[name = tensor("op_5801_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_5801_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_5812_axes_0 = const()[name = tensor("op_5812_axes_0"), val = tensor([2])]; + tensor var_5812_cast_fp16 = squeeze(axes = var_5812_axes_0, x = x_cast_fp16)[name = tensor("op_5812_cast_fp16")]; + tensor var_5815_perm_0 = const()[name = tensor("op_5815_perm_0"), val = tensor([0, 2, 1])]; + tensor var_5815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5815_cast_fp16 = transpose(perm = var_5815_perm_0, x = var_5812_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_5815_cast_fp16_to_fp32_dtype_0, x = var_5815_cast_fp16)[name = tensor("cast_99")]; + } -> (output); +} \ No newline at end of file diff --git a/medium/ggml-medium-encoder.mlmodelc/weights/weight.bin b/medium/ggml-medium-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..37209ce1111af094d83ca044d0079af339fc9142 --- /dev/null +++ b/medium/ggml-medium-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2972798300feedfddbf8f3f898b8492320d538e8ecddebbfdf0d03a82e00b83 +size 614456320 diff --git a/medium/ggml-medium.bin b/medium/ggml-medium.bin new file mode 100644 index 0000000000000000000000000000000000000000..be775c464024bdcd9e834c2f9ed42aae6b708d2b --- /dev/null +++ b/medium/ggml-medium.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c14d5adee5f86394037b4e4e8b59f1673b6cee10e3cf0b11bbdbee79c156208 +size 1533763059 diff --git a/small.en/.DS_Store b/small.en/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..3ed91b3729b5e763cfdf15a9dc0a94436869230e Binary files /dev/null and b/small.en/.DS_Store differ diff --git a/small.en/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin b/small.en/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b15b26f7f6a3c8a8718cc69c7c4ff81a9990881e --- /dev/null +++ b/small.en/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ea4d5727e7b1c8762c548f4b80a30447740200e986075b056cf3da16359da56 +size 243 diff --git a/small.en/ggml-small.en-encoder.mlmodelc/coremldata.bin b/small.en/ggml-small.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2aecfa7c477de9fb63dc789d720d5e2fb0e72d4a --- /dev/null +++ b/small.en/ggml-small.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a9554f8fae587d40c808e5a6411d79ce20ab6b473d7a3a88df2cd8123978fb4 +size 320 diff --git a/small.en/ggml-small.en-encoder.mlmodelc/metadata.json b/small.en/ggml-small.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..63a51cf95f700377fe5aa4a5285ac15eb1d85214 --- /dev/null +++ b/small.en/ggml-small.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 768)", + "shortDescription" : "", + "shape" : "[1, 1500, 768]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 12, + "Gelu" : 14, + "LayerNorm" : 25, + "Transpose" : 13, + "Softmax" : 144, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 25, + "Einsum" : 288, + "ExpandDims" : 1, + "Split" : 36, + "Conv" : 74 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.source" : "torch==2.2.2" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_small_en", + "method" : "predict" + } +] \ No newline at end of file diff --git a/small.en/ggml-small.en-encoder.mlmodelc/model.mil b/small.en/ggml-small.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0c32f029586d9600dccb30b0eeece3bf02323585 --- /dev/null +++ b/small.en/ggml-small.en-encoder.mlmodelc/model.mil @@ -0,0 +1,1663 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_44_pad_type_0 = const()[name = tensor("op_44_pad_type_0"), val = tensor("custom")]; + tensor var_44_pad_0 = const()[name = tensor("op_44_pad_0"), val = tensor([1, 1])]; + tensor var_44_strides_0 = const()[name = tensor("op_44_strides_0"), val = tensor([1])]; + tensor var_44_dilations_0 = const()[name = tensor("op_44_dilations_0"), val = tensor([1])]; + tensor var_44_groups_0 = const()[name = tensor("op_44_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368768)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_52")]; + tensor var_44_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_44_dilations_0, groups = var_44_groups_0, pad = var_44_pad_0, pad_type = var_44_pad_type_0, strides = var_44_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_44_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_44_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_62_pad_type_0 = const()[name = tensor("op_62_pad_type_0"), val = tensor("custom")]; + tensor var_62_pad_0 = const()[name = tensor("op_62_pad_0"), val = tensor([1, 1])]; + tensor var_62_strides_0 = const()[name = tensor("op_62_strides_0"), val = tensor([2])]; + tensor var_62_dilations_0 = const()[name = tensor("op_62_dilations_0"), val = tensor([1])]; + tensor var_62_groups_0 = const()[name = tensor("op_62_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370368)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3909376)))]; + tensor var_62_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_62_dilations_0, groups = var_62_groups_0, pad = var_62_pad_0, pad_type = var_62_pad_type_0, strides = var_62_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_62_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_62_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_67_to_fp16 = const()[name = tensor("op_67_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3910976)))]; + tensor var_69_cast_fp16 = add(x = x_3_cast_fp16, y = var_67_to_fp16)[name = tensor("op_69_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_69_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6215040)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6216640)))]; + tensor var_100_to_fp16 = const()[name = tensor("op_100_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_100_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_135_weight_0_to_fp16 = const()[name = tensor("op_135_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6218240)))]; + tensor var_135_bias_0_to_fp16 = const()[name = tensor("op_135_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397952)))]; + tensor var_135_cast_fp16 = conv(bias = var_135_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_135_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_135_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7399552)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_133_pad_type_0 = const()[name = tensor("op_133_pad_type_0"), val = tensor("valid")]; + tensor var_133_strides_0 = const()[name = tensor("op_133_strides_0"), val = tensor([1, 1])]; + tensor var_133_pad_0 = const()[name = tensor("op_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_133_dilations_0 = const()[name = tensor("op_133_dilations_0"), val = tensor([1, 1])]; + tensor var_133_groups_0 = const()[name = tensor("op_133_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8579264)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9758976)))]; + tensor var_133_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_133_dilations_0, groups = var_133_groups_0, pad = var_133_pad_0, pad_type = var_133_pad_type_0, strides = var_133_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_133_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_136_axis_0 = const()[name = tensor("op_136_axis_0"), val = tensor(1)]; + tensor var_136_cast_fp16_0, tensor var_136_cast_fp16_1, tensor var_136_cast_fp16_2, tensor var_136_cast_fp16_3, tensor var_136_cast_fp16_4, tensor var_136_cast_fp16_5, tensor var_136_cast_fp16_6, tensor var_136_cast_fp16_7, tensor var_136_cast_fp16_8, tensor var_136_cast_fp16_9, tensor var_136_cast_fp16_10, tensor var_136_cast_fp16_11 = split(axis = var_136_axis_0, split_sizes = tile_0, x = var_135_cast_fp16)[name = tensor("op_136_cast_fp16")]; + tensor var_149_perm_0 = const()[name = tensor("op_149_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_150_axis_0 = const()[name = tensor("op_150_axis_0"), val = tensor(3)]; + tensor var_149_cast_fp16 = transpose(perm = var_149_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_12")]; + tensor var_150_cast_fp16_0, tensor var_150_cast_fp16_1, tensor var_150_cast_fp16_2, tensor var_150_cast_fp16_3, tensor var_150_cast_fp16_4, tensor var_150_cast_fp16_5, tensor var_150_cast_fp16_6, tensor var_150_cast_fp16_7, tensor var_150_cast_fp16_8, tensor var_150_cast_fp16_9, tensor var_150_cast_fp16_10, tensor var_150_cast_fp16_11 = split(axis = var_150_axis_0, split_sizes = tile_1, x = var_149_cast_fp16)[name = tensor("op_150_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_163_axis_0 = const()[name = tensor("op_163_axis_0"), val = tensor(1)]; + tensor var_163_cast_fp16_0, tensor var_163_cast_fp16_1, tensor var_163_cast_fp16_2, tensor var_163_cast_fp16_3, tensor var_163_cast_fp16_4, tensor var_163_cast_fp16_5, tensor var_163_cast_fp16_6, tensor var_163_cast_fp16_7, tensor var_163_cast_fp16_8, tensor var_163_cast_fp16_9, tensor var_163_cast_fp16_10, tensor var_163_cast_fp16_11 = split(axis = var_163_axis_0, split_sizes = tile_2, x = var_133_cast_fp16)[name = tensor("op_163_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_150_cast_fp16_0, var_136_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_150_cast_fp16_1, var_136_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_150_cast_fp16_2, var_136_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_150_cast_fp16_3, var_136_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_150_cast_fp16_4, var_136_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_150_cast_fp16_5, var_136_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_150_cast_fp16_6, var_136_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_150_cast_fp16_7, var_136_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_150_cast_fp16_8, var_136_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_150_cast_fp16_9, var_136_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_150_cast_fp16_10, var_136_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_150_cast_fp16_11, var_136_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor var_200_cast_fp16 = softmax(axis = var_84, x = aw_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = softmax(axis = var_84, x = aw_3_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor var_202_cast_fp16 = softmax(axis = var_84, x = aw_5_cast_fp16)[name = tensor("op_202_cast_fp16")]; + tensor var_203_cast_fp16 = softmax(axis = var_84, x = aw_7_cast_fp16)[name = tensor("op_203_cast_fp16")]; + tensor var_204_cast_fp16 = softmax(axis = var_84, x = aw_9_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = softmax(axis = var_84, x = aw_11_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor var_206_cast_fp16 = softmax(axis = var_84, x = aw_13_cast_fp16)[name = tensor("op_206_cast_fp16")]; + tensor var_207_cast_fp16 = softmax(axis = var_84, x = aw_15_cast_fp16)[name = tensor("op_207_cast_fp16")]; + tensor var_208_cast_fp16 = softmax(axis = var_84, x = aw_17_cast_fp16)[name = tensor("op_208_cast_fp16")]; + tensor var_209_cast_fp16 = softmax(axis = var_84, x = aw_19_cast_fp16)[name = tensor("op_209_cast_fp16")]; + tensor var_210_cast_fp16 = softmax(axis = var_84, x = aw_21_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor var_211_cast_fp16 = softmax(axis = var_84, x = aw_23_cast_fp16)[name = tensor("op_211_cast_fp16")]; + tensor var_213_equation_0 = const()[name = tensor("op_213_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_213_cast_fp16 = einsum(equation = var_213_equation_0, values = (var_163_cast_fp16_0, var_200_cast_fp16))[name = tensor("op_213_cast_fp16")]; + tensor var_215_equation_0 = const()[name = tensor("op_215_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_215_cast_fp16 = einsum(equation = var_215_equation_0, values = (var_163_cast_fp16_1, var_201_cast_fp16))[name = tensor("op_215_cast_fp16")]; + tensor var_217_equation_0 = const()[name = tensor("op_217_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_217_cast_fp16 = einsum(equation = var_217_equation_0, values = (var_163_cast_fp16_2, var_202_cast_fp16))[name = tensor("op_217_cast_fp16")]; + tensor var_219_equation_0 = const()[name = tensor("op_219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_219_cast_fp16 = einsum(equation = var_219_equation_0, values = (var_163_cast_fp16_3, var_203_cast_fp16))[name = tensor("op_219_cast_fp16")]; + tensor var_221_equation_0 = const()[name = tensor("op_221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_221_cast_fp16 = einsum(equation = var_221_equation_0, values = (var_163_cast_fp16_4, var_204_cast_fp16))[name = tensor("op_221_cast_fp16")]; + tensor var_223_equation_0 = const()[name = tensor("op_223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_223_cast_fp16 = einsum(equation = var_223_equation_0, values = (var_163_cast_fp16_5, var_205_cast_fp16))[name = tensor("op_223_cast_fp16")]; + tensor var_225_equation_0 = const()[name = tensor("op_225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_225_cast_fp16 = einsum(equation = var_225_equation_0, values = (var_163_cast_fp16_6, var_206_cast_fp16))[name = tensor("op_225_cast_fp16")]; + tensor var_227_equation_0 = const()[name = tensor("op_227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_227_cast_fp16 = einsum(equation = var_227_equation_0, values = (var_163_cast_fp16_7, var_207_cast_fp16))[name = tensor("op_227_cast_fp16")]; + tensor var_229_equation_0 = const()[name = tensor("op_229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_229_cast_fp16 = einsum(equation = var_229_equation_0, values = (var_163_cast_fp16_8, var_208_cast_fp16))[name = tensor("op_229_cast_fp16")]; + tensor var_231_equation_0 = const()[name = tensor("op_231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_231_cast_fp16 = einsum(equation = var_231_equation_0, values = (var_163_cast_fp16_9, var_209_cast_fp16))[name = tensor("op_231_cast_fp16")]; + tensor var_233_equation_0 = const()[name = tensor("op_233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_233_cast_fp16 = einsum(equation = var_233_equation_0, values = (var_163_cast_fp16_10, var_210_cast_fp16))[name = tensor("op_233_cast_fp16")]; + tensor var_235_equation_0 = const()[name = tensor("op_235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_235_cast_fp16 = einsum(equation = var_235_equation_0, values = (var_163_cast_fp16_11, var_211_cast_fp16))[name = tensor("op_235_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_84, interleave = input_5_interleave_0, values = (var_213_cast_fp16, var_215_cast_fp16, var_217_cast_fp16, var_219_cast_fp16, var_221_cast_fp16, var_223_cast_fp16, var_225_cast_fp16, var_227_cast_fp16, var_229_cast_fp16, var_231_cast_fp16, var_233_cast_fp16, var_235_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_244_pad_type_0 = const()[name = tensor("op_244_pad_type_0"), val = tensor("valid")]; + tensor var_244_strides_0 = const()[name = tensor("op_244_strides_0"), val = tensor([1, 1])]; + tensor var_244_pad_0 = const()[name = tensor("op_244_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_244_dilations_0 = const()[name = tensor("op_244_dilations_0"), val = tensor([1, 1])]; + tensor var_244_groups_0 = const()[name = tensor("op_244_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9760576)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10940288)))]; + tensor var_244_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_244_dilations_0, groups = var_244_groups_0, pad = var_244_pad_0, pad_type = var_244_pad_type_0, strides = var_244_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_244_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10941888)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10943488)))]; + tensor var_254_to_fp16 = const()[name = tensor("op_254_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_254_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10945088)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15663744)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_280_pad_type_0 = const()[name = tensor("op_280_pad_type_0"), val = tensor("valid")]; + tensor var_280_strides_0 = const()[name = tensor("op_280_strides_0"), val = tensor([1, 1])]; + tensor var_280_pad_0 = const()[name = tensor("op_280_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_280_dilations_0 = const()[name = tensor("op_280_dilations_0"), val = tensor([1, 1])]; + tensor var_280_groups_0 = const()[name = tensor("op_280_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15669952)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20388608)))]; + tensor var_280_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_280_dilations_0, groups = var_280_groups_0, pad = var_280_pad_0, pad_type = var_280_pad_type_0, strides = var_280_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_280_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_280_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_289 = const()[name = tensor("op_289"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20390208)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20391808)))]; + tensor var_305_to_fp16 = const()[name = tensor("op_305_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_305_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_340_weight_0_to_fp16 = const()[name = tensor("op_340_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20393408)))]; + tensor var_340_bias_0_to_fp16 = const()[name = tensor("op_340_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21573120)))]; + tensor var_340_cast_fp16 = conv(bias = var_340_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_340_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21574720)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_338_pad_type_0 = const()[name = tensor("op_338_pad_type_0"), val = tensor("valid")]; + tensor var_338_strides_0 = const()[name = tensor("op_338_strides_0"), val = tensor([1, 1])]; + tensor var_338_pad_0 = const()[name = tensor("op_338_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_338_dilations_0 = const()[name = tensor("op_338_dilations_0"), val = tensor([1, 1])]; + tensor var_338_groups_0 = const()[name = tensor("op_338_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22754432)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23934144)))]; + tensor var_338_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_338_dilations_0, groups = var_338_groups_0, pad = var_338_pad_0, pad_type = var_338_pad_type_0, strides = var_338_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_338_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_341_axis_0 = const()[name = tensor("op_341_axis_0"), val = tensor(1)]; + tensor var_341_cast_fp16_0, tensor var_341_cast_fp16_1, tensor var_341_cast_fp16_2, tensor var_341_cast_fp16_3, tensor var_341_cast_fp16_4, tensor var_341_cast_fp16_5, tensor var_341_cast_fp16_6, tensor var_341_cast_fp16_7, tensor var_341_cast_fp16_8, tensor var_341_cast_fp16_9, tensor var_341_cast_fp16_10, tensor var_341_cast_fp16_11 = split(axis = var_341_axis_0, split_sizes = tile_3, x = var_340_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor var_354_perm_0 = const()[name = tensor("op_354_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_355_axis_0 = const()[name = tensor("op_355_axis_0"), val = tensor(3)]; + tensor var_354_cast_fp16 = transpose(perm = var_354_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_11")]; + tensor var_355_cast_fp16_0, tensor var_355_cast_fp16_1, tensor var_355_cast_fp16_2, tensor var_355_cast_fp16_3, tensor var_355_cast_fp16_4, tensor var_355_cast_fp16_5, tensor var_355_cast_fp16_6, tensor var_355_cast_fp16_7, tensor var_355_cast_fp16_8, tensor var_355_cast_fp16_9, tensor var_355_cast_fp16_10, tensor var_355_cast_fp16_11 = split(axis = var_355_axis_0, split_sizes = tile_4, x = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_368_axis_0 = const()[name = tensor("op_368_axis_0"), val = tensor(1)]; + tensor var_368_cast_fp16_0, tensor var_368_cast_fp16_1, tensor var_368_cast_fp16_2, tensor var_368_cast_fp16_3, tensor var_368_cast_fp16_4, tensor var_368_cast_fp16_5, tensor var_368_cast_fp16_6, tensor var_368_cast_fp16_7, tensor var_368_cast_fp16_8, tensor var_368_cast_fp16_9, tensor var_368_cast_fp16_10, tensor var_368_cast_fp16_11 = split(axis = var_368_axis_0, split_sizes = tile_5, x = var_338_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_355_cast_fp16_0, var_341_cast_fp16_0))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_355_cast_fp16_1, var_341_cast_fp16_1))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_355_cast_fp16_2, var_341_cast_fp16_2))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_355_cast_fp16_3, var_341_cast_fp16_3))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_355_cast_fp16_4, var_341_cast_fp16_4))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_355_cast_fp16_5, var_341_cast_fp16_5))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_355_cast_fp16_6, var_341_cast_fp16_6))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_355_cast_fp16_7, var_341_cast_fp16_7))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_355_cast_fp16_8, var_341_cast_fp16_8))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_355_cast_fp16_9, var_341_cast_fp16_9))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_355_cast_fp16_10, var_341_cast_fp16_10))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_355_cast_fp16_11, var_341_cast_fp16_11))[name = tensor("aw_47_cast_fp16")]; + tensor var_405_cast_fp16 = softmax(axis = var_289, x = aw_25_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor var_406_cast_fp16 = softmax(axis = var_289, x = aw_27_cast_fp16)[name = tensor("op_406_cast_fp16")]; + tensor var_407_cast_fp16 = softmax(axis = var_289, x = aw_29_cast_fp16)[name = tensor("op_407_cast_fp16")]; + tensor var_408_cast_fp16 = softmax(axis = var_289, x = aw_31_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_409_cast_fp16 = softmax(axis = var_289, x = aw_33_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor var_410_cast_fp16 = softmax(axis = var_289, x = aw_35_cast_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = softmax(axis = var_289, x = aw_37_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor var_412_cast_fp16 = softmax(axis = var_289, x = aw_39_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = softmax(axis = var_289, x = aw_41_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor var_414_cast_fp16 = softmax(axis = var_289, x = aw_43_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_cast_fp16 = softmax(axis = var_289, x = aw_45_cast_fp16)[name = tensor("op_415_cast_fp16")]; + tensor var_416_cast_fp16 = softmax(axis = var_289, x = aw_47_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_418_equation_0 = const()[name = tensor("op_418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_418_cast_fp16 = einsum(equation = var_418_equation_0, values = (var_368_cast_fp16_0, var_405_cast_fp16))[name = tensor("op_418_cast_fp16")]; + tensor var_420_equation_0 = const()[name = tensor("op_420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_420_cast_fp16 = einsum(equation = var_420_equation_0, values = (var_368_cast_fp16_1, var_406_cast_fp16))[name = tensor("op_420_cast_fp16")]; + tensor var_422_equation_0 = const()[name = tensor("op_422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_422_cast_fp16 = einsum(equation = var_422_equation_0, values = (var_368_cast_fp16_2, var_407_cast_fp16))[name = tensor("op_422_cast_fp16")]; + tensor var_424_equation_0 = const()[name = tensor("op_424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_424_cast_fp16 = einsum(equation = var_424_equation_0, values = (var_368_cast_fp16_3, var_408_cast_fp16))[name = tensor("op_424_cast_fp16")]; + tensor var_426_equation_0 = const()[name = tensor("op_426_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_426_cast_fp16 = einsum(equation = var_426_equation_0, values = (var_368_cast_fp16_4, var_409_cast_fp16))[name = tensor("op_426_cast_fp16")]; + tensor var_428_equation_0 = const()[name = tensor("op_428_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_428_cast_fp16 = einsum(equation = var_428_equation_0, values = (var_368_cast_fp16_5, var_410_cast_fp16))[name = tensor("op_428_cast_fp16")]; + tensor var_430_equation_0 = const()[name = tensor("op_430_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_430_cast_fp16 = einsum(equation = var_430_equation_0, values = (var_368_cast_fp16_6, var_411_cast_fp16))[name = tensor("op_430_cast_fp16")]; + tensor var_432_equation_0 = const()[name = tensor("op_432_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_432_cast_fp16 = einsum(equation = var_432_equation_0, values = (var_368_cast_fp16_7, var_412_cast_fp16))[name = tensor("op_432_cast_fp16")]; + tensor var_434_equation_0 = const()[name = tensor("op_434_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_434_cast_fp16 = einsum(equation = var_434_equation_0, values = (var_368_cast_fp16_8, var_413_cast_fp16))[name = tensor("op_434_cast_fp16")]; + tensor var_436_equation_0 = const()[name = tensor("op_436_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_436_cast_fp16 = einsum(equation = var_436_equation_0, values = (var_368_cast_fp16_9, var_414_cast_fp16))[name = tensor("op_436_cast_fp16")]; + tensor var_438_equation_0 = const()[name = tensor("op_438_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_438_cast_fp16 = einsum(equation = var_438_equation_0, values = (var_368_cast_fp16_10, var_415_cast_fp16))[name = tensor("op_438_cast_fp16")]; + tensor var_440_equation_0 = const()[name = tensor("op_440_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_440_cast_fp16 = einsum(equation = var_440_equation_0, values = (var_368_cast_fp16_11, var_416_cast_fp16))[name = tensor("op_440_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_289, interleave = input_15_interleave_0, values = (var_418_cast_fp16, var_420_cast_fp16, var_422_cast_fp16, var_424_cast_fp16, var_426_cast_fp16, var_428_cast_fp16, var_430_cast_fp16, var_432_cast_fp16, var_434_cast_fp16, var_436_cast_fp16, var_438_cast_fp16, var_440_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_449_pad_type_0 = const()[name = tensor("op_449_pad_type_0"), val = tensor("valid")]; + tensor var_449_strides_0 = const()[name = tensor("op_449_strides_0"), val = tensor([1, 1])]; + tensor var_449_pad_0 = const()[name = tensor("op_449_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_449_dilations_0 = const()[name = tensor("op_449_dilations_0"), val = tensor([1, 1])]; + tensor var_449_groups_0 = const()[name = tensor("op_449_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23935744)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25115456)))]; + tensor var_449_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_449_dilations_0, groups = var_449_groups_0, pad = var_449_pad_0, pad_type = var_449_pad_type_0, strides = var_449_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_449_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25117056)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25118656)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_459_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25120256)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29838912)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_485_pad_type_0 = const()[name = tensor("op_485_pad_type_0"), val = tensor("valid")]; + tensor var_485_strides_0 = const()[name = tensor("op_485_strides_0"), val = tensor([1, 1])]; + tensor var_485_pad_0 = const()[name = tensor("op_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_485_dilations_0 = const()[name = tensor("op_485_dilations_0"), val = tensor([1, 1])]; + tensor var_485_groups_0 = const()[name = tensor("op_485_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29845120)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34563776)))]; + tensor var_485_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_485_dilations_0, groups = var_485_groups_0, pad = var_485_pad_0, pad_type = var_485_pad_type_0, strides = var_485_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_485_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_494 = const()[name = tensor("op_494"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34565376)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34566976)))]; + tensor var_510_to_fp16 = const()[name = tensor("op_510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_510_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_545_weight_0_to_fp16 = const()[name = tensor("op_545_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34568576)))]; + tensor var_545_bias_0_to_fp16 = const()[name = tensor("op_545_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35748288)))]; + tensor var_545_cast_fp16 = conv(bias = var_545_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_545_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_545_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35749888)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_543_pad_type_0 = const()[name = tensor("op_543_pad_type_0"), val = tensor("valid")]; + tensor var_543_strides_0 = const()[name = tensor("op_543_strides_0"), val = tensor([1, 1])]; + tensor var_543_pad_0 = const()[name = tensor("op_543_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_543_dilations_0 = const()[name = tensor("op_543_dilations_0"), val = tensor([1, 1])]; + tensor var_543_groups_0 = const()[name = tensor("op_543_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36929600)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38109312)))]; + tensor var_543_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_543_dilations_0, groups = var_543_groups_0, pad = var_543_pad_0, pad_type = var_543_pad_type_0, strides = var_543_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_543_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_546_axis_0 = const()[name = tensor("op_546_axis_0"), val = tensor(1)]; + tensor var_546_cast_fp16_0, tensor var_546_cast_fp16_1, tensor var_546_cast_fp16_2, tensor var_546_cast_fp16_3, tensor var_546_cast_fp16_4, tensor var_546_cast_fp16_5, tensor var_546_cast_fp16_6, tensor var_546_cast_fp16_7, tensor var_546_cast_fp16_8, tensor var_546_cast_fp16_9, tensor var_546_cast_fp16_10, tensor var_546_cast_fp16_11 = split(axis = var_546_axis_0, split_sizes = tile_6, x = var_545_cast_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_559_perm_0 = const()[name = tensor("op_559_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_560_axis_0 = const()[name = tensor("op_560_axis_0"), val = tensor(3)]; + tensor var_559_cast_fp16 = transpose(perm = var_559_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_10")]; + tensor var_560_cast_fp16_0, tensor var_560_cast_fp16_1, tensor var_560_cast_fp16_2, tensor var_560_cast_fp16_3, tensor var_560_cast_fp16_4, tensor var_560_cast_fp16_5, tensor var_560_cast_fp16_6, tensor var_560_cast_fp16_7, tensor var_560_cast_fp16_8, tensor var_560_cast_fp16_9, tensor var_560_cast_fp16_10, tensor var_560_cast_fp16_11 = split(axis = var_560_axis_0, split_sizes = tile_7, x = var_559_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_573_axis_0 = const()[name = tensor("op_573_axis_0"), val = tensor(1)]; + tensor var_573_cast_fp16_0, tensor var_573_cast_fp16_1, tensor var_573_cast_fp16_2, tensor var_573_cast_fp16_3, tensor var_573_cast_fp16_4, tensor var_573_cast_fp16_5, tensor var_573_cast_fp16_6, tensor var_573_cast_fp16_7, tensor var_573_cast_fp16_8, tensor var_573_cast_fp16_9, tensor var_573_cast_fp16_10, tensor var_573_cast_fp16_11 = split(axis = var_573_axis_0, split_sizes = tile_8, x = var_543_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_560_cast_fp16_0, var_546_cast_fp16_0))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_560_cast_fp16_1, var_546_cast_fp16_1))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_560_cast_fp16_2, var_546_cast_fp16_2))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_560_cast_fp16_3, var_546_cast_fp16_3))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_560_cast_fp16_4, var_546_cast_fp16_4))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_560_cast_fp16_5, var_546_cast_fp16_5))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_560_cast_fp16_6, var_546_cast_fp16_6))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_560_cast_fp16_7, var_546_cast_fp16_7))[name = tensor("aw_63_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_560_cast_fp16_8, var_546_cast_fp16_8))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_560_cast_fp16_9, var_546_cast_fp16_9))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_560_cast_fp16_10, var_546_cast_fp16_10))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_560_cast_fp16_11, var_546_cast_fp16_11))[name = tensor("aw_71_cast_fp16")]; + tensor var_610_cast_fp16 = softmax(axis = var_494, x = aw_49_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = softmax(axis = var_494, x = aw_51_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor var_612_cast_fp16 = softmax(axis = var_494, x = aw_53_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_cast_fp16 = softmax(axis = var_494, x = aw_55_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor var_614_cast_fp16 = softmax(axis = var_494, x = aw_57_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = softmax(axis = var_494, x = aw_59_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor var_616_cast_fp16 = softmax(axis = var_494, x = aw_61_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_617_cast_fp16 = softmax(axis = var_494, x = aw_63_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_618_cast_fp16 = softmax(axis = var_494, x = aw_65_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = softmax(axis = var_494, x = aw_67_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor var_620_cast_fp16 = softmax(axis = var_494, x = aw_69_cast_fp16)[name = tensor("op_620_cast_fp16")]; + tensor var_621_cast_fp16 = softmax(axis = var_494, x = aw_71_cast_fp16)[name = tensor("op_621_cast_fp16")]; + tensor var_623_equation_0 = const()[name = tensor("op_623_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_623_cast_fp16 = einsum(equation = var_623_equation_0, values = (var_573_cast_fp16_0, var_610_cast_fp16))[name = tensor("op_623_cast_fp16")]; + tensor var_625_equation_0 = const()[name = tensor("op_625_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_625_cast_fp16 = einsum(equation = var_625_equation_0, values = (var_573_cast_fp16_1, var_611_cast_fp16))[name = tensor("op_625_cast_fp16")]; + tensor var_627_equation_0 = const()[name = tensor("op_627_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_627_cast_fp16 = einsum(equation = var_627_equation_0, values = (var_573_cast_fp16_2, var_612_cast_fp16))[name = tensor("op_627_cast_fp16")]; + tensor var_629_equation_0 = const()[name = tensor("op_629_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_629_cast_fp16 = einsum(equation = var_629_equation_0, values = (var_573_cast_fp16_3, var_613_cast_fp16))[name = tensor("op_629_cast_fp16")]; + tensor var_631_equation_0 = const()[name = tensor("op_631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_631_cast_fp16 = einsum(equation = var_631_equation_0, values = (var_573_cast_fp16_4, var_614_cast_fp16))[name = tensor("op_631_cast_fp16")]; + tensor var_633_equation_0 = const()[name = tensor("op_633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_633_cast_fp16 = einsum(equation = var_633_equation_0, values = (var_573_cast_fp16_5, var_615_cast_fp16))[name = tensor("op_633_cast_fp16")]; + tensor var_635_equation_0 = const()[name = tensor("op_635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_635_cast_fp16 = einsum(equation = var_635_equation_0, values = (var_573_cast_fp16_6, var_616_cast_fp16))[name = tensor("op_635_cast_fp16")]; + tensor var_637_equation_0 = const()[name = tensor("op_637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_637_cast_fp16 = einsum(equation = var_637_equation_0, values = (var_573_cast_fp16_7, var_617_cast_fp16))[name = tensor("op_637_cast_fp16")]; + tensor var_639_equation_0 = const()[name = tensor("op_639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_639_cast_fp16 = einsum(equation = var_639_equation_0, values = (var_573_cast_fp16_8, var_618_cast_fp16))[name = tensor("op_639_cast_fp16")]; + tensor var_641_equation_0 = const()[name = tensor("op_641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_641_cast_fp16 = einsum(equation = var_641_equation_0, values = (var_573_cast_fp16_9, var_619_cast_fp16))[name = tensor("op_641_cast_fp16")]; + tensor var_643_equation_0 = const()[name = tensor("op_643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_643_cast_fp16 = einsum(equation = var_643_equation_0, values = (var_573_cast_fp16_10, var_620_cast_fp16))[name = tensor("op_643_cast_fp16")]; + tensor var_645_equation_0 = const()[name = tensor("op_645_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_645_cast_fp16 = einsum(equation = var_645_equation_0, values = (var_573_cast_fp16_11, var_621_cast_fp16))[name = tensor("op_645_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_494, interleave = input_25_interleave_0, values = (var_623_cast_fp16, var_625_cast_fp16, var_627_cast_fp16, var_629_cast_fp16, var_631_cast_fp16, var_633_cast_fp16, var_635_cast_fp16, var_637_cast_fp16, var_639_cast_fp16, var_641_cast_fp16, var_643_cast_fp16, var_645_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_654_pad_type_0 = const()[name = tensor("op_654_pad_type_0"), val = tensor("valid")]; + tensor var_654_strides_0 = const()[name = tensor("op_654_strides_0"), val = tensor([1, 1])]; + tensor var_654_pad_0 = const()[name = tensor("op_654_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_654_dilations_0 = const()[name = tensor("op_654_dilations_0"), val = tensor([1, 1])]; + tensor var_654_groups_0 = const()[name = tensor("op_654_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38110912)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39290624)))]; + tensor var_654_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_654_dilations_0, groups = var_654_groups_0, pad = var_654_pad_0, pad_type = var_654_pad_type_0, strides = var_654_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_654_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_654_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39292224)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39293824)))]; + tensor var_664_to_fp16 = const()[name = tensor("op_664_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_664_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39295424)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44014080)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_690_pad_type_0 = const()[name = tensor("op_690_pad_type_0"), val = tensor("valid")]; + tensor var_690_strides_0 = const()[name = tensor("op_690_strides_0"), val = tensor([1, 1])]; + tensor var_690_pad_0 = const()[name = tensor("op_690_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_690_dilations_0 = const()[name = tensor("op_690_dilations_0"), val = tensor([1, 1])]; + tensor var_690_groups_0 = const()[name = tensor("op_690_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44020288)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48738944)))]; + tensor var_690_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_690_dilations_0, groups = var_690_groups_0, pad = var_690_pad_0, pad_type = var_690_pad_type_0, strides = var_690_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_690_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48740544)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48742144)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_715_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_750_weight_0_to_fp16 = const()[name = tensor("op_750_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48743744)))]; + tensor var_750_bias_0_to_fp16 = const()[name = tensor("op_750_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49923456)))]; + tensor var_750_cast_fp16 = conv(bias = var_750_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_750_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_750_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49925056)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_748_pad_type_0 = const()[name = tensor("op_748_pad_type_0"), val = tensor("valid")]; + tensor var_748_strides_0 = const()[name = tensor("op_748_strides_0"), val = tensor([1, 1])]; + tensor var_748_pad_0 = const()[name = tensor("op_748_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_748_dilations_0 = const()[name = tensor("op_748_dilations_0"), val = tensor([1, 1])]; + tensor var_748_groups_0 = const()[name = tensor("op_748_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51104768)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52284480)))]; + tensor var_748_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_748_dilations_0, groups = var_748_groups_0, pad = var_748_pad_0, pad_type = var_748_pad_type_0, strides = var_748_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_cast_fp16_0, tensor var_751_cast_fp16_1, tensor var_751_cast_fp16_2, tensor var_751_cast_fp16_3, tensor var_751_cast_fp16_4, tensor var_751_cast_fp16_5, tensor var_751_cast_fp16_6, tensor var_751_cast_fp16_7, tensor var_751_cast_fp16_8, tensor var_751_cast_fp16_9, tensor var_751_cast_fp16_10, tensor var_751_cast_fp16_11 = split(axis = var_751_axis_0, split_sizes = tile_9, x = var_750_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_764_perm_0 = const()[name = tensor("op_764_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_765_axis_0 = const()[name = tensor("op_765_axis_0"), val = tensor(3)]; + tensor var_764_cast_fp16 = transpose(perm = var_764_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_9")]; + tensor var_765_cast_fp16_0, tensor var_765_cast_fp16_1, tensor var_765_cast_fp16_2, tensor var_765_cast_fp16_3, tensor var_765_cast_fp16_4, tensor var_765_cast_fp16_5, tensor var_765_cast_fp16_6, tensor var_765_cast_fp16_7, tensor var_765_cast_fp16_8, tensor var_765_cast_fp16_9, tensor var_765_cast_fp16_10, tensor var_765_cast_fp16_11 = split(axis = var_765_axis_0, split_sizes = tile_10, x = var_764_cast_fp16)[name = tensor("op_765_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_778_axis_0 = const()[name = tensor("op_778_axis_0"), val = tensor(1)]; + tensor var_778_cast_fp16_0, tensor var_778_cast_fp16_1, tensor var_778_cast_fp16_2, tensor var_778_cast_fp16_3, tensor var_778_cast_fp16_4, tensor var_778_cast_fp16_5, tensor var_778_cast_fp16_6, tensor var_778_cast_fp16_7, tensor var_778_cast_fp16_8, tensor var_778_cast_fp16_9, tensor var_778_cast_fp16_10, tensor var_778_cast_fp16_11 = split(axis = var_778_axis_0, split_sizes = tile_11, x = var_748_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_765_cast_fp16_0, var_751_cast_fp16_0))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_765_cast_fp16_1, var_751_cast_fp16_1))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_765_cast_fp16_2, var_751_cast_fp16_2))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_765_cast_fp16_3, var_751_cast_fp16_3))[name = tensor("aw_79_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_765_cast_fp16_4, var_751_cast_fp16_4))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_765_cast_fp16_5, var_751_cast_fp16_5))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_765_cast_fp16_6, var_751_cast_fp16_6))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_765_cast_fp16_7, var_751_cast_fp16_7))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_765_cast_fp16_8, var_751_cast_fp16_8))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_765_cast_fp16_9, var_751_cast_fp16_9))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_765_cast_fp16_10, var_751_cast_fp16_10))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_765_cast_fp16_11, var_751_cast_fp16_11))[name = tensor("aw_95_cast_fp16")]; + tensor var_815_cast_fp16 = softmax(axis = var_699, x = aw_73_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor var_816_cast_fp16 = softmax(axis = var_699, x = aw_75_cast_fp16)[name = tensor("op_816_cast_fp16")]; + tensor var_817_cast_fp16 = softmax(axis = var_699, x = aw_77_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = softmax(axis = var_699, x = aw_79_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_819_cast_fp16 = softmax(axis = var_699, x = aw_81_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820_cast_fp16 = softmax(axis = var_699, x = aw_83_cast_fp16)[name = tensor("op_820_cast_fp16")]; + tensor var_821_cast_fp16 = softmax(axis = var_699, x = aw_85_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = softmax(axis = var_699, x = aw_87_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor var_823_cast_fp16 = softmax(axis = var_699, x = aw_89_cast_fp16)[name = tensor("op_823_cast_fp16")]; + tensor var_824_cast_fp16 = softmax(axis = var_699, x = aw_91_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor var_825_cast_fp16 = softmax(axis = var_699, x = aw_93_cast_fp16)[name = tensor("op_825_cast_fp16")]; + tensor var_826_cast_fp16 = softmax(axis = var_699, x = aw_95_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_828_equation_0 = const()[name = tensor("op_828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_828_cast_fp16 = einsum(equation = var_828_equation_0, values = (var_778_cast_fp16_0, var_815_cast_fp16))[name = tensor("op_828_cast_fp16")]; + tensor var_830_equation_0 = const()[name = tensor("op_830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_830_cast_fp16 = einsum(equation = var_830_equation_0, values = (var_778_cast_fp16_1, var_816_cast_fp16))[name = tensor("op_830_cast_fp16")]; + tensor var_832_equation_0 = const()[name = tensor("op_832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_832_cast_fp16 = einsum(equation = var_832_equation_0, values = (var_778_cast_fp16_2, var_817_cast_fp16))[name = tensor("op_832_cast_fp16")]; + tensor var_834_equation_0 = const()[name = tensor("op_834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_834_cast_fp16 = einsum(equation = var_834_equation_0, values = (var_778_cast_fp16_3, var_818_cast_fp16))[name = tensor("op_834_cast_fp16")]; + tensor var_836_equation_0 = const()[name = tensor("op_836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_836_cast_fp16 = einsum(equation = var_836_equation_0, values = (var_778_cast_fp16_4, var_819_cast_fp16))[name = tensor("op_836_cast_fp16")]; + tensor var_838_equation_0 = const()[name = tensor("op_838_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_838_cast_fp16 = einsum(equation = var_838_equation_0, values = (var_778_cast_fp16_5, var_820_cast_fp16))[name = tensor("op_838_cast_fp16")]; + tensor var_840_equation_0 = const()[name = tensor("op_840_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_840_cast_fp16 = einsum(equation = var_840_equation_0, values = (var_778_cast_fp16_6, var_821_cast_fp16))[name = tensor("op_840_cast_fp16")]; + tensor var_842_equation_0 = const()[name = tensor("op_842_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_842_cast_fp16 = einsum(equation = var_842_equation_0, values = (var_778_cast_fp16_7, var_822_cast_fp16))[name = tensor("op_842_cast_fp16")]; + tensor var_844_equation_0 = const()[name = tensor("op_844_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_844_cast_fp16 = einsum(equation = var_844_equation_0, values = (var_778_cast_fp16_8, var_823_cast_fp16))[name = tensor("op_844_cast_fp16")]; + tensor var_846_equation_0 = const()[name = tensor("op_846_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_846_cast_fp16 = einsum(equation = var_846_equation_0, values = (var_778_cast_fp16_9, var_824_cast_fp16))[name = tensor("op_846_cast_fp16")]; + tensor var_848_equation_0 = const()[name = tensor("op_848_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_848_cast_fp16 = einsum(equation = var_848_equation_0, values = (var_778_cast_fp16_10, var_825_cast_fp16))[name = tensor("op_848_cast_fp16")]; + tensor var_850_equation_0 = const()[name = tensor("op_850_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_850_cast_fp16 = einsum(equation = var_850_equation_0, values = (var_778_cast_fp16_11, var_826_cast_fp16))[name = tensor("op_850_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_699, interleave = input_35_interleave_0, values = (var_828_cast_fp16, var_830_cast_fp16, var_832_cast_fp16, var_834_cast_fp16, var_836_cast_fp16, var_838_cast_fp16, var_840_cast_fp16, var_842_cast_fp16, var_844_cast_fp16, var_846_cast_fp16, var_848_cast_fp16, var_850_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_859_pad_type_0 = const()[name = tensor("op_859_pad_type_0"), val = tensor("valid")]; + tensor var_859_strides_0 = const()[name = tensor("op_859_strides_0"), val = tensor([1, 1])]; + tensor var_859_pad_0 = const()[name = tensor("op_859_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_859_dilations_0 = const()[name = tensor("op_859_dilations_0"), val = tensor([1, 1])]; + tensor var_859_groups_0 = const()[name = tensor("op_859_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52286080)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53465792)))]; + tensor var_859_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_859_dilations_0, groups = var_859_groups_0, pad = var_859_pad_0, pad_type = var_859_pad_type_0, strides = var_859_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_859_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_859_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53467392)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53468992)))]; + tensor var_869_to_fp16 = const()[name = tensor("op_869_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_869_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53470592)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58189248)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_895_pad_type_0 = const()[name = tensor("op_895_pad_type_0"), val = tensor("valid")]; + tensor var_895_strides_0 = const()[name = tensor("op_895_strides_0"), val = tensor([1, 1])]; + tensor var_895_pad_0 = const()[name = tensor("op_895_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_895_dilations_0 = const()[name = tensor("op_895_dilations_0"), val = tensor([1, 1])]; + tensor var_895_groups_0 = const()[name = tensor("op_895_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58195456)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62914112)))]; + tensor var_895_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_895_dilations_0, groups = var_895_groups_0, pad = var_895_pad_0, pad_type = var_895_pad_type_0, strides = var_895_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_895_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_895_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62915712)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62917312)))]; + tensor var_920_to_fp16 = const()[name = tensor("op_920_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_920_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_955_weight_0_to_fp16 = const()[name = tensor("op_955_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62918912)))]; + tensor var_955_bias_0_to_fp16 = const()[name = tensor("op_955_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64098624)))]; + tensor var_955_cast_fp16 = conv(bias = var_955_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_955_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_955_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64100224)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_953_pad_type_0 = const()[name = tensor("op_953_pad_type_0"), val = tensor("valid")]; + tensor var_953_strides_0 = const()[name = tensor("op_953_strides_0"), val = tensor([1, 1])]; + tensor var_953_pad_0 = const()[name = tensor("op_953_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_953_dilations_0 = const()[name = tensor("op_953_dilations_0"), val = tensor([1, 1])]; + tensor var_953_groups_0 = const()[name = tensor("op_953_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65279936)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66459648)))]; + tensor var_953_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_953_dilations_0, groups = var_953_groups_0, pad = var_953_pad_0, pad_type = var_953_pad_type_0, strides = var_953_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_953_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_956_axis_0 = const()[name = tensor("op_956_axis_0"), val = tensor(1)]; + tensor var_956_cast_fp16_0, tensor var_956_cast_fp16_1, tensor var_956_cast_fp16_2, tensor var_956_cast_fp16_3, tensor var_956_cast_fp16_4, tensor var_956_cast_fp16_5, tensor var_956_cast_fp16_6, tensor var_956_cast_fp16_7, tensor var_956_cast_fp16_8, tensor var_956_cast_fp16_9, tensor var_956_cast_fp16_10, tensor var_956_cast_fp16_11 = split(axis = var_956_axis_0, split_sizes = tile_12, x = var_955_cast_fp16)[name = tensor("op_956_cast_fp16")]; + tensor var_969_perm_0 = const()[name = tensor("op_969_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_970_axis_0 = const()[name = tensor("op_970_axis_0"), val = tensor(3)]; + tensor var_969_cast_fp16 = transpose(perm = var_969_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_8")]; + tensor var_970_cast_fp16_0, tensor var_970_cast_fp16_1, tensor var_970_cast_fp16_2, tensor var_970_cast_fp16_3, tensor var_970_cast_fp16_4, tensor var_970_cast_fp16_5, tensor var_970_cast_fp16_6, tensor var_970_cast_fp16_7, tensor var_970_cast_fp16_8, tensor var_970_cast_fp16_9, tensor var_970_cast_fp16_10, tensor var_970_cast_fp16_11 = split(axis = var_970_axis_0, split_sizes = tile_13, x = var_969_cast_fp16)[name = tensor("op_970_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_983_axis_0 = const()[name = tensor("op_983_axis_0"), val = tensor(1)]; + tensor var_983_cast_fp16_0, tensor var_983_cast_fp16_1, tensor var_983_cast_fp16_2, tensor var_983_cast_fp16_3, tensor var_983_cast_fp16_4, tensor var_983_cast_fp16_5, tensor var_983_cast_fp16_6, tensor var_983_cast_fp16_7, tensor var_983_cast_fp16_8, tensor var_983_cast_fp16_9, tensor var_983_cast_fp16_10, tensor var_983_cast_fp16_11 = split(axis = var_983_axis_0, split_sizes = tile_14, x = var_953_cast_fp16)[name = tensor("op_983_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_970_cast_fp16_0, var_956_cast_fp16_0))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_970_cast_fp16_1, var_956_cast_fp16_1))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_970_cast_fp16_2, var_956_cast_fp16_2))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_970_cast_fp16_3, var_956_cast_fp16_3))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_970_cast_fp16_4, var_956_cast_fp16_4))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_970_cast_fp16_5, var_956_cast_fp16_5))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_970_cast_fp16_6, var_956_cast_fp16_6))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_970_cast_fp16_7, var_956_cast_fp16_7))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_970_cast_fp16_8, var_956_cast_fp16_8))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_970_cast_fp16_9, var_956_cast_fp16_9))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_970_cast_fp16_10, var_956_cast_fp16_10))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_970_cast_fp16_11, var_956_cast_fp16_11))[name = tensor("aw_119_cast_fp16")]; + tensor var_1020_cast_fp16 = softmax(axis = var_904, x = aw_97_cast_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1021_cast_fp16 = softmax(axis = var_904, x = aw_99_cast_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1022_cast_fp16 = softmax(axis = var_904, x = aw_101_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = softmax(axis = var_904, x = aw_103_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = softmax(axis = var_904, x = aw_105_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor var_1025_cast_fp16 = softmax(axis = var_904, x = aw_107_cast_fp16)[name = tensor("op_1025_cast_fp16")]; + tensor var_1026_cast_fp16 = softmax(axis = var_904, x = aw_109_cast_fp16)[name = tensor("op_1026_cast_fp16")]; + tensor var_1027_cast_fp16 = softmax(axis = var_904, x = aw_111_cast_fp16)[name = tensor("op_1027_cast_fp16")]; + tensor var_1028_cast_fp16 = softmax(axis = var_904, x = aw_113_cast_fp16)[name = tensor("op_1028_cast_fp16")]; + tensor var_1029_cast_fp16 = softmax(axis = var_904, x = aw_115_cast_fp16)[name = tensor("op_1029_cast_fp16")]; + tensor var_1030_cast_fp16 = softmax(axis = var_904, x = aw_117_cast_fp16)[name = tensor("op_1030_cast_fp16")]; + tensor var_1031_cast_fp16 = softmax(axis = var_904, x = aw_119_cast_fp16)[name = tensor("op_1031_cast_fp16")]; + tensor var_1033_equation_0 = const()[name = tensor("op_1033_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1033_cast_fp16 = einsum(equation = var_1033_equation_0, values = (var_983_cast_fp16_0, var_1020_cast_fp16))[name = tensor("op_1033_cast_fp16")]; + tensor var_1035_equation_0 = const()[name = tensor("op_1035_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1035_cast_fp16 = einsum(equation = var_1035_equation_0, values = (var_983_cast_fp16_1, var_1021_cast_fp16))[name = tensor("op_1035_cast_fp16")]; + tensor var_1037_equation_0 = const()[name = tensor("op_1037_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1037_cast_fp16 = einsum(equation = var_1037_equation_0, values = (var_983_cast_fp16_2, var_1022_cast_fp16))[name = tensor("op_1037_cast_fp16")]; + tensor var_1039_equation_0 = const()[name = tensor("op_1039_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1039_cast_fp16 = einsum(equation = var_1039_equation_0, values = (var_983_cast_fp16_3, var_1023_cast_fp16))[name = tensor("op_1039_cast_fp16")]; + tensor var_1041_equation_0 = const()[name = tensor("op_1041_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1041_cast_fp16 = einsum(equation = var_1041_equation_0, values = (var_983_cast_fp16_4, var_1024_cast_fp16))[name = tensor("op_1041_cast_fp16")]; + tensor var_1043_equation_0 = const()[name = tensor("op_1043_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1043_cast_fp16 = einsum(equation = var_1043_equation_0, values = (var_983_cast_fp16_5, var_1025_cast_fp16))[name = tensor("op_1043_cast_fp16")]; + tensor var_1045_equation_0 = const()[name = tensor("op_1045_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1045_cast_fp16 = einsum(equation = var_1045_equation_0, values = (var_983_cast_fp16_6, var_1026_cast_fp16))[name = tensor("op_1045_cast_fp16")]; + tensor var_1047_equation_0 = const()[name = tensor("op_1047_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1047_cast_fp16 = einsum(equation = var_1047_equation_0, values = (var_983_cast_fp16_7, var_1027_cast_fp16))[name = tensor("op_1047_cast_fp16")]; + tensor var_1049_equation_0 = const()[name = tensor("op_1049_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1049_cast_fp16 = einsum(equation = var_1049_equation_0, values = (var_983_cast_fp16_8, var_1028_cast_fp16))[name = tensor("op_1049_cast_fp16")]; + tensor var_1051_equation_0 = const()[name = tensor("op_1051_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1051_cast_fp16 = einsum(equation = var_1051_equation_0, values = (var_983_cast_fp16_9, var_1029_cast_fp16))[name = tensor("op_1051_cast_fp16")]; + tensor var_1053_equation_0 = const()[name = tensor("op_1053_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1053_cast_fp16 = einsum(equation = var_1053_equation_0, values = (var_983_cast_fp16_10, var_1030_cast_fp16))[name = tensor("op_1053_cast_fp16")]; + tensor var_1055_equation_0 = const()[name = tensor("op_1055_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1055_cast_fp16 = einsum(equation = var_1055_equation_0, values = (var_983_cast_fp16_11, var_1031_cast_fp16))[name = tensor("op_1055_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_904, interleave = input_45_interleave_0, values = (var_1033_cast_fp16, var_1035_cast_fp16, var_1037_cast_fp16, var_1039_cast_fp16, var_1041_cast_fp16, var_1043_cast_fp16, var_1045_cast_fp16, var_1047_cast_fp16, var_1049_cast_fp16, var_1051_cast_fp16, var_1053_cast_fp16, var_1055_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1064_pad_type_0 = const()[name = tensor("op_1064_pad_type_0"), val = tensor("valid")]; + tensor var_1064_strides_0 = const()[name = tensor("op_1064_strides_0"), val = tensor([1, 1])]; + tensor var_1064_pad_0 = const()[name = tensor("op_1064_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1064_dilations_0 = const()[name = tensor("op_1064_dilations_0"), val = tensor([1, 1])]; + tensor var_1064_groups_0 = const()[name = tensor("op_1064_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66461248)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67640960)))]; + tensor var_1064_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1064_dilations_0, groups = var_1064_groups_0, pad = var_1064_pad_0, pad_type = var_1064_pad_type_0, strides = var_1064_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1064_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1064_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67642560)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67644160)))]; + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1074_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67645760)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72364416)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1100_pad_type_0 = const()[name = tensor("op_1100_pad_type_0"), val = tensor("valid")]; + tensor var_1100_strides_0 = const()[name = tensor("op_1100_strides_0"), val = tensor([1, 1])]; + tensor var_1100_pad_0 = const()[name = tensor("op_1100_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1100_dilations_0 = const()[name = tensor("op_1100_dilations_0"), val = tensor([1, 1])]; + tensor var_1100_groups_0 = const()[name = tensor("op_1100_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72370624)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77089280)))]; + tensor var_1100_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1100_dilations_0, groups = var_1100_groups_0, pad = var_1100_pad_0, pad_type = var_1100_pad_type_0, strides = var_1100_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1100_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77090880)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77092480)))]; + tensor var_1125_to_fp16 = const()[name = tensor("op_1125_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1125_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1160_weight_0_to_fp16 = const()[name = tensor("op_1160_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77094080)))]; + tensor var_1160_bias_0_to_fp16 = const()[name = tensor("op_1160_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78273792)))]; + tensor var_1160_cast_fp16 = conv(bias = var_1160_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1160_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1160_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78275392)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1158_pad_type_0 = const()[name = tensor("op_1158_pad_type_0"), val = tensor("valid")]; + tensor var_1158_strides_0 = const()[name = tensor("op_1158_strides_0"), val = tensor([1, 1])]; + tensor var_1158_pad_0 = const()[name = tensor("op_1158_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1158_dilations_0 = const()[name = tensor("op_1158_dilations_0"), val = tensor([1, 1])]; + tensor var_1158_groups_0 = const()[name = tensor("op_1158_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79455104)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80634816)))]; + tensor var_1158_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1158_dilations_0, groups = var_1158_groups_0, pad = var_1158_pad_0, pad_type = var_1158_pad_type_0, strides = var_1158_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1, tensor var_1161_cast_fp16_2, tensor var_1161_cast_fp16_3, tensor var_1161_cast_fp16_4, tensor var_1161_cast_fp16_5, tensor var_1161_cast_fp16_6, tensor var_1161_cast_fp16_7, tensor var_1161_cast_fp16_8, tensor var_1161_cast_fp16_9, tensor var_1161_cast_fp16_10, tensor var_1161_cast_fp16_11 = split(axis = var_1161_axis_0, split_sizes = tile_15, x = var_1160_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor var_1174_perm_0 = const()[name = tensor("op_1174_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1175_axis_0 = const()[name = tensor("op_1175_axis_0"), val = tensor(3)]; + tensor var_1174_cast_fp16 = transpose(perm = var_1174_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_7")]; + tensor var_1175_cast_fp16_0, tensor var_1175_cast_fp16_1, tensor var_1175_cast_fp16_2, tensor var_1175_cast_fp16_3, tensor var_1175_cast_fp16_4, tensor var_1175_cast_fp16_5, tensor var_1175_cast_fp16_6, tensor var_1175_cast_fp16_7, tensor var_1175_cast_fp16_8, tensor var_1175_cast_fp16_9, tensor var_1175_cast_fp16_10, tensor var_1175_cast_fp16_11 = split(axis = var_1175_axis_0, split_sizes = tile_16, x = var_1174_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1188_axis_0 = const()[name = tensor("op_1188_axis_0"), val = tensor(1)]; + tensor var_1188_cast_fp16_0, tensor var_1188_cast_fp16_1, tensor var_1188_cast_fp16_2, tensor var_1188_cast_fp16_3, tensor var_1188_cast_fp16_4, tensor var_1188_cast_fp16_5, tensor var_1188_cast_fp16_6, tensor var_1188_cast_fp16_7, tensor var_1188_cast_fp16_8, tensor var_1188_cast_fp16_9, tensor var_1188_cast_fp16_10, tensor var_1188_cast_fp16_11 = split(axis = var_1188_axis_0, split_sizes = tile_17, x = var_1158_cast_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_1175_cast_fp16_0, var_1161_cast_fp16_0))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_1175_cast_fp16_1, var_1161_cast_fp16_1))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_1175_cast_fp16_2, var_1161_cast_fp16_2))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_1175_cast_fp16_3, var_1161_cast_fp16_3))[name = tensor("aw_127_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1175_cast_fp16_4, var_1161_cast_fp16_4))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1175_cast_fp16_5, var_1161_cast_fp16_5))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1175_cast_fp16_6, var_1161_cast_fp16_6))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1175_cast_fp16_7, var_1161_cast_fp16_7))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1175_cast_fp16_8, var_1161_cast_fp16_8))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1175_cast_fp16_9, var_1161_cast_fp16_9))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1175_cast_fp16_10, var_1161_cast_fp16_10))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1175_cast_fp16_11, var_1161_cast_fp16_11))[name = tensor("aw_143_cast_fp16")]; + tensor var_1225_cast_fp16 = softmax(axis = var_1109, x = aw_121_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor var_1226_cast_fp16 = softmax(axis = var_1109, x = aw_123_cast_fp16)[name = tensor("op_1226_cast_fp16")]; + tensor var_1227_cast_fp16 = softmax(axis = var_1109, x = aw_125_cast_fp16)[name = tensor("op_1227_cast_fp16")]; + tensor var_1228_cast_fp16 = softmax(axis = var_1109, x = aw_127_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = softmax(axis = var_1109, x = aw_129_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor var_1230_cast_fp16 = softmax(axis = var_1109, x = aw_131_cast_fp16)[name = tensor("op_1230_cast_fp16")]; + tensor var_1231_cast_fp16 = softmax(axis = var_1109, x = aw_133_cast_fp16)[name = tensor("op_1231_cast_fp16")]; + tensor var_1232_cast_fp16 = softmax(axis = var_1109, x = aw_135_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = softmax(axis = var_1109, x = aw_137_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor var_1234_cast_fp16 = softmax(axis = var_1109, x = aw_139_cast_fp16)[name = tensor("op_1234_cast_fp16")]; + tensor var_1235_cast_fp16 = softmax(axis = var_1109, x = aw_141_cast_fp16)[name = tensor("op_1235_cast_fp16")]; + tensor var_1236_cast_fp16 = softmax(axis = var_1109, x = aw_143_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1238_equation_0 = const()[name = tensor("op_1238_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1238_cast_fp16 = einsum(equation = var_1238_equation_0, values = (var_1188_cast_fp16_0, var_1225_cast_fp16))[name = tensor("op_1238_cast_fp16")]; + tensor var_1240_equation_0 = const()[name = tensor("op_1240_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1240_cast_fp16 = einsum(equation = var_1240_equation_0, values = (var_1188_cast_fp16_1, var_1226_cast_fp16))[name = tensor("op_1240_cast_fp16")]; + tensor var_1242_equation_0 = const()[name = tensor("op_1242_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1242_cast_fp16 = einsum(equation = var_1242_equation_0, values = (var_1188_cast_fp16_2, var_1227_cast_fp16))[name = tensor("op_1242_cast_fp16")]; + tensor var_1244_equation_0 = const()[name = tensor("op_1244_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1244_cast_fp16 = einsum(equation = var_1244_equation_0, values = (var_1188_cast_fp16_3, var_1228_cast_fp16))[name = tensor("op_1244_cast_fp16")]; + tensor var_1246_equation_0 = const()[name = tensor("op_1246_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1246_cast_fp16 = einsum(equation = var_1246_equation_0, values = (var_1188_cast_fp16_4, var_1229_cast_fp16))[name = tensor("op_1246_cast_fp16")]; + tensor var_1248_equation_0 = const()[name = tensor("op_1248_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1248_cast_fp16 = einsum(equation = var_1248_equation_0, values = (var_1188_cast_fp16_5, var_1230_cast_fp16))[name = tensor("op_1248_cast_fp16")]; + tensor var_1250_equation_0 = const()[name = tensor("op_1250_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1250_cast_fp16 = einsum(equation = var_1250_equation_0, values = (var_1188_cast_fp16_6, var_1231_cast_fp16))[name = tensor("op_1250_cast_fp16")]; + tensor var_1252_equation_0 = const()[name = tensor("op_1252_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1252_cast_fp16 = einsum(equation = var_1252_equation_0, values = (var_1188_cast_fp16_7, var_1232_cast_fp16))[name = tensor("op_1252_cast_fp16")]; + tensor var_1254_equation_0 = const()[name = tensor("op_1254_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1254_cast_fp16 = einsum(equation = var_1254_equation_0, values = (var_1188_cast_fp16_8, var_1233_cast_fp16))[name = tensor("op_1254_cast_fp16")]; + tensor var_1256_equation_0 = const()[name = tensor("op_1256_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1256_cast_fp16 = einsum(equation = var_1256_equation_0, values = (var_1188_cast_fp16_9, var_1234_cast_fp16))[name = tensor("op_1256_cast_fp16")]; + tensor var_1258_equation_0 = const()[name = tensor("op_1258_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1258_cast_fp16 = einsum(equation = var_1258_equation_0, values = (var_1188_cast_fp16_10, var_1235_cast_fp16))[name = tensor("op_1258_cast_fp16")]; + tensor var_1260_equation_0 = const()[name = tensor("op_1260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1260_cast_fp16 = einsum(equation = var_1260_equation_0, values = (var_1188_cast_fp16_11, var_1236_cast_fp16))[name = tensor("op_1260_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1109, interleave = input_55_interleave_0, values = (var_1238_cast_fp16, var_1240_cast_fp16, var_1242_cast_fp16, var_1244_cast_fp16, var_1246_cast_fp16, var_1248_cast_fp16, var_1250_cast_fp16, var_1252_cast_fp16, var_1254_cast_fp16, var_1256_cast_fp16, var_1258_cast_fp16, var_1260_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1269_pad_type_0 = const()[name = tensor("op_1269_pad_type_0"), val = tensor("valid")]; + tensor var_1269_strides_0 = const()[name = tensor("op_1269_strides_0"), val = tensor([1, 1])]; + tensor var_1269_pad_0 = const()[name = tensor("op_1269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1269_dilations_0 = const()[name = tensor("op_1269_dilations_0"), val = tensor([1, 1])]; + tensor var_1269_groups_0 = const()[name = tensor("op_1269_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80636416)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81816128)))]; + tensor var_1269_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1269_dilations_0, groups = var_1269_groups_0, pad = var_1269_pad_0, pad_type = var_1269_pad_type_0, strides = var_1269_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1269_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1269_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81817728)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81819328)))]; + tensor var_1279_to_fp16 = const()[name = tensor("op_1279_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1279_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81820928)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86539584)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1305_pad_type_0 = const()[name = tensor("op_1305_pad_type_0"), val = tensor("valid")]; + tensor var_1305_strides_0 = const()[name = tensor("op_1305_strides_0"), val = tensor([1, 1])]; + tensor var_1305_pad_0 = const()[name = tensor("op_1305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1305_dilations_0 = const()[name = tensor("op_1305_dilations_0"), val = tensor([1, 1])]; + tensor var_1305_groups_0 = const()[name = tensor("op_1305_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86545792)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91264448)))]; + tensor var_1305_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1305_dilations_0, groups = var_1305_groups_0, pad = var_1305_pad_0, pad_type = var_1305_pad_type_0, strides = var_1305_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1305_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1305_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91266048)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91267648)))]; + tensor var_1330_to_fp16 = const()[name = tensor("op_1330_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1330_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1365_weight_0_to_fp16 = const()[name = tensor("op_1365_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91269248)))]; + tensor var_1365_bias_0_to_fp16 = const()[name = tensor("op_1365_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92448960)))]; + tensor var_1365_cast_fp16 = conv(bias = var_1365_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1365_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1365_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92450560)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1363_pad_type_0 = const()[name = tensor("op_1363_pad_type_0"), val = tensor("valid")]; + tensor var_1363_strides_0 = const()[name = tensor("op_1363_strides_0"), val = tensor([1, 1])]; + tensor var_1363_pad_0 = const()[name = tensor("op_1363_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1363_dilations_0 = const()[name = tensor("op_1363_dilations_0"), val = tensor([1, 1])]; + tensor var_1363_groups_0 = const()[name = tensor("op_1363_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93630272)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94809984)))]; + tensor var_1363_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1363_dilations_0, groups = var_1363_groups_0, pad = var_1363_pad_0, pad_type = var_1363_pad_type_0, strides = var_1363_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1366_axis_0 = const()[name = tensor("op_1366_axis_0"), val = tensor(1)]; + tensor var_1366_cast_fp16_0, tensor var_1366_cast_fp16_1, tensor var_1366_cast_fp16_2, tensor var_1366_cast_fp16_3, tensor var_1366_cast_fp16_4, tensor var_1366_cast_fp16_5, tensor var_1366_cast_fp16_6, tensor var_1366_cast_fp16_7, tensor var_1366_cast_fp16_8, tensor var_1366_cast_fp16_9, tensor var_1366_cast_fp16_10, tensor var_1366_cast_fp16_11 = split(axis = var_1366_axis_0, split_sizes = tile_18, x = var_1365_cast_fp16)[name = tensor("op_1366_cast_fp16")]; + tensor var_1379_perm_0 = const()[name = tensor("op_1379_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(3)]; + tensor var_1379_cast_fp16 = transpose(perm = var_1379_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_6")]; + tensor var_1380_cast_fp16_0, tensor var_1380_cast_fp16_1, tensor var_1380_cast_fp16_2, tensor var_1380_cast_fp16_3, tensor var_1380_cast_fp16_4, tensor var_1380_cast_fp16_5, tensor var_1380_cast_fp16_6, tensor var_1380_cast_fp16_7, tensor var_1380_cast_fp16_8, tensor var_1380_cast_fp16_9, tensor var_1380_cast_fp16_10, tensor var_1380_cast_fp16_11 = split(axis = var_1380_axis_0, split_sizes = tile_19, x = var_1379_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1393_axis_0 = const()[name = tensor("op_1393_axis_0"), val = tensor(1)]; + tensor var_1393_cast_fp16_0, tensor var_1393_cast_fp16_1, tensor var_1393_cast_fp16_2, tensor var_1393_cast_fp16_3, tensor var_1393_cast_fp16_4, tensor var_1393_cast_fp16_5, tensor var_1393_cast_fp16_6, tensor var_1393_cast_fp16_7, tensor var_1393_cast_fp16_8, tensor var_1393_cast_fp16_9, tensor var_1393_cast_fp16_10, tensor var_1393_cast_fp16_11 = split(axis = var_1393_axis_0, split_sizes = tile_20, x = var_1363_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1380_cast_fp16_0, var_1366_cast_fp16_0))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1380_cast_fp16_1, var_1366_cast_fp16_1))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1380_cast_fp16_2, var_1366_cast_fp16_2))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1380_cast_fp16_3, var_1366_cast_fp16_3))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1380_cast_fp16_4, var_1366_cast_fp16_4))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1380_cast_fp16_5, var_1366_cast_fp16_5))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1380_cast_fp16_6, var_1366_cast_fp16_6))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1380_cast_fp16_7, var_1366_cast_fp16_7))[name = tensor("aw_159_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1380_cast_fp16_8, var_1366_cast_fp16_8))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1380_cast_fp16_9, var_1366_cast_fp16_9))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1380_cast_fp16_10, var_1366_cast_fp16_10))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1380_cast_fp16_11, var_1366_cast_fp16_11))[name = tensor("aw_167_cast_fp16")]; + tensor var_1430_cast_fp16 = softmax(axis = var_1314, x = aw_145_cast_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor var_1431_cast_fp16 = softmax(axis = var_1314, x = aw_147_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1432_cast_fp16 = softmax(axis = var_1314, x = aw_149_cast_fp16)[name = tensor("op_1432_cast_fp16")]; + tensor var_1433_cast_fp16 = softmax(axis = var_1314, x = aw_151_cast_fp16)[name = tensor("op_1433_cast_fp16")]; + tensor var_1434_cast_fp16 = softmax(axis = var_1314, x = aw_153_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor var_1435_cast_fp16 = softmax(axis = var_1314, x = aw_155_cast_fp16)[name = tensor("op_1435_cast_fp16")]; + tensor var_1436_cast_fp16 = softmax(axis = var_1314, x = aw_157_cast_fp16)[name = tensor("op_1436_cast_fp16")]; + tensor var_1437_cast_fp16 = softmax(axis = var_1314, x = aw_159_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = softmax(axis = var_1314, x = aw_161_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor var_1439_cast_fp16 = softmax(axis = var_1314, x = aw_163_cast_fp16)[name = tensor("op_1439_cast_fp16")]; + tensor var_1440_cast_fp16 = softmax(axis = var_1314, x = aw_165_cast_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1441_cast_fp16 = softmax(axis = var_1314, x = aw_167_cast_fp16)[name = tensor("op_1441_cast_fp16")]; + tensor var_1443_equation_0 = const()[name = tensor("op_1443_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1443_cast_fp16 = einsum(equation = var_1443_equation_0, values = (var_1393_cast_fp16_0, var_1430_cast_fp16))[name = tensor("op_1443_cast_fp16")]; + tensor var_1445_equation_0 = const()[name = tensor("op_1445_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1445_cast_fp16 = einsum(equation = var_1445_equation_0, values = (var_1393_cast_fp16_1, var_1431_cast_fp16))[name = tensor("op_1445_cast_fp16")]; + tensor var_1447_equation_0 = const()[name = tensor("op_1447_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1447_cast_fp16 = einsum(equation = var_1447_equation_0, values = (var_1393_cast_fp16_2, var_1432_cast_fp16))[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_equation_0 = const()[name = tensor("op_1449_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1449_cast_fp16 = einsum(equation = var_1449_equation_0, values = (var_1393_cast_fp16_3, var_1433_cast_fp16))[name = tensor("op_1449_cast_fp16")]; + tensor var_1451_equation_0 = const()[name = tensor("op_1451_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1451_cast_fp16 = einsum(equation = var_1451_equation_0, values = (var_1393_cast_fp16_4, var_1434_cast_fp16))[name = tensor("op_1451_cast_fp16")]; + tensor var_1453_equation_0 = const()[name = tensor("op_1453_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1453_cast_fp16 = einsum(equation = var_1453_equation_0, values = (var_1393_cast_fp16_5, var_1435_cast_fp16))[name = tensor("op_1453_cast_fp16")]; + tensor var_1455_equation_0 = const()[name = tensor("op_1455_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1455_cast_fp16 = einsum(equation = var_1455_equation_0, values = (var_1393_cast_fp16_6, var_1436_cast_fp16))[name = tensor("op_1455_cast_fp16")]; + tensor var_1457_equation_0 = const()[name = tensor("op_1457_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1457_cast_fp16 = einsum(equation = var_1457_equation_0, values = (var_1393_cast_fp16_7, var_1437_cast_fp16))[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_equation_0 = const()[name = tensor("op_1459_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1459_cast_fp16 = einsum(equation = var_1459_equation_0, values = (var_1393_cast_fp16_8, var_1438_cast_fp16))[name = tensor("op_1459_cast_fp16")]; + tensor var_1461_equation_0 = const()[name = tensor("op_1461_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1461_cast_fp16 = einsum(equation = var_1461_equation_0, values = (var_1393_cast_fp16_9, var_1439_cast_fp16))[name = tensor("op_1461_cast_fp16")]; + tensor var_1463_equation_0 = const()[name = tensor("op_1463_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1463_cast_fp16 = einsum(equation = var_1463_equation_0, values = (var_1393_cast_fp16_10, var_1440_cast_fp16))[name = tensor("op_1463_cast_fp16")]; + tensor var_1465_equation_0 = const()[name = tensor("op_1465_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1465_cast_fp16 = einsum(equation = var_1465_equation_0, values = (var_1393_cast_fp16_11, var_1441_cast_fp16))[name = tensor("op_1465_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1314, interleave = input_65_interleave_0, values = (var_1443_cast_fp16, var_1445_cast_fp16, var_1447_cast_fp16, var_1449_cast_fp16, var_1451_cast_fp16, var_1453_cast_fp16, var_1455_cast_fp16, var_1457_cast_fp16, var_1459_cast_fp16, var_1461_cast_fp16, var_1463_cast_fp16, var_1465_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1474_pad_type_0 = const()[name = tensor("op_1474_pad_type_0"), val = tensor("valid")]; + tensor var_1474_strides_0 = const()[name = tensor("op_1474_strides_0"), val = tensor([1, 1])]; + tensor var_1474_pad_0 = const()[name = tensor("op_1474_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1474_dilations_0 = const()[name = tensor("op_1474_dilations_0"), val = tensor([1, 1])]; + tensor var_1474_groups_0 = const()[name = tensor("op_1474_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94811584)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95991296)))]; + tensor var_1474_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1474_dilations_0, groups = var_1474_groups_0, pad = var_1474_pad_0, pad_type = var_1474_pad_type_0, strides = var_1474_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1474_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1474_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95992896)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95994496)))]; + tensor var_1484_to_fp16 = const()[name = tensor("op_1484_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1484_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95996096)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100714752)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1510_pad_type_0 = const()[name = tensor("op_1510_pad_type_0"), val = tensor("valid")]; + tensor var_1510_strides_0 = const()[name = tensor("op_1510_strides_0"), val = tensor([1, 1])]; + tensor var_1510_pad_0 = const()[name = tensor("op_1510_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1510_dilations_0 = const()[name = tensor("op_1510_dilations_0"), val = tensor([1, 1])]; + tensor var_1510_groups_0 = const()[name = tensor("op_1510_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100720960)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105439616)))]; + tensor var_1510_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1510_dilations_0, groups = var_1510_groups_0, pad = var_1510_pad_0, pad_type = var_1510_pad_type_0, strides = var_1510_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1510_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105441216)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105442816)))]; + tensor var_1535_to_fp16 = const()[name = tensor("op_1535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_1535_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_1570_weight_0_to_fp16 = const()[name = tensor("op_1570_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105444416)))]; + tensor var_1570_bias_0_to_fp16 = const()[name = tensor("op_1570_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106624128)))]; + tensor var_1570_cast_fp16 = conv(bias = var_1570_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_1570_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106625728)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_1568_pad_type_0 = const()[name = tensor("op_1568_pad_type_0"), val = tensor("valid")]; + tensor var_1568_strides_0 = const()[name = tensor("op_1568_strides_0"), val = tensor([1, 1])]; + tensor var_1568_pad_0 = const()[name = tensor("op_1568_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1568_dilations_0 = const()[name = tensor("op_1568_dilations_0"), val = tensor([1, 1])]; + tensor var_1568_groups_0 = const()[name = tensor("op_1568_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107805440)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108985152)))]; + tensor var_1568_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_1568_dilations_0, groups = var_1568_groups_0, pad = var_1568_pad_0, pad_type = var_1568_pad_type_0, strides = var_1568_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1571_axis_0 = const()[name = tensor("op_1571_axis_0"), val = tensor(1)]; + tensor var_1571_cast_fp16_0, tensor var_1571_cast_fp16_1, tensor var_1571_cast_fp16_2, tensor var_1571_cast_fp16_3, tensor var_1571_cast_fp16_4, tensor var_1571_cast_fp16_5, tensor var_1571_cast_fp16_6, tensor var_1571_cast_fp16_7, tensor var_1571_cast_fp16_8, tensor var_1571_cast_fp16_9, tensor var_1571_cast_fp16_10, tensor var_1571_cast_fp16_11 = split(axis = var_1571_axis_0, split_sizes = tile_21, x = var_1570_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1584_perm_0 = const()[name = tensor("op_1584_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1585_axis_0 = const()[name = tensor("op_1585_axis_0"), val = tensor(3)]; + tensor var_1584_cast_fp16 = transpose(perm = var_1584_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_5")]; + tensor var_1585_cast_fp16_0, tensor var_1585_cast_fp16_1, tensor var_1585_cast_fp16_2, tensor var_1585_cast_fp16_3, tensor var_1585_cast_fp16_4, tensor var_1585_cast_fp16_5, tensor var_1585_cast_fp16_6, tensor var_1585_cast_fp16_7, tensor var_1585_cast_fp16_8, tensor var_1585_cast_fp16_9, tensor var_1585_cast_fp16_10, tensor var_1585_cast_fp16_11 = split(axis = var_1585_axis_0, split_sizes = tile_22, x = var_1584_cast_fp16)[name = tensor("op_1585_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1598_axis_0 = const()[name = tensor("op_1598_axis_0"), val = tensor(1)]; + tensor var_1598_cast_fp16_0, tensor var_1598_cast_fp16_1, tensor var_1598_cast_fp16_2, tensor var_1598_cast_fp16_3, tensor var_1598_cast_fp16_4, tensor var_1598_cast_fp16_5, tensor var_1598_cast_fp16_6, tensor var_1598_cast_fp16_7, tensor var_1598_cast_fp16_8, tensor var_1598_cast_fp16_9, tensor var_1598_cast_fp16_10, tensor var_1598_cast_fp16_11 = split(axis = var_1598_axis_0, split_sizes = tile_23, x = var_1568_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1585_cast_fp16_0, var_1571_cast_fp16_0))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1585_cast_fp16_1, var_1571_cast_fp16_1))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1585_cast_fp16_2, var_1571_cast_fp16_2))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1585_cast_fp16_3, var_1571_cast_fp16_3))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1585_cast_fp16_4, var_1571_cast_fp16_4))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1585_cast_fp16_5, var_1571_cast_fp16_5))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1585_cast_fp16_6, var_1571_cast_fp16_6))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1585_cast_fp16_7, var_1571_cast_fp16_7))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1585_cast_fp16_8, var_1571_cast_fp16_8))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1585_cast_fp16_9, var_1571_cast_fp16_9))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1585_cast_fp16_10, var_1571_cast_fp16_10))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1585_cast_fp16_11, var_1571_cast_fp16_11))[name = tensor("aw_191_cast_fp16")]; + tensor var_1635_cast_fp16 = softmax(axis = var_1519, x = aw_169_cast_fp16)[name = tensor("op_1635_cast_fp16")]; + tensor var_1636_cast_fp16 = softmax(axis = var_1519, x = aw_171_cast_fp16)[name = tensor("op_1636_cast_fp16")]; + tensor var_1637_cast_fp16 = softmax(axis = var_1519, x = aw_173_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor var_1638_cast_fp16 = softmax(axis = var_1519, x = aw_175_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_cast_fp16 = softmax(axis = var_1519, x = aw_177_cast_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1640_cast_fp16 = softmax(axis = var_1519, x = aw_179_cast_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor var_1641_cast_fp16 = softmax(axis = var_1519, x = aw_181_cast_fp16)[name = tensor("op_1641_cast_fp16")]; + tensor var_1642_cast_fp16 = softmax(axis = var_1519, x = aw_183_cast_fp16)[name = tensor("op_1642_cast_fp16")]; + tensor var_1643_cast_fp16 = softmax(axis = var_1519, x = aw_185_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor var_1644_cast_fp16 = softmax(axis = var_1519, x = aw_187_cast_fp16)[name = tensor("op_1644_cast_fp16")]; + tensor var_1645_cast_fp16 = softmax(axis = var_1519, x = aw_189_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_cast_fp16 = softmax(axis = var_1519, x = aw_191_cast_fp16)[name = tensor("op_1646_cast_fp16")]; + tensor var_1648_equation_0 = const()[name = tensor("op_1648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1648_cast_fp16 = einsum(equation = var_1648_equation_0, values = (var_1598_cast_fp16_0, var_1635_cast_fp16))[name = tensor("op_1648_cast_fp16")]; + tensor var_1650_equation_0 = const()[name = tensor("op_1650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1650_cast_fp16 = einsum(equation = var_1650_equation_0, values = (var_1598_cast_fp16_1, var_1636_cast_fp16))[name = tensor("op_1650_cast_fp16")]; + tensor var_1652_equation_0 = const()[name = tensor("op_1652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1652_cast_fp16 = einsum(equation = var_1652_equation_0, values = (var_1598_cast_fp16_2, var_1637_cast_fp16))[name = tensor("op_1652_cast_fp16")]; + tensor var_1654_equation_0 = const()[name = tensor("op_1654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1654_cast_fp16 = einsum(equation = var_1654_equation_0, values = (var_1598_cast_fp16_3, var_1638_cast_fp16))[name = tensor("op_1654_cast_fp16")]; + tensor var_1656_equation_0 = const()[name = tensor("op_1656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1656_cast_fp16 = einsum(equation = var_1656_equation_0, values = (var_1598_cast_fp16_4, var_1639_cast_fp16))[name = tensor("op_1656_cast_fp16")]; + tensor var_1658_equation_0 = const()[name = tensor("op_1658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1658_cast_fp16 = einsum(equation = var_1658_equation_0, values = (var_1598_cast_fp16_5, var_1640_cast_fp16))[name = tensor("op_1658_cast_fp16")]; + tensor var_1660_equation_0 = const()[name = tensor("op_1660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1660_cast_fp16 = einsum(equation = var_1660_equation_0, values = (var_1598_cast_fp16_6, var_1641_cast_fp16))[name = tensor("op_1660_cast_fp16")]; + tensor var_1662_equation_0 = const()[name = tensor("op_1662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1662_cast_fp16 = einsum(equation = var_1662_equation_0, values = (var_1598_cast_fp16_7, var_1642_cast_fp16))[name = tensor("op_1662_cast_fp16")]; + tensor var_1664_equation_0 = const()[name = tensor("op_1664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1664_cast_fp16 = einsum(equation = var_1664_equation_0, values = (var_1598_cast_fp16_8, var_1643_cast_fp16))[name = tensor("op_1664_cast_fp16")]; + tensor var_1666_equation_0 = const()[name = tensor("op_1666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1666_cast_fp16 = einsum(equation = var_1666_equation_0, values = (var_1598_cast_fp16_9, var_1644_cast_fp16))[name = tensor("op_1666_cast_fp16")]; + tensor var_1668_equation_0 = const()[name = tensor("op_1668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1668_cast_fp16 = einsum(equation = var_1668_equation_0, values = (var_1598_cast_fp16_10, var_1645_cast_fp16))[name = tensor("op_1668_cast_fp16")]; + tensor var_1670_equation_0 = const()[name = tensor("op_1670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1670_cast_fp16 = einsum(equation = var_1670_equation_0, values = (var_1598_cast_fp16_11, var_1646_cast_fp16))[name = tensor("op_1670_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_1519, interleave = input_75_interleave_0, values = (var_1648_cast_fp16, var_1650_cast_fp16, var_1652_cast_fp16, var_1654_cast_fp16, var_1656_cast_fp16, var_1658_cast_fp16, var_1660_cast_fp16, var_1662_cast_fp16, var_1664_cast_fp16, var_1666_cast_fp16, var_1668_cast_fp16, var_1670_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_1679_pad_type_0 = const()[name = tensor("op_1679_pad_type_0"), val = tensor("valid")]; + tensor var_1679_strides_0 = const()[name = tensor("op_1679_strides_0"), val = tensor([1, 1])]; + tensor var_1679_pad_0 = const()[name = tensor("op_1679_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1679_dilations_0 = const()[name = tensor("op_1679_dilations_0"), val = tensor([1, 1])]; + tensor var_1679_groups_0 = const()[name = tensor("op_1679_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108986752)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110166464)))]; + tensor var_1679_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_1679_dilations_0, groups = var_1679_groups_0, pad = var_1679_pad_0, pad_type = var_1679_pad_type_0, strides = var_1679_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_1679_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_1679_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110168064)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110169664)))]; + tensor var_1689_to_fp16 = const()[name = tensor("op_1689_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_1689_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110171264)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114889920)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1715_pad_type_0 = const()[name = tensor("op_1715_pad_type_0"), val = tensor("valid")]; + tensor var_1715_strides_0 = const()[name = tensor("op_1715_strides_0"), val = tensor([1, 1])]; + tensor var_1715_pad_0 = const()[name = tensor("op_1715_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1715_dilations_0 = const()[name = tensor("op_1715_dilations_0"), val = tensor([1, 1])]; + tensor var_1715_groups_0 = const()[name = tensor("op_1715_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114896128)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119614784)))]; + tensor var_1715_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_1715_dilations_0, groups = var_1715_groups_0, pad = var_1715_pad_0, pad_type = var_1715_pad_type_0, strides = var_1715_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1715_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1724 = const()[name = tensor("op_1724"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119616384)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119617984)))]; + tensor var_1740_to_fp16 = const()[name = tensor("op_1740_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_1740_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_1775_weight_0_to_fp16 = const()[name = tensor("op_1775_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119619584)))]; + tensor var_1775_bias_0_to_fp16 = const()[name = tensor("op_1775_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120799296)))]; + tensor var_1775_cast_fp16 = conv(bias = var_1775_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_1775_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_1775_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120800896)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_1773_pad_type_0 = const()[name = tensor("op_1773_pad_type_0"), val = tensor("valid")]; + tensor var_1773_strides_0 = const()[name = tensor("op_1773_strides_0"), val = tensor([1, 1])]; + tensor var_1773_pad_0 = const()[name = tensor("op_1773_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1773_dilations_0 = const()[name = tensor("op_1773_dilations_0"), val = tensor([1, 1])]; + tensor var_1773_groups_0 = const()[name = tensor("op_1773_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121980608)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123160320)))]; + tensor var_1773_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_1773_dilations_0, groups = var_1773_groups_0, pad = var_1773_pad_0, pad_type = var_1773_pad_type_0, strides = var_1773_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1776_axis_0 = const()[name = tensor("op_1776_axis_0"), val = tensor(1)]; + tensor var_1776_cast_fp16_0, tensor var_1776_cast_fp16_1, tensor var_1776_cast_fp16_2, tensor var_1776_cast_fp16_3, tensor var_1776_cast_fp16_4, tensor var_1776_cast_fp16_5, tensor var_1776_cast_fp16_6, tensor var_1776_cast_fp16_7, tensor var_1776_cast_fp16_8, tensor var_1776_cast_fp16_9, tensor var_1776_cast_fp16_10, tensor var_1776_cast_fp16_11 = split(axis = var_1776_axis_0, split_sizes = tile_24, x = var_1775_cast_fp16)[name = tensor("op_1776_cast_fp16")]; + tensor var_1789_perm_0 = const()[name = tensor("op_1789_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1790_axis_0 = const()[name = tensor("op_1790_axis_0"), val = tensor(3)]; + tensor var_1789_cast_fp16 = transpose(perm = var_1789_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_4")]; + tensor var_1790_cast_fp16_0, tensor var_1790_cast_fp16_1, tensor var_1790_cast_fp16_2, tensor var_1790_cast_fp16_3, tensor var_1790_cast_fp16_4, tensor var_1790_cast_fp16_5, tensor var_1790_cast_fp16_6, tensor var_1790_cast_fp16_7, tensor var_1790_cast_fp16_8, tensor var_1790_cast_fp16_9, tensor var_1790_cast_fp16_10, tensor var_1790_cast_fp16_11 = split(axis = var_1790_axis_0, split_sizes = tile_25, x = var_1789_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1803_axis_0 = const()[name = tensor("op_1803_axis_0"), val = tensor(1)]; + tensor var_1803_cast_fp16_0, tensor var_1803_cast_fp16_1, tensor var_1803_cast_fp16_2, tensor var_1803_cast_fp16_3, tensor var_1803_cast_fp16_4, tensor var_1803_cast_fp16_5, tensor var_1803_cast_fp16_6, tensor var_1803_cast_fp16_7, tensor var_1803_cast_fp16_8, tensor var_1803_cast_fp16_9, tensor var_1803_cast_fp16_10, tensor var_1803_cast_fp16_11 = split(axis = var_1803_axis_0, split_sizes = tile_26, x = var_1773_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1790_cast_fp16_0, var_1776_cast_fp16_0))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1790_cast_fp16_1, var_1776_cast_fp16_1))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1790_cast_fp16_2, var_1776_cast_fp16_2))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1790_cast_fp16_3, var_1776_cast_fp16_3))[name = tensor("aw_199_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1790_cast_fp16_4, var_1776_cast_fp16_4))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1790_cast_fp16_5, var_1776_cast_fp16_5))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1790_cast_fp16_6, var_1776_cast_fp16_6))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1790_cast_fp16_7, var_1776_cast_fp16_7))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1790_cast_fp16_8, var_1776_cast_fp16_8))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1790_cast_fp16_9, var_1776_cast_fp16_9))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1790_cast_fp16_10, var_1776_cast_fp16_10))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1790_cast_fp16_11, var_1776_cast_fp16_11))[name = tensor("aw_215_cast_fp16")]; + tensor var_1840_cast_fp16 = softmax(axis = var_1724, x = aw_193_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1841_cast_fp16 = softmax(axis = var_1724, x = aw_195_cast_fp16)[name = tensor("op_1841_cast_fp16")]; + tensor var_1842_cast_fp16 = softmax(axis = var_1724, x = aw_197_cast_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = softmax(axis = var_1724, x = aw_199_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = softmax(axis = var_1724, x = aw_201_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor var_1845_cast_fp16 = softmax(axis = var_1724, x = aw_203_cast_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor var_1846_cast_fp16 = softmax(axis = var_1724, x = aw_205_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1847_cast_fp16 = softmax(axis = var_1724, x = aw_207_cast_fp16)[name = tensor("op_1847_cast_fp16")]; + tensor var_1848_cast_fp16 = softmax(axis = var_1724, x = aw_209_cast_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = softmax(axis = var_1724, x = aw_211_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1850_cast_fp16 = softmax(axis = var_1724, x = aw_213_cast_fp16)[name = tensor("op_1850_cast_fp16")]; + tensor var_1851_cast_fp16 = softmax(axis = var_1724, x = aw_215_cast_fp16)[name = tensor("op_1851_cast_fp16")]; + tensor var_1853_equation_0 = const()[name = tensor("op_1853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1853_cast_fp16 = einsum(equation = var_1853_equation_0, values = (var_1803_cast_fp16_0, var_1840_cast_fp16))[name = tensor("op_1853_cast_fp16")]; + tensor var_1855_equation_0 = const()[name = tensor("op_1855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1855_cast_fp16 = einsum(equation = var_1855_equation_0, values = (var_1803_cast_fp16_1, var_1841_cast_fp16))[name = tensor("op_1855_cast_fp16")]; + tensor var_1857_equation_0 = const()[name = tensor("op_1857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1857_cast_fp16 = einsum(equation = var_1857_equation_0, values = (var_1803_cast_fp16_2, var_1842_cast_fp16))[name = tensor("op_1857_cast_fp16")]; + tensor var_1859_equation_0 = const()[name = tensor("op_1859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1859_cast_fp16 = einsum(equation = var_1859_equation_0, values = (var_1803_cast_fp16_3, var_1843_cast_fp16))[name = tensor("op_1859_cast_fp16")]; + tensor var_1861_equation_0 = const()[name = tensor("op_1861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1861_cast_fp16 = einsum(equation = var_1861_equation_0, values = (var_1803_cast_fp16_4, var_1844_cast_fp16))[name = tensor("op_1861_cast_fp16")]; + tensor var_1863_equation_0 = const()[name = tensor("op_1863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1863_cast_fp16 = einsum(equation = var_1863_equation_0, values = (var_1803_cast_fp16_5, var_1845_cast_fp16))[name = tensor("op_1863_cast_fp16")]; + tensor var_1865_equation_0 = const()[name = tensor("op_1865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1865_cast_fp16 = einsum(equation = var_1865_equation_0, values = (var_1803_cast_fp16_6, var_1846_cast_fp16))[name = tensor("op_1865_cast_fp16")]; + tensor var_1867_equation_0 = const()[name = tensor("op_1867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1867_cast_fp16 = einsum(equation = var_1867_equation_0, values = (var_1803_cast_fp16_7, var_1847_cast_fp16))[name = tensor("op_1867_cast_fp16")]; + tensor var_1869_equation_0 = const()[name = tensor("op_1869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1869_cast_fp16 = einsum(equation = var_1869_equation_0, values = (var_1803_cast_fp16_8, var_1848_cast_fp16))[name = tensor("op_1869_cast_fp16")]; + tensor var_1871_equation_0 = const()[name = tensor("op_1871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1871_cast_fp16 = einsum(equation = var_1871_equation_0, values = (var_1803_cast_fp16_9, var_1849_cast_fp16))[name = tensor("op_1871_cast_fp16")]; + tensor var_1873_equation_0 = const()[name = tensor("op_1873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1873_cast_fp16 = einsum(equation = var_1873_equation_0, values = (var_1803_cast_fp16_10, var_1850_cast_fp16))[name = tensor("op_1873_cast_fp16")]; + tensor var_1875_equation_0 = const()[name = tensor("op_1875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1875_cast_fp16 = einsum(equation = var_1875_equation_0, values = (var_1803_cast_fp16_11, var_1851_cast_fp16))[name = tensor("op_1875_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_1724, interleave = input_85_interleave_0, values = (var_1853_cast_fp16, var_1855_cast_fp16, var_1857_cast_fp16, var_1859_cast_fp16, var_1861_cast_fp16, var_1863_cast_fp16, var_1865_cast_fp16, var_1867_cast_fp16, var_1869_cast_fp16, var_1871_cast_fp16, var_1873_cast_fp16, var_1875_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_1884_pad_type_0 = const()[name = tensor("op_1884_pad_type_0"), val = tensor("valid")]; + tensor var_1884_strides_0 = const()[name = tensor("op_1884_strides_0"), val = tensor([1, 1])]; + tensor var_1884_pad_0 = const()[name = tensor("op_1884_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1884_dilations_0 = const()[name = tensor("op_1884_dilations_0"), val = tensor([1, 1])]; + tensor var_1884_groups_0 = const()[name = tensor("op_1884_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123161920)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124341632)))]; + tensor var_1884_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_1884_dilations_0, groups = var_1884_groups_0, pad = var_1884_pad_0, pad_type = var_1884_pad_type_0, strides = var_1884_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_1884_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124343232)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124344832)))]; + tensor var_1894_to_fp16 = const()[name = tensor("op_1894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_1894_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124346432)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129065088)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_1920_pad_type_0 = const()[name = tensor("op_1920_pad_type_0"), val = tensor("valid")]; + tensor var_1920_strides_0 = const()[name = tensor("op_1920_strides_0"), val = tensor([1, 1])]; + tensor var_1920_pad_0 = const()[name = tensor("op_1920_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1920_dilations_0 = const()[name = tensor("op_1920_dilations_0"), val = tensor([1, 1])]; + tensor var_1920_groups_0 = const()[name = tensor("op_1920_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129071296)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133789952)))]; + tensor var_1920_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_1920_dilations_0, groups = var_1920_groups_0, pad = var_1920_pad_0, pad_type = var_1920_pad_type_0, strides = var_1920_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_1920_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133791552)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133793152)))]; + tensor var_1945_to_fp16 = const()[name = tensor("op_1945_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_1945_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_1980_weight_0_to_fp16 = const()[name = tensor("op_1980_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133794752)))]; + tensor var_1980_bias_0_to_fp16 = const()[name = tensor("op_1980_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134974464)))]; + tensor var_1980_cast_fp16 = conv(bias = var_1980_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_1980_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134976064)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_1978_pad_type_0 = const()[name = tensor("op_1978_pad_type_0"), val = tensor("valid")]; + tensor var_1978_strides_0 = const()[name = tensor("op_1978_strides_0"), val = tensor([1, 1])]; + tensor var_1978_pad_0 = const()[name = tensor("op_1978_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1978_dilations_0 = const()[name = tensor("op_1978_dilations_0"), val = tensor([1, 1])]; + tensor var_1978_groups_0 = const()[name = tensor("op_1978_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136155776)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137335488)))]; + tensor var_1978_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_1978_dilations_0, groups = var_1978_groups_0, pad = var_1978_pad_0, pad_type = var_1978_pad_type_0, strides = var_1978_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1981_axis_0 = const()[name = tensor("op_1981_axis_0"), val = tensor(1)]; + tensor var_1981_cast_fp16_0, tensor var_1981_cast_fp16_1, tensor var_1981_cast_fp16_2, tensor var_1981_cast_fp16_3, tensor var_1981_cast_fp16_4, tensor var_1981_cast_fp16_5, tensor var_1981_cast_fp16_6, tensor var_1981_cast_fp16_7, tensor var_1981_cast_fp16_8, tensor var_1981_cast_fp16_9, tensor var_1981_cast_fp16_10, tensor var_1981_cast_fp16_11 = split(axis = var_1981_axis_0, split_sizes = tile_27, x = var_1980_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor var_1994_perm_0 = const()[name = tensor("op_1994_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1995_axis_0 = const()[name = tensor("op_1995_axis_0"), val = tensor(3)]; + tensor var_1994_cast_fp16 = transpose(perm = var_1994_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_3")]; + tensor var_1995_cast_fp16_0, tensor var_1995_cast_fp16_1, tensor var_1995_cast_fp16_2, tensor var_1995_cast_fp16_3, tensor var_1995_cast_fp16_4, tensor var_1995_cast_fp16_5, tensor var_1995_cast_fp16_6, tensor var_1995_cast_fp16_7, tensor var_1995_cast_fp16_8, tensor var_1995_cast_fp16_9, tensor var_1995_cast_fp16_10, tensor var_1995_cast_fp16_11 = split(axis = var_1995_axis_0, split_sizes = tile_28, x = var_1994_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_cast_fp16_0, tensor var_2008_cast_fp16_1, tensor var_2008_cast_fp16_2, tensor var_2008_cast_fp16_3, tensor var_2008_cast_fp16_4, tensor var_2008_cast_fp16_5, tensor var_2008_cast_fp16_6, tensor var_2008_cast_fp16_7, tensor var_2008_cast_fp16_8, tensor var_2008_cast_fp16_9, tensor var_2008_cast_fp16_10, tensor var_2008_cast_fp16_11 = split(axis = var_2008_axis_0, split_sizes = tile_29, x = var_1978_cast_fp16)[name = tensor("op_2008_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1995_cast_fp16_0, var_1981_cast_fp16_0))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1995_cast_fp16_1, var_1981_cast_fp16_1))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1995_cast_fp16_2, var_1981_cast_fp16_2))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1995_cast_fp16_3, var_1981_cast_fp16_3))[name = tensor("aw_223_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1995_cast_fp16_4, var_1981_cast_fp16_4))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1995_cast_fp16_5, var_1981_cast_fp16_5))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1995_cast_fp16_6, var_1981_cast_fp16_6))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1995_cast_fp16_7, var_1981_cast_fp16_7))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1995_cast_fp16_8, var_1981_cast_fp16_8))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1995_cast_fp16_9, var_1981_cast_fp16_9))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1995_cast_fp16_10, var_1981_cast_fp16_10))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1995_cast_fp16_11, var_1981_cast_fp16_11))[name = tensor("aw_239_cast_fp16")]; + tensor var_2045_cast_fp16 = softmax(axis = var_1929, x = aw_217_cast_fp16)[name = tensor("op_2045_cast_fp16")]; + tensor var_2046_cast_fp16 = softmax(axis = var_1929, x = aw_219_cast_fp16)[name = tensor("op_2046_cast_fp16")]; + tensor var_2047_cast_fp16 = softmax(axis = var_1929, x = aw_221_cast_fp16)[name = tensor("op_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = softmax(axis = var_1929, x = aw_223_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_cast_fp16 = softmax(axis = var_1929, x = aw_225_cast_fp16)[name = tensor("op_2049_cast_fp16")]; + tensor var_2050_cast_fp16 = softmax(axis = var_1929, x = aw_227_cast_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2051_cast_fp16 = softmax(axis = var_1929, x = aw_229_cast_fp16)[name = tensor("op_2051_cast_fp16")]; + tensor var_2052_cast_fp16 = softmax(axis = var_1929, x = aw_231_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor var_2053_cast_fp16 = softmax(axis = var_1929, x = aw_233_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2054_cast_fp16 = softmax(axis = var_1929, x = aw_235_cast_fp16)[name = tensor("op_2054_cast_fp16")]; + tensor var_2055_cast_fp16 = softmax(axis = var_1929, x = aw_237_cast_fp16)[name = tensor("op_2055_cast_fp16")]; + tensor var_2056_cast_fp16 = softmax(axis = var_1929, x = aw_239_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor var_2058_equation_0 = const()[name = tensor("op_2058_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2058_cast_fp16 = einsum(equation = var_2058_equation_0, values = (var_2008_cast_fp16_0, var_2045_cast_fp16))[name = tensor("op_2058_cast_fp16")]; + tensor var_2060_equation_0 = const()[name = tensor("op_2060_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2060_cast_fp16 = einsum(equation = var_2060_equation_0, values = (var_2008_cast_fp16_1, var_2046_cast_fp16))[name = tensor("op_2060_cast_fp16")]; + tensor var_2062_equation_0 = const()[name = tensor("op_2062_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2062_cast_fp16 = einsum(equation = var_2062_equation_0, values = (var_2008_cast_fp16_2, var_2047_cast_fp16))[name = tensor("op_2062_cast_fp16")]; + tensor var_2064_equation_0 = const()[name = tensor("op_2064_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2064_cast_fp16 = einsum(equation = var_2064_equation_0, values = (var_2008_cast_fp16_3, var_2048_cast_fp16))[name = tensor("op_2064_cast_fp16")]; + tensor var_2066_equation_0 = const()[name = tensor("op_2066_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2066_cast_fp16 = einsum(equation = var_2066_equation_0, values = (var_2008_cast_fp16_4, var_2049_cast_fp16))[name = tensor("op_2066_cast_fp16")]; + tensor var_2068_equation_0 = const()[name = tensor("op_2068_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2068_cast_fp16 = einsum(equation = var_2068_equation_0, values = (var_2008_cast_fp16_5, var_2050_cast_fp16))[name = tensor("op_2068_cast_fp16")]; + tensor var_2070_equation_0 = const()[name = tensor("op_2070_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2070_cast_fp16 = einsum(equation = var_2070_equation_0, values = (var_2008_cast_fp16_6, var_2051_cast_fp16))[name = tensor("op_2070_cast_fp16")]; + tensor var_2072_equation_0 = const()[name = tensor("op_2072_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2072_cast_fp16 = einsum(equation = var_2072_equation_0, values = (var_2008_cast_fp16_7, var_2052_cast_fp16))[name = tensor("op_2072_cast_fp16")]; + tensor var_2074_equation_0 = const()[name = tensor("op_2074_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2074_cast_fp16 = einsum(equation = var_2074_equation_0, values = (var_2008_cast_fp16_8, var_2053_cast_fp16))[name = tensor("op_2074_cast_fp16")]; + tensor var_2076_equation_0 = const()[name = tensor("op_2076_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2076_cast_fp16 = einsum(equation = var_2076_equation_0, values = (var_2008_cast_fp16_9, var_2054_cast_fp16))[name = tensor("op_2076_cast_fp16")]; + tensor var_2078_equation_0 = const()[name = tensor("op_2078_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2078_cast_fp16 = einsum(equation = var_2078_equation_0, values = (var_2008_cast_fp16_10, var_2055_cast_fp16))[name = tensor("op_2078_cast_fp16")]; + tensor var_2080_equation_0 = const()[name = tensor("op_2080_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2080_cast_fp16 = einsum(equation = var_2080_equation_0, values = (var_2008_cast_fp16_11, var_2056_cast_fp16))[name = tensor("op_2080_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_1929, interleave = input_95_interleave_0, values = (var_2058_cast_fp16, var_2060_cast_fp16, var_2062_cast_fp16, var_2064_cast_fp16, var_2066_cast_fp16, var_2068_cast_fp16, var_2070_cast_fp16, var_2072_cast_fp16, var_2074_cast_fp16, var_2076_cast_fp16, var_2078_cast_fp16, var_2080_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2089_pad_type_0 = const()[name = tensor("op_2089_pad_type_0"), val = tensor("valid")]; + tensor var_2089_strides_0 = const()[name = tensor("op_2089_strides_0"), val = tensor([1, 1])]; + tensor var_2089_pad_0 = const()[name = tensor("op_2089_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2089_dilations_0 = const()[name = tensor("op_2089_dilations_0"), val = tensor([1, 1])]; + tensor var_2089_groups_0 = const()[name = tensor("op_2089_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137337088)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138516800)))]; + tensor var_2089_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2089_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2089_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138518400)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138520000)))]; + tensor var_2099_to_fp16 = const()[name = tensor("op_2099_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2099_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138521600)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143240256)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2125_pad_type_0 = const()[name = tensor("op_2125_pad_type_0"), val = tensor("valid")]; + tensor var_2125_strides_0 = const()[name = tensor("op_2125_strides_0"), val = tensor([1, 1])]; + tensor var_2125_pad_0 = const()[name = tensor("op_2125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2125_dilations_0 = const()[name = tensor("op_2125_dilations_0"), val = tensor([1, 1])]; + tensor var_2125_groups_0 = const()[name = tensor("op_2125_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143246464)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147965120)))]; + tensor var_2125_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2125_dilations_0, groups = var_2125_groups_0, pad = var_2125_pad_0, pad_type = var_2125_pad_type_0, strides = var_2125_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2125_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147966720)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147968320)))]; + tensor var_2150_to_fp16 = const()[name = tensor("op_2150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2150_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2185_weight_0_to_fp16 = const()[name = tensor("op_2185_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147969920)))]; + tensor var_2185_bias_0_to_fp16 = const()[name = tensor("op_2185_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149149632)))]; + tensor var_2185_cast_fp16 = conv(bias = var_2185_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2185_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149151232)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2183_pad_type_0 = const()[name = tensor("op_2183_pad_type_0"), val = tensor("valid")]; + tensor var_2183_strides_0 = const()[name = tensor("op_2183_strides_0"), val = tensor([1, 1])]; + tensor var_2183_pad_0 = const()[name = tensor("op_2183_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2183_dilations_0 = const()[name = tensor("op_2183_dilations_0"), val = tensor([1, 1])]; + tensor var_2183_groups_0 = const()[name = tensor("op_2183_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150330944)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151510656)))]; + tensor var_2183_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2183_dilations_0, groups = var_2183_groups_0, pad = var_2183_pad_0, pad_type = var_2183_pad_type_0, strides = var_2183_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2183_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2186_axis_0 = const()[name = tensor("op_2186_axis_0"), val = tensor(1)]; + tensor var_2186_cast_fp16_0, tensor var_2186_cast_fp16_1, tensor var_2186_cast_fp16_2, tensor var_2186_cast_fp16_3, tensor var_2186_cast_fp16_4, tensor var_2186_cast_fp16_5, tensor var_2186_cast_fp16_6, tensor var_2186_cast_fp16_7, tensor var_2186_cast_fp16_8, tensor var_2186_cast_fp16_9, tensor var_2186_cast_fp16_10, tensor var_2186_cast_fp16_11 = split(axis = var_2186_axis_0, split_sizes = tile_30, x = var_2185_cast_fp16)[name = tensor("op_2186_cast_fp16")]; + tensor var_2199_perm_0 = const()[name = tensor("op_2199_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2200_axis_0 = const()[name = tensor("op_2200_axis_0"), val = tensor(3)]; + tensor var_2199_cast_fp16 = transpose(perm = var_2199_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_2")]; + tensor var_2200_cast_fp16_0, tensor var_2200_cast_fp16_1, tensor var_2200_cast_fp16_2, tensor var_2200_cast_fp16_3, tensor var_2200_cast_fp16_4, tensor var_2200_cast_fp16_5, tensor var_2200_cast_fp16_6, tensor var_2200_cast_fp16_7, tensor var_2200_cast_fp16_8, tensor var_2200_cast_fp16_9, tensor var_2200_cast_fp16_10, tensor var_2200_cast_fp16_11 = split(axis = var_2200_axis_0, split_sizes = tile_31, x = var_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2213_axis_0 = const()[name = tensor("op_2213_axis_0"), val = tensor(1)]; + tensor var_2213_cast_fp16_0, tensor var_2213_cast_fp16_1, tensor var_2213_cast_fp16_2, tensor var_2213_cast_fp16_3, tensor var_2213_cast_fp16_4, tensor var_2213_cast_fp16_5, tensor var_2213_cast_fp16_6, tensor var_2213_cast_fp16_7, tensor var_2213_cast_fp16_8, tensor var_2213_cast_fp16_9, tensor var_2213_cast_fp16_10, tensor var_2213_cast_fp16_11 = split(axis = var_2213_axis_0, split_sizes = tile_32, x = var_2183_cast_fp16)[name = tensor("op_2213_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_2200_cast_fp16_0, var_2186_cast_fp16_0))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_2200_cast_fp16_1, var_2186_cast_fp16_1))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_2200_cast_fp16_2, var_2186_cast_fp16_2))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_2200_cast_fp16_3, var_2186_cast_fp16_3))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_2200_cast_fp16_4, var_2186_cast_fp16_4))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_2200_cast_fp16_5, var_2186_cast_fp16_5))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_2200_cast_fp16_6, var_2186_cast_fp16_6))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_2200_cast_fp16_7, var_2186_cast_fp16_7))[name = tensor("aw_255_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_2200_cast_fp16_8, var_2186_cast_fp16_8))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_2200_cast_fp16_9, var_2186_cast_fp16_9))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_2200_cast_fp16_10, var_2186_cast_fp16_10))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_2200_cast_fp16_11, var_2186_cast_fp16_11))[name = tensor("aw_263_cast_fp16")]; + tensor var_2250_cast_fp16 = softmax(axis = var_2134, x = aw_241_cast_fp16)[name = tensor("op_2250_cast_fp16")]; + tensor var_2251_cast_fp16 = softmax(axis = var_2134, x = aw_243_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor var_2252_cast_fp16 = softmax(axis = var_2134, x = aw_245_cast_fp16)[name = tensor("op_2252_cast_fp16")]; + tensor var_2253_cast_fp16 = softmax(axis = var_2134, x = aw_247_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor var_2254_cast_fp16 = softmax(axis = var_2134, x = aw_249_cast_fp16)[name = tensor("op_2254_cast_fp16")]; + tensor var_2255_cast_fp16 = softmax(axis = var_2134, x = aw_251_cast_fp16)[name = tensor("op_2255_cast_fp16")]; + tensor var_2256_cast_fp16 = softmax(axis = var_2134, x = aw_253_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor var_2257_cast_fp16 = softmax(axis = var_2134, x = aw_255_cast_fp16)[name = tensor("op_2257_cast_fp16")]; + tensor var_2258_cast_fp16 = softmax(axis = var_2134, x = aw_257_cast_fp16)[name = tensor("op_2258_cast_fp16")]; + tensor var_2259_cast_fp16 = softmax(axis = var_2134, x = aw_259_cast_fp16)[name = tensor("op_2259_cast_fp16")]; + tensor var_2260_cast_fp16 = softmax(axis = var_2134, x = aw_261_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = softmax(axis = var_2134, x = aw_263_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor var_2263_equation_0 = const()[name = tensor("op_2263_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2263_cast_fp16 = einsum(equation = var_2263_equation_0, values = (var_2213_cast_fp16_0, var_2250_cast_fp16))[name = tensor("op_2263_cast_fp16")]; + tensor var_2265_equation_0 = const()[name = tensor("op_2265_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2265_cast_fp16 = einsum(equation = var_2265_equation_0, values = (var_2213_cast_fp16_1, var_2251_cast_fp16))[name = tensor("op_2265_cast_fp16")]; + tensor var_2267_equation_0 = const()[name = tensor("op_2267_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2267_cast_fp16 = einsum(equation = var_2267_equation_0, values = (var_2213_cast_fp16_2, var_2252_cast_fp16))[name = tensor("op_2267_cast_fp16")]; + tensor var_2269_equation_0 = const()[name = tensor("op_2269_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2269_cast_fp16 = einsum(equation = var_2269_equation_0, values = (var_2213_cast_fp16_3, var_2253_cast_fp16))[name = tensor("op_2269_cast_fp16")]; + tensor var_2271_equation_0 = const()[name = tensor("op_2271_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2271_cast_fp16 = einsum(equation = var_2271_equation_0, values = (var_2213_cast_fp16_4, var_2254_cast_fp16))[name = tensor("op_2271_cast_fp16")]; + tensor var_2273_equation_0 = const()[name = tensor("op_2273_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2273_cast_fp16 = einsum(equation = var_2273_equation_0, values = (var_2213_cast_fp16_5, var_2255_cast_fp16))[name = tensor("op_2273_cast_fp16")]; + tensor var_2275_equation_0 = const()[name = tensor("op_2275_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2275_cast_fp16 = einsum(equation = var_2275_equation_0, values = (var_2213_cast_fp16_6, var_2256_cast_fp16))[name = tensor("op_2275_cast_fp16")]; + tensor var_2277_equation_0 = const()[name = tensor("op_2277_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2277_cast_fp16 = einsum(equation = var_2277_equation_0, values = (var_2213_cast_fp16_7, var_2257_cast_fp16))[name = tensor("op_2277_cast_fp16")]; + tensor var_2279_equation_0 = const()[name = tensor("op_2279_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2279_cast_fp16 = einsum(equation = var_2279_equation_0, values = (var_2213_cast_fp16_8, var_2258_cast_fp16))[name = tensor("op_2279_cast_fp16")]; + tensor var_2281_equation_0 = const()[name = tensor("op_2281_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2281_cast_fp16 = einsum(equation = var_2281_equation_0, values = (var_2213_cast_fp16_9, var_2259_cast_fp16))[name = tensor("op_2281_cast_fp16")]; + tensor var_2283_equation_0 = const()[name = tensor("op_2283_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2283_cast_fp16 = einsum(equation = var_2283_equation_0, values = (var_2213_cast_fp16_10, var_2260_cast_fp16))[name = tensor("op_2283_cast_fp16")]; + tensor var_2285_equation_0 = const()[name = tensor("op_2285_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2285_cast_fp16 = einsum(equation = var_2285_equation_0, values = (var_2213_cast_fp16_11, var_2261_cast_fp16))[name = tensor("op_2285_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2134, interleave = input_105_interleave_0, values = (var_2263_cast_fp16, var_2265_cast_fp16, var_2267_cast_fp16, var_2269_cast_fp16, var_2271_cast_fp16, var_2273_cast_fp16, var_2275_cast_fp16, var_2277_cast_fp16, var_2279_cast_fp16, var_2281_cast_fp16, var_2283_cast_fp16, var_2285_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_2294_pad_type_0 = const()[name = tensor("op_2294_pad_type_0"), val = tensor("valid")]; + tensor var_2294_strides_0 = const()[name = tensor("op_2294_strides_0"), val = tensor([1, 1])]; + tensor var_2294_pad_0 = const()[name = tensor("op_2294_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2294_dilations_0 = const()[name = tensor("op_2294_dilations_0"), val = tensor([1, 1])]; + tensor var_2294_groups_0 = const()[name = tensor("op_2294_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151512256)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152691968)))]; + tensor var_2294_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_2294_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152693568)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152695168)))]; + tensor var_2304_to_fp16 = const()[name = tensor("op_2304_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_2304_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152696768)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157415424)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_2330_pad_type_0 = const()[name = tensor("op_2330_pad_type_0"), val = tensor("valid")]; + tensor var_2330_strides_0 = const()[name = tensor("op_2330_strides_0"), val = tensor([1, 1])]; + tensor var_2330_pad_0 = const()[name = tensor("op_2330_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2330_dilations_0 = const()[name = tensor("op_2330_dilations_0"), val = tensor([1, 1])]; + tensor var_2330_groups_0 = const()[name = tensor("op_2330_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157421632)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162140288)))]; + tensor var_2330_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_2330_dilations_0, groups = var_2330_groups_0, pad = var_2330_pad_0, pad_type = var_2330_pad_type_0, strides = var_2330_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_2330_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_2330_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_2339 = const()[name = tensor("op_2339"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162141888)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162143488)))]; + tensor var_2355_to_fp16 = const()[name = tensor("op_2355_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_2355_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_2390_weight_0_to_fp16 = const()[name = tensor("op_2390_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162145088)))]; + tensor var_2390_bias_0_to_fp16 = const()[name = tensor("op_2390_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163324800)))]; + tensor var_2390_cast_fp16 = conv(bias = var_2390_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_2390_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163326400)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_2388_pad_type_0 = const()[name = tensor("op_2388_pad_type_0"), val = tensor("valid")]; + tensor var_2388_strides_0 = const()[name = tensor("op_2388_strides_0"), val = tensor([1, 1])]; + tensor var_2388_pad_0 = const()[name = tensor("op_2388_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2388_dilations_0 = const()[name = tensor("op_2388_dilations_0"), val = tensor([1, 1])]; + tensor var_2388_groups_0 = const()[name = tensor("op_2388_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164506112)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165685824)))]; + tensor var_2388_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_2388_dilations_0, groups = var_2388_groups_0, pad = var_2388_pad_0, pad_type = var_2388_pad_type_0, strides = var_2388_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2391_axis_0 = const()[name = tensor("op_2391_axis_0"), val = tensor(1)]; + tensor var_2391_cast_fp16_0, tensor var_2391_cast_fp16_1, tensor var_2391_cast_fp16_2, tensor var_2391_cast_fp16_3, tensor var_2391_cast_fp16_4, tensor var_2391_cast_fp16_5, tensor var_2391_cast_fp16_6, tensor var_2391_cast_fp16_7, tensor var_2391_cast_fp16_8, tensor var_2391_cast_fp16_9, tensor var_2391_cast_fp16_10, tensor var_2391_cast_fp16_11 = split(axis = var_2391_axis_0, split_sizes = tile_33, x = var_2390_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2404_perm_0 = const()[name = tensor("op_2404_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2405_axis_0 = const()[name = tensor("op_2405_axis_0"), val = tensor(3)]; + tensor var_2404_cast_fp16 = transpose(perm = var_2404_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_2405_cast_fp16_0, tensor var_2405_cast_fp16_1, tensor var_2405_cast_fp16_2, tensor var_2405_cast_fp16_3, tensor var_2405_cast_fp16_4, tensor var_2405_cast_fp16_5, tensor var_2405_cast_fp16_6, tensor var_2405_cast_fp16_7, tensor var_2405_cast_fp16_8, tensor var_2405_cast_fp16_9, tensor var_2405_cast_fp16_10, tensor var_2405_cast_fp16_11 = split(axis = var_2405_axis_0, split_sizes = tile_34, x = var_2404_cast_fp16)[name = tensor("op_2405_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2418_axis_0 = const()[name = tensor("op_2418_axis_0"), val = tensor(1)]; + tensor var_2418_cast_fp16_0, tensor var_2418_cast_fp16_1, tensor var_2418_cast_fp16_2, tensor var_2418_cast_fp16_3, tensor var_2418_cast_fp16_4, tensor var_2418_cast_fp16_5, tensor var_2418_cast_fp16_6, tensor var_2418_cast_fp16_7, tensor var_2418_cast_fp16_8, tensor var_2418_cast_fp16_9, tensor var_2418_cast_fp16_10, tensor var_2418_cast_fp16_11 = split(axis = var_2418_axis_0, split_sizes = tile_35, x = var_2388_cast_fp16)[name = tensor("op_2418_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_2405_cast_fp16_0, var_2391_cast_fp16_0))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_2405_cast_fp16_1, var_2391_cast_fp16_1))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_2405_cast_fp16_2, var_2391_cast_fp16_2))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_2405_cast_fp16_3, var_2391_cast_fp16_3))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_2405_cast_fp16_4, var_2391_cast_fp16_4))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_2405_cast_fp16_5, var_2391_cast_fp16_5))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_2405_cast_fp16_6, var_2391_cast_fp16_6))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_2405_cast_fp16_7, var_2391_cast_fp16_7))[name = tensor("aw_279_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2405_cast_fp16_8, var_2391_cast_fp16_8))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2405_cast_fp16_9, var_2391_cast_fp16_9))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2405_cast_fp16_10, var_2391_cast_fp16_10))[name = tensor("aw_285_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_2405_cast_fp16_11, var_2391_cast_fp16_11))[name = tensor("aw_cast_fp16")]; + tensor var_2455_cast_fp16 = softmax(axis = var_2339, x = aw_265_cast_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor var_2456_cast_fp16 = softmax(axis = var_2339, x = aw_267_cast_fp16)[name = tensor("op_2456_cast_fp16")]; + tensor var_2457_cast_fp16 = softmax(axis = var_2339, x = aw_269_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor var_2458_cast_fp16 = softmax(axis = var_2339, x = aw_271_cast_fp16)[name = tensor("op_2458_cast_fp16")]; + tensor var_2459_cast_fp16 = softmax(axis = var_2339, x = aw_273_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2460_cast_fp16 = softmax(axis = var_2339, x = aw_275_cast_fp16)[name = tensor("op_2460_cast_fp16")]; + tensor var_2461_cast_fp16 = softmax(axis = var_2339, x = aw_277_cast_fp16)[name = tensor("op_2461_cast_fp16")]; + tensor var_2462_cast_fp16 = softmax(axis = var_2339, x = aw_279_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = softmax(axis = var_2339, x = aw_281_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor var_2464_cast_fp16 = softmax(axis = var_2339, x = aw_283_cast_fp16)[name = tensor("op_2464_cast_fp16")]; + tensor var_2465_cast_fp16 = softmax(axis = var_2339, x = aw_285_cast_fp16)[name = tensor("op_2465_cast_fp16")]; + tensor var_2466_cast_fp16 = softmax(axis = var_2339, x = aw_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2468_equation_0 = const()[name = tensor("op_2468_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2468_cast_fp16 = einsum(equation = var_2468_equation_0, values = (var_2418_cast_fp16_0, var_2455_cast_fp16))[name = tensor("op_2468_cast_fp16")]; + tensor var_2470_equation_0 = const()[name = tensor("op_2470_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2470_cast_fp16 = einsum(equation = var_2470_equation_0, values = (var_2418_cast_fp16_1, var_2456_cast_fp16))[name = tensor("op_2470_cast_fp16")]; + tensor var_2472_equation_0 = const()[name = tensor("op_2472_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2472_cast_fp16 = einsum(equation = var_2472_equation_0, values = (var_2418_cast_fp16_2, var_2457_cast_fp16))[name = tensor("op_2472_cast_fp16")]; + tensor var_2474_equation_0 = const()[name = tensor("op_2474_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2474_cast_fp16 = einsum(equation = var_2474_equation_0, values = (var_2418_cast_fp16_3, var_2458_cast_fp16))[name = tensor("op_2474_cast_fp16")]; + tensor var_2476_equation_0 = const()[name = tensor("op_2476_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2476_cast_fp16 = einsum(equation = var_2476_equation_0, values = (var_2418_cast_fp16_4, var_2459_cast_fp16))[name = tensor("op_2476_cast_fp16")]; + tensor var_2478_equation_0 = const()[name = tensor("op_2478_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2478_cast_fp16 = einsum(equation = var_2478_equation_0, values = (var_2418_cast_fp16_5, var_2460_cast_fp16))[name = tensor("op_2478_cast_fp16")]; + tensor var_2480_equation_0 = const()[name = tensor("op_2480_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2480_cast_fp16 = einsum(equation = var_2480_equation_0, values = (var_2418_cast_fp16_6, var_2461_cast_fp16))[name = tensor("op_2480_cast_fp16")]; + tensor var_2482_equation_0 = const()[name = tensor("op_2482_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2482_cast_fp16 = einsum(equation = var_2482_equation_0, values = (var_2418_cast_fp16_7, var_2462_cast_fp16))[name = tensor("op_2482_cast_fp16")]; + tensor var_2484_equation_0 = const()[name = tensor("op_2484_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2484_cast_fp16 = einsum(equation = var_2484_equation_0, values = (var_2418_cast_fp16_8, var_2463_cast_fp16))[name = tensor("op_2484_cast_fp16")]; + tensor var_2486_equation_0 = const()[name = tensor("op_2486_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2486_cast_fp16 = einsum(equation = var_2486_equation_0, values = (var_2418_cast_fp16_9, var_2464_cast_fp16))[name = tensor("op_2486_cast_fp16")]; + tensor var_2488_equation_0 = const()[name = tensor("op_2488_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2488_cast_fp16 = einsum(equation = var_2488_equation_0, values = (var_2418_cast_fp16_10, var_2465_cast_fp16))[name = tensor("op_2488_cast_fp16")]; + tensor var_2490_equation_0 = const()[name = tensor("op_2490_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2490_cast_fp16 = einsum(equation = var_2490_equation_0, values = (var_2418_cast_fp16_11, var_2466_cast_fp16))[name = tensor("op_2490_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_2339, interleave = input_115_interleave_0, values = (var_2468_cast_fp16, var_2470_cast_fp16, var_2472_cast_fp16, var_2474_cast_fp16, var_2476_cast_fp16, var_2478_cast_fp16, var_2480_cast_fp16, var_2482_cast_fp16, var_2484_cast_fp16, var_2486_cast_fp16, var_2488_cast_fp16, var_2490_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_2499_pad_type_0 = const()[name = tensor("op_2499_pad_type_0"), val = tensor("valid")]; + tensor var_2499_strides_0 = const()[name = tensor("op_2499_strides_0"), val = tensor([1, 1])]; + tensor var_2499_pad_0 = const()[name = tensor("op_2499_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2499_dilations_0 = const()[name = tensor("op_2499_dilations_0"), val = tensor([1, 1])]; + tensor var_2499_groups_0 = const()[name = tensor("op_2499_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165687424)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166867136)))]; + tensor var_2499_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_2499_dilations_0, groups = var_2499_groups_0, pad = var_2499_pad_0, pad_type = var_2499_pad_type_0, strides = var_2499_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_2499_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166868736)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166870336)))]; + tensor var_2509_to_fp16 = const()[name = tensor("op_2509_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_2509_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166871936)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171590592)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_119_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_2535_pad_type_0 = const()[name = tensor("op_2535_pad_type_0"), val = tensor("valid")]; + tensor var_2535_strides_0 = const()[name = tensor("op_2535_strides_0"), val = tensor([1, 1])]; + tensor var_2535_pad_0 = const()[name = tensor("op_2535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2535_dilations_0 = const()[name = tensor("op_2535_dilations_0"), val = tensor([1, 1])]; + tensor var_2535_groups_0 = const()[name = tensor("op_2535_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171596800)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176315456)))]; + tensor var_2535_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_2535_dilations_0, groups = var_2535_groups_0, pad = var_2535_pad_0, pad_type = var_2535_pad_type_0, strides = var_2535_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_2535_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_2535_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176317056)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176318656)))]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_2549_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_2560_axes_0 = const()[name = tensor("op_2560_axes_0"), val = tensor([2])]; + tensor var_2560_cast_fp16 = squeeze(axes = var_2560_axes_0, x = x_cast_fp16)[name = tensor("op_2560_cast_fp16")]; + tensor var_2563_perm_0 = const()[name = tensor("op_2563_perm_0"), val = tensor([0, 2, 1])]; + tensor var_2563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2563_cast_fp16 = transpose(perm = var_2563_perm_0, x = var_2560_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_2563_cast_fp16_to_fp32_dtype_0, x = var_2563_cast_fp16)[name = tensor("cast_51")]; + } -> (output); +} \ No newline at end of file diff --git a/small.en/ggml-small.en-encoder.mlmodelc/weights/weight.bin b/small.en/ggml-small.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f1d3a28e11bdb37e8dcd5967ca78c5e0b5b2a3e1 --- /dev/null +++ b/small.en/ggml-small.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57d5901ae064655c83da5a7281e643d56974fd875e986e6bd7997038e65e428a +size 176320256 diff --git a/small.en/ggml-small.en.bin b/small.en/ggml-small.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..eaeeb6d63378cf6515ff2c1cb4e33486ae6bcc2f --- /dev/null +++ b/small.en/ggml-small.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6138d6d58ecc8322097e0f987c32f1be8bb0a18532a3f88f734d1bbf9c41e5d +size 487614201 diff --git a/small/.DS_Store b/small/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..43ac60a654a7916477d24dfceb59df85052f2077 Binary files /dev/null and b/small/.DS_Store differ diff --git a/small/ggml-small-encoder.mlmodelc/analytics/coremldata.bin b/small/ggml-small-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b15b26f7f6a3c8a8718cc69c7c4ff81a9990881e --- /dev/null +++ b/small/ggml-small-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ea4d5727e7b1c8762c548f4b80a30447740200e986075b056cf3da16359da56 +size 243 diff --git a/small/ggml-small-encoder.mlmodelc/coremldata.bin b/small/ggml-small-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2aecfa7c477de9fb63dc789d720d5e2fb0e72d4a --- /dev/null +++ b/small/ggml-small-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a9554f8fae587d40c808e5a6411d79ce20ab6b473d7a3a88df2cd8123978fb4 +size 320 diff --git a/small/ggml-small-encoder.mlmodelc/metadata.json b/small/ggml-small-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..51befae09ac573051582aa395e8d1449589dbc6b --- /dev/null +++ b/small/ggml-small-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 768)", + "shortDescription" : "", + "shape" : "[1, 1500, 768]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 12, + "Gelu" : 14, + "LayerNorm" : 25, + "Transpose" : 13, + "Softmax" : 144, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 25, + "Einsum" : 288, + "ExpandDims" : 1, + "Split" : 36, + "Conv" : 74 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_small", + "method" : "predict" + } +] \ No newline at end of file diff --git a/small/ggml-small-encoder.mlmodelc/model.mil b/small/ggml-small-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..0c32f029586d9600dccb30b0eeece3bf02323585 --- /dev/null +++ b/small/ggml-small-encoder.mlmodelc/model.mil @@ -0,0 +1,1663 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_44_pad_type_0 = const()[name = tensor("op_44_pad_type_0"), val = tensor("custom")]; + tensor var_44_pad_0 = const()[name = tensor("op_44_pad_0"), val = tensor([1, 1])]; + tensor var_44_strides_0 = const()[name = tensor("op_44_strides_0"), val = tensor([1])]; + tensor var_44_dilations_0 = const()[name = tensor("op_44_dilations_0"), val = tensor([1])]; + tensor var_44_groups_0 = const()[name = tensor("op_44_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368768)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_52")]; + tensor var_44_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_44_dilations_0, groups = var_44_groups_0, pad = var_44_pad_0, pad_type = var_44_pad_type_0, strides = var_44_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_44_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_44_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_62_pad_type_0 = const()[name = tensor("op_62_pad_type_0"), val = tensor("custom")]; + tensor var_62_pad_0 = const()[name = tensor("op_62_pad_0"), val = tensor([1, 1])]; + tensor var_62_strides_0 = const()[name = tensor("op_62_strides_0"), val = tensor([2])]; + tensor var_62_dilations_0 = const()[name = tensor("op_62_dilations_0"), val = tensor([1])]; + tensor var_62_groups_0 = const()[name = tensor("op_62_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370368)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3909376)))]; + tensor var_62_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_62_dilations_0, groups = var_62_groups_0, pad = var_62_pad_0, pad_type = var_62_pad_type_0, strides = var_62_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_62_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_62_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_67_to_fp16 = const()[name = tensor("op_67_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3910976)))]; + tensor var_69_cast_fp16 = add(x = x_3_cast_fp16, y = var_67_to_fp16)[name = tensor("op_69_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_69_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6215040)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6216640)))]; + tensor var_100_to_fp16 = const()[name = tensor("op_100_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_100_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_135_weight_0_to_fp16 = const()[name = tensor("op_135_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6218240)))]; + tensor var_135_bias_0_to_fp16 = const()[name = tensor("op_135_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397952)))]; + tensor var_135_cast_fp16 = conv(bias = var_135_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_135_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_135_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7399552)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_133_pad_type_0 = const()[name = tensor("op_133_pad_type_0"), val = tensor("valid")]; + tensor var_133_strides_0 = const()[name = tensor("op_133_strides_0"), val = tensor([1, 1])]; + tensor var_133_pad_0 = const()[name = tensor("op_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_133_dilations_0 = const()[name = tensor("op_133_dilations_0"), val = tensor([1, 1])]; + tensor var_133_groups_0 = const()[name = tensor("op_133_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8579264)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9758976)))]; + tensor var_133_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_133_dilations_0, groups = var_133_groups_0, pad = var_133_pad_0, pad_type = var_133_pad_type_0, strides = var_133_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_133_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_136_axis_0 = const()[name = tensor("op_136_axis_0"), val = tensor(1)]; + tensor var_136_cast_fp16_0, tensor var_136_cast_fp16_1, tensor var_136_cast_fp16_2, tensor var_136_cast_fp16_3, tensor var_136_cast_fp16_4, tensor var_136_cast_fp16_5, tensor var_136_cast_fp16_6, tensor var_136_cast_fp16_7, tensor var_136_cast_fp16_8, tensor var_136_cast_fp16_9, tensor var_136_cast_fp16_10, tensor var_136_cast_fp16_11 = split(axis = var_136_axis_0, split_sizes = tile_0, x = var_135_cast_fp16)[name = tensor("op_136_cast_fp16")]; + tensor var_149_perm_0 = const()[name = tensor("op_149_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_150_axis_0 = const()[name = tensor("op_150_axis_0"), val = tensor(3)]; + tensor var_149_cast_fp16 = transpose(perm = var_149_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_12")]; + tensor var_150_cast_fp16_0, tensor var_150_cast_fp16_1, tensor var_150_cast_fp16_2, tensor var_150_cast_fp16_3, tensor var_150_cast_fp16_4, tensor var_150_cast_fp16_5, tensor var_150_cast_fp16_6, tensor var_150_cast_fp16_7, tensor var_150_cast_fp16_8, tensor var_150_cast_fp16_9, tensor var_150_cast_fp16_10, tensor var_150_cast_fp16_11 = split(axis = var_150_axis_0, split_sizes = tile_1, x = var_149_cast_fp16)[name = tensor("op_150_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_163_axis_0 = const()[name = tensor("op_163_axis_0"), val = tensor(1)]; + tensor var_163_cast_fp16_0, tensor var_163_cast_fp16_1, tensor var_163_cast_fp16_2, tensor var_163_cast_fp16_3, tensor var_163_cast_fp16_4, tensor var_163_cast_fp16_5, tensor var_163_cast_fp16_6, tensor var_163_cast_fp16_7, tensor var_163_cast_fp16_8, tensor var_163_cast_fp16_9, tensor var_163_cast_fp16_10, tensor var_163_cast_fp16_11 = split(axis = var_163_axis_0, split_sizes = tile_2, x = var_133_cast_fp16)[name = tensor("op_163_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_150_cast_fp16_0, var_136_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_150_cast_fp16_1, var_136_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_150_cast_fp16_2, var_136_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_150_cast_fp16_3, var_136_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_150_cast_fp16_4, var_136_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_150_cast_fp16_5, var_136_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_150_cast_fp16_6, var_136_cast_fp16_6))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_150_cast_fp16_7, var_136_cast_fp16_7))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_150_cast_fp16_8, var_136_cast_fp16_8))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_150_cast_fp16_9, var_136_cast_fp16_9))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_150_cast_fp16_10, var_136_cast_fp16_10))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_150_cast_fp16_11, var_136_cast_fp16_11))[name = tensor("aw_23_cast_fp16")]; + tensor var_200_cast_fp16 = softmax(axis = var_84, x = aw_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = softmax(axis = var_84, x = aw_3_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor var_202_cast_fp16 = softmax(axis = var_84, x = aw_5_cast_fp16)[name = tensor("op_202_cast_fp16")]; + tensor var_203_cast_fp16 = softmax(axis = var_84, x = aw_7_cast_fp16)[name = tensor("op_203_cast_fp16")]; + tensor var_204_cast_fp16 = softmax(axis = var_84, x = aw_9_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = softmax(axis = var_84, x = aw_11_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor var_206_cast_fp16 = softmax(axis = var_84, x = aw_13_cast_fp16)[name = tensor("op_206_cast_fp16")]; + tensor var_207_cast_fp16 = softmax(axis = var_84, x = aw_15_cast_fp16)[name = tensor("op_207_cast_fp16")]; + tensor var_208_cast_fp16 = softmax(axis = var_84, x = aw_17_cast_fp16)[name = tensor("op_208_cast_fp16")]; + tensor var_209_cast_fp16 = softmax(axis = var_84, x = aw_19_cast_fp16)[name = tensor("op_209_cast_fp16")]; + tensor var_210_cast_fp16 = softmax(axis = var_84, x = aw_21_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor var_211_cast_fp16 = softmax(axis = var_84, x = aw_23_cast_fp16)[name = tensor("op_211_cast_fp16")]; + tensor var_213_equation_0 = const()[name = tensor("op_213_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_213_cast_fp16 = einsum(equation = var_213_equation_0, values = (var_163_cast_fp16_0, var_200_cast_fp16))[name = tensor("op_213_cast_fp16")]; + tensor var_215_equation_0 = const()[name = tensor("op_215_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_215_cast_fp16 = einsum(equation = var_215_equation_0, values = (var_163_cast_fp16_1, var_201_cast_fp16))[name = tensor("op_215_cast_fp16")]; + tensor var_217_equation_0 = const()[name = tensor("op_217_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_217_cast_fp16 = einsum(equation = var_217_equation_0, values = (var_163_cast_fp16_2, var_202_cast_fp16))[name = tensor("op_217_cast_fp16")]; + tensor var_219_equation_0 = const()[name = tensor("op_219_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_219_cast_fp16 = einsum(equation = var_219_equation_0, values = (var_163_cast_fp16_3, var_203_cast_fp16))[name = tensor("op_219_cast_fp16")]; + tensor var_221_equation_0 = const()[name = tensor("op_221_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_221_cast_fp16 = einsum(equation = var_221_equation_0, values = (var_163_cast_fp16_4, var_204_cast_fp16))[name = tensor("op_221_cast_fp16")]; + tensor var_223_equation_0 = const()[name = tensor("op_223_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_223_cast_fp16 = einsum(equation = var_223_equation_0, values = (var_163_cast_fp16_5, var_205_cast_fp16))[name = tensor("op_223_cast_fp16")]; + tensor var_225_equation_0 = const()[name = tensor("op_225_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_225_cast_fp16 = einsum(equation = var_225_equation_0, values = (var_163_cast_fp16_6, var_206_cast_fp16))[name = tensor("op_225_cast_fp16")]; + tensor var_227_equation_0 = const()[name = tensor("op_227_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_227_cast_fp16 = einsum(equation = var_227_equation_0, values = (var_163_cast_fp16_7, var_207_cast_fp16))[name = tensor("op_227_cast_fp16")]; + tensor var_229_equation_0 = const()[name = tensor("op_229_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_229_cast_fp16 = einsum(equation = var_229_equation_0, values = (var_163_cast_fp16_8, var_208_cast_fp16))[name = tensor("op_229_cast_fp16")]; + tensor var_231_equation_0 = const()[name = tensor("op_231_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_231_cast_fp16 = einsum(equation = var_231_equation_0, values = (var_163_cast_fp16_9, var_209_cast_fp16))[name = tensor("op_231_cast_fp16")]; + tensor var_233_equation_0 = const()[name = tensor("op_233_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_233_cast_fp16 = einsum(equation = var_233_equation_0, values = (var_163_cast_fp16_10, var_210_cast_fp16))[name = tensor("op_233_cast_fp16")]; + tensor var_235_equation_0 = const()[name = tensor("op_235_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_235_cast_fp16 = einsum(equation = var_235_equation_0, values = (var_163_cast_fp16_11, var_211_cast_fp16))[name = tensor("op_235_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_84, interleave = input_5_interleave_0, values = (var_213_cast_fp16, var_215_cast_fp16, var_217_cast_fp16, var_219_cast_fp16, var_221_cast_fp16, var_223_cast_fp16, var_225_cast_fp16, var_227_cast_fp16, var_229_cast_fp16, var_231_cast_fp16, var_233_cast_fp16, var_235_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_244_pad_type_0 = const()[name = tensor("op_244_pad_type_0"), val = tensor("valid")]; + tensor var_244_strides_0 = const()[name = tensor("op_244_strides_0"), val = tensor([1, 1])]; + tensor var_244_pad_0 = const()[name = tensor("op_244_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_244_dilations_0 = const()[name = tensor("op_244_dilations_0"), val = tensor([1, 1])]; + tensor var_244_groups_0 = const()[name = tensor("op_244_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9760576)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10940288)))]; + tensor var_244_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_244_dilations_0, groups = var_244_groups_0, pad = var_244_pad_0, pad_type = var_244_pad_type_0, strides = var_244_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_244_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_244_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10941888)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10943488)))]; + tensor var_254_to_fp16 = const()[name = tensor("op_254_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_254_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10945088)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15663744)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_280_pad_type_0 = const()[name = tensor("op_280_pad_type_0"), val = tensor("valid")]; + tensor var_280_strides_0 = const()[name = tensor("op_280_strides_0"), val = tensor([1, 1])]; + tensor var_280_pad_0 = const()[name = tensor("op_280_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_280_dilations_0 = const()[name = tensor("op_280_dilations_0"), val = tensor([1, 1])]; + tensor var_280_groups_0 = const()[name = tensor("op_280_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15669952)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20388608)))]; + tensor var_280_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_280_dilations_0, groups = var_280_groups_0, pad = var_280_pad_0, pad_type = var_280_pad_type_0, strides = var_280_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_280_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_280_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_289 = const()[name = tensor("op_289"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20390208)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20391808)))]; + tensor var_305_to_fp16 = const()[name = tensor("op_305_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_305_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_340_weight_0_to_fp16 = const()[name = tensor("op_340_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20393408)))]; + tensor var_340_bias_0_to_fp16 = const()[name = tensor("op_340_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21573120)))]; + tensor var_340_cast_fp16 = conv(bias = var_340_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_340_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21574720)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_338_pad_type_0 = const()[name = tensor("op_338_pad_type_0"), val = tensor("valid")]; + tensor var_338_strides_0 = const()[name = tensor("op_338_strides_0"), val = tensor([1, 1])]; + tensor var_338_pad_0 = const()[name = tensor("op_338_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_338_dilations_0 = const()[name = tensor("op_338_dilations_0"), val = tensor([1, 1])]; + tensor var_338_groups_0 = const()[name = tensor("op_338_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22754432)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23934144)))]; + tensor var_338_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_338_dilations_0, groups = var_338_groups_0, pad = var_338_pad_0, pad_type = var_338_pad_type_0, strides = var_338_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_338_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_341_axis_0 = const()[name = tensor("op_341_axis_0"), val = tensor(1)]; + tensor var_341_cast_fp16_0, tensor var_341_cast_fp16_1, tensor var_341_cast_fp16_2, tensor var_341_cast_fp16_3, tensor var_341_cast_fp16_4, tensor var_341_cast_fp16_5, tensor var_341_cast_fp16_6, tensor var_341_cast_fp16_7, tensor var_341_cast_fp16_8, tensor var_341_cast_fp16_9, tensor var_341_cast_fp16_10, tensor var_341_cast_fp16_11 = split(axis = var_341_axis_0, split_sizes = tile_3, x = var_340_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor var_354_perm_0 = const()[name = tensor("op_354_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_355_axis_0 = const()[name = tensor("op_355_axis_0"), val = tensor(3)]; + tensor var_354_cast_fp16 = transpose(perm = var_354_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_11")]; + tensor var_355_cast_fp16_0, tensor var_355_cast_fp16_1, tensor var_355_cast_fp16_2, tensor var_355_cast_fp16_3, tensor var_355_cast_fp16_4, tensor var_355_cast_fp16_5, tensor var_355_cast_fp16_6, tensor var_355_cast_fp16_7, tensor var_355_cast_fp16_8, tensor var_355_cast_fp16_9, tensor var_355_cast_fp16_10, tensor var_355_cast_fp16_11 = split(axis = var_355_axis_0, split_sizes = tile_4, x = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_368_axis_0 = const()[name = tensor("op_368_axis_0"), val = tensor(1)]; + tensor var_368_cast_fp16_0, tensor var_368_cast_fp16_1, tensor var_368_cast_fp16_2, tensor var_368_cast_fp16_3, tensor var_368_cast_fp16_4, tensor var_368_cast_fp16_5, tensor var_368_cast_fp16_6, tensor var_368_cast_fp16_7, tensor var_368_cast_fp16_8, tensor var_368_cast_fp16_9, tensor var_368_cast_fp16_10, tensor var_368_cast_fp16_11 = split(axis = var_368_axis_0, split_sizes = tile_5, x = var_338_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_355_cast_fp16_0, var_341_cast_fp16_0))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_355_cast_fp16_1, var_341_cast_fp16_1))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_355_cast_fp16_2, var_341_cast_fp16_2))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_355_cast_fp16_3, var_341_cast_fp16_3))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_355_cast_fp16_4, var_341_cast_fp16_4))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_355_cast_fp16_5, var_341_cast_fp16_5))[name = tensor("aw_35_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_355_cast_fp16_6, var_341_cast_fp16_6))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_355_cast_fp16_7, var_341_cast_fp16_7))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_355_cast_fp16_8, var_341_cast_fp16_8))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_355_cast_fp16_9, var_341_cast_fp16_9))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_355_cast_fp16_10, var_341_cast_fp16_10))[name = tensor("aw_45_cast_fp16")]; + tensor aw_47_equation_0 = const()[name = tensor("aw_47_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_47_cast_fp16 = einsum(equation = aw_47_equation_0, values = (var_355_cast_fp16_11, var_341_cast_fp16_11))[name = tensor("aw_47_cast_fp16")]; + tensor var_405_cast_fp16 = softmax(axis = var_289, x = aw_25_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor var_406_cast_fp16 = softmax(axis = var_289, x = aw_27_cast_fp16)[name = tensor("op_406_cast_fp16")]; + tensor var_407_cast_fp16 = softmax(axis = var_289, x = aw_29_cast_fp16)[name = tensor("op_407_cast_fp16")]; + tensor var_408_cast_fp16 = softmax(axis = var_289, x = aw_31_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_409_cast_fp16 = softmax(axis = var_289, x = aw_33_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor var_410_cast_fp16 = softmax(axis = var_289, x = aw_35_cast_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = softmax(axis = var_289, x = aw_37_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor var_412_cast_fp16 = softmax(axis = var_289, x = aw_39_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = softmax(axis = var_289, x = aw_41_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor var_414_cast_fp16 = softmax(axis = var_289, x = aw_43_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_cast_fp16 = softmax(axis = var_289, x = aw_45_cast_fp16)[name = tensor("op_415_cast_fp16")]; + tensor var_416_cast_fp16 = softmax(axis = var_289, x = aw_47_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_418_equation_0 = const()[name = tensor("op_418_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_418_cast_fp16 = einsum(equation = var_418_equation_0, values = (var_368_cast_fp16_0, var_405_cast_fp16))[name = tensor("op_418_cast_fp16")]; + tensor var_420_equation_0 = const()[name = tensor("op_420_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_420_cast_fp16 = einsum(equation = var_420_equation_0, values = (var_368_cast_fp16_1, var_406_cast_fp16))[name = tensor("op_420_cast_fp16")]; + tensor var_422_equation_0 = const()[name = tensor("op_422_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_422_cast_fp16 = einsum(equation = var_422_equation_0, values = (var_368_cast_fp16_2, var_407_cast_fp16))[name = tensor("op_422_cast_fp16")]; + tensor var_424_equation_0 = const()[name = tensor("op_424_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_424_cast_fp16 = einsum(equation = var_424_equation_0, values = (var_368_cast_fp16_3, var_408_cast_fp16))[name = tensor("op_424_cast_fp16")]; + tensor var_426_equation_0 = const()[name = tensor("op_426_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_426_cast_fp16 = einsum(equation = var_426_equation_0, values = (var_368_cast_fp16_4, var_409_cast_fp16))[name = tensor("op_426_cast_fp16")]; + tensor var_428_equation_0 = const()[name = tensor("op_428_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_428_cast_fp16 = einsum(equation = var_428_equation_0, values = (var_368_cast_fp16_5, var_410_cast_fp16))[name = tensor("op_428_cast_fp16")]; + tensor var_430_equation_0 = const()[name = tensor("op_430_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_430_cast_fp16 = einsum(equation = var_430_equation_0, values = (var_368_cast_fp16_6, var_411_cast_fp16))[name = tensor("op_430_cast_fp16")]; + tensor var_432_equation_0 = const()[name = tensor("op_432_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_432_cast_fp16 = einsum(equation = var_432_equation_0, values = (var_368_cast_fp16_7, var_412_cast_fp16))[name = tensor("op_432_cast_fp16")]; + tensor var_434_equation_0 = const()[name = tensor("op_434_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_434_cast_fp16 = einsum(equation = var_434_equation_0, values = (var_368_cast_fp16_8, var_413_cast_fp16))[name = tensor("op_434_cast_fp16")]; + tensor var_436_equation_0 = const()[name = tensor("op_436_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_436_cast_fp16 = einsum(equation = var_436_equation_0, values = (var_368_cast_fp16_9, var_414_cast_fp16))[name = tensor("op_436_cast_fp16")]; + tensor var_438_equation_0 = const()[name = tensor("op_438_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_438_cast_fp16 = einsum(equation = var_438_equation_0, values = (var_368_cast_fp16_10, var_415_cast_fp16))[name = tensor("op_438_cast_fp16")]; + tensor var_440_equation_0 = const()[name = tensor("op_440_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_440_cast_fp16 = einsum(equation = var_440_equation_0, values = (var_368_cast_fp16_11, var_416_cast_fp16))[name = tensor("op_440_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_289, interleave = input_15_interleave_0, values = (var_418_cast_fp16, var_420_cast_fp16, var_422_cast_fp16, var_424_cast_fp16, var_426_cast_fp16, var_428_cast_fp16, var_430_cast_fp16, var_432_cast_fp16, var_434_cast_fp16, var_436_cast_fp16, var_438_cast_fp16, var_440_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_449_pad_type_0 = const()[name = tensor("op_449_pad_type_0"), val = tensor("valid")]; + tensor var_449_strides_0 = const()[name = tensor("op_449_strides_0"), val = tensor([1, 1])]; + tensor var_449_pad_0 = const()[name = tensor("op_449_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_449_dilations_0 = const()[name = tensor("op_449_dilations_0"), val = tensor([1, 1])]; + tensor var_449_groups_0 = const()[name = tensor("op_449_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23935744)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25115456)))]; + tensor var_449_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_449_dilations_0, groups = var_449_groups_0, pad = var_449_pad_0, pad_type = var_449_pad_type_0, strides = var_449_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_449_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25117056)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25118656)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_459_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25120256)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29838912)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_485_pad_type_0 = const()[name = tensor("op_485_pad_type_0"), val = tensor("valid")]; + tensor var_485_strides_0 = const()[name = tensor("op_485_strides_0"), val = tensor([1, 1])]; + tensor var_485_pad_0 = const()[name = tensor("op_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_485_dilations_0 = const()[name = tensor("op_485_dilations_0"), val = tensor([1, 1])]; + tensor var_485_groups_0 = const()[name = tensor("op_485_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29845120)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34563776)))]; + tensor var_485_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_485_dilations_0, groups = var_485_groups_0, pad = var_485_pad_0, pad_type = var_485_pad_type_0, strides = var_485_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_485_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_494 = const()[name = tensor("op_494"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34565376)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34566976)))]; + tensor var_510_to_fp16 = const()[name = tensor("op_510_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_510_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_545_weight_0_to_fp16 = const()[name = tensor("op_545_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34568576)))]; + tensor var_545_bias_0_to_fp16 = const()[name = tensor("op_545_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35748288)))]; + tensor var_545_cast_fp16 = conv(bias = var_545_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_545_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_545_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35749888)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_543_pad_type_0 = const()[name = tensor("op_543_pad_type_0"), val = tensor("valid")]; + tensor var_543_strides_0 = const()[name = tensor("op_543_strides_0"), val = tensor([1, 1])]; + tensor var_543_pad_0 = const()[name = tensor("op_543_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_543_dilations_0 = const()[name = tensor("op_543_dilations_0"), val = tensor([1, 1])]; + tensor var_543_groups_0 = const()[name = tensor("op_543_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36929600)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38109312)))]; + tensor var_543_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_543_dilations_0, groups = var_543_groups_0, pad = var_543_pad_0, pad_type = var_543_pad_type_0, strides = var_543_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_543_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_546_axis_0 = const()[name = tensor("op_546_axis_0"), val = tensor(1)]; + tensor var_546_cast_fp16_0, tensor var_546_cast_fp16_1, tensor var_546_cast_fp16_2, tensor var_546_cast_fp16_3, tensor var_546_cast_fp16_4, tensor var_546_cast_fp16_5, tensor var_546_cast_fp16_6, tensor var_546_cast_fp16_7, tensor var_546_cast_fp16_8, tensor var_546_cast_fp16_9, tensor var_546_cast_fp16_10, tensor var_546_cast_fp16_11 = split(axis = var_546_axis_0, split_sizes = tile_6, x = var_545_cast_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_559_perm_0 = const()[name = tensor("op_559_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_560_axis_0 = const()[name = tensor("op_560_axis_0"), val = tensor(3)]; + tensor var_559_cast_fp16 = transpose(perm = var_559_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_10")]; + tensor var_560_cast_fp16_0, tensor var_560_cast_fp16_1, tensor var_560_cast_fp16_2, tensor var_560_cast_fp16_3, tensor var_560_cast_fp16_4, tensor var_560_cast_fp16_5, tensor var_560_cast_fp16_6, tensor var_560_cast_fp16_7, tensor var_560_cast_fp16_8, tensor var_560_cast_fp16_9, tensor var_560_cast_fp16_10, tensor var_560_cast_fp16_11 = split(axis = var_560_axis_0, split_sizes = tile_7, x = var_559_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_573_axis_0 = const()[name = tensor("op_573_axis_0"), val = tensor(1)]; + tensor var_573_cast_fp16_0, tensor var_573_cast_fp16_1, tensor var_573_cast_fp16_2, tensor var_573_cast_fp16_3, tensor var_573_cast_fp16_4, tensor var_573_cast_fp16_5, tensor var_573_cast_fp16_6, tensor var_573_cast_fp16_7, tensor var_573_cast_fp16_8, tensor var_573_cast_fp16_9, tensor var_573_cast_fp16_10, tensor var_573_cast_fp16_11 = split(axis = var_573_axis_0, split_sizes = tile_8, x = var_543_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor aw_49_equation_0 = const()[name = tensor("aw_49_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_49_cast_fp16 = einsum(equation = aw_49_equation_0, values = (var_560_cast_fp16_0, var_546_cast_fp16_0))[name = tensor("aw_49_cast_fp16")]; + tensor aw_51_equation_0 = const()[name = tensor("aw_51_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_51_cast_fp16 = einsum(equation = aw_51_equation_0, values = (var_560_cast_fp16_1, var_546_cast_fp16_1))[name = tensor("aw_51_cast_fp16")]; + tensor aw_53_equation_0 = const()[name = tensor("aw_53_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_53_cast_fp16 = einsum(equation = aw_53_equation_0, values = (var_560_cast_fp16_2, var_546_cast_fp16_2))[name = tensor("aw_53_cast_fp16")]; + tensor aw_55_equation_0 = const()[name = tensor("aw_55_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_55_cast_fp16 = einsum(equation = aw_55_equation_0, values = (var_560_cast_fp16_3, var_546_cast_fp16_3))[name = tensor("aw_55_cast_fp16")]; + tensor aw_57_equation_0 = const()[name = tensor("aw_57_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_57_cast_fp16 = einsum(equation = aw_57_equation_0, values = (var_560_cast_fp16_4, var_546_cast_fp16_4))[name = tensor("aw_57_cast_fp16")]; + tensor aw_59_equation_0 = const()[name = tensor("aw_59_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_59_cast_fp16 = einsum(equation = aw_59_equation_0, values = (var_560_cast_fp16_5, var_546_cast_fp16_5))[name = tensor("aw_59_cast_fp16")]; + tensor aw_61_equation_0 = const()[name = tensor("aw_61_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_61_cast_fp16 = einsum(equation = aw_61_equation_0, values = (var_560_cast_fp16_6, var_546_cast_fp16_6))[name = tensor("aw_61_cast_fp16")]; + tensor aw_63_equation_0 = const()[name = tensor("aw_63_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_63_cast_fp16 = einsum(equation = aw_63_equation_0, values = (var_560_cast_fp16_7, var_546_cast_fp16_7))[name = tensor("aw_63_cast_fp16")]; + tensor aw_65_equation_0 = const()[name = tensor("aw_65_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_65_cast_fp16 = einsum(equation = aw_65_equation_0, values = (var_560_cast_fp16_8, var_546_cast_fp16_8))[name = tensor("aw_65_cast_fp16")]; + tensor aw_67_equation_0 = const()[name = tensor("aw_67_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_67_cast_fp16 = einsum(equation = aw_67_equation_0, values = (var_560_cast_fp16_9, var_546_cast_fp16_9))[name = tensor("aw_67_cast_fp16")]; + tensor aw_69_equation_0 = const()[name = tensor("aw_69_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_69_cast_fp16 = einsum(equation = aw_69_equation_0, values = (var_560_cast_fp16_10, var_546_cast_fp16_10))[name = tensor("aw_69_cast_fp16")]; + tensor aw_71_equation_0 = const()[name = tensor("aw_71_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_71_cast_fp16 = einsum(equation = aw_71_equation_0, values = (var_560_cast_fp16_11, var_546_cast_fp16_11))[name = tensor("aw_71_cast_fp16")]; + tensor var_610_cast_fp16 = softmax(axis = var_494, x = aw_49_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = softmax(axis = var_494, x = aw_51_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor var_612_cast_fp16 = softmax(axis = var_494, x = aw_53_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_cast_fp16 = softmax(axis = var_494, x = aw_55_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor var_614_cast_fp16 = softmax(axis = var_494, x = aw_57_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = softmax(axis = var_494, x = aw_59_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor var_616_cast_fp16 = softmax(axis = var_494, x = aw_61_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_617_cast_fp16 = softmax(axis = var_494, x = aw_63_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_618_cast_fp16 = softmax(axis = var_494, x = aw_65_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = softmax(axis = var_494, x = aw_67_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor var_620_cast_fp16 = softmax(axis = var_494, x = aw_69_cast_fp16)[name = tensor("op_620_cast_fp16")]; + tensor var_621_cast_fp16 = softmax(axis = var_494, x = aw_71_cast_fp16)[name = tensor("op_621_cast_fp16")]; + tensor var_623_equation_0 = const()[name = tensor("op_623_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_623_cast_fp16 = einsum(equation = var_623_equation_0, values = (var_573_cast_fp16_0, var_610_cast_fp16))[name = tensor("op_623_cast_fp16")]; + tensor var_625_equation_0 = const()[name = tensor("op_625_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_625_cast_fp16 = einsum(equation = var_625_equation_0, values = (var_573_cast_fp16_1, var_611_cast_fp16))[name = tensor("op_625_cast_fp16")]; + tensor var_627_equation_0 = const()[name = tensor("op_627_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_627_cast_fp16 = einsum(equation = var_627_equation_0, values = (var_573_cast_fp16_2, var_612_cast_fp16))[name = tensor("op_627_cast_fp16")]; + tensor var_629_equation_0 = const()[name = tensor("op_629_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_629_cast_fp16 = einsum(equation = var_629_equation_0, values = (var_573_cast_fp16_3, var_613_cast_fp16))[name = tensor("op_629_cast_fp16")]; + tensor var_631_equation_0 = const()[name = tensor("op_631_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_631_cast_fp16 = einsum(equation = var_631_equation_0, values = (var_573_cast_fp16_4, var_614_cast_fp16))[name = tensor("op_631_cast_fp16")]; + tensor var_633_equation_0 = const()[name = tensor("op_633_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_633_cast_fp16 = einsum(equation = var_633_equation_0, values = (var_573_cast_fp16_5, var_615_cast_fp16))[name = tensor("op_633_cast_fp16")]; + tensor var_635_equation_0 = const()[name = tensor("op_635_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_635_cast_fp16 = einsum(equation = var_635_equation_0, values = (var_573_cast_fp16_6, var_616_cast_fp16))[name = tensor("op_635_cast_fp16")]; + tensor var_637_equation_0 = const()[name = tensor("op_637_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_637_cast_fp16 = einsum(equation = var_637_equation_0, values = (var_573_cast_fp16_7, var_617_cast_fp16))[name = tensor("op_637_cast_fp16")]; + tensor var_639_equation_0 = const()[name = tensor("op_639_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_639_cast_fp16 = einsum(equation = var_639_equation_0, values = (var_573_cast_fp16_8, var_618_cast_fp16))[name = tensor("op_639_cast_fp16")]; + tensor var_641_equation_0 = const()[name = tensor("op_641_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_641_cast_fp16 = einsum(equation = var_641_equation_0, values = (var_573_cast_fp16_9, var_619_cast_fp16))[name = tensor("op_641_cast_fp16")]; + tensor var_643_equation_0 = const()[name = tensor("op_643_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_643_cast_fp16 = einsum(equation = var_643_equation_0, values = (var_573_cast_fp16_10, var_620_cast_fp16))[name = tensor("op_643_cast_fp16")]; + tensor var_645_equation_0 = const()[name = tensor("op_645_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_645_cast_fp16 = einsum(equation = var_645_equation_0, values = (var_573_cast_fp16_11, var_621_cast_fp16))[name = tensor("op_645_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_494, interleave = input_25_interleave_0, values = (var_623_cast_fp16, var_625_cast_fp16, var_627_cast_fp16, var_629_cast_fp16, var_631_cast_fp16, var_633_cast_fp16, var_635_cast_fp16, var_637_cast_fp16, var_639_cast_fp16, var_641_cast_fp16, var_643_cast_fp16, var_645_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_654_pad_type_0 = const()[name = tensor("op_654_pad_type_0"), val = tensor("valid")]; + tensor var_654_strides_0 = const()[name = tensor("op_654_strides_0"), val = tensor([1, 1])]; + tensor var_654_pad_0 = const()[name = tensor("op_654_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_654_dilations_0 = const()[name = tensor("op_654_dilations_0"), val = tensor([1, 1])]; + tensor var_654_groups_0 = const()[name = tensor("op_654_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38110912)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39290624)))]; + tensor var_654_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_654_dilations_0, groups = var_654_groups_0, pad = var_654_pad_0, pad_type = var_654_pad_type_0, strides = var_654_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_654_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_654_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39292224)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39293824)))]; + tensor var_664_to_fp16 = const()[name = tensor("op_664_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_664_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39295424)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44014080)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_690_pad_type_0 = const()[name = tensor("op_690_pad_type_0"), val = tensor("valid")]; + tensor var_690_strides_0 = const()[name = tensor("op_690_strides_0"), val = tensor([1, 1])]; + tensor var_690_pad_0 = const()[name = tensor("op_690_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_690_dilations_0 = const()[name = tensor("op_690_dilations_0"), val = tensor([1, 1])]; + tensor var_690_groups_0 = const()[name = tensor("op_690_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44020288)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48738944)))]; + tensor var_690_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_690_dilations_0, groups = var_690_groups_0, pad = var_690_pad_0, pad_type = var_690_pad_type_0, strides = var_690_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_690_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48740544)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48742144)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_715_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_7_pad_type_0 = const()[name = tensor("q_7_pad_type_0"), val = tensor("valid")]; + tensor q_7_strides_0 = const()[name = tensor("q_7_strides_0"), val = tensor([1, 1])]; + tensor q_7_pad_0 = const()[name = tensor("q_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_dilations_0 = const()[name = tensor("q_7_dilations_0"), val = tensor([1, 1])]; + tensor q_7_groups_0 = const()[name = tensor("q_7_groups_0"), val = tensor(1)]; + tensor var_750_weight_0_to_fp16 = const()[name = tensor("op_750_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48743744)))]; + tensor var_750_bias_0_to_fp16 = const()[name = tensor("op_750_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49923456)))]; + tensor var_750_cast_fp16 = conv(bias = var_750_bias_0_to_fp16, dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = var_750_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_750_cast_fp16")]; + tensor k_7_pad_type_0 = const()[name = tensor("k_7_pad_type_0"), val = tensor("valid")]; + tensor k_7_strides_0 = const()[name = tensor("k_7_strides_0"), val = tensor([1, 1])]; + tensor k_7_pad_0 = const()[name = tensor("k_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_7_dilations_0 = const()[name = tensor("k_7_dilations_0"), val = tensor([1, 1])]; + tensor k_7_groups_0 = const()[name = tensor("k_7_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49925056)))]; + tensor k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_748_pad_type_0 = const()[name = tensor("op_748_pad_type_0"), val = tensor("valid")]; + tensor var_748_strides_0 = const()[name = tensor("op_748_strides_0"), val = tensor([1, 1])]; + tensor var_748_pad_0 = const()[name = tensor("op_748_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_748_dilations_0 = const()[name = tensor("op_748_dilations_0"), val = tensor([1, 1])]; + tensor var_748_groups_0 = const()[name = tensor("op_748_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51104768)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52284480)))]; + tensor var_748_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_748_dilations_0, groups = var_748_groups_0, pad = var_748_pad_0, pad_type = var_748_pad_type_0, strides = var_748_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_cast_fp16_0, tensor var_751_cast_fp16_1, tensor var_751_cast_fp16_2, tensor var_751_cast_fp16_3, tensor var_751_cast_fp16_4, tensor var_751_cast_fp16_5, tensor var_751_cast_fp16_6, tensor var_751_cast_fp16_7, tensor var_751_cast_fp16_8, tensor var_751_cast_fp16_9, tensor var_751_cast_fp16_10, tensor var_751_cast_fp16_11 = split(axis = var_751_axis_0, split_sizes = tile_9, x = var_750_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_764_perm_0 = const()[name = tensor("op_764_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_765_axis_0 = const()[name = tensor("op_765_axis_0"), val = tensor(3)]; + tensor var_764_cast_fp16 = transpose(perm = var_764_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_9")]; + tensor var_765_cast_fp16_0, tensor var_765_cast_fp16_1, tensor var_765_cast_fp16_2, tensor var_765_cast_fp16_3, tensor var_765_cast_fp16_4, tensor var_765_cast_fp16_5, tensor var_765_cast_fp16_6, tensor var_765_cast_fp16_7, tensor var_765_cast_fp16_8, tensor var_765_cast_fp16_9, tensor var_765_cast_fp16_10, tensor var_765_cast_fp16_11 = split(axis = var_765_axis_0, split_sizes = tile_10, x = var_764_cast_fp16)[name = tensor("op_765_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_778_axis_0 = const()[name = tensor("op_778_axis_0"), val = tensor(1)]; + tensor var_778_cast_fp16_0, tensor var_778_cast_fp16_1, tensor var_778_cast_fp16_2, tensor var_778_cast_fp16_3, tensor var_778_cast_fp16_4, tensor var_778_cast_fp16_5, tensor var_778_cast_fp16_6, tensor var_778_cast_fp16_7, tensor var_778_cast_fp16_8, tensor var_778_cast_fp16_9, tensor var_778_cast_fp16_10, tensor var_778_cast_fp16_11 = split(axis = var_778_axis_0, split_sizes = tile_11, x = var_748_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor aw_73_equation_0 = const()[name = tensor("aw_73_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_73_cast_fp16 = einsum(equation = aw_73_equation_0, values = (var_765_cast_fp16_0, var_751_cast_fp16_0))[name = tensor("aw_73_cast_fp16")]; + tensor aw_75_equation_0 = const()[name = tensor("aw_75_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_75_cast_fp16 = einsum(equation = aw_75_equation_0, values = (var_765_cast_fp16_1, var_751_cast_fp16_1))[name = tensor("aw_75_cast_fp16")]; + tensor aw_77_equation_0 = const()[name = tensor("aw_77_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_77_cast_fp16 = einsum(equation = aw_77_equation_0, values = (var_765_cast_fp16_2, var_751_cast_fp16_2))[name = tensor("aw_77_cast_fp16")]; + tensor aw_79_equation_0 = const()[name = tensor("aw_79_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_79_cast_fp16 = einsum(equation = aw_79_equation_0, values = (var_765_cast_fp16_3, var_751_cast_fp16_3))[name = tensor("aw_79_cast_fp16")]; + tensor aw_81_equation_0 = const()[name = tensor("aw_81_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_81_cast_fp16 = einsum(equation = aw_81_equation_0, values = (var_765_cast_fp16_4, var_751_cast_fp16_4))[name = tensor("aw_81_cast_fp16")]; + tensor aw_83_equation_0 = const()[name = tensor("aw_83_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_83_cast_fp16 = einsum(equation = aw_83_equation_0, values = (var_765_cast_fp16_5, var_751_cast_fp16_5))[name = tensor("aw_83_cast_fp16")]; + tensor aw_85_equation_0 = const()[name = tensor("aw_85_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_85_cast_fp16 = einsum(equation = aw_85_equation_0, values = (var_765_cast_fp16_6, var_751_cast_fp16_6))[name = tensor("aw_85_cast_fp16")]; + tensor aw_87_equation_0 = const()[name = tensor("aw_87_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_87_cast_fp16 = einsum(equation = aw_87_equation_0, values = (var_765_cast_fp16_7, var_751_cast_fp16_7))[name = tensor("aw_87_cast_fp16")]; + tensor aw_89_equation_0 = const()[name = tensor("aw_89_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_89_cast_fp16 = einsum(equation = aw_89_equation_0, values = (var_765_cast_fp16_8, var_751_cast_fp16_8))[name = tensor("aw_89_cast_fp16")]; + tensor aw_91_equation_0 = const()[name = tensor("aw_91_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_91_cast_fp16 = einsum(equation = aw_91_equation_0, values = (var_765_cast_fp16_9, var_751_cast_fp16_9))[name = tensor("aw_91_cast_fp16")]; + tensor aw_93_equation_0 = const()[name = tensor("aw_93_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_93_cast_fp16 = einsum(equation = aw_93_equation_0, values = (var_765_cast_fp16_10, var_751_cast_fp16_10))[name = tensor("aw_93_cast_fp16")]; + tensor aw_95_equation_0 = const()[name = tensor("aw_95_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_95_cast_fp16 = einsum(equation = aw_95_equation_0, values = (var_765_cast_fp16_11, var_751_cast_fp16_11))[name = tensor("aw_95_cast_fp16")]; + tensor var_815_cast_fp16 = softmax(axis = var_699, x = aw_73_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor var_816_cast_fp16 = softmax(axis = var_699, x = aw_75_cast_fp16)[name = tensor("op_816_cast_fp16")]; + tensor var_817_cast_fp16 = softmax(axis = var_699, x = aw_77_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = softmax(axis = var_699, x = aw_79_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_819_cast_fp16 = softmax(axis = var_699, x = aw_81_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820_cast_fp16 = softmax(axis = var_699, x = aw_83_cast_fp16)[name = tensor("op_820_cast_fp16")]; + tensor var_821_cast_fp16 = softmax(axis = var_699, x = aw_85_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = softmax(axis = var_699, x = aw_87_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor var_823_cast_fp16 = softmax(axis = var_699, x = aw_89_cast_fp16)[name = tensor("op_823_cast_fp16")]; + tensor var_824_cast_fp16 = softmax(axis = var_699, x = aw_91_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor var_825_cast_fp16 = softmax(axis = var_699, x = aw_93_cast_fp16)[name = tensor("op_825_cast_fp16")]; + tensor var_826_cast_fp16 = softmax(axis = var_699, x = aw_95_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_828_equation_0 = const()[name = tensor("op_828_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_828_cast_fp16 = einsum(equation = var_828_equation_0, values = (var_778_cast_fp16_0, var_815_cast_fp16))[name = tensor("op_828_cast_fp16")]; + tensor var_830_equation_0 = const()[name = tensor("op_830_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_830_cast_fp16 = einsum(equation = var_830_equation_0, values = (var_778_cast_fp16_1, var_816_cast_fp16))[name = tensor("op_830_cast_fp16")]; + tensor var_832_equation_0 = const()[name = tensor("op_832_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_832_cast_fp16 = einsum(equation = var_832_equation_0, values = (var_778_cast_fp16_2, var_817_cast_fp16))[name = tensor("op_832_cast_fp16")]; + tensor var_834_equation_0 = const()[name = tensor("op_834_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_834_cast_fp16 = einsum(equation = var_834_equation_0, values = (var_778_cast_fp16_3, var_818_cast_fp16))[name = tensor("op_834_cast_fp16")]; + tensor var_836_equation_0 = const()[name = tensor("op_836_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_836_cast_fp16 = einsum(equation = var_836_equation_0, values = (var_778_cast_fp16_4, var_819_cast_fp16))[name = tensor("op_836_cast_fp16")]; + tensor var_838_equation_0 = const()[name = tensor("op_838_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_838_cast_fp16 = einsum(equation = var_838_equation_0, values = (var_778_cast_fp16_5, var_820_cast_fp16))[name = tensor("op_838_cast_fp16")]; + tensor var_840_equation_0 = const()[name = tensor("op_840_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_840_cast_fp16 = einsum(equation = var_840_equation_0, values = (var_778_cast_fp16_6, var_821_cast_fp16))[name = tensor("op_840_cast_fp16")]; + tensor var_842_equation_0 = const()[name = tensor("op_842_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_842_cast_fp16 = einsum(equation = var_842_equation_0, values = (var_778_cast_fp16_7, var_822_cast_fp16))[name = tensor("op_842_cast_fp16")]; + tensor var_844_equation_0 = const()[name = tensor("op_844_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_844_cast_fp16 = einsum(equation = var_844_equation_0, values = (var_778_cast_fp16_8, var_823_cast_fp16))[name = tensor("op_844_cast_fp16")]; + tensor var_846_equation_0 = const()[name = tensor("op_846_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_846_cast_fp16 = einsum(equation = var_846_equation_0, values = (var_778_cast_fp16_9, var_824_cast_fp16))[name = tensor("op_846_cast_fp16")]; + tensor var_848_equation_0 = const()[name = tensor("op_848_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_848_cast_fp16 = einsum(equation = var_848_equation_0, values = (var_778_cast_fp16_10, var_825_cast_fp16))[name = tensor("op_848_cast_fp16")]; + tensor var_850_equation_0 = const()[name = tensor("op_850_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_850_cast_fp16 = einsum(equation = var_850_equation_0, values = (var_778_cast_fp16_11, var_826_cast_fp16))[name = tensor("op_850_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_699, interleave = input_35_interleave_0, values = (var_828_cast_fp16, var_830_cast_fp16, var_832_cast_fp16, var_834_cast_fp16, var_836_cast_fp16, var_838_cast_fp16, var_840_cast_fp16, var_842_cast_fp16, var_844_cast_fp16, var_846_cast_fp16, var_848_cast_fp16, var_850_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_859_pad_type_0 = const()[name = tensor("op_859_pad_type_0"), val = tensor("valid")]; + tensor var_859_strides_0 = const()[name = tensor("op_859_strides_0"), val = tensor([1, 1])]; + tensor var_859_pad_0 = const()[name = tensor("op_859_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_859_dilations_0 = const()[name = tensor("op_859_dilations_0"), val = tensor([1, 1])]; + tensor var_859_groups_0 = const()[name = tensor("op_859_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52286080)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53465792)))]; + tensor var_859_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_859_dilations_0, groups = var_859_groups_0, pad = var_859_pad_0, pad_type = var_859_pad_type_0, strides = var_859_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_859_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_859_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53467392)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53468992)))]; + tensor var_869_to_fp16 = const()[name = tensor("op_869_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_869_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53470592)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58189248)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_895_pad_type_0 = const()[name = tensor("op_895_pad_type_0"), val = tensor("valid")]; + tensor var_895_strides_0 = const()[name = tensor("op_895_strides_0"), val = tensor([1, 1])]; + tensor var_895_pad_0 = const()[name = tensor("op_895_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_895_dilations_0 = const()[name = tensor("op_895_dilations_0"), val = tensor([1, 1])]; + tensor var_895_groups_0 = const()[name = tensor("op_895_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58195456)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62914112)))]; + tensor var_895_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_895_dilations_0, groups = var_895_groups_0, pad = var_895_pad_0, pad_type = var_895_pad_type_0, strides = var_895_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("op_895_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_895_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor(1)]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([1])]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62915712)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62917312)))]; + tensor var_920_to_fp16 = const()[name = tensor("op_920_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_920_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor q_9_pad_type_0 = const()[name = tensor("q_9_pad_type_0"), val = tensor("valid")]; + tensor q_9_strides_0 = const()[name = tensor("q_9_strides_0"), val = tensor([1, 1])]; + tensor q_9_pad_0 = const()[name = tensor("q_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_9_dilations_0 = const()[name = tensor("q_9_dilations_0"), val = tensor([1, 1])]; + tensor q_9_groups_0 = const()[name = tensor("q_9_groups_0"), val = tensor(1)]; + tensor var_955_weight_0_to_fp16 = const()[name = tensor("op_955_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62918912)))]; + tensor var_955_bias_0_to_fp16 = const()[name = tensor("op_955_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64098624)))]; + tensor var_955_cast_fp16 = conv(bias = var_955_bias_0_to_fp16, dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = var_955_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("op_955_cast_fp16")]; + tensor k_9_pad_type_0 = const()[name = tensor("k_9_pad_type_0"), val = tensor("valid")]; + tensor k_9_strides_0 = const()[name = tensor("k_9_strides_0"), val = tensor([1, 1])]; + tensor k_9_pad_0 = const()[name = tensor("k_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_9_dilations_0 = const()[name = tensor("k_9_dilations_0"), val = tensor([1, 1])]; + tensor k_9_groups_0 = const()[name = tensor("k_9_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_key_weight_to_fp16 = const()[name = tensor("blocks_4_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64100224)))]; + tensor k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = blocks_4_attn_key_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor var_953_pad_type_0 = const()[name = tensor("op_953_pad_type_0"), val = tensor("valid")]; + tensor var_953_strides_0 = const()[name = tensor("op_953_strides_0"), val = tensor([1, 1])]; + tensor var_953_pad_0 = const()[name = tensor("op_953_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_953_dilations_0 = const()[name = tensor("op_953_dilations_0"), val = tensor([1, 1])]; + tensor var_953_groups_0 = const()[name = tensor("op_953_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_value_weight_to_fp16 = const()[name = tensor("blocks_4_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65279936)))]; + tensor blocks_4_attn_value_bias_to_fp16 = const()[name = tensor("blocks_4_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66459648)))]; + tensor var_953_cast_fp16 = conv(bias = blocks_4_attn_value_bias_to_fp16, dilations = var_953_dilations_0, groups = var_953_groups_0, pad = var_953_pad_0, pad_type = var_953_pad_type_0, strides = var_953_strides_0, weight = blocks_4_attn_value_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("op_953_cast_fp16")]; + tensor tile_12 = const()[name = tensor("tile_12"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_956_axis_0 = const()[name = tensor("op_956_axis_0"), val = tensor(1)]; + tensor var_956_cast_fp16_0, tensor var_956_cast_fp16_1, tensor var_956_cast_fp16_2, tensor var_956_cast_fp16_3, tensor var_956_cast_fp16_4, tensor var_956_cast_fp16_5, tensor var_956_cast_fp16_6, tensor var_956_cast_fp16_7, tensor var_956_cast_fp16_8, tensor var_956_cast_fp16_9, tensor var_956_cast_fp16_10, tensor var_956_cast_fp16_11 = split(axis = var_956_axis_0, split_sizes = tile_12, x = var_955_cast_fp16)[name = tensor("op_956_cast_fp16")]; + tensor var_969_perm_0 = const()[name = tensor("op_969_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_13 = const()[name = tensor("tile_13"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_970_axis_0 = const()[name = tensor("op_970_axis_0"), val = tensor(3)]; + tensor var_969_cast_fp16 = transpose(perm = var_969_perm_0, x = k_9_cast_fp16)[name = tensor("transpose_8")]; + tensor var_970_cast_fp16_0, tensor var_970_cast_fp16_1, tensor var_970_cast_fp16_2, tensor var_970_cast_fp16_3, tensor var_970_cast_fp16_4, tensor var_970_cast_fp16_5, tensor var_970_cast_fp16_6, tensor var_970_cast_fp16_7, tensor var_970_cast_fp16_8, tensor var_970_cast_fp16_9, tensor var_970_cast_fp16_10, tensor var_970_cast_fp16_11 = split(axis = var_970_axis_0, split_sizes = tile_13, x = var_969_cast_fp16)[name = tensor("op_970_cast_fp16")]; + tensor tile_14 = const()[name = tensor("tile_14"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_983_axis_0 = const()[name = tensor("op_983_axis_0"), val = tensor(1)]; + tensor var_983_cast_fp16_0, tensor var_983_cast_fp16_1, tensor var_983_cast_fp16_2, tensor var_983_cast_fp16_3, tensor var_983_cast_fp16_4, tensor var_983_cast_fp16_5, tensor var_983_cast_fp16_6, tensor var_983_cast_fp16_7, tensor var_983_cast_fp16_8, tensor var_983_cast_fp16_9, tensor var_983_cast_fp16_10, tensor var_983_cast_fp16_11 = split(axis = var_983_axis_0, split_sizes = tile_14, x = var_953_cast_fp16)[name = tensor("op_983_cast_fp16")]; + tensor aw_97_equation_0 = const()[name = tensor("aw_97_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_97_cast_fp16 = einsum(equation = aw_97_equation_0, values = (var_970_cast_fp16_0, var_956_cast_fp16_0))[name = tensor("aw_97_cast_fp16")]; + tensor aw_99_equation_0 = const()[name = tensor("aw_99_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_99_cast_fp16 = einsum(equation = aw_99_equation_0, values = (var_970_cast_fp16_1, var_956_cast_fp16_1))[name = tensor("aw_99_cast_fp16")]; + tensor aw_101_equation_0 = const()[name = tensor("aw_101_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_101_cast_fp16 = einsum(equation = aw_101_equation_0, values = (var_970_cast_fp16_2, var_956_cast_fp16_2))[name = tensor("aw_101_cast_fp16")]; + tensor aw_103_equation_0 = const()[name = tensor("aw_103_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_103_cast_fp16 = einsum(equation = aw_103_equation_0, values = (var_970_cast_fp16_3, var_956_cast_fp16_3))[name = tensor("aw_103_cast_fp16")]; + tensor aw_105_equation_0 = const()[name = tensor("aw_105_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_105_cast_fp16 = einsum(equation = aw_105_equation_0, values = (var_970_cast_fp16_4, var_956_cast_fp16_4))[name = tensor("aw_105_cast_fp16")]; + tensor aw_107_equation_0 = const()[name = tensor("aw_107_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_107_cast_fp16 = einsum(equation = aw_107_equation_0, values = (var_970_cast_fp16_5, var_956_cast_fp16_5))[name = tensor("aw_107_cast_fp16")]; + tensor aw_109_equation_0 = const()[name = tensor("aw_109_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_109_cast_fp16 = einsum(equation = aw_109_equation_0, values = (var_970_cast_fp16_6, var_956_cast_fp16_6))[name = tensor("aw_109_cast_fp16")]; + tensor aw_111_equation_0 = const()[name = tensor("aw_111_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_111_cast_fp16 = einsum(equation = aw_111_equation_0, values = (var_970_cast_fp16_7, var_956_cast_fp16_7))[name = tensor("aw_111_cast_fp16")]; + tensor aw_113_equation_0 = const()[name = tensor("aw_113_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_113_cast_fp16 = einsum(equation = aw_113_equation_0, values = (var_970_cast_fp16_8, var_956_cast_fp16_8))[name = tensor("aw_113_cast_fp16")]; + tensor aw_115_equation_0 = const()[name = tensor("aw_115_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_115_cast_fp16 = einsum(equation = aw_115_equation_0, values = (var_970_cast_fp16_9, var_956_cast_fp16_9))[name = tensor("aw_115_cast_fp16")]; + tensor aw_117_equation_0 = const()[name = tensor("aw_117_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_117_cast_fp16 = einsum(equation = aw_117_equation_0, values = (var_970_cast_fp16_10, var_956_cast_fp16_10))[name = tensor("aw_117_cast_fp16")]; + tensor aw_119_equation_0 = const()[name = tensor("aw_119_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_119_cast_fp16 = einsum(equation = aw_119_equation_0, values = (var_970_cast_fp16_11, var_956_cast_fp16_11))[name = tensor("aw_119_cast_fp16")]; + tensor var_1020_cast_fp16 = softmax(axis = var_904, x = aw_97_cast_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1021_cast_fp16 = softmax(axis = var_904, x = aw_99_cast_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1022_cast_fp16 = softmax(axis = var_904, x = aw_101_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = softmax(axis = var_904, x = aw_103_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = softmax(axis = var_904, x = aw_105_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor var_1025_cast_fp16 = softmax(axis = var_904, x = aw_107_cast_fp16)[name = tensor("op_1025_cast_fp16")]; + tensor var_1026_cast_fp16 = softmax(axis = var_904, x = aw_109_cast_fp16)[name = tensor("op_1026_cast_fp16")]; + tensor var_1027_cast_fp16 = softmax(axis = var_904, x = aw_111_cast_fp16)[name = tensor("op_1027_cast_fp16")]; + tensor var_1028_cast_fp16 = softmax(axis = var_904, x = aw_113_cast_fp16)[name = tensor("op_1028_cast_fp16")]; + tensor var_1029_cast_fp16 = softmax(axis = var_904, x = aw_115_cast_fp16)[name = tensor("op_1029_cast_fp16")]; + tensor var_1030_cast_fp16 = softmax(axis = var_904, x = aw_117_cast_fp16)[name = tensor("op_1030_cast_fp16")]; + tensor var_1031_cast_fp16 = softmax(axis = var_904, x = aw_119_cast_fp16)[name = tensor("op_1031_cast_fp16")]; + tensor var_1033_equation_0 = const()[name = tensor("op_1033_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1033_cast_fp16 = einsum(equation = var_1033_equation_0, values = (var_983_cast_fp16_0, var_1020_cast_fp16))[name = tensor("op_1033_cast_fp16")]; + tensor var_1035_equation_0 = const()[name = tensor("op_1035_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1035_cast_fp16 = einsum(equation = var_1035_equation_0, values = (var_983_cast_fp16_1, var_1021_cast_fp16))[name = tensor("op_1035_cast_fp16")]; + tensor var_1037_equation_0 = const()[name = tensor("op_1037_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1037_cast_fp16 = einsum(equation = var_1037_equation_0, values = (var_983_cast_fp16_2, var_1022_cast_fp16))[name = tensor("op_1037_cast_fp16")]; + tensor var_1039_equation_0 = const()[name = tensor("op_1039_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1039_cast_fp16 = einsum(equation = var_1039_equation_0, values = (var_983_cast_fp16_3, var_1023_cast_fp16))[name = tensor("op_1039_cast_fp16")]; + tensor var_1041_equation_0 = const()[name = tensor("op_1041_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1041_cast_fp16 = einsum(equation = var_1041_equation_0, values = (var_983_cast_fp16_4, var_1024_cast_fp16))[name = tensor("op_1041_cast_fp16")]; + tensor var_1043_equation_0 = const()[name = tensor("op_1043_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1043_cast_fp16 = einsum(equation = var_1043_equation_0, values = (var_983_cast_fp16_5, var_1025_cast_fp16))[name = tensor("op_1043_cast_fp16")]; + tensor var_1045_equation_0 = const()[name = tensor("op_1045_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1045_cast_fp16 = einsum(equation = var_1045_equation_0, values = (var_983_cast_fp16_6, var_1026_cast_fp16))[name = tensor("op_1045_cast_fp16")]; + tensor var_1047_equation_0 = const()[name = tensor("op_1047_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1047_cast_fp16 = einsum(equation = var_1047_equation_0, values = (var_983_cast_fp16_7, var_1027_cast_fp16))[name = tensor("op_1047_cast_fp16")]; + tensor var_1049_equation_0 = const()[name = tensor("op_1049_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1049_cast_fp16 = einsum(equation = var_1049_equation_0, values = (var_983_cast_fp16_8, var_1028_cast_fp16))[name = tensor("op_1049_cast_fp16")]; + tensor var_1051_equation_0 = const()[name = tensor("op_1051_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1051_cast_fp16 = einsum(equation = var_1051_equation_0, values = (var_983_cast_fp16_9, var_1029_cast_fp16))[name = tensor("op_1051_cast_fp16")]; + tensor var_1053_equation_0 = const()[name = tensor("op_1053_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1053_cast_fp16 = einsum(equation = var_1053_equation_0, values = (var_983_cast_fp16_10, var_1030_cast_fp16))[name = tensor("op_1053_cast_fp16")]; + tensor var_1055_equation_0 = const()[name = tensor("op_1055_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1055_cast_fp16 = einsum(equation = var_1055_equation_0, values = (var_983_cast_fp16_11, var_1031_cast_fp16))[name = tensor("op_1055_cast_fp16")]; + tensor input_45_interleave_0 = const()[name = tensor("input_45_interleave_0"), val = tensor(false)]; + tensor input_45_cast_fp16 = concat(axis = var_904, interleave = input_45_interleave_0, values = (var_1033_cast_fp16, var_1035_cast_fp16, var_1037_cast_fp16, var_1039_cast_fp16, var_1041_cast_fp16, var_1043_cast_fp16, var_1045_cast_fp16, var_1047_cast_fp16, var_1049_cast_fp16, var_1051_cast_fp16, var_1053_cast_fp16, var_1055_cast_fp16))[name = tensor("input_45_cast_fp16")]; + tensor var_1064_pad_type_0 = const()[name = tensor("op_1064_pad_type_0"), val = tensor("valid")]; + tensor var_1064_strides_0 = const()[name = tensor("op_1064_strides_0"), val = tensor([1, 1])]; + tensor var_1064_pad_0 = const()[name = tensor("op_1064_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1064_dilations_0 = const()[name = tensor("op_1064_dilations_0"), val = tensor([1, 1])]; + tensor var_1064_groups_0 = const()[name = tensor("op_1064_groups_0"), val = tensor(1)]; + tensor blocks_4_attn_out_weight_to_fp16 = const()[name = tensor("blocks_4_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66461248)))]; + tensor blocks_4_attn_out_bias_to_fp16 = const()[name = tensor("blocks_4_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67640960)))]; + tensor var_1064_cast_fp16 = conv(bias = blocks_4_attn_out_bias_to_fp16, dilations = var_1064_dilations_0, groups = var_1064_groups_0, pad = var_1064_pad_0, pad_type = var_1064_pad_type_0, strides = var_1064_strides_0, weight = blocks_4_attn_out_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("op_1064_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_1064_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([1])]; + tensor input_47_gamma_0_to_fp16 = const()[name = tensor("input_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67642560)))]; + tensor input_47_beta_0_to_fp16 = const()[name = tensor("input_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67644160)))]; + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_1074_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs_19_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; + tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; + tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67645760)))]; + tensor blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72364416)))]; + tensor input_49_cast_fp16 = conv(bias = blocks_4_mlp_0_bias_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = blocks_4_mlp_0_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1100_pad_type_0 = const()[name = tensor("op_1100_pad_type_0"), val = tensor("valid")]; + tensor var_1100_strides_0 = const()[name = tensor("op_1100_strides_0"), val = tensor([1, 1])]; + tensor var_1100_pad_0 = const()[name = tensor("op_1100_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1100_dilations_0 = const()[name = tensor("op_1100_dilations_0"), val = tensor([1, 1])]; + tensor var_1100_groups_0 = const()[name = tensor("op_1100_groups_0"), val = tensor(1)]; + tensor blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72370624)))]; + tensor blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77089280)))]; + tensor var_1100_cast_fp16 = conv(bias = blocks_4_mlp_2_bias_to_fp16, dilations = var_1100_dilations_0, groups = var_1100_groups_0, pad = var_1100_pad_0, pad_type = var_1100_pad_type_0, strides = var_1100_strides_0, weight = blocks_4_mlp_2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = var_1100_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor(1)]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([1])]; + tensor input_53_gamma_0_to_fp16 = const()[name = tensor("input_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77090880)))]; + tensor input_53_beta_0_to_fp16 = const()[name = tensor("input_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77092480)))]; + tensor var_1125_to_fp16 = const()[name = tensor("op_1125_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1125_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor q_11_pad_type_0 = const()[name = tensor("q_11_pad_type_0"), val = tensor("valid")]; + tensor q_11_strides_0 = const()[name = tensor("q_11_strides_0"), val = tensor([1, 1])]; + tensor q_11_pad_0 = const()[name = tensor("q_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_dilations_0 = const()[name = tensor("q_11_dilations_0"), val = tensor([1, 1])]; + tensor q_11_groups_0 = const()[name = tensor("q_11_groups_0"), val = tensor(1)]; + tensor var_1160_weight_0_to_fp16 = const()[name = tensor("op_1160_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77094080)))]; + tensor var_1160_bias_0_to_fp16 = const()[name = tensor("op_1160_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78273792)))]; + tensor var_1160_cast_fp16 = conv(bias = var_1160_bias_0_to_fp16, dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = var_1160_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1160_cast_fp16")]; + tensor k_11_pad_type_0 = const()[name = tensor("k_11_pad_type_0"), val = tensor("valid")]; + tensor k_11_strides_0 = const()[name = tensor("k_11_strides_0"), val = tensor([1, 1])]; + tensor k_11_pad_0 = const()[name = tensor("k_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_11_dilations_0 = const()[name = tensor("k_11_dilations_0"), val = tensor([1, 1])]; + tensor k_11_groups_0 = const()[name = tensor("k_11_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_key_weight_to_fp16 = const()[name = tensor("blocks_5_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78275392)))]; + tensor k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = blocks_5_attn_key_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_1158_pad_type_0 = const()[name = tensor("op_1158_pad_type_0"), val = tensor("valid")]; + tensor var_1158_strides_0 = const()[name = tensor("op_1158_strides_0"), val = tensor([1, 1])]; + tensor var_1158_pad_0 = const()[name = tensor("op_1158_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1158_dilations_0 = const()[name = tensor("op_1158_dilations_0"), val = tensor([1, 1])]; + tensor var_1158_groups_0 = const()[name = tensor("op_1158_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_value_weight_to_fp16 = const()[name = tensor("blocks_5_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79455104)))]; + tensor blocks_5_attn_value_bias_to_fp16 = const()[name = tensor("blocks_5_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80634816)))]; + tensor var_1158_cast_fp16 = conv(bias = blocks_5_attn_value_bias_to_fp16, dilations = var_1158_dilations_0, groups = var_1158_groups_0, pad = var_1158_pad_0, pad_type = var_1158_pad_type_0, strides = var_1158_strides_0, weight = blocks_5_attn_value_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor tile_15 = const()[name = tensor("tile_15"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1, tensor var_1161_cast_fp16_2, tensor var_1161_cast_fp16_3, tensor var_1161_cast_fp16_4, tensor var_1161_cast_fp16_5, tensor var_1161_cast_fp16_6, tensor var_1161_cast_fp16_7, tensor var_1161_cast_fp16_8, tensor var_1161_cast_fp16_9, tensor var_1161_cast_fp16_10, tensor var_1161_cast_fp16_11 = split(axis = var_1161_axis_0, split_sizes = tile_15, x = var_1160_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor var_1174_perm_0 = const()[name = tensor("op_1174_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_16 = const()[name = tensor("tile_16"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1175_axis_0 = const()[name = tensor("op_1175_axis_0"), val = tensor(3)]; + tensor var_1174_cast_fp16 = transpose(perm = var_1174_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_7")]; + tensor var_1175_cast_fp16_0, tensor var_1175_cast_fp16_1, tensor var_1175_cast_fp16_2, tensor var_1175_cast_fp16_3, tensor var_1175_cast_fp16_4, tensor var_1175_cast_fp16_5, tensor var_1175_cast_fp16_6, tensor var_1175_cast_fp16_7, tensor var_1175_cast_fp16_8, tensor var_1175_cast_fp16_9, tensor var_1175_cast_fp16_10, tensor var_1175_cast_fp16_11 = split(axis = var_1175_axis_0, split_sizes = tile_16, x = var_1174_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor tile_17 = const()[name = tensor("tile_17"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1188_axis_0 = const()[name = tensor("op_1188_axis_0"), val = tensor(1)]; + tensor var_1188_cast_fp16_0, tensor var_1188_cast_fp16_1, tensor var_1188_cast_fp16_2, tensor var_1188_cast_fp16_3, tensor var_1188_cast_fp16_4, tensor var_1188_cast_fp16_5, tensor var_1188_cast_fp16_6, tensor var_1188_cast_fp16_7, tensor var_1188_cast_fp16_8, tensor var_1188_cast_fp16_9, tensor var_1188_cast_fp16_10, tensor var_1188_cast_fp16_11 = split(axis = var_1188_axis_0, split_sizes = tile_17, x = var_1158_cast_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor aw_121_equation_0 = const()[name = tensor("aw_121_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_121_cast_fp16 = einsum(equation = aw_121_equation_0, values = (var_1175_cast_fp16_0, var_1161_cast_fp16_0))[name = tensor("aw_121_cast_fp16")]; + tensor aw_123_equation_0 = const()[name = tensor("aw_123_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_123_cast_fp16 = einsum(equation = aw_123_equation_0, values = (var_1175_cast_fp16_1, var_1161_cast_fp16_1))[name = tensor("aw_123_cast_fp16")]; + tensor aw_125_equation_0 = const()[name = tensor("aw_125_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_125_cast_fp16 = einsum(equation = aw_125_equation_0, values = (var_1175_cast_fp16_2, var_1161_cast_fp16_2))[name = tensor("aw_125_cast_fp16")]; + tensor aw_127_equation_0 = const()[name = tensor("aw_127_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_127_cast_fp16 = einsum(equation = aw_127_equation_0, values = (var_1175_cast_fp16_3, var_1161_cast_fp16_3))[name = tensor("aw_127_cast_fp16")]; + tensor aw_129_equation_0 = const()[name = tensor("aw_129_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_129_cast_fp16 = einsum(equation = aw_129_equation_0, values = (var_1175_cast_fp16_4, var_1161_cast_fp16_4))[name = tensor("aw_129_cast_fp16")]; + tensor aw_131_equation_0 = const()[name = tensor("aw_131_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_131_cast_fp16 = einsum(equation = aw_131_equation_0, values = (var_1175_cast_fp16_5, var_1161_cast_fp16_5))[name = tensor("aw_131_cast_fp16")]; + tensor aw_133_equation_0 = const()[name = tensor("aw_133_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_133_cast_fp16 = einsum(equation = aw_133_equation_0, values = (var_1175_cast_fp16_6, var_1161_cast_fp16_6))[name = tensor("aw_133_cast_fp16")]; + tensor aw_135_equation_0 = const()[name = tensor("aw_135_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_135_cast_fp16 = einsum(equation = aw_135_equation_0, values = (var_1175_cast_fp16_7, var_1161_cast_fp16_7))[name = tensor("aw_135_cast_fp16")]; + tensor aw_137_equation_0 = const()[name = tensor("aw_137_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_137_cast_fp16 = einsum(equation = aw_137_equation_0, values = (var_1175_cast_fp16_8, var_1161_cast_fp16_8))[name = tensor("aw_137_cast_fp16")]; + tensor aw_139_equation_0 = const()[name = tensor("aw_139_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_139_cast_fp16 = einsum(equation = aw_139_equation_0, values = (var_1175_cast_fp16_9, var_1161_cast_fp16_9))[name = tensor("aw_139_cast_fp16")]; + tensor aw_141_equation_0 = const()[name = tensor("aw_141_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_141_cast_fp16 = einsum(equation = aw_141_equation_0, values = (var_1175_cast_fp16_10, var_1161_cast_fp16_10))[name = tensor("aw_141_cast_fp16")]; + tensor aw_143_equation_0 = const()[name = tensor("aw_143_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_143_cast_fp16 = einsum(equation = aw_143_equation_0, values = (var_1175_cast_fp16_11, var_1161_cast_fp16_11))[name = tensor("aw_143_cast_fp16")]; + tensor var_1225_cast_fp16 = softmax(axis = var_1109, x = aw_121_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor var_1226_cast_fp16 = softmax(axis = var_1109, x = aw_123_cast_fp16)[name = tensor("op_1226_cast_fp16")]; + tensor var_1227_cast_fp16 = softmax(axis = var_1109, x = aw_125_cast_fp16)[name = tensor("op_1227_cast_fp16")]; + tensor var_1228_cast_fp16 = softmax(axis = var_1109, x = aw_127_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = softmax(axis = var_1109, x = aw_129_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor var_1230_cast_fp16 = softmax(axis = var_1109, x = aw_131_cast_fp16)[name = tensor("op_1230_cast_fp16")]; + tensor var_1231_cast_fp16 = softmax(axis = var_1109, x = aw_133_cast_fp16)[name = tensor("op_1231_cast_fp16")]; + tensor var_1232_cast_fp16 = softmax(axis = var_1109, x = aw_135_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = softmax(axis = var_1109, x = aw_137_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor var_1234_cast_fp16 = softmax(axis = var_1109, x = aw_139_cast_fp16)[name = tensor("op_1234_cast_fp16")]; + tensor var_1235_cast_fp16 = softmax(axis = var_1109, x = aw_141_cast_fp16)[name = tensor("op_1235_cast_fp16")]; + tensor var_1236_cast_fp16 = softmax(axis = var_1109, x = aw_143_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1238_equation_0 = const()[name = tensor("op_1238_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1238_cast_fp16 = einsum(equation = var_1238_equation_0, values = (var_1188_cast_fp16_0, var_1225_cast_fp16))[name = tensor("op_1238_cast_fp16")]; + tensor var_1240_equation_0 = const()[name = tensor("op_1240_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1240_cast_fp16 = einsum(equation = var_1240_equation_0, values = (var_1188_cast_fp16_1, var_1226_cast_fp16))[name = tensor("op_1240_cast_fp16")]; + tensor var_1242_equation_0 = const()[name = tensor("op_1242_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1242_cast_fp16 = einsum(equation = var_1242_equation_0, values = (var_1188_cast_fp16_2, var_1227_cast_fp16))[name = tensor("op_1242_cast_fp16")]; + tensor var_1244_equation_0 = const()[name = tensor("op_1244_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1244_cast_fp16 = einsum(equation = var_1244_equation_0, values = (var_1188_cast_fp16_3, var_1228_cast_fp16))[name = tensor("op_1244_cast_fp16")]; + tensor var_1246_equation_0 = const()[name = tensor("op_1246_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1246_cast_fp16 = einsum(equation = var_1246_equation_0, values = (var_1188_cast_fp16_4, var_1229_cast_fp16))[name = tensor("op_1246_cast_fp16")]; + tensor var_1248_equation_0 = const()[name = tensor("op_1248_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1248_cast_fp16 = einsum(equation = var_1248_equation_0, values = (var_1188_cast_fp16_5, var_1230_cast_fp16))[name = tensor("op_1248_cast_fp16")]; + tensor var_1250_equation_0 = const()[name = tensor("op_1250_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1250_cast_fp16 = einsum(equation = var_1250_equation_0, values = (var_1188_cast_fp16_6, var_1231_cast_fp16))[name = tensor("op_1250_cast_fp16")]; + tensor var_1252_equation_0 = const()[name = tensor("op_1252_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1252_cast_fp16 = einsum(equation = var_1252_equation_0, values = (var_1188_cast_fp16_7, var_1232_cast_fp16))[name = tensor("op_1252_cast_fp16")]; + tensor var_1254_equation_0 = const()[name = tensor("op_1254_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1254_cast_fp16 = einsum(equation = var_1254_equation_0, values = (var_1188_cast_fp16_8, var_1233_cast_fp16))[name = tensor("op_1254_cast_fp16")]; + tensor var_1256_equation_0 = const()[name = tensor("op_1256_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1256_cast_fp16 = einsum(equation = var_1256_equation_0, values = (var_1188_cast_fp16_9, var_1234_cast_fp16))[name = tensor("op_1256_cast_fp16")]; + tensor var_1258_equation_0 = const()[name = tensor("op_1258_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1258_cast_fp16 = einsum(equation = var_1258_equation_0, values = (var_1188_cast_fp16_10, var_1235_cast_fp16))[name = tensor("op_1258_cast_fp16")]; + tensor var_1260_equation_0 = const()[name = tensor("op_1260_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1260_cast_fp16 = einsum(equation = var_1260_equation_0, values = (var_1188_cast_fp16_11, var_1236_cast_fp16))[name = tensor("op_1260_cast_fp16")]; + tensor input_55_interleave_0 = const()[name = tensor("input_55_interleave_0"), val = tensor(false)]; + tensor input_55_cast_fp16 = concat(axis = var_1109, interleave = input_55_interleave_0, values = (var_1238_cast_fp16, var_1240_cast_fp16, var_1242_cast_fp16, var_1244_cast_fp16, var_1246_cast_fp16, var_1248_cast_fp16, var_1250_cast_fp16, var_1252_cast_fp16, var_1254_cast_fp16, var_1256_cast_fp16, var_1258_cast_fp16, var_1260_cast_fp16))[name = tensor("input_55_cast_fp16")]; + tensor var_1269_pad_type_0 = const()[name = tensor("op_1269_pad_type_0"), val = tensor("valid")]; + tensor var_1269_strides_0 = const()[name = tensor("op_1269_strides_0"), val = tensor([1, 1])]; + tensor var_1269_pad_0 = const()[name = tensor("op_1269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1269_dilations_0 = const()[name = tensor("op_1269_dilations_0"), val = tensor([1, 1])]; + tensor var_1269_groups_0 = const()[name = tensor("op_1269_groups_0"), val = tensor(1)]; + tensor blocks_5_attn_out_weight_to_fp16 = const()[name = tensor("blocks_5_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80636416)))]; + tensor blocks_5_attn_out_bias_to_fp16 = const()[name = tensor("blocks_5_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81816128)))]; + tensor var_1269_cast_fp16 = conv(bias = blocks_5_attn_out_bias_to_fp16, dilations = var_1269_dilations_0, groups = var_1269_groups_0, pad = var_1269_pad_0, pad_type = var_1269_pad_type_0, strides = var_1269_strides_0, weight = blocks_5_attn_out_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("op_1269_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_1269_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([1])]; + tensor input_57_gamma_0_to_fp16 = const()[name = tensor("input_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81817728)))]; + tensor input_57_beta_0_to_fp16 = const()[name = tensor("input_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81819328)))]; + tensor var_1279_to_fp16 = const()[name = tensor("op_1279_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1279_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs_23_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; + tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; + tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81820928)))]; + tensor blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86539584)))]; + tensor input_59_cast_fp16 = conv(bias = blocks_5_mlp_0_bias_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1305_pad_type_0 = const()[name = tensor("op_1305_pad_type_0"), val = tensor("valid")]; + tensor var_1305_strides_0 = const()[name = tensor("op_1305_strides_0"), val = tensor([1, 1])]; + tensor var_1305_pad_0 = const()[name = tensor("op_1305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1305_dilations_0 = const()[name = tensor("op_1305_dilations_0"), val = tensor([1, 1])]; + tensor var_1305_groups_0 = const()[name = tensor("op_1305_groups_0"), val = tensor(1)]; + tensor blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86545792)))]; + tensor blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91264448)))]; + tensor var_1305_cast_fp16 = conv(bias = blocks_5_mlp_2_bias_to_fp16, dilations = var_1305_dilations_0, groups = var_1305_groups_0, pad = var_1305_pad_0, pad_type = var_1305_pad_type_0, strides = var_1305_strides_0, weight = blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("op_1305_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = var_1305_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor(1)]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([1])]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91266048)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91267648)))]; + tensor var_1330_to_fp16 = const()[name = tensor("op_1330_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = input_63_beta_0_to_fp16, epsilon = var_1330_to_fp16, gamma = input_63_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor q_13_pad_type_0 = const()[name = tensor("q_13_pad_type_0"), val = tensor("valid")]; + tensor q_13_strides_0 = const()[name = tensor("q_13_strides_0"), val = tensor([1, 1])]; + tensor q_13_pad_0 = const()[name = tensor("q_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_dilations_0 = const()[name = tensor("q_13_dilations_0"), val = tensor([1, 1])]; + tensor q_13_groups_0 = const()[name = tensor("q_13_groups_0"), val = tensor(1)]; + tensor var_1365_weight_0_to_fp16 = const()[name = tensor("op_1365_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91269248)))]; + tensor var_1365_bias_0_to_fp16 = const()[name = tensor("op_1365_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92448960)))]; + tensor var_1365_cast_fp16 = conv(bias = var_1365_bias_0_to_fp16, dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = var_1365_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1365_cast_fp16")]; + tensor k_13_pad_type_0 = const()[name = tensor("k_13_pad_type_0"), val = tensor("valid")]; + tensor k_13_strides_0 = const()[name = tensor("k_13_strides_0"), val = tensor([1, 1])]; + tensor k_13_pad_0 = const()[name = tensor("k_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_13_dilations_0 = const()[name = tensor("k_13_dilations_0"), val = tensor([1, 1])]; + tensor k_13_groups_0 = const()[name = tensor("k_13_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_key_weight_to_fp16 = const()[name = tensor("blocks_6_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92450560)))]; + tensor k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = blocks_6_attn_key_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor var_1363_pad_type_0 = const()[name = tensor("op_1363_pad_type_0"), val = tensor("valid")]; + tensor var_1363_strides_0 = const()[name = tensor("op_1363_strides_0"), val = tensor([1, 1])]; + tensor var_1363_pad_0 = const()[name = tensor("op_1363_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1363_dilations_0 = const()[name = tensor("op_1363_dilations_0"), val = tensor([1, 1])]; + tensor var_1363_groups_0 = const()[name = tensor("op_1363_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_value_weight_to_fp16 = const()[name = tensor("blocks_6_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93630272)))]; + tensor blocks_6_attn_value_bias_to_fp16 = const()[name = tensor("blocks_6_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94809984)))]; + tensor var_1363_cast_fp16 = conv(bias = blocks_6_attn_value_bias_to_fp16, dilations = var_1363_dilations_0, groups = var_1363_groups_0, pad = var_1363_pad_0, pad_type = var_1363_pad_type_0, strides = var_1363_strides_0, weight = blocks_6_attn_value_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor tile_18 = const()[name = tensor("tile_18"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1366_axis_0 = const()[name = tensor("op_1366_axis_0"), val = tensor(1)]; + tensor var_1366_cast_fp16_0, tensor var_1366_cast_fp16_1, tensor var_1366_cast_fp16_2, tensor var_1366_cast_fp16_3, tensor var_1366_cast_fp16_4, tensor var_1366_cast_fp16_5, tensor var_1366_cast_fp16_6, tensor var_1366_cast_fp16_7, tensor var_1366_cast_fp16_8, tensor var_1366_cast_fp16_9, tensor var_1366_cast_fp16_10, tensor var_1366_cast_fp16_11 = split(axis = var_1366_axis_0, split_sizes = tile_18, x = var_1365_cast_fp16)[name = tensor("op_1366_cast_fp16")]; + tensor var_1379_perm_0 = const()[name = tensor("op_1379_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_19 = const()[name = tensor("tile_19"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(3)]; + tensor var_1379_cast_fp16 = transpose(perm = var_1379_perm_0, x = k_13_cast_fp16)[name = tensor("transpose_6")]; + tensor var_1380_cast_fp16_0, tensor var_1380_cast_fp16_1, tensor var_1380_cast_fp16_2, tensor var_1380_cast_fp16_3, tensor var_1380_cast_fp16_4, tensor var_1380_cast_fp16_5, tensor var_1380_cast_fp16_6, tensor var_1380_cast_fp16_7, tensor var_1380_cast_fp16_8, tensor var_1380_cast_fp16_9, tensor var_1380_cast_fp16_10, tensor var_1380_cast_fp16_11 = split(axis = var_1380_axis_0, split_sizes = tile_19, x = var_1379_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor tile_20 = const()[name = tensor("tile_20"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1393_axis_0 = const()[name = tensor("op_1393_axis_0"), val = tensor(1)]; + tensor var_1393_cast_fp16_0, tensor var_1393_cast_fp16_1, tensor var_1393_cast_fp16_2, tensor var_1393_cast_fp16_3, tensor var_1393_cast_fp16_4, tensor var_1393_cast_fp16_5, tensor var_1393_cast_fp16_6, tensor var_1393_cast_fp16_7, tensor var_1393_cast_fp16_8, tensor var_1393_cast_fp16_9, tensor var_1393_cast_fp16_10, tensor var_1393_cast_fp16_11 = split(axis = var_1393_axis_0, split_sizes = tile_20, x = var_1363_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor aw_145_equation_0 = const()[name = tensor("aw_145_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_145_cast_fp16 = einsum(equation = aw_145_equation_0, values = (var_1380_cast_fp16_0, var_1366_cast_fp16_0))[name = tensor("aw_145_cast_fp16")]; + tensor aw_147_equation_0 = const()[name = tensor("aw_147_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_147_cast_fp16 = einsum(equation = aw_147_equation_0, values = (var_1380_cast_fp16_1, var_1366_cast_fp16_1))[name = tensor("aw_147_cast_fp16")]; + tensor aw_149_equation_0 = const()[name = tensor("aw_149_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_149_cast_fp16 = einsum(equation = aw_149_equation_0, values = (var_1380_cast_fp16_2, var_1366_cast_fp16_2))[name = tensor("aw_149_cast_fp16")]; + tensor aw_151_equation_0 = const()[name = tensor("aw_151_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_151_cast_fp16 = einsum(equation = aw_151_equation_0, values = (var_1380_cast_fp16_3, var_1366_cast_fp16_3))[name = tensor("aw_151_cast_fp16")]; + tensor aw_153_equation_0 = const()[name = tensor("aw_153_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_153_cast_fp16 = einsum(equation = aw_153_equation_0, values = (var_1380_cast_fp16_4, var_1366_cast_fp16_4))[name = tensor("aw_153_cast_fp16")]; + tensor aw_155_equation_0 = const()[name = tensor("aw_155_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_155_cast_fp16 = einsum(equation = aw_155_equation_0, values = (var_1380_cast_fp16_5, var_1366_cast_fp16_5))[name = tensor("aw_155_cast_fp16")]; + tensor aw_157_equation_0 = const()[name = tensor("aw_157_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_157_cast_fp16 = einsum(equation = aw_157_equation_0, values = (var_1380_cast_fp16_6, var_1366_cast_fp16_6))[name = tensor("aw_157_cast_fp16")]; + tensor aw_159_equation_0 = const()[name = tensor("aw_159_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_159_cast_fp16 = einsum(equation = aw_159_equation_0, values = (var_1380_cast_fp16_7, var_1366_cast_fp16_7))[name = tensor("aw_159_cast_fp16")]; + tensor aw_161_equation_0 = const()[name = tensor("aw_161_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_161_cast_fp16 = einsum(equation = aw_161_equation_0, values = (var_1380_cast_fp16_8, var_1366_cast_fp16_8))[name = tensor("aw_161_cast_fp16")]; + tensor aw_163_equation_0 = const()[name = tensor("aw_163_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_163_cast_fp16 = einsum(equation = aw_163_equation_0, values = (var_1380_cast_fp16_9, var_1366_cast_fp16_9))[name = tensor("aw_163_cast_fp16")]; + tensor aw_165_equation_0 = const()[name = tensor("aw_165_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_165_cast_fp16 = einsum(equation = aw_165_equation_0, values = (var_1380_cast_fp16_10, var_1366_cast_fp16_10))[name = tensor("aw_165_cast_fp16")]; + tensor aw_167_equation_0 = const()[name = tensor("aw_167_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_167_cast_fp16 = einsum(equation = aw_167_equation_0, values = (var_1380_cast_fp16_11, var_1366_cast_fp16_11))[name = tensor("aw_167_cast_fp16")]; + tensor var_1430_cast_fp16 = softmax(axis = var_1314, x = aw_145_cast_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor var_1431_cast_fp16 = softmax(axis = var_1314, x = aw_147_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1432_cast_fp16 = softmax(axis = var_1314, x = aw_149_cast_fp16)[name = tensor("op_1432_cast_fp16")]; + tensor var_1433_cast_fp16 = softmax(axis = var_1314, x = aw_151_cast_fp16)[name = tensor("op_1433_cast_fp16")]; + tensor var_1434_cast_fp16 = softmax(axis = var_1314, x = aw_153_cast_fp16)[name = tensor("op_1434_cast_fp16")]; + tensor var_1435_cast_fp16 = softmax(axis = var_1314, x = aw_155_cast_fp16)[name = tensor("op_1435_cast_fp16")]; + tensor var_1436_cast_fp16 = softmax(axis = var_1314, x = aw_157_cast_fp16)[name = tensor("op_1436_cast_fp16")]; + tensor var_1437_cast_fp16 = softmax(axis = var_1314, x = aw_159_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = softmax(axis = var_1314, x = aw_161_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor var_1439_cast_fp16 = softmax(axis = var_1314, x = aw_163_cast_fp16)[name = tensor("op_1439_cast_fp16")]; + tensor var_1440_cast_fp16 = softmax(axis = var_1314, x = aw_165_cast_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1441_cast_fp16 = softmax(axis = var_1314, x = aw_167_cast_fp16)[name = tensor("op_1441_cast_fp16")]; + tensor var_1443_equation_0 = const()[name = tensor("op_1443_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1443_cast_fp16 = einsum(equation = var_1443_equation_0, values = (var_1393_cast_fp16_0, var_1430_cast_fp16))[name = tensor("op_1443_cast_fp16")]; + tensor var_1445_equation_0 = const()[name = tensor("op_1445_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1445_cast_fp16 = einsum(equation = var_1445_equation_0, values = (var_1393_cast_fp16_1, var_1431_cast_fp16))[name = tensor("op_1445_cast_fp16")]; + tensor var_1447_equation_0 = const()[name = tensor("op_1447_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1447_cast_fp16 = einsum(equation = var_1447_equation_0, values = (var_1393_cast_fp16_2, var_1432_cast_fp16))[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_equation_0 = const()[name = tensor("op_1449_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1449_cast_fp16 = einsum(equation = var_1449_equation_0, values = (var_1393_cast_fp16_3, var_1433_cast_fp16))[name = tensor("op_1449_cast_fp16")]; + tensor var_1451_equation_0 = const()[name = tensor("op_1451_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1451_cast_fp16 = einsum(equation = var_1451_equation_0, values = (var_1393_cast_fp16_4, var_1434_cast_fp16))[name = tensor("op_1451_cast_fp16")]; + tensor var_1453_equation_0 = const()[name = tensor("op_1453_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1453_cast_fp16 = einsum(equation = var_1453_equation_0, values = (var_1393_cast_fp16_5, var_1435_cast_fp16))[name = tensor("op_1453_cast_fp16")]; + tensor var_1455_equation_0 = const()[name = tensor("op_1455_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1455_cast_fp16 = einsum(equation = var_1455_equation_0, values = (var_1393_cast_fp16_6, var_1436_cast_fp16))[name = tensor("op_1455_cast_fp16")]; + tensor var_1457_equation_0 = const()[name = tensor("op_1457_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1457_cast_fp16 = einsum(equation = var_1457_equation_0, values = (var_1393_cast_fp16_7, var_1437_cast_fp16))[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_equation_0 = const()[name = tensor("op_1459_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1459_cast_fp16 = einsum(equation = var_1459_equation_0, values = (var_1393_cast_fp16_8, var_1438_cast_fp16))[name = tensor("op_1459_cast_fp16")]; + tensor var_1461_equation_0 = const()[name = tensor("op_1461_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1461_cast_fp16 = einsum(equation = var_1461_equation_0, values = (var_1393_cast_fp16_9, var_1439_cast_fp16))[name = tensor("op_1461_cast_fp16")]; + tensor var_1463_equation_0 = const()[name = tensor("op_1463_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1463_cast_fp16 = einsum(equation = var_1463_equation_0, values = (var_1393_cast_fp16_10, var_1440_cast_fp16))[name = tensor("op_1463_cast_fp16")]; + tensor var_1465_equation_0 = const()[name = tensor("op_1465_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1465_cast_fp16 = einsum(equation = var_1465_equation_0, values = (var_1393_cast_fp16_11, var_1441_cast_fp16))[name = tensor("op_1465_cast_fp16")]; + tensor input_65_interleave_0 = const()[name = tensor("input_65_interleave_0"), val = tensor(false)]; + tensor input_65_cast_fp16 = concat(axis = var_1314, interleave = input_65_interleave_0, values = (var_1443_cast_fp16, var_1445_cast_fp16, var_1447_cast_fp16, var_1449_cast_fp16, var_1451_cast_fp16, var_1453_cast_fp16, var_1455_cast_fp16, var_1457_cast_fp16, var_1459_cast_fp16, var_1461_cast_fp16, var_1463_cast_fp16, var_1465_cast_fp16))[name = tensor("input_65_cast_fp16")]; + tensor var_1474_pad_type_0 = const()[name = tensor("op_1474_pad_type_0"), val = tensor("valid")]; + tensor var_1474_strides_0 = const()[name = tensor("op_1474_strides_0"), val = tensor([1, 1])]; + tensor var_1474_pad_0 = const()[name = tensor("op_1474_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1474_dilations_0 = const()[name = tensor("op_1474_dilations_0"), val = tensor([1, 1])]; + tensor var_1474_groups_0 = const()[name = tensor("op_1474_groups_0"), val = tensor(1)]; + tensor blocks_6_attn_out_weight_to_fp16 = const()[name = tensor("blocks_6_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94811584)))]; + tensor blocks_6_attn_out_bias_to_fp16 = const()[name = tensor("blocks_6_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95991296)))]; + tensor var_1474_cast_fp16 = conv(bias = blocks_6_attn_out_bias_to_fp16, dilations = var_1474_dilations_0, groups = var_1474_groups_0, pad = var_1474_pad_0, pad_type = var_1474_pad_type_0, strides = var_1474_strides_0, weight = blocks_6_attn_out_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("op_1474_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = var_1474_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor input_67_axes_0 = const()[name = tensor("input_67_axes_0"), val = tensor([1])]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95992896)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95994496)))]; + tensor var_1484_to_fp16 = const()[name = tensor("op_1484_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1484_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_27_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95996096)))]; + tensor blocks_6_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100714752)))]; + tensor input_69_cast_fp16 = conv(bias = blocks_6_mlp_0_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = blocks_6_mlp_0_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1510_pad_type_0 = const()[name = tensor("op_1510_pad_type_0"), val = tensor("valid")]; + tensor var_1510_strides_0 = const()[name = tensor("op_1510_strides_0"), val = tensor([1, 1])]; + tensor var_1510_pad_0 = const()[name = tensor("op_1510_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1510_dilations_0 = const()[name = tensor("op_1510_dilations_0"), val = tensor([1, 1])]; + tensor var_1510_groups_0 = const()[name = tensor("op_1510_groups_0"), val = tensor(1)]; + tensor blocks_6_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100720960)))]; + tensor blocks_6_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105439616)))]; + tensor var_1510_cast_fp16 = conv(bias = blocks_6_mlp_2_bias_to_fp16, dilations = var_1510_dilations_0, groups = var_1510_groups_0, pad = var_1510_pad_0, pad_type = var_1510_pad_type_0, strides = var_1510_strides_0, weight = blocks_6_mlp_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_1510_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor(1)]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([1])]; + tensor input_73_gamma_0_to_fp16 = const()[name = tensor("input_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105441216)))]; + tensor input_73_beta_0_to_fp16 = const()[name = tensor("input_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105442816)))]; + tensor var_1535_to_fp16 = const()[name = tensor("op_1535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = input_73_beta_0_to_fp16, epsilon = var_1535_to_fp16, gamma = input_73_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor q_15_pad_type_0 = const()[name = tensor("q_15_pad_type_0"), val = tensor("valid")]; + tensor q_15_strides_0 = const()[name = tensor("q_15_strides_0"), val = tensor([1, 1])]; + tensor q_15_pad_0 = const()[name = tensor("q_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_15_dilations_0 = const()[name = tensor("q_15_dilations_0"), val = tensor([1, 1])]; + tensor q_15_groups_0 = const()[name = tensor("q_15_groups_0"), val = tensor(1)]; + tensor var_1570_weight_0_to_fp16 = const()[name = tensor("op_1570_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105444416)))]; + tensor var_1570_bias_0_to_fp16 = const()[name = tensor("op_1570_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106624128)))]; + tensor var_1570_cast_fp16 = conv(bias = var_1570_bias_0_to_fp16, dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = var_1570_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor k_15_pad_type_0 = const()[name = tensor("k_15_pad_type_0"), val = tensor("valid")]; + tensor k_15_strides_0 = const()[name = tensor("k_15_strides_0"), val = tensor([1, 1])]; + tensor k_15_pad_0 = const()[name = tensor("k_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_15_dilations_0 = const()[name = tensor("k_15_dilations_0"), val = tensor([1, 1])]; + tensor k_15_groups_0 = const()[name = tensor("k_15_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_key_weight_to_fp16 = const()[name = tensor("blocks_7_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106625728)))]; + tensor k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = blocks_7_attn_key_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_1568_pad_type_0 = const()[name = tensor("op_1568_pad_type_0"), val = tensor("valid")]; + tensor var_1568_strides_0 = const()[name = tensor("op_1568_strides_0"), val = tensor([1, 1])]; + tensor var_1568_pad_0 = const()[name = tensor("op_1568_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1568_dilations_0 = const()[name = tensor("op_1568_dilations_0"), val = tensor([1, 1])]; + tensor var_1568_groups_0 = const()[name = tensor("op_1568_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_value_weight_to_fp16 = const()[name = tensor("blocks_7_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107805440)))]; + tensor blocks_7_attn_value_bias_to_fp16 = const()[name = tensor("blocks_7_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108985152)))]; + tensor var_1568_cast_fp16 = conv(bias = blocks_7_attn_value_bias_to_fp16, dilations = var_1568_dilations_0, groups = var_1568_groups_0, pad = var_1568_pad_0, pad_type = var_1568_pad_type_0, strides = var_1568_strides_0, weight = blocks_7_attn_value_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor tile_21 = const()[name = tensor("tile_21"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1571_axis_0 = const()[name = tensor("op_1571_axis_0"), val = tensor(1)]; + tensor var_1571_cast_fp16_0, tensor var_1571_cast_fp16_1, tensor var_1571_cast_fp16_2, tensor var_1571_cast_fp16_3, tensor var_1571_cast_fp16_4, tensor var_1571_cast_fp16_5, tensor var_1571_cast_fp16_6, tensor var_1571_cast_fp16_7, tensor var_1571_cast_fp16_8, tensor var_1571_cast_fp16_9, tensor var_1571_cast_fp16_10, tensor var_1571_cast_fp16_11 = split(axis = var_1571_axis_0, split_sizes = tile_21, x = var_1570_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1584_perm_0 = const()[name = tensor("op_1584_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_22 = const()[name = tensor("tile_22"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1585_axis_0 = const()[name = tensor("op_1585_axis_0"), val = tensor(3)]; + tensor var_1584_cast_fp16 = transpose(perm = var_1584_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_5")]; + tensor var_1585_cast_fp16_0, tensor var_1585_cast_fp16_1, tensor var_1585_cast_fp16_2, tensor var_1585_cast_fp16_3, tensor var_1585_cast_fp16_4, tensor var_1585_cast_fp16_5, tensor var_1585_cast_fp16_6, tensor var_1585_cast_fp16_7, tensor var_1585_cast_fp16_8, tensor var_1585_cast_fp16_9, tensor var_1585_cast_fp16_10, tensor var_1585_cast_fp16_11 = split(axis = var_1585_axis_0, split_sizes = tile_22, x = var_1584_cast_fp16)[name = tensor("op_1585_cast_fp16")]; + tensor tile_23 = const()[name = tensor("tile_23"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1598_axis_0 = const()[name = tensor("op_1598_axis_0"), val = tensor(1)]; + tensor var_1598_cast_fp16_0, tensor var_1598_cast_fp16_1, tensor var_1598_cast_fp16_2, tensor var_1598_cast_fp16_3, tensor var_1598_cast_fp16_4, tensor var_1598_cast_fp16_5, tensor var_1598_cast_fp16_6, tensor var_1598_cast_fp16_7, tensor var_1598_cast_fp16_8, tensor var_1598_cast_fp16_9, tensor var_1598_cast_fp16_10, tensor var_1598_cast_fp16_11 = split(axis = var_1598_axis_0, split_sizes = tile_23, x = var_1568_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor aw_169_equation_0 = const()[name = tensor("aw_169_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_169_cast_fp16 = einsum(equation = aw_169_equation_0, values = (var_1585_cast_fp16_0, var_1571_cast_fp16_0))[name = tensor("aw_169_cast_fp16")]; + tensor aw_171_equation_0 = const()[name = tensor("aw_171_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_171_cast_fp16 = einsum(equation = aw_171_equation_0, values = (var_1585_cast_fp16_1, var_1571_cast_fp16_1))[name = tensor("aw_171_cast_fp16")]; + tensor aw_173_equation_0 = const()[name = tensor("aw_173_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_173_cast_fp16 = einsum(equation = aw_173_equation_0, values = (var_1585_cast_fp16_2, var_1571_cast_fp16_2))[name = tensor("aw_173_cast_fp16")]; + tensor aw_175_equation_0 = const()[name = tensor("aw_175_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_175_cast_fp16 = einsum(equation = aw_175_equation_0, values = (var_1585_cast_fp16_3, var_1571_cast_fp16_3))[name = tensor("aw_175_cast_fp16")]; + tensor aw_177_equation_0 = const()[name = tensor("aw_177_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_177_cast_fp16 = einsum(equation = aw_177_equation_0, values = (var_1585_cast_fp16_4, var_1571_cast_fp16_4))[name = tensor("aw_177_cast_fp16")]; + tensor aw_179_equation_0 = const()[name = tensor("aw_179_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_179_cast_fp16 = einsum(equation = aw_179_equation_0, values = (var_1585_cast_fp16_5, var_1571_cast_fp16_5))[name = tensor("aw_179_cast_fp16")]; + tensor aw_181_equation_0 = const()[name = tensor("aw_181_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_181_cast_fp16 = einsum(equation = aw_181_equation_0, values = (var_1585_cast_fp16_6, var_1571_cast_fp16_6))[name = tensor("aw_181_cast_fp16")]; + tensor aw_183_equation_0 = const()[name = tensor("aw_183_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_183_cast_fp16 = einsum(equation = aw_183_equation_0, values = (var_1585_cast_fp16_7, var_1571_cast_fp16_7))[name = tensor("aw_183_cast_fp16")]; + tensor aw_185_equation_0 = const()[name = tensor("aw_185_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_185_cast_fp16 = einsum(equation = aw_185_equation_0, values = (var_1585_cast_fp16_8, var_1571_cast_fp16_8))[name = tensor("aw_185_cast_fp16")]; + tensor aw_187_equation_0 = const()[name = tensor("aw_187_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_187_cast_fp16 = einsum(equation = aw_187_equation_0, values = (var_1585_cast_fp16_9, var_1571_cast_fp16_9))[name = tensor("aw_187_cast_fp16")]; + tensor aw_189_equation_0 = const()[name = tensor("aw_189_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_189_cast_fp16 = einsum(equation = aw_189_equation_0, values = (var_1585_cast_fp16_10, var_1571_cast_fp16_10))[name = tensor("aw_189_cast_fp16")]; + tensor aw_191_equation_0 = const()[name = tensor("aw_191_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_191_cast_fp16 = einsum(equation = aw_191_equation_0, values = (var_1585_cast_fp16_11, var_1571_cast_fp16_11))[name = tensor("aw_191_cast_fp16")]; + tensor var_1635_cast_fp16 = softmax(axis = var_1519, x = aw_169_cast_fp16)[name = tensor("op_1635_cast_fp16")]; + tensor var_1636_cast_fp16 = softmax(axis = var_1519, x = aw_171_cast_fp16)[name = tensor("op_1636_cast_fp16")]; + tensor var_1637_cast_fp16 = softmax(axis = var_1519, x = aw_173_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor var_1638_cast_fp16 = softmax(axis = var_1519, x = aw_175_cast_fp16)[name = tensor("op_1638_cast_fp16")]; + tensor var_1639_cast_fp16 = softmax(axis = var_1519, x = aw_177_cast_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1640_cast_fp16 = softmax(axis = var_1519, x = aw_179_cast_fp16)[name = tensor("op_1640_cast_fp16")]; + tensor var_1641_cast_fp16 = softmax(axis = var_1519, x = aw_181_cast_fp16)[name = tensor("op_1641_cast_fp16")]; + tensor var_1642_cast_fp16 = softmax(axis = var_1519, x = aw_183_cast_fp16)[name = tensor("op_1642_cast_fp16")]; + tensor var_1643_cast_fp16 = softmax(axis = var_1519, x = aw_185_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor var_1644_cast_fp16 = softmax(axis = var_1519, x = aw_187_cast_fp16)[name = tensor("op_1644_cast_fp16")]; + tensor var_1645_cast_fp16 = softmax(axis = var_1519, x = aw_189_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_cast_fp16 = softmax(axis = var_1519, x = aw_191_cast_fp16)[name = tensor("op_1646_cast_fp16")]; + tensor var_1648_equation_0 = const()[name = tensor("op_1648_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1648_cast_fp16 = einsum(equation = var_1648_equation_0, values = (var_1598_cast_fp16_0, var_1635_cast_fp16))[name = tensor("op_1648_cast_fp16")]; + tensor var_1650_equation_0 = const()[name = tensor("op_1650_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1650_cast_fp16 = einsum(equation = var_1650_equation_0, values = (var_1598_cast_fp16_1, var_1636_cast_fp16))[name = tensor("op_1650_cast_fp16")]; + tensor var_1652_equation_0 = const()[name = tensor("op_1652_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1652_cast_fp16 = einsum(equation = var_1652_equation_0, values = (var_1598_cast_fp16_2, var_1637_cast_fp16))[name = tensor("op_1652_cast_fp16")]; + tensor var_1654_equation_0 = const()[name = tensor("op_1654_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1654_cast_fp16 = einsum(equation = var_1654_equation_0, values = (var_1598_cast_fp16_3, var_1638_cast_fp16))[name = tensor("op_1654_cast_fp16")]; + tensor var_1656_equation_0 = const()[name = tensor("op_1656_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1656_cast_fp16 = einsum(equation = var_1656_equation_0, values = (var_1598_cast_fp16_4, var_1639_cast_fp16))[name = tensor("op_1656_cast_fp16")]; + tensor var_1658_equation_0 = const()[name = tensor("op_1658_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1658_cast_fp16 = einsum(equation = var_1658_equation_0, values = (var_1598_cast_fp16_5, var_1640_cast_fp16))[name = tensor("op_1658_cast_fp16")]; + tensor var_1660_equation_0 = const()[name = tensor("op_1660_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1660_cast_fp16 = einsum(equation = var_1660_equation_0, values = (var_1598_cast_fp16_6, var_1641_cast_fp16))[name = tensor("op_1660_cast_fp16")]; + tensor var_1662_equation_0 = const()[name = tensor("op_1662_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1662_cast_fp16 = einsum(equation = var_1662_equation_0, values = (var_1598_cast_fp16_7, var_1642_cast_fp16))[name = tensor("op_1662_cast_fp16")]; + tensor var_1664_equation_0 = const()[name = tensor("op_1664_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1664_cast_fp16 = einsum(equation = var_1664_equation_0, values = (var_1598_cast_fp16_8, var_1643_cast_fp16))[name = tensor("op_1664_cast_fp16")]; + tensor var_1666_equation_0 = const()[name = tensor("op_1666_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1666_cast_fp16 = einsum(equation = var_1666_equation_0, values = (var_1598_cast_fp16_9, var_1644_cast_fp16))[name = tensor("op_1666_cast_fp16")]; + tensor var_1668_equation_0 = const()[name = tensor("op_1668_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1668_cast_fp16 = einsum(equation = var_1668_equation_0, values = (var_1598_cast_fp16_10, var_1645_cast_fp16))[name = tensor("op_1668_cast_fp16")]; + tensor var_1670_equation_0 = const()[name = tensor("op_1670_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1670_cast_fp16 = einsum(equation = var_1670_equation_0, values = (var_1598_cast_fp16_11, var_1646_cast_fp16))[name = tensor("op_1670_cast_fp16")]; + tensor input_75_interleave_0 = const()[name = tensor("input_75_interleave_0"), val = tensor(false)]; + tensor input_75_cast_fp16 = concat(axis = var_1519, interleave = input_75_interleave_0, values = (var_1648_cast_fp16, var_1650_cast_fp16, var_1652_cast_fp16, var_1654_cast_fp16, var_1656_cast_fp16, var_1658_cast_fp16, var_1660_cast_fp16, var_1662_cast_fp16, var_1664_cast_fp16, var_1666_cast_fp16, var_1668_cast_fp16, var_1670_cast_fp16))[name = tensor("input_75_cast_fp16")]; + tensor var_1679_pad_type_0 = const()[name = tensor("op_1679_pad_type_0"), val = tensor("valid")]; + tensor var_1679_strides_0 = const()[name = tensor("op_1679_strides_0"), val = tensor([1, 1])]; + tensor var_1679_pad_0 = const()[name = tensor("op_1679_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1679_dilations_0 = const()[name = tensor("op_1679_dilations_0"), val = tensor([1, 1])]; + tensor var_1679_groups_0 = const()[name = tensor("op_1679_groups_0"), val = tensor(1)]; + tensor blocks_7_attn_out_weight_to_fp16 = const()[name = tensor("blocks_7_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108986752)))]; + tensor blocks_7_attn_out_bias_to_fp16 = const()[name = tensor("blocks_7_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110166464)))]; + tensor var_1679_cast_fp16 = conv(bias = blocks_7_attn_out_bias_to_fp16, dilations = var_1679_dilations_0, groups = var_1679_groups_0, pad = var_1679_pad_0, pad_type = var_1679_pad_type_0, strides = var_1679_strides_0, weight = blocks_7_attn_out_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("op_1679_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = var_1679_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([1])]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110168064)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110169664)))]; + tensor var_1689_to_fp16 = const()[name = tensor("op_1689_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = input_77_beta_0_to_fp16, epsilon = var_1689_to_fp16, gamma = input_77_gamma_0_to_fp16, x = inputs_31_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110171264)))]; + tensor blocks_7_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114889920)))]; + tensor input_79_cast_fp16 = conv(bias = blocks_7_mlp_0_bias_to_fp16, dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = blocks_7_mlp_0_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1715_pad_type_0 = const()[name = tensor("op_1715_pad_type_0"), val = tensor("valid")]; + tensor var_1715_strides_0 = const()[name = tensor("op_1715_strides_0"), val = tensor([1, 1])]; + tensor var_1715_pad_0 = const()[name = tensor("op_1715_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1715_dilations_0 = const()[name = tensor("op_1715_dilations_0"), val = tensor([1, 1])]; + tensor var_1715_groups_0 = const()[name = tensor("op_1715_groups_0"), val = tensor(1)]; + tensor blocks_7_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114896128)))]; + tensor blocks_7_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119614784)))]; + tensor var_1715_cast_fp16 = conv(bias = blocks_7_mlp_2_bias_to_fp16, dilations = var_1715_dilations_0, groups = var_1715_groups_0, pad = var_1715_pad_0, pad_type = var_1715_pad_type_0, strides = var_1715_strides_0, weight = blocks_7_mlp_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1715_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1724 = const()[name = tensor("op_1724"), val = tensor(1)]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([1])]; + tensor input_83_gamma_0_to_fp16 = const()[name = tensor("input_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119616384)))]; + tensor input_83_beta_0_to_fp16 = const()[name = tensor("input_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119617984)))]; + tensor var_1740_to_fp16 = const()[name = tensor("op_1740_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_1740_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs_33_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor q_17_pad_type_0 = const()[name = tensor("q_17_pad_type_0"), val = tensor("valid")]; + tensor q_17_strides_0 = const()[name = tensor("q_17_strides_0"), val = tensor([1, 1])]; + tensor q_17_pad_0 = const()[name = tensor("q_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_17_dilations_0 = const()[name = tensor("q_17_dilations_0"), val = tensor([1, 1])]; + tensor q_17_groups_0 = const()[name = tensor("q_17_groups_0"), val = tensor(1)]; + tensor var_1775_weight_0_to_fp16 = const()[name = tensor("op_1775_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119619584)))]; + tensor var_1775_bias_0_to_fp16 = const()[name = tensor("op_1775_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120799296)))]; + tensor var_1775_cast_fp16 = conv(bias = var_1775_bias_0_to_fp16, dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = var_1775_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("op_1775_cast_fp16")]; + tensor k_17_pad_type_0 = const()[name = tensor("k_17_pad_type_0"), val = tensor("valid")]; + tensor k_17_strides_0 = const()[name = tensor("k_17_strides_0"), val = tensor([1, 1])]; + tensor k_17_pad_0 = const()[name = tensor("k_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_17_dilations_0 = const()[name = tensor("k_17_dilations_0"), val = tensor([1, 1])]; + tensor k_17_groups_0 = const()[name = tensor("k_17_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_key_weight_to_fp16 = const()[name = tensor("blocks_8_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120800896)))]; + tensor k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = blocks_8_attn_key_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_1773_pad_type_0 = const()[name = tensor("op_1773_pad_type_0"), val = tensor("valid")]; + tensor var_1773_strides_0 = const()[name = tensor("op_1773_strides_0"), val = tensor([1, 1])]; + tensor var_1773_pad_0 = const()[name = tensor("op_1773_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1773_dilations_0 = const()[name = tensor("op_1773_dilations_0"), val = tensor([1, 1])]; + tensor var_1773_groups_0 = const()[name = tensor("op_1773_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_value_weight_to_fp16 = const()[name = tensor("blocks_8_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121980608)))]; + tensor blocks_8_attn_value_bias_to_fp16 = const()[name = tensor("blocks_8_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123160320)))]; + tensor var_1773_cast_fp16 = conv(bias = blocks_8_attn_value_bias_to_fp16, dilations = var_1773_dilations_0, groups = var_1773_groups_0, pad = var_1773_pad_0, pad_type = var_1773_pad_type_0, strides = var_1773_strides_0, weight = blocks_8_attn_value_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor tile_24 = const()[name = tensor("tile_24"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1776_axis_0 = const()[name = tensor("op_1776_axis_0"), val = tensor(1)]; + tensor var_1776_cast_fp16_0, tensor var_1776_cast_fp16_1, tensor var_1776_cast_fp16_2, tensor var_1776_cast_fp16_3, tensor var_1776_cast_fp16_4, tensor var_1776_cast_fp16_5, tensor var_1776_cast_fp16_6, tensor var_1776_cast_fp16_7, tensor var_1776_cast_fp16_8, tensor var_1776_cast_fp16_9, tensor var_1776_cast_fp16_10, tensor var_1776_cast_fp16_11 = split(axis = var_1776_axis_0, split_sizes = tile_24, x = var_1775_cast_fp16)[name = tensor("op_1776_cast_fp16")]; + tensor var_1789_perm_0 = const()[name = tensor("op_1789_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_25 = const()[name = tensor("tile_25"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1790_axis_0 = const()[name = tensor("op_1790_axis_0"), val = tensor(3)]; + tensor var_1789_cast_fp16 = transpose(perm = var_1789_perm_0, x = k_17_cast_fp16)[name = tensor("transpose_4")]; + tensor var_1790_cast_fp16_0, tensor var_1790_cast_fp16_1, tensor var_1790_cast_fp16_2, tensor var_1790_cast_fp16_3, tensor var_1790_cast_fp16_4, tensor var_1790_cast_fp16_5, tensor var_1790_cast_fp16_6, tensor var_1790_cast_fp16_7, tensor var_1790_cast_fp16_8, tensor var_1790_cast_fp16_9, tensor var_1790_cast_fp16_10, tensor var_1790_cast_fp16_11 = split(axis = var_1790_axis_0, split_sizes = tile_25, x = var_1789_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor tile_26 = const()[name = tensor("tile_26"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1803_axis_0 = const()[name = tensor("op_1803_axis_0"), val = tensor(1)]; + tensor var_1803_cast_fp16_0, tensor var_1803_cast_fp16_1, tensor var_1803_cast_fp16_2, tensor var_1803_cast_fp16_3, tensor var_1803_cast_fp16_4, tensor var_1803_cast_fp16_5, tensor var_1803_cast_fp16_6, tensor var_1803_cast_fp16_7, tensor var_1803_cast_fp16_8, tensor var_1803_cast_fp16_9, tensor var_1803_cast_fp16_10, tensor var_1803_cast_fp16_11 = split(axis = var_1803_axis_0, split_sizes = tile_26, x = var_1773_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor aw_193_equation_0 = const()[name = tensor("aw_193_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_193_cast_fp16 = einsum(equation = aw_193_equation_0, values = (var_1790_cast_fp16_0, var_1776_cast_fp16_0))[name = tensor("aw_193_cast_fp16")]; + tensor aw_195_equation_0 = const()[name = tensor("aw_195_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_195_cast_fp16 = einsum(equation = aw_195_equation_0, values = (var_1790_cast_fp16_1, var_1776_cast_fp16_1))[name = tensor("aw_195_cast_fp16")]; + tensor aw_197_equation_0 = const()[name = tensor("aw_197_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_197_cast_fp16 = einsum(equation = aw_197_equation_0, values = (var_1790_cast_fp16_2, var_1776_cast_fp16_2))[name = tensor("aw_197_cast_fp16")]; + tensor aw_199_equation_0 = const()[name = tensor("aw_199_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_199_cast_fp16 = einsum(equation = aw_199_equation_0, values = (var_1790_cast_fp16_3, var_1776_cast_fp16_3))[name = tensor("aw_199_cast_fp16")]; + tensor aw_201_equation_0 = const()[name = tensor("aw_201_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_201_cast_fp16 = einsum(equation = aw_201_equation_0, values = (var_1790_cast_fp16_4, var_1776_cast_fp16_4))[name = tensor("aw_201_cast_fp16")]; + tensor aw_203_equation_0 = const()[name = tensor("aw_203_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_203_cast_fp16 = einsum(equation = aw_203_equation_0, values = (var_1790_cast_fp16_5, var_1776_cast_fp16_5))[name = tensor("aw_203_cast_fp16")]; + tensor aw_205_equation_0 = const()[name = tensor("aw_205_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_205_cast_fp16 = einsum(equation = aw_205_equation_0, values = (var_1790_cast_fp16_6, var_1776_cast_fp16_6))[name = tensor("aw_205_cast_fp16")]; + tensor aw_207_equation_0 = const()[name = tensor("aw_207_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_207_cast_fp16 = einsum(equation = aw_207_equation_0, values = (var_1790_cast_fp16_7, var_1776_cast_fp16_7))[name = tensor("aw_207_cast_fp16")]; + tensor aw_209_equation_0 = const()[name = tensor("aw_209_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_209_cast_fp16 = einsum(equation = aw_209_equation_0, values = (var_1790_cast_fp16_8, var_1776_cast_fp16_8))[name = tensor("aw_209_cast_fp16")]; + tensor aw_211_equation_0 = const()[name = tensor("aw_211_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_211_cast_fp16 = einsum(equation = aw_211_equation_0, values = (var_1790_cast_fp16_9, var_1776_cast_fp16_9))[name = tensor("aw_211_cast_fp16")]; + tensor aw_213_equation_0 = const()[name = tensor("aw_213_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_213_cast_fp16 = einsum(equation = aw_213_equation_0, values = (var_1790_cast_fp16_10, var_1776_cast_fp16_10))[name = tensor("aw_213_cast_fp16")]; + tensor aw_215_equation_0 = const()[name = tensor("aw_215_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_215_cast_fp16 = einsum(equation = aw_215_equation_0, values = (var_1790_cast_fp16_11, var_1776_cast_fp16_11))[name = tensor("aw_215_cast_fp16")]; + tensor var_1840_cast_fp16 = softmax(axis = var_1724, x = aw_193_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1841_cast_fp16 = softmax(axis = var_1724, x = aw_195_cast_fp16)[name = tensor("op_1841_cast_fp16")]; + tensor var_1842_cast_fp16 = softmax(axis = var_1724, x = aw_197_cast_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = softmax(axis = var_1724, x = aw_199_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = softmax(axis = var_1724, x = aw_201_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor var_1845_cast_fp16 = softmax(axis = var_1724, x = aw_203_cast_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor var_1846_cast_fp16 = softmax(axis = var_1724, x = aw_205_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1847_cast_fp16 = softmax(axis = var_1724, x = aw_207_cast_fp16)[name = tensor("op_1847_cast_fp16")]; + tensor var_1848_cast_fp16 = softmax(axis = var_1724, x = aw_209_cast_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = softmax(axis = var_1724, x = aw_211_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1850_cast_fp16 = softmax(axis = var_1724, x = aw_213_cast_fp16)[name = tensor("op_1850_cast_fp16")]; + tensor var_1851_cast_fp16 = softmax(axis = var_1724, x = aw_215_cast_fp16)[name = tensor("op_1851_cast_fp16")]; + tensor var_1853_equation_0 = const()[name = tensor("op_1853_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1853_cast_fp16 = einsum(equation = var_1853_equation_0, values = (var_1803_cast_fp16_0, var_1840_cast_fp16))[name = tensor("op_1853_cast_fp16")]; + tensor var_1855_equation_0 = const()[name = tensor("op_1855_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1855_cast_fp16 = einsum(equation = var_1855_equation_0, values = (var_1803_cast_fp16_1, var_1841_cast_fp16))[name = tensor("op_1855_cast_fp16")]; + tensor var_1857_equation_0 = const()[name = tensor("op_1857_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1857_cast_fp16 = einsum(equation = var_1857_equation_0, values = (var_1803_cast_fp16_2, var_1842_cast_fp16))[name = tensor("op_1857_cast_fp16")]; + tensor var_1859_equation_0 = const()[name = tensor("op_1859_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1859_cast_fp16 = einsum(equation = var_1859_equation_0, values = (var_1803_cast_fp16_3, var_1843_cast_fp16))[name = tensor("op_1859_cast_fp16")]; + tensor var_1861_equation_0 = const()[name = tensor("op_1861_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1861_cast_fp16 = einsum(equation = var_1861_equation_0, values = (var_1803_cast_fp16_4, var_1844_cast_fp16))[name = tensor("op_1861_cast_fp16")]; + tensor var_1863_equation_0 = const()[name = tensor("op_1863_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1863_cast_fp16 = einsum(equation = var_1863_equation_0, values = (var_1803_cast_fp16_5, var_1845_cast_fp16))[name = tensor("op_1863_cast_fp16")]; + tensor var_1865_equation_0 = const()[name = tensor("op_1865_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1865_cast_fp16 = einsum(equation = var_1865_equation_0, values = (var_1803_cast_fp16_6, var_1846_cast_fp16))[name = tensor("op_1865_cast_fp16")]; + tensor var_1867_equation_0 = const()[name = tensor("op_1867_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1867_cast_fp16 = einsum(equation = var_1867_equation_0, values = (var_1803_cast_fp16_7, var_1847_cast_fp16))[name = tensor("op_1867_cast_fp16")]; + tensor var_1869_equation_0 = const()[name = tensor("op_1869_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1869_cast_fp16 = einsum(equation = var_1869_equation_0, values = (var_1803_cast_fp16_8, var_1848_cast_fp16))[name = tensor("op_1869_cast_fp16")]; + tensor var_1871_equation_0 = const()[name = tensor("op_1871_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1871_cast_fp16 = einsum(equation = var_1871_equation_0, values = (var_1803_cast_fp16_9, var_1849_cast_fp16))[name = tensor("op_1871_cast_fp16")]; + tensor var_1873_equation_0 = const()[name = tensor("op_1873_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1873_cast_fp16 = einsum(equation = var_1873_equation_0, values = (var_1803_cast_fp16_10, var_1850_cast_fp16))[name = tensor("op_1873_cast_fp16")]; + tensor var_1875_equation_0 = const()[name = tensor("op_1875_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_1875_cast_fp16 = einsum(equation = var_1875_equation_0, values = (var_1803_cast_fp16_11, var_1851_cast_fp16))[name = tensor("op_1875_cast_fp16")]; + tensor input_85_interleave_0 = const()[name = tensor("input_85_interleave_0"), val = tensor(false)]; + tensor input_85_cast_fp16 = concat(axis = var_1724, interleave = input_85_interleave_0, values = (var_1853_cast_fp16, var_1855_cast_fp16, var_1857_cast_fp16, var_1859_cast_fp16, var_1861_cast_fp16, var_1863_cast_fp16, var_1865_cast_fp16, var_1867_cast_fp16, var_1869_cast_fp16, var_1871_cast_fp16, var_1873_cast_fp16, var_1875_cast_fp16))[name = tensor("input_85_cast_fp16")]; + tensor var_1884_pad_type_0 = const()[name = tensor("op_1884_pad_type_0"), val = tensor("valid")]; + tensor var_1884_strides_0 = const()[name = tensor("op_1884_strides_0"), val = tensor([1, 1])]; + tensor var_1884_pad_0 = const()[name = tensor("op_1884_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1884_dilations_0 = const()[name = tensor("op_1884_dilations_0"), val = tensor([1, 1])]; + tensor var_1884_groups_0 = const()[name = tensor("op_1884_groups_0"), val = tensor(1)]; + tensor blocks_8_attn_out_weight_to_fp16 = const()[name = tensor("blocks_8_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123161920)))]; + tensor blocks_8_attn_out_bias_to_fp16 = const()[name = tensor("blocks_8_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124341632)))]; + tensor var_1884_cast_fp16 = conv(bias = blocks_8_attn_out_bias_to_fp16, dilations = var_1884_dilations_0, groups = var_1884_groups_0, pad = var_1884_pad_0, pad_type = var_1884_pad_type_0, strides = var_1884_strides_0, weight = blocks_8_attn_out_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = var_1884_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([1])]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124343232)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124344832)))]; + tensor var_1894_to_fp16 = const()[name = tensor("op_1894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = input_87_beta_0_to_fp16, epsilon = var_1894_to_fp16, gamma = input_87_gamma_0_to_fp16, x = inputs_35_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124346432)))]; + tensor blocks_8_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129065088)))]; + tensor input_89_cast_fp16 = conv(bias = blocks_8_mlp_0_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = blocks_8_mlp_0_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_1920_pad_type_0 = const()[name = tensor("op_1920_pad_type_0"), val = tensor("valid")]; + tensor var_1920_strides_0 = const()[name = tensor("op_1920_strides_0"), val = tensor([1, 1])]; + tensor var_1920_pad_0 = const()[name = tensor("op_1920_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1920_dilations_0 = const()[name = tensor("op_1920_dilations_0"), val = tensor([1, 1])]; + tensor var_1920_groups_0 = const()[name = tensor("op_1920_groups_0"), val = tensor(1)]; + tensor blocks_8_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129071296)))]; + tensor blocks_8_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133789952)))]; + tensor var_1920_cast_fp16 = conv(bias = blocks_8_mlp_2_bias_to_fp16, dilations = var_1920_dilations_0, groups = var_1920_groups_0, pad = var_1920_pad_0, pad_type = var_1920_pad_type_0, strides = var_1920_strides_0, weight = blocks_8_mlp_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = var_1920_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor(1)]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([1])]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133791552)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133793152)))]; + tensor var_1945_to_fp16 = const()[name = tensor("op_1945_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = input_93_beta_0_to_fp16, epsilon = var_1945_to_fp16, gamma = input_93_gamma_0_to_fp16, x = inputs_37_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor q_19_pad_type_0 = const()[name = tensor("q_19_pad_type_0"), val = tensor("valid")]; + tensor q_19_strides_0 = const()[name = tensor("q_19_strides_0"), val = tensor([1, 1])]; + tensor q_19_pad_0 = const()[name = tensor("q_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_dilations_0 = const()[name = tensor("q_19_dilations_0"), val = tensor([1, 1])]; + tensor q_19_groups_0 = const()[name = tensor("q_19_groups_0"), val = tensor(1)]; + tensor var_1980_weight_0_to_fp16 = const()[name = tensor("op_1980_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133794752)))]; + tensor var_1980_bias_0_to_fp16 = const()[name = tensor("op_1980_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134974464)))]; + tensor var_1980_cast_fp16 = conv(bias = var_1980_bias_0_to_fp16, dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = var_1980_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor k_19_pad_type_0 = const()[name = tensor("k_19_pad_type_0"), val = tensor("valid")]; + tensor k_19_strides_0 = const()[name = tensor("k_19_strides_0"), val = tensor([1, 1])]; + tensor k_19_pad_0 = const()[name = tensor("k_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_19_dilations_0 = const()[name = tensor("k_19_dilations_0"), val = tensor([1, 1])]; + tensor k_19_groups_0 = const()[name = tensor("k_19_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_key_weight_to_fp16 = const()[name = tensor("blocks_9_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134976064)))]; + tensor k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = blocks_9_attn_key_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_1978_pad_type_0 = const()[name = tensor("op_1978_pad_type_0"), val = tensor("valid")]; + tensor var_1978_strides_0 = const()[name = tensor("op_1978_strides_0"), val = tensor([1, 1])]; + tensor var_1978_pad_0 = const()[name = tensor("op_1978_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1978_dilations_0 = const()[name = tensor("op_1978_dilations_0"), val = tensor([1, 1])]; + tensor var_1978_groups_0 = const()[name = tensor("op_1978_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_value_weight_to_fp16 = const()[name = tensor("blocks_9_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136155776)))]; + tensor blocks_9_attn_value_bias_to_fp16 = const()[name = tensor("blocks_9_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137335488)))]; + tensor var_1978_cast_fp16 = conv(bias = blocks_9_attn_value_bias_to_fp16, dilations = var_1978_dilations_0, groups = var_1978_groups_0, pad = var_1978_pad_0, pad_type = var_1978_pad_type_0, strides = var_1978_strides_0, weight = blocks_9_attn_value_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor tile_27 = const()[name = tensor("tile_27"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1981_axis_0 = const()[name = tensor("op_1981_axis_0"), val = tensor(1)]; + tensor var_1981_cast_fp16_0, tensor var_1981_cast_fp16_1, tensor var_1981_cast_fp16_2, tensor var_1981_cast_fp16_3, tensor var_1981_cast_fp16_4, tensor var_1981_cast_fp16_5, tensor var_1981_cast_fp16_6, tensor var_1981_cast_fp16_7, tensor var_1981_cast_fp16_8, tensor var_1981_cast_fp16_9, tensor var_1981_cast_fp16_10, tensor var_1981_cast_fp16_11 = split(axis = var_1981_axis_0, split_sizes = tile_27, x = var_1980_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor var_1994_perm_0 = const()[name = tensor("op_1994_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_28 = const()[name = tensor("tile_28"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_1995_axis_0 = const()[name = tensor("op_1995_axis_0"), val = tensor(3)]; + tensor var_1994_cast_fp16 = transpose(perm = var_1994_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_3")]; + tensor var_1995_cast_fp16_0, tensor var_1995_cast_fp16_1, tensor var_1995_cast_fp16_2, tensor var_1995_cast_fp16_3, tensor var_1995_cast_fp16_4, tensor var_1995_cast_fp16_5, tensor var_1995_cast_fp16_6, tensor var_1995_cast_fp16_7, tensor var_1995_cast_fp16_8, tensor var_1995_cast_fp16_9, tensor var_1995_cast_fp16_10, tensor var_1995_cast_fp16_11 = split(axis = var_1995_axis_0, split_sizes = tile_28, x = var_1994_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor tile_29 = const()[name = tensor("tile_29"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_cast_fp16_0, tensor var_2008_cast_fp16_1, tensor var_2008_cast_fp16_2, tensor var_2008_cast_fp16_3, tensor var_2008_cast_fp16_4, tensor var_2008_cast_fp16_5, tensor var_2008_cast_fp16_6, tensor var_2008_cast_fp16_7, tensor var_2008_cast_fp16_8, tensor var_2008_cast_fp16_9, tensor var_2008_cast_fp16_10, tensor var_2008_cast_fp16_11 = split(axis = var_2008_axis_0, split_sizes = tile_29, x = var_1978_cast_fp16)[name = tensor("op_2008_cast_fp16")]; + tensor aw_217_equation_0 = const()[name = tensor("aw_217_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_217_cast_fp16 = einsum(equation = aw_217_equation_0, values = (var_1995_cast_fp16_0, var_1981_cast_fp16_0))[name = tensor("aw_217_cast_fp16")]; + tensor aw_219_equation_0 = const()[name = tensor("aw_219_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_219_cast_fp16 = einsum(equation = aw_219_equation_0, values = (var_1995_cast_fp16_1, var_1981_cast_fp16_1))[name = tensor("aw_219_cast_fp16")]; + tensor aw_221_equation_0 = const()[name = tensor("aw_221_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_221_cast_fp16 = einsum(equation = aw_221_equation_0, values = (var_1995_cast_fp16_2, var_1981_cast_fp16_2))[name = tensor("aw_221_cast_fp16")]; + tensor aw_223_equation_0 = const()[name = tensor("aw_223_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_223_cast_fp16 = einsum(equation = aw_223_equation_0, values = (var_1995_cast_fp16_3, var_1981_cast_fp16_3))[name = tensor("aw_223_cast_fp16")]; + tensor aw_225_equation_0 = const()[name = tensor("aw_225_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_225_cast_fp16 = einsum(equation = aw_225_equation_0, values = (var_1995_cast_fp16_4, var_1981_cast_fp16_4))[name = tensor("aw_225_cast_fp16")]; + tensor aw_227_equation_0 = const()[name = tensor("aw_227_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_227_cast_fp16 = einsum(equation = aw_227_equation_0, values = (var_1995_cast_fp16_5, var_1981_cast_fp16_5))[name = tensor("aw_227_cast_fp16")]; + tensor aw_229_equation_0 = const()[name = tensor("aw_229_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_229_cast_fp16 = einsum(equation = aw_229_equation_0, values = (var_1995_cast_fp16_6, var_1981_cast_fp16_6))[name = tensor("aw_229_cast_fp16")]; + tensor aw_231_equation_0 = const()[name = tensor("aw_231_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_231_cast_fp16 = einsum(equation = aw_231_equation_0, values = (var_1995_cast_fp16_7, var_1981_cast_fp16_7))[name = tensor("aw_231_cast_fp16")]; + tensor aw_233_equation_0 = const()[name = tensor("aw_233_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_233_cast_fp16 = einsum(equation = aw_233_equation_0, values = (var_1995_cast_fp16_8, var_1981_cast_fp16_8))[name = tensor("aw_233_cast_fp16")]; + tensor aw_235_equation_0 = const()[name = tensor("aw_235_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_235_cast_fp16 = einsum(equation = aw_235_equation_0, values = (var_1995_cast_fp16_9, var_1981_cast_fp16_9))[name = tensor("aw_235_cast_fp16")]; + tensor aw_237_equation_0 = const()[name = tensor("aw_237_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_237_cast_fp16 = einsum(equation = aw_237_equation_0, values = (var_1995_cast_fp16_10, var_1981_cast_fp16_10))[name = tensor("aw_237_cast_fp16")]; + tensor aw_239_equation_0 = const()[name = tensor("aw_239_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_239_cast_fp16 = einsum(equation = aw_239_equation_0, values = (var_1995_cast_fp16_11, var_1981_cast_fp16_11))[name = tensor("aw_239_cast_fp16")]; + tensor var_2045_cast_fp16 = softmax(axis = var_1929, x = aw_217_cast_fp16)[name = tensor("op_2045_cast_fp16")]; + tensor var_2046_cast_fp16 = softmax(axis = var_1929, x = aw_219_cast_fp16)[name = tensor("op_2046_cast_fp16")]; + tensor var_2047_cast_fp16 = softmax(axis = var_1929, x = aw_221_cast_fp16)[name = tensor("op_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = softmax(axis = var_1929, x = aw_223_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_cast_fp16 = softmax(axis = var_1929, x = aw_225_cast_fp16)[name = tensor("op_2049_cast_fp16")]; + tensor var_2050_cast_fp16 = softmax(axis = var_1929, x = aw_227_cast_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2051_cast_fp16 = softmax(axis = var_1929, x = aw_229_cast_fp16)[name = tensor("op_2051_cast_fp16")]; + tensor var_2052_cast_fp16 = softmax(axis = var_1929, x = aw_231_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor var_2053_cast_fp16 = softmax(axis = var_1929, x = aw_233_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2054_cast_fp16 = softmax(axis = var_1929, x = aw_235_cast_fp16)[name = tensor("op_2054_cast_fp16")]; + tensor var_2055_cast_fp16 = softmax(axis = var_1929, x = aw_237_cast_fp16)[name = tensor("op_2055_cast_fp16")]; + tensor var_2056_cast_fp16 = softmax(axis = var_1929, x = aw_239_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor var_2058_equation_0 = const()[name = tensor("op_2058_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2058_cast_fp16 = einsum(equation = var_2058_equation_0, values = (var_2008_cast_fp16_0, var_2045_cast_fp16))[name = tensor("op_2058_cast_fp16")]; + tensor var_2060_equation_0 = const()[name = tensor("op_2060_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2060_cast_fp16 = einsum(equation = var_2060_equation_0, values = (var_2008_cast_fp16_1, var_2046_cast_fp16))[name = tensor("op_2060_cast_fp16")]; + tensor var_2062_equation_0 = const()[name = tensor("op_2062_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2062_cast_fp16 = einsum(equation = var_2062_equation_0, values = (var_2008_cast_fp16_2, var_2047_cast_fp16))[name = tensor("op_2062_cast_fp16")]; + tensor var_2064_equation_0 = const()[name = tensor("op_2064_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2064_cast_fp16 = einsum(equation = var_2064_equation_0, values = (var_2008_cast_fp16_3, var_2048_cast_fp16))[name = tensor("op_2064_cast_fp16")]; + tensor var_2066_equation_0 = const()[name = tensor("op_2066_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2066_cast_fp16 = einsum(equation = var_2066_equation_0, values = (var_2008_cast_fp16_4, var_2049_cast_fp16))[name = tensor("op_2066_cast_fp16")]; + tensor var_2068_equation_0 = const()[name = tensor("op_2068_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2068_cast_fp16 = einsum(equation = var_2068_equation_0, values = (var_2008_cast_fp16_5, var_2050_cast_fp16))[name = tensor("op_2068_cast_fp16")]; + tensor var_2070_equation_0 = const()[name = tensor("op_2070_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2070_cast_fp16 = einsum(equation = var_2070_equation_0, values = (var_2008_cast_fp16_6, var_2051_cast_fp16))[name = tensor("op_2070_cast_fp16")]; + tensor var_2072_equation_0 = const()[name = tensor("op_2072_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2072_cast_fp16 = einsum(equation = var_2072_equation_0, values = (var_2008_cast_fp16_7, var_2052_cast_fp16))[name = tensor("op_2072_cast_fp16")]; + tensor var_2074_equation_0 = const()[name = tensor("op_2074_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2074_cast_fp16 = einsum(equation = var_2074_equation_0, values = (var_2008_cast_fp16_8, var_2053_cast_fp16))[name = tensor("op_2074_cast_fp16")]; + tensor var_2076_equation_0 = const()[name = tensor("op_2076_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2076_cast_fp16 = einsum(equation = var_2076_equation_0, values = (var_2008_cast_fp16_9, var_2054_cast_fp16))[name = tensor("op_2076_cast_fp16")]; + tensor var_2078_equation_0 = const()[name = tensor("op_2078_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2078_cast_fp16 = einsum(equation = var_2078_equation_0, values = (var_2008_cast_fp16_10, var_2055_cast_fp16))[name = tensor("op_2078_cast_fp16")]; + tensor var_2080_equation_0 = const()[name = tensor("op_2080_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2080_cast_fp16 = einsum(equation = var_2080_equation_0, values = (var_2008_cast_fp16_11, var_2056_cast_fp16))[name = tensor("op_2080_cast_fp16")]; + tensor input_95_interleave_0 = const()[name = tensor("input_95_interleave_0"), val = tensor(false)]; + tensor input_95_cast_fp16 = concat(axis = var_1929, interleave = input_95_interleave_0, values = (var_2058_cast_fp16, var_2060_cast_fp16, var_2062_cast_fp16, var_2064_cast_fp16, var_2066_cast_fp16, var_2068_cast_fp16, var_2070_cast_fp16, var_2072_cast_fp16, var_2074_cast_fp16, var_2076_cast_fp16, var_2078_cast_fp16, var_2080_cast_fp16))[name = tensor("input_95_cast_fp16")]; + tensor var_2089_pad_type_0 = const()[name = tensor("op_2089_pad_type_0"), val = tensor("valid")]; + tensor var_2089_strides_0 = const()[name = tensor("op_2089_strides_0"), val = tensor([1, 1])]; + tensor var_2089_pad_0 = const()[name = tensor("op_2089_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2089_dilations_0 = const()[name = tensor("op_2089_dilations_0"), val = tensor([1, 1])]; + tensor var_2089_groups_0 = const()[name = tensor("op_2089_groups_0"), val = tensor(1)]; + tensor blocks_9_attn_out_weight_to_fp16 = const()[name = tensor("blocks_9_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137337088)))]; + tensor blocks_9_attn_out_bias_to_fp16 = const()[name = tensor("blocks_9_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138516800)))]; + tensor var_2089_cast_fp16 = conv(bias = blocks_9_attn_out_bias_to_fp16, dilations = var_2089_dilations_0, groups = var_2089_groups_0, pad = var_2089_pad_0, pad_type = var_2089_pad_type_0, strides = var_2089_strides_0, weight = blocks_9_attn_out_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("op_2089_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_2089_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor input_97_axes_0 = const()[name = tensor("input_97_axes_0"), val = tensor([1])]; + tensor input_97_gamma_0_to_fp16 = const()[name = tensor("input_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138518400)))]; + tensor input_97_beta_0_to_fp16 = const()[name = tensor("input_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138520000)))]; + tensor var_2099_to_fp16 = const()[name = tensor("op_2099_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = input_97_beta_0_to_fp16, epsilon = var_2099_to_fp16, gamma = input_97_gamma_0_to_fp16, x = inputs_39_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; + tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; + tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138521600)))]; + tensor blocks_9_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143240256)))]; + tensor input_99_cast_fp16 = conv(bias = blocks_9_mlp_0_bias_to_fp16, dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = blocks_9_mlp_0_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_2125_pad_type_0 = const()[name = tensor("op_2125_pad_type_0"), val = tensor("valid")]; + tensor var_2125_strides_0 = const()[name = tensor("op_2125_strides_0"), val = tensor([1, 1])]; + tensor var_2125_pad_0 = const()[name = tensor("op_2125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2125_dilations_0 = const()[name = tensor("op_2125_dilations_0"), val = tensor([1, 1])]; + tensor var_2125_groups_0 = const()[name = tensor("op_2125_groups_0"), val = tensor(1)]; + tensor blocks_9_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143246464)))]; + tensor blocks_9_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147965120)))]; + tensor var_2125_cast_fp16 = conv(bias = blocks_9_mlp_2_bias_to_fp16, dilations = var_2125_dilations_0, groups = var_2125_groups_0, pad = var_2125_pad_0, pad_type = var_2125_pad_type_0, strides = var_2125_strides_0, weight = blocks_9_mlp_2_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = var_2125_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor(1)]; + tensor input_103_axes_0 = const()[name = tensor("input_103_axes_0"), val = tensor([1])]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147966720)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147968320)))]; + tensor var_2150_to_fp16 = const()[name = tensor("op_2150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2150_to_fp16, gamma = input_103_gamma_0_to_fp16, x = inputs_41_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor q_21_pad_type_0 = const()[name = tensor("q_21_pad_type_0"), val = tensor("valid")]; + tensor q_21_strides_0 = const()[name = tensor("q_21_strides_0"), val = tensor([1, 1])]; + tensor q_21_pad_0 = const()[name = tensor("q_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_21_dilations_0 = const()[name = tensor("q_21_dilations_0"), val = tensor([1, 1])]; + tensor q_21_groups_0 = const()[name = tensor("q_21_groups_0"), val = tensor(1)]; + tensor var_2185_weight_0_to_fp16 = const()[name = tensor("op_2185_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147969920)))]; + tensor var_2185_bias_0_to_fp16 = const()[name = tensor("op_2185_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149149632)))]; + tensor var_2185_cast_fp16 = conv(bias = var_2185_bias_0_to_fp16, dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = var_2185_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor k_21_pad_type_0 = const()[name = tensor("k_21_pad_type_0"), val = tensor("valid")]; + tensor k_21_strides_0 = const()[name = tensor("k_21_strides_0"), val = tensor([1, 1])]; + tensor k_21_pad_0 = const()[name = tensor("k_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_21_dilations_0 = const()[name = tensor("k_21_dilations_0"), val = tensor([1, 1])]; + tensor k_21_groups_0 = const()[name = tensor("k_21_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_key_weight_to_fp16 = const()[name = tensor("blocks_10_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149151232)))]; + tensor k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = blocks_10_attn_key_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor var_2183_pad_type_0 = const()[name = tensor("op_2183_pad_type_0"), val = tensor("valid")]; + tensor var_2183_strides_0 = const()[name = tensor("op_2183_strides_0"), val = tensor([1, 1])]; + tensor var_2183_pad_0 = const()[name = tensor("op_2183_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2183_dilations_0 = const()[name = tensor("op_2183_dilations_0"), val = tensor([1, 1])]; + tensor var_2183_groups_0 = const()[name = tensor("op_2183_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_value_weight_to_fp16 = const()[name = tensor("blocks_10_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150330944)))]; + tensor blocks_10_attn_value_bias_to_fp16 = const()[name = tensor("blocks_10_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151510656)))]; + tensor var_2183_cast_fp16 = conv(bias = blocks_10_attn_value_bias_to_fp16, dilations = var_2183_dilations_0, groups = var_2183_groups_0, pad = var_2183_pad_0, pad_type = var_2183_pad_type_0, strides = var_2183_strides_0, weight = blocks_10_attn_value_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("op_2183_cast_fp16")]; + tensor tile_30 = const()[name = tensor("tile_30"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2186_axis_0 = const()[name = tensor("op_2186_axis_0"), val = tensor(1)]; + tensor var_2186_cast_fp16_0, tensor var_2186_cast_fp16_1, tensor var_2186_cast_fp16_2, tensor var_2186_cast_fp16_3, tensor var_2186_cast_fp16_4, tensor var_2186_cast_fp16_5, tensor var_2186_cast_fp16_6, tensor var_2186_cast_fp16_7, tensor var_2186_cast_fp16_8, tensor var_2186_cast_fp16_9, tensor var_2186_cast_fp16_10, tensor var_2186_cast_fp16_11 = split(axis = var_2186_axis_0, split_sizes = tile_30, x = var_2185_cast_fp16)[name = tensor("op_2186_cast_fp16")]; + tensor var_2199_perm_0 = const()[name = tensor("op_2199_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_31 = const()[name = tensor("tile_31"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2200_axis_0 = const()[name = tensor("op_2200_axis_0"), val = tensor(3)]; + tensor var_2199_cast_fp16 = transpose(perm = var_2199_perm_0, x = k_21_cast_fp16)[name = tensor("transpose_2")]; + tensor var_2200_cast_fp16_0, tensor var_2200_cast_fp16_1, tensor var_2200_cast_fp16_2, tensor var_2200_cast_fp16_3, tensor var_2200_cast_fp16_4, tensor var_2200_cast_fp16_5, tensor var_2200_cast_fp16_6, tensor var_2200_cast_fp16_7, tensor var_2200_cast_fp16_8, tensor var_2200_cast_fp16_9, tensor var_2200_cast_fp16_10, tensor var_2200_cast_fp16_11 = split(axis = var_2200_axis_0, split_sizes = tile_31, x = var_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor tile_32 = const()[name = tensor("tile_32"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2213_axis_0 = const()[name = tensor("op_2213_axis_0"), val = tensor(1)]; + tensor var_2213_cast_fp16_0, tensor var_2213_cast_fp16_1, tensor var_2213_cast_fp16_2, tensor var_2213_cast_fp16_3, tensor var_2213_cast_fp16_4, tensor var_2213_cast_fp16_5, tensor var_2213_cast_fp16_6, tensor var_2213_cast_fp16_7, tensor var_2213_cast_fp16_8, tensor var_2213_cast_fp16_9, tensor var_2213_cast_fp16_10, tensor var_2213_cast_fp16_11 = split(axis = var_2213_axis_0, split_sizes = tile_32, x = var_2183_cast_fp16)[name = tensor("op_2213_cast_fp16")]; + tensor aw_241_equation_0 = const()[name = tensor("aw_241_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_241_cast_fp16 = einsum(equation = aw_241_equation_0, values = (var_2200_cast_fp16_0, var_2186_cast_fp16_0))[name = tensor("aw_241_cast_fp16")]; + tensor aw_243_equation_0 = const()[name = tensor("aw_243_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_243_cast_fp16 = einsum(equation = aw_243_equation_0, values = (var_2200_cast_fp16_1, var_2186_cast_fp16_1))[name = tensor("aw_243_cast_fp16")]; + tensor aw_245_equation_0 = const()[name = tensor("aw_245_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_245_cast_fp16 = einsum(equation = aw_245_equation_0, values = (var_2200_cast_fp16_2, var_2186_cast_fp16_2))[name = tensor("aw_245_cast_fp16")]; + tensor aw_247_equation_0 = const()[name = tensor("aw_247_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_247_cast_fp16 = einsum(equation = aw_247_equation_0, values = (var_2200_cast_fp16_3, var_2186_cast_fp16_3))[name = tensor("aw_247_cast_fp16")]; + tensor aw_249_equation_0 = const()[name = tensor("aw_249_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_249_cast_fp16 = einsum(equation = aw_249_equation_0, values = (var_2200_cast_fp16_4, var_2186_cast_fp16_4))[name = tensor("aw_249_cast_fp16")]; + tensor aw_251_equation_0 = const()[name = tensor("aw_251_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_251_cast_fp16 = einsum(equation = aw_251_equation_0, values = (var_2200_cast_fp16_5, var_2186_cast_fp16_5))[name = tensor("aw_251_cast_fp16")]; + tensor aw_253_equation_0 = const()[name = tensor("aw_253_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_253_cast_fp16 = einsum(equation = aw_253_equation_0, values = (var_2200_cast_fp16_6, var_2186_cast_fp16_6))[name = tensor("aw_253_cast_fp16")]; + tensor aw_255_equation_0 = const()[name = tensor("aw_255_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_255_cast_fp16 = einsum(equation = aw_255_equation_0, values = (var_2200_cast_fp16_7, var_2186_cast_fp16_7))[name = tensor("aw_255_cast_fp16")]; + tensor aw_257_equation_0 = const()[name = tensor("aw_257_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_257_cast_fp16 = einsum(equation = aw_257_equation_0, values = (var_2200_cast_fp16_8, var_2186_cast_fp16_8))[name = tensor("aw_257_cast_fp16")]; + tensor aw_259_equation_0 = const()[name = tensor("aw_259_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_259_cast_fp16 = einsum(equation = aw_259_equation_0, values = (var_2200_cast_fp16_9, var_2186_cast_fp16_9))[name = tensor("aw_259_cast_fp16")]; + tensor aw_261_equation_0 = const()[name = tensor("aw_261_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_261_cast_fp16 = einsum(equation = aw_261_equation_0, values = (var_2200_cast_fp16_10, var_2186_cast_fp16_10))[name = tensor("aw_261_cast_fp16")]; + tensor aw_263_equation_0 = const()[name = tensor("aw_263_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_263_cast_fp16 = einsum(equation = aw_263_equation_0, values = (var_2200_cast_fp16_11, var_2186_cast_fp16_11))[name = tensor("aw_263_cast_fp16")]; + tensor var_2250_cast_fp16 = softmax(axis = var_2134, x = aw_241_cast_fp16)[name = tensor("op_2250_cast_fp16")]; + tensor var_2251_cast_fp16 = softmax(axis = var_2134, x = aw_243_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor var_2252_cast_fp16 = softmax(axis = var_2134, x = aw_245_cast_fp16)[name = tensor("op_2252_cast_fp16")]; + tensor var_2253_cast_fp16 = softmax(axis = var_2134, x = aw_247_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor var_2254_cast_fp16 = softmax(axis = var_2134, x = aw_249_cast_fp16)[name = tensor("op_2254_cast_fp16")]; + tensor var_2255_cast_fp16 = softmax(axis = var_2134, x = aw_251_cast_fp16)[name = tensor("op_2255_cast_fp16")]; + tensor var_2256_cast_fp16 = softmax(axis = var_2134, x = aw_253_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor var_2257_cast_fp16 = softmax(axis = var_2134, x = aw_255_cast_fp16)[name = tensor("op_2257_cast_fp16")]; + tensor var_2258_cast_fp16 = softmax(axis = var_2134, x = aw_257_cast_fp16)[name = tensor("op_2258_cast_fp16")]; + tensor var_2259_cast_fp16 = softmax(axis = var_2134, x = aw_259_cast_fp16)[name = tensor("op_2259_cast_fp16")]; + tensor var_2260_cast_fp16 = softmax(axis = var_2134, x = aw_261_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = softmax(axis = var_2134, x = aw_263_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor var_2263_equation_0 = const()[name = tensor("op_2263_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2263_cast_fp16 = einsum(equation = var_2263_equation_0, values = (var_2213_cast_fp16_0, var_2250_cast_fp16))[name = tensor("op_2263_cast_fp16")]; + tensor var_2265_equation_0 = const()[name = tensor("op_2265_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2265_cast_fp16 = einsum(equation = var_2265_equation_0, values = (var_2213_cast_fp16_1, var_2251_cast_fp16))[name = tensor("op_2265_cast_fp16")]; + tensor var_2267_equation_0 = const()[name = tensor("op_2267_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2267_cast_fp16 = einsum(equation = var_2267_equation_0, values = (var_2213_cast_fp16_2, var_2252_cast_fp16))[name = tensor("op_2267_cast_fp16")]; + tensor var_2269_equation_0 = const()[name = tensor("op_2269_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2269_cast_fp16 = einsum(equation = var_2269_equation_0, values = (var_2213_cast_fp16_3, var_2253_cast_fp16))[name = tensor("op_2269_cast_fp16")]; + tensor var_2271_equation_0 = const()[name = tensor("op_2271_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2271_cast_fp16 = einsum(equation = var_2271_equation_0, values = (var_2213_cast_fp16_4, var_2254_cast_fp16))[name = tensor("op_2271_cast_fp16")]; + tensor var_2273_equation_0 = const()[name = tensor("op_2273_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2273_cast_fp16 = einsum(equation = var_2273_equation_0, values = (var_2213_cast_fp16_5, var_2255_cast_fp16))[name = tensor("op_2273_cast_fp16")]; + tensor var_2275_equation_0 = const()[name = tensor("op_2275_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2275_cast_fp16 = einsum(equation = var_2275_equation_0, values = (var_2213_cast_fp16_6, var_2256_cast_fp16))[name = tensor("op_2275_cast_fp16")]; + tensor var_2277_equation_0 = const()[name = tensor("op_2277_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2277_cast_fp16 = einsum(equation = var_2277_equation_0, values = (var_2213_cast_fp16_7, var_2257_cast_fp16))[name = tensor("op_2277_cast_fp16")]; + tensor var_2279_equation_0 = const()[name = tensor("op_2279_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2279_cast_fp16 = einsum(equation = var_2279_equation_0, values = (var_2213_cast_fp16_8, var_2258_cast_fp16))[name = tensor("op_2279_cast_fp16")]; + tensor var_2281_equation_0 = const()[name = tensor("op_2281_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2281_cast_fp16 = einsum(equation = var_2281_equation_0, values = (var_2213_cast_fp16_9, var_2259_cast_fp16))[name = tensor("op_2281_cast_fp16")]; + tensor var_2283_equation_0 = const()[name = tensor("op_2283_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2283_cast_fp16 = einsum(equation = var_2283_equation_0, values = (var_2213_cast_fp16_10, var_2260_cast_fp16))[name = tensor("op_2283_cast_fp16")]; + tensor var_2285_equation_0 = const()[name = tensor("op_2285_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2285_cast_fp16 = einsum(equation = var_2285_equation_0, values = (var_2213_cast_fp16_11, var_2261_cast_fp16))[name = tensor("op_2285_cast_fp16")]; + tensor input_105_interleave_0 = const()[name = tensor("input_105_interleave_0"), val = tensor(false)]; + tensor input_105_cast_fp16 = concat(axis = var_2134, interleave = input_105_interleave_0, values = (var_2263_cast_fp16, var_2265_cast_fp16, var_2267_cast_fp16, var_2269_cast_fp16, var_2271_cast_fp16, var_2273_cast_fp16, var_2275_cast_fp16, var_2277_cast_fp16, var_2279_cast_fp16, var_2281_cast_fp16, var_2283_cast_fp16, var_2285_cast_fp16))[name = tensor("input_105_cast_fp16")]; + tensor var_2294_pad_type_0 = const()[name = tensor("op_2294_pad_type_0"), val = tensor("valid")]; + tensor var_2294_strides_0 = const()[name = tensor("op_2294_strides_0"), val = tensor([1, 1])]; + tensor var_2294_pad_0 = const()[name = tensor("op_2294_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2294_dilations_0 = const()[name = tensor("op_2294_dilations_0"), val = tensor([1, 1])]; + tensor var_2294_groups_0 = const()[name = tensor("op_2294_groups_0"), val = tensor(1)]; + tensor blocks_10_attn_out_weight_to_fp16 = const()[name = tensor("blocks_10_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151512256)))]; + tensor blocks_10_attn_out_bias_to_fp16 = const()[name = tensor("blocks_10_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152691968)))]; + tensor var_2294_cast_fp16 = conv(bias = blocks_10_attn_out_bias_to_fp16, dilations = var_2294_dilations_0, groups = var_2294_groups_0, pad = var_2294_pad_0, pad_type = var_2294_pad_type_0, strides = var_2294_strides_0, weight = blocks_10_attn_out_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_2294_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([1])]; + tensor input_107_gamma_0_to_fp16 = const()[name = tensor("input_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152693568)))]; + tensor input_107_beta_0_to_fp16 = const()[name = tensor("input_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152695168)))]; + tensor var_2304_to_fp16 = const()[name = tensor("op_2304_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = input_107_beta_0_to_fp16, epsilon = var_2304_to_fp16, gamma = input_107_gamma_0_to_fp16, x = inputs_43_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152696768)))]; + tensor blocks_10_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157415424)))]; + tensor input_109_cast_fp16 = conv(bias = blocks_10_mlp_0_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = blocks_10_mlp_0_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_2330_pad_type_0 = const()[name = tensor("op_2330_pad_type_0"), val = tensor("valid")]; + tensor var_2330_strides_0 = const()[name = tensor("op_2330_strides_0"), val = tensor([1, 1])]; + tensor var_2330_pad_0 = const()[name = tensor("op_2330_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2330_dilations_0 = const()[name = tensor("op_2330_dilations_0"), val = tensor([1, 1])]; + tensor var_2330_groups_0 = const()[name = tensor("op_2330_groups_0"), val = tensor(1)]; + tensor blocks_10_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157421632)))]; + tensor blocks_10_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162140288)))]; + tensor var_2330_cast_fp16 = conv(bias = blocks_10_mlp_2_bias_to_fp16, dilations = var_2330_dilations_0, groups = var_2330_groups_0, pad = var_2330_pad_0, pad_type = var_2330_pad_type_0, strides = var_2330_strides_0, weight = blocks_10_mlp_2_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("op_2330_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = var_2330_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_2339 = const()[name = tensor("op_2339"), val = tensor(1)]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([1])]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162141888)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162143488)))]; + tensor var_2355_to_fp16 = const()[name = tensor("op_2355_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = input_113_beta_0_to_fp16, epsilon = var_2355_to_fp16, gamma = input_113_gamma_0_to_fp16, x = inputs_45_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_2390_weight_0_to_fp16 = const()[name = tensor("op_2390_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162145088)))]; + tensor var_2390_bias_0_to_fp16 = const()[name = tensor("op_2390_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163324800)))]; + tensor var_2390_cast_fp16 = conv(bias = var_2390_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_2390_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2390_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_key_weight_to_fp16 = const()[name = tensor("blocks_11_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163326400)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_11_attn_key_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_2388_pad_type_0 = const()[name = tensor("op_2388_pad_type_0"), val = tensor("valid")]; + tensor var_2388_strides_0 = const()[name = tensor("op_2388_strides_0"), val = tensor([1, 1])]; + tensor var_2388_pad_0 = const()[name = tensor("op_2388_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2388_dilations_0 = const()[name = tensor("op_2388_dilations_0"), val = tensor([1, 1])]; + tensor var_2388_groups_0 = const()[name = tensor("op_2388_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_value_weight_to_fp16 = const()[name = tensor("blocks_11_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164506112)))]; + tensor blocks_11_attn_value_bias_to_fp16 = const()[name = tensor("blocks_11_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165685824)))]; + tensor var_2388_cast_fp16 = conv(bias = blocks_11_attn_value_bias_to_fp16, dilations = var_2388_dilations_0, groups = var_2388_groups_0, pad = var_2388_pad_0, pad_type = var_2388_pad_type_0, strides = var_2388_strides_0, weight = blocks_11_attn_value_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor tile_33 = const()[name = tensor("tile_33"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2391_axis_0 = const()[name = tensor("op_2391_axis_0"), val = tensor(1)]; + tensor var_2391_cast_fp16_0, tensor var_2391_cast_fp16_1, tensor var_2391_cast_fp16_2, tensor var_2391_cast_fp16_3, tensor var_2391_cast_fp16_4, tensor var_2391_cast_fp16_5, tensor var_2391_cast_fp16_6, tensor var_2391_cast_fp16_7, tensor var_2391_cast_fp16_8, tensor var_2391_cast_fp16_9, tensor var_2391_cast_fp16_10, tensor var_2391_cast_fp16_11 = split(axis = var_2391_axis_0, split_sizes = tile_33, x = var_2390_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2404_perm_0 = const()[name = tensor("op_2404_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_34 = const()[name = tensor("tile_34"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2405_axis_0 = const()[name = tensor("op_2405_axis_0"), val = tensor(3)]; + tensor var_2404_cast_fp16 = transpose(perm = var_2404_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_2405_cast_fp16_0, tensor var_2405_cast_fp16_1, tensor var_2405_cast_fp16_2, tensor var_2405_cast_fp16_3, tensor var_2405_cast_fp16_4, tensor var_2405_cast_fp16_5, tensor var_2405_cast_fp16_6, tensor var_2405_cast_fp16_7, tensor var_2405_cast_fp16_8, tensor var_2405_cast_fp16_9, tensor var_2405_cast_fp16_10, tensor var_2405_cast_fp16_11 = split(axis = var_2405_axis_0, split_sizes = tile_34, x = var_2404_cast_fp16)[name = tensor("op_2405_cast_fp16")]; + tensor tile_35 = const()[name = tensor("tile_35"), val = tensor([64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64])]; + tensor var_2418_axis_0 = const()[name = tensor("op_2418_axis_0"), val = tensor(1)]; + tensor var_2418_cast_fp16_0, tensor var_2418_cast_fp16_1, tensor var_2418_cast_fp16_2, tensor var_2418_cast_fp16_3, tensor var_2418_cast_fp16_4, tensor var_2418_cast_fp16_5, tensor var_2418_cast_fp16_6, tensor var_2418_cast_fp16_7, tensor var_2418_cast_fp16_8, tensor var_2418_cast_fp16_9, tensor var_2418_cast_fp16_10, tensor var_2418_cast_fp16_11 = split(axis = var_2418_axis_0, split_sizes = tile_35, x = var_2388_cast_fp16)[name = tensor("op_2418_cast_fp16")]; + tensor aw_265_equation_0 = const()[name = tensor("aw_265_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_265_cast_fp16 = einsum(equation = aw_265_equation_0, values = (var_2405_cast_fp16_0, var_2391_cast_fp16_0))[name = tensor("aw_265_cast_fp16")]; + tensor aw_267_equation_0 = const()[name = tensor("aw_267_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_267_cast_fp16 = einsum(equation = aw_267_equation_0, values = (var_2405_cast_fp16_1, var_2391_cast_fp16_1))[name = tensor("aw_267_cast_fp16")]; + tensor aw_269_equation_0 = const()[name = tensor("aw_269_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_269_cast_fp16 = einsum(equation = aw_269_equation_0, values = (var_2405_cast_fp16_2, var_2391_cast_fp16_2))[name = tensor("aw_269_cast_fp16")]; + tensor aw_271_equation_0 = const()[name = tensor("aw_271_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_271_cast_fp16 = einsum(equation = aw_271_equation_0, values = (var_2405_cast_fp16_3, var_2391_cast_fp16_3))[name = tensor("aw_271_cast_fp16")]; + tensor aw_273_equation_0 = const()[name = tensor("aw_273_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_273_cast_fp16 = einsum(equation = aw_273_equation_0, values = (var_2405_cast_fp16_4, var_2391_cast_fp16_4))[name = tensor("aw_273_cast_fp16")]; + tensor aw_275_equation_0 = const()[name = tensor("aw_275_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_275_cast_fp16 = einsum(equation = aw_275_equation_0, values = (var_2405_cast_fp16_5, var_2391_cast_fp16_5))[name = tensor("aw_275_cast_fp16")]; + tensor aw_277_equation_0 = const()[name = tensor("aw_277_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_277_cast_fp16 = einsum(equation = aw_277_equation_0, values = (var_2405_cast_fp16_6, var_2391_cast_fp16_6))[name = tensor("aw_277_cast_fp16")]; + tensor aw_279_equation_0 = const()[name = tensor("aw_279_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_279_cast_fp16 = einsum(equation = aw_279_equation_0, values = (var_2405_cast_fp16_7, var_2391_cast_fp16_7))[name = tensor("aw_279_cast_fp16")]; + tensor aw_281_equation_0 = const()[name = tensor("aw_281_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_281_cast_fp16 = einsum(equation = aw_281_equation_0, values = (var_2405_cast_fp16_8, var_2391_cast_fp16_8))[name = tensor("aw_281_cast_fp16")]; + tensor aw_283_equation_0 = const()[name = tensor("aw_283_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_283_cast_fp16 = einsum(equation = aw_283_equation_0, values = (var_2405_cast_fp16_9, var_2391_cast_fp16_9))[name = tensor("aw_283_cast_fp16")]; + tensor aw_285_equation_0 = const()[name = tensor("aw_285_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_285_cast_fp16 = einsum(equation = aw_285_equation_0, values = (var_2405_cast_fp16_10, var_2391_cast_fp16_10))[name = tensor("aw_285_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_2405_cast_fp16_11, var_2391_cast_fp16_11))[name = tensor("aw_cast_fp16")]; + tensor var_2455_cast_fp16 = softmax(axis = var_2339, x = aw_265_cast_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor var_2456_cast_fp16 = softmax(axis = var_2339, x = aw_267_cast_fp16)[name = tensor("op_2456_cast_fp16")]; + tensor var_2457_cast_fp16 = softmax(axis = var_2339, x = aw_269_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor var_2458_cast_fp16 = softmax(axis = var_2339, x = aw_271_cast_fp16)[name = tensor("op_2458_cast_fp16")]; + tensor var_2459_cast_fp16 = softmax(axis = var_2339, x = aw_273_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2460_cast_fp16 = softmax(axis = var_2339, x = aw_275_cast_fp16)[name = tensor("op_2460_cast_fp16")]; + tensor var_2461_cast_fp16 = softmax(axis = var_2339, x = aw_277_cast_fp16)[name = tensor("op_2461_cast_fp16")]; + tensor var_2462_cast_fp16 = softmax(axis = var_2339, x = aw_279_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = softmax(axis = var_2339, x = aw_281_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor var_2464_cast_fp16 = softmax(axis = var_2339, x = aw_283_cast_fp16)[name = tensor("op_2464_cast_fp16")]; + tensor var_2465_cast_fp16 = softmax(axis = var_2339, x = aw_285_cast_fp16)[name = tensor("op_2465_cast_fp16")]; + tensor var_2466_cast_fp16 = softmax(axis = var_2339, x = aw_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2468_equation_0 = const()[name = tensor("op_2468_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2468_cast_fp16 = einsum(equation = var_2468_equation_0, values = (var_2418_cast_fp16_0, var_2455_cast_fp16))[name = tensor("op_2468_cast_fp16")]; + tensor var_2470_equation_0 = const()[name = tensor("op_2470_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2470_cast_fp16 = einsum(equation = var_2470_equation_0, values = (var_2418_cast_fp16_1, var_2456_cast_fp16))[name = tensor("op_2470_cast_fp16")]; + tensor var_2472_equation_0 = const()[name = tensor("op_2472_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2472_cast_fp16 = einsum(equation = var_2472_equation_0, values = (var_2418_cast_fp16_2, var_2457_cast_fp16))[name = tensor("op_2472_cast_fp16")]; + tensor var_2474_equation_0 = const()[name = tensor("op_2474_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2474_cast_fp16 = einsum(equation = var_2474_equation_0, values = (var_2418_cast_fp16_3, var_2458_cast_fp16))[name = tensor("op_2474_cast_fp16")]; + tensor var_2476_equation_0 = const()[name = tensor("op_2476_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2476_cast_fp16 = einsum(equation = var_2476_equation_0, values = (var_2418_cast_fp16_4, var_2459_cast_fp16))[name = tensor("op_2476_cast_fp16")]; + tensor var_2478_equation_0 = const()[name = tensor("op_2478_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2478_cast_fp16 = einsum(equation = var_2478_equation_0, values = (var_2418_cast_fp16_5, var_2460_cast_fp16))[name = tensor("op_2478_cast_fp16")]; + tensor var_2480_equation_0 = const()[name = tensor("op_2480_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2480_cast_fp16 = einsum(equation = var_2480_equation_0, values = (var_2418_cast_fp16_6, var_2461_cast_fp16))[name = tensor("op_2480_cast_fp16")]; + tensor var_2482_equation_0 = const()[name = tensor("op_2482_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2482_cast_fp16 = einsum(equation = var_2482_equation_0, values = (var_2418_cast_fp16_7, var_2462_cast_fp16))[name = tensor("op_2482_cast_fp16")]; + tensor var_2484_equation_0 = const()[name = tensor("op_2484_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2484_cast_fp16 = einsum(equation = var_2484_equation_0, values = (var_2418_cast_fp16_8, var_2463_cast_fp16))[name = tensor("op_2484_cast_fp16")]; + tensor var_2486_equation_0 = const()[name = tensor("op_2486_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2486_cast_fp16 = einsum(equation = var_2486_equation_0, values = (var_2418_cast_fp16_9, var_2464_cast_fp16))[name = tensor("op_2486_cast_fp16")]; + tensor var_2488_equation_0 = const()[name = tensor("op_2488_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2488_cast_fp16 = einsum(equation = var_2488_equation_0, values = (var_2418_cast_fp16_10, var_2465_cast_fp16))[name = tensor("op_2488_cast_fp16")]; + tensor var_2490_equation_0 = const()[name = tensor("op_2490_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_2490_cast_fp16 = einsum(equation = var_2490_equation_0, values = (var_2418_cast_fp16_11, var_2466_cast_fp16))[name = tensor("op_2490_cast_fp16")]; + tensor input_115_interleave_0 = const()[name = tensor("input_115_interleave_0"), val = tensor(false)]; + tensor input_115_cast_fp16 = concat(axis = var_2339, interleave = input_115_interleave_0, values = (var_2468_cast_fp16, var_2470_cast_fp16, var_2472_cast_fp16, var_2474_cast_fp16, var_2476_cast_fp16, var_2478_cast_fp16, var_2480_cast_fp16, var_2482_cast_fp16, var_2484_cast_fp16, var_2486_cast_fp16, var_2488_cast_fp16, var_2490_cast_fp16))[name = tensor("input_115_cast_fp16")]; + tensor var_2499_pad_type_0 = const()[name = tensor("op_2499_pad_type_0"), val = tensor("valid")]; + tensor var_2499_strides_0 = const()[name = tensor("op_2499_strides_0"), val = tensor([1, 1])]; + tensor var_2499_pad_0 = const()[name = tensor("op_2499_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2499_dilations_0 = const()[name = tensor("op_2499_dilations_0"), val = tensor([1, 1])]; + tensor var_2499_groups_0 = const()[name = tensor("op_2499_groups_0"), val = tensor(1)]; + tensor blocks_11_attn_out_weight_to_fp16 = const()[name = tensor("blocks_11_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165687424)))]; + tensor blocks_11_attn_out_bias_to_fp16 = const()[name = tensor("blocks_11_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166867136)))]; + tensor var_2499_cast_fp16 = conv(bias = blocks_11_attn_out_bias_to_fp16, dilations = var_2499_dilations_0, groups = var_2499_groups_0, pad = var_2499_pad_0, pad_type = var_2499_pad_type_0, strides = var_2499_strides_0, weight = blocks_11_attn_out_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = var_2499_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([1])]; + tensor input_117_gamma_0_to_fp16 = const()[name = tensor("input_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166868736)))]; + tensor input_117_beta_0_to_fp16 = const()[name = tensor("input_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166870336)))]; + tensor var_2509_to_fp16 = const()[name = tensor("op_2509_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = input_117_beta_0_to_fp16, epsilon = var_2509_to_fp16, gamma = input_117_gamma_0_to_fp16, x = inputs_47_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("valid")]; + tensor input_119_strides_0 = const()[name = tensor("input_119_strides_0"), val = tensor([1, 1])]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_119_dilations_0 = const()[name = tensor("input_119_dilations_0"), val = tensor([1, 1])]; + tensor input_119_groups_0 = const()[name = tensor("input_119_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166871936)))]; + tensor blocks_11_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171590592)))]; + tensor input_119_cast_fp16 = conv(bias = blocks_11_mlp_0_bias_to_fp16, dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = blocks_11_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_119_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_2535_pad_type_0 = const()[name = tensor("op_2535_pad_type_0"), val = tensor("valid")]; + tensor var_2535_strides_0 = const()[name = tensor("op_2535_strides_0"), val = tensor([1, 1])]; + tensor var_2535_pad_0 = const()[name = tensor("op_2535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2535_dilations_0 = const()[name = tensor("op_2535_dilations_0"), val = tensor([1, 1])]; + tensor var_2535_groups_0 = const()[name = tensor("op_2535_groups_0"), val = tensor(1)]; + tensor blocks_11_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171596800)))]; + tensor blocks_11_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176315456)))]; + tensor var_2535_cast_fp16 = conv(bias = blocks_11_mlp_2_bias_to_fp16, dilations = var_2535_dilations_0, groups = var_2535_groups_0, pad = var_2535_pad_0, pad_type = var_2535_pad_type_0, strides = var_2535_strides_0, weight = blocks_11_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_2535_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_2535_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176317056)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176318656)))]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_2549_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_2560_axes_0 = const()[name = tensor("op_2560_axes_0"), val = tensor([2])]; + tensor var_2560_cast_fp16 = squeeze(axes = var_2560_axes_0, x = x_cast_fp16)[name = tensor("op_2560_cast_fp16")]; + tensor var_2563_perm_0 = const()[name = tensor("op_2563_perm_0"), val = tensor([0, 2, 1])]; + tensor var_2563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2563_cast_fp16 = transpose(perm = var_2563_perm_0, x = var_2560_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_2563_cast_fp16_to_fp32_dtype_0, x = var_2563_cast_fp16)[name = tensor("cast_51")]; + } -> (output); +} \ No newline at end of file diff --git a/small/ggml-small-encoder.mlmodelc/weights/weight.bin b/small/ggml-small-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d8790e3bd8bd730f26d3cc409f1739d02e2c09e5 --- /dev/null +++ b/small/ggml-small-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd99222f556ad6a561a7286c4dc94d14fee3d2b60bdeb4e838bfed437f501b1e +size 176320256 diff --git a/small/ggml-small.bin b/small/ggml-small.bin new file mode 100644 index 0000000000000000000000000000000000000000..bc4c4d96b528654fb067e2b243bdc438cfaf0072 --- /dev/null +++ b/small/ggml-small.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1be3a9b2063867b937e64e2ec7483364a79917e157fa98c5d94b5c1fffea987b +size 487601967 diff --git a/tiny.en/.DS_Store b/tiny.en/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..60a0ba4058482b15937c2750a7ce25181c753fdc Binary files /dev/null and b/tiny.en/.DS_Store differ diff --git a/tiny.en/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin b/tiny.en/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fb65fe66c88d9f34cbfc6d72c42eee5a375a471f --- /dev/null +++ b/tiny.en/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c811ab131cbe7fef6230b32b61cb04cc99fb5990e5fb70ab5d7ec907a4a124b2 +size 243 diff --git a/tiny.en/ggml-tiny.en-encoder.mlmodelc/coremldata.bin b/tiny.en/ggml-tiny.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..5335642d44c0b6e3110bafc93373a63a26ae73d5 --- /dev/null +++ b/tiny.en/ggml-tiny.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93254f2adaadd6535328df3b1200d4922c2f701f84df2c9b37c9a5df9ae4f4b7 +size 320 diff --git a/tiny.en/ggml-tiny.en-encoder.mlmodelc/metadata.json b/tiny.en/ggml-tiny.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..d3ba041cf378d66de0a2f58c407a1f810a237045 --- /dev/null +++ b/tiny.en/ggml-tiny.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 384)", + "shortDescription" : "", + "shape" : "[1, 1500, 384]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 4, + "Gelu" : 6, + "LayerNorm" : 9, + "Transpose" : 5, + "Softmax" : 24, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 9, + "Einsum" : 48, + "ExpandDims" : 1, + "Split" : 12, + "Conv" : 26 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_tiny_en", + "method" : "predict" + } +] \ No newline at end of file diff --git a/tiny.en/ggml-tiny.en-encoder.mlmodelc/model.mil b/tiny.en/ggml-tiny.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..705ff46e2edf3b81a1e15e29babaefdf0ed49ab9 --- /dev/null +++ b/tiny.en/ggml-tiny.en-encoder.mlmodelc/model.mil @@ -0,0 +1,463 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_28_pad_type_0 = const()[name = tensor("op_28_pad_type_0"), val = tensor("custom")]; + tensor var_28_pad_0 = const()[name = tensor("op_28_pad_0"), val = tensor([1, 1])]; + tensor var_28_strides_0 = const()[name = tensor("op_28_strides_0"), val = tensor([1])]; + tensor var_28_dilations_0 = const()[name = tensor("op_28_dilations_0"), val = tensor([1])]; + tensor var_28_groups_0 = const()[name = tensor("op_28_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184448)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_20")]; + tensor var_28_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_28_dilations_0, groups = var_28_groups_0, pad = var_28_pad_0, pad_type = var_28_pad_type_0, strides = var_28_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_28_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_28_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_46_pad_type_0 = const()[name = tensor("op_46_pad_type_0"), val = tensor("custom")]; + tensor var_46_pad_0 = const()[name = tensor("op_46_pad_0"), val = tensor([1, 1])]; + tensor var_46_strides_0 = const()[name = tensor("op_46_strides_0"), val = tensor([2])]; + tensor var_46_dilations_0 = const()[name = tensor("op_46_dilations_0"), val = tensor([1])]; + tensor var_46_groups_0 = const()[name = tensor("op_46_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185280)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070080)))]; + tensor var_46_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_46_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_46_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_51_to_fp16 = const()[name = tensor("op_51_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070912)))]; + tensor var_53_cast_fp16 = add(x = x_3_cast_fp16, y = var_51_to_fp16)[name = tensor("op_53_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_53_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2222976)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2223808)))]; + tensor var_84_to_fp16 = const()[name = tensor("op_84_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_84_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_119_weight_0_to_fp16 = const()[name = tensor("op_119_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2224640)))]; + tensor var_119_bias_0_to_fp16 = const()[name = tensor("op_119_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2519616)))]; + tensor var_119_cast_fp16 = conv(bias = var_119_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_119_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_119_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2520448)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_117_pad_type_0 = const()[name = tensor("op_117_pad_type_0"), val = tensor("valid")]; + tensor var_117_strides_0 = const()[name = tensor("op_117_strides_0"), val = tensor([1, 1])]; + tensor var_117_pad_0 = const()[name = tensor("op_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_117_dilations_0 = const()[name = tensor("op_117_dilations_0"), val = tensor([1, 1])]; + tensor var_117_groups_0 = const()[name = tensor("op_117_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2815424)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3110400)))]; + tensor var_117_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_117_dilations_0, groups = var_117_groups_0, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_117_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_117_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_120_axis_0 = const()[name = tensor("op_120_axis_0"), val = tensor(1)]; + tensor var_120_cast_fp16_0, tensor var_120_cast_fp16_1, tensor var_120_cast_fp16_2, tensor var_120_cast_fp16_3, tensor var_120_cast_fp16_4, tensor var_120_cast_fp16_5 = split(axis = var_120_axis_0, split_sizes = tile_0, x = var_119_cast_fp16)[name = tensor("op_120_cast_fp16")]; + tensor var_127_perm_0 = const()[name = tensor("op_127_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_128_axis_0 = const()[name = tensor("op_128_axis_0"), val = tensor(3)]; + tensor var_127_cast_fp16 = transpose(perm = var_127_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_4")]; + tensor var_128_cast_fp16_0, tensor var_128_cast_fp16_1, tensor var_128_cast_fp16_2, tensor var_128_cast_fp16_3, tensor var_128_cast_fp16_4, tensor var_128_cast_fp16_5 = split(axis = var_128_axis_0, split_sizes = tile_1, x = var_127_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_135_axis_0 = const()[name = tensor("op_135_axis_0"), val = tensor(1)]; + tensor var_135_cast_fp16_0, tensor var_135_cast_fp16_1, tensor var_135_cast_fp16_2, tensor var_135_cast_fp16_3, tensor var_135_cast_fp16_4, tensor var_135_cast_fp16_5 = split(axis = var_135_axis_0, split_sizes = tile_2, x = var_117_cast_fp16)[name = tensor("op_135_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_128_cast_fp16_0, var_120_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_128_cast_fp16_1, var_120_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_128_cast_fp16_2, var_120_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_128_cast_fp16_3, var_120_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_128_cast_fp16_4, var_120_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_128_cast_fp16_5, var_120_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor var_154_cast_fp16 = softmax(axis = var_68, x = aw_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_155_cast_fp16 = softmax(axis = var_68, x = aw_3_cast_fp16)[name = tensor("op_155_cast_fp16")]; + tensor var_156_cast_fp16 = softmax(axis = var_68, x = aw_5_cast_fp16)[name = tensor("op_156_cast_fp16")]; + tensor var_157_cast_fp16 = softmax(axis = var_68, x = aw_7_cast_fp16)[name = tensor("op_157_cast_fp16")]; + tensor var_158_cast_fp16 = softmax(axis = var_68, x = aw_9_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_159_cast_fp16 = softmax(axis = var_68, x = aw_11_cast_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_161_equation_0 = const()[name = tensor("op_161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_161_cast_fp16 = einsum(equation = var_161_equation_0, values = (var_135_cast_fp16_0, var_154_cast_fp16))[name = tensor("op_161_cast_fp16")]; + tensor var_163_equation_0 = const()[name = tensor("op_163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_163_cast_fp16 = einsum(equation = var_163_equation_0, values = (var_135_cast_fp16_1, var_155_cast_fp16))[name = tensor("op_163_cast_fp16")]; + tensor var_165_equation_0 = const()[name = tensor("op_165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_165_cast_fp16 = einsum(equation = var_165_equation_0, values = (var_135_cast_fp16_2, var_156_cast_fp16))[name = tensor("op_165_cast_fp16")]; + tensor var_167_equation_0 = const()[name = tensor("op_167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_167_cast_fp16 = einsum(equation = var_167_equation_0, values = (var_135_cast_fp16_3, var_157_cast_fp16))[name = tensor("op_167_cast_fp16")]; + tensor var_169_equation_0 = const()[name = tensor("op_169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_169_cast_fp16 = einsum(equation = var_169_equation_0, values = (var_135_cast_fp16_4, var_158_cast_fp16))[name = tensor("op_169_cast_fp16")]; + tensor var_171_equation_0 = const()[name = tensor("op_171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_171_cast_fp16 = einsum(equation = var_171_equation_0, values = (var_135_cast_fp16_5, var_159_cast_fp16))[name = tensor("op_171_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_68, interleave = input_5_interleave_0, values = (var_161_cast_fp16, var_163_cast_fp16, var_165_cast_fp16, var_167_cast_fp16, var_169_cast_fp16, var_171_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_180_pad_type_0 = const()[name = tensor("op_180_pad_type_0"), val = tensor("valid")]; + tensor var_180_strides_0 = const()[name = tensor("op_180_strides_0"), val = tensor([1, 1])]; + tensor var_180_pad_0 = const()[name = tensor("op_180_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_180_dilations_0 = const()[name = tensor("op_180_dilations_0"), val = tensor([1, 1])]; + tensor var_180_groups_0 = const()[name = tensor("op_180_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3111232)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3406208)))]; + tensor var_180_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_180_dilations_0, groups = var_180_groups_0, pad = var_180_pad_0, pad_type = var_180_pad_type_0, strides = var_180_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_180_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407040)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407872)))]; + tensor var_190_to_fp16 = const()[name = tensor("op_190_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_190_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3408704)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4588416)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_216_pad_type_0 = const()[name = tensor("op_216_pad_type_0"), val = tensor("valid")]; + tensor var_216_strides_0 = const()[name = tensor("op_216_strides_0"), val = tensor([1, 1])]; + tensor var_216_pad_0 = const()[name = tensor("op_216_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_216_dilations_0 = const()[name = tensor("op_216_dilations_0"), val = tensor([1, 1])]; + tensor var_216_groups_0 = const()[name = tensor("op_216_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4591552)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5771264)))]; + tensor var_216_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_216_dilations_0, groups = var_216_groups_0, pad = var_216_pad_0, pad_type = var_216_pad_type_0, strides = var_216_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_216_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_216_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772096)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772928)))]; + tensor var_241_to_fp16 = const()[name = tensor("op_241_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_241_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_276_weight_0_to_fp16 = const()[name = tensor("op_276_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5773760)))]; + tensor var_276_bias_0_to_fp16 = const()[name = tensor("op_276_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6068736)))]; + tensor var_276_cast_fp16 = conv(bias = var_276_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_276_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_276_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6069568)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_274_pad_type_0 = const()[name = tensor("op_274_pad_type_0"), val = tensor("valid")]; + tensor var_274_strides_0 = const()[name = tensor("op_274_strides_0"), val = tensor([1, 1])]; + tensor var_274_pad_0 = const()[name = tensor("op_274_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_274_dilations_0 = const()[name = tensor("op_274_dilations_0"), val = tensor([1, 1])]; + tensor var_274_groups_0 = const()[name = tensor("op_274_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6364544)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6659520)))]; + tensor var_274_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_274_dilations_0, groups = var_274_groups_0, pad = var_274_pad_0, pad_type = var_274_pad_type_0, strides = var_274_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_274_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_277_axis_0 = const()[name = tensor("op_277_axis_0"), val = tensor(1)]; + tensor var_277_cast_fp16_0, tensor var_277_cast_fp16_1, tensor var_277_cast_fp16_2, tensor var_277_cast_fp16_3, tensor var_277_cast_fp16_4, tensor var_277_cast_fp16_5 = split(axis = var_277_axis_0, split_sizes = tile_3, x = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_285_axis_0 = const()[name = tensor("op_285_axis_0"), val = tensor(3)]; + tensor var_284_cast_fp16 = transpose(perm = var_284_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_3")]; + tensor var_285_cast_fp16_0, tensor var_285_cast_fp16_1, tensor var_285_cast_fp16_2, tensor var_285_cast_fp16_3, tensor var_285_cast_fp16_4, tensor var_285_cast_fp16_5 = split(axis = var_285_axis_0, split_sizes = tile_4, x = var_284_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_292_axis_0 = const()[name = tensor("op_292_axis_0"), val = tensor(1)]; + tensor var_292_cast_fp16_0, tensor var_292_cast_fp16_1, tensor var_292_cast_fp16_2, tensor var_292_cast_fp16_3, tensor var_292_cast_fp16_4, tensor var_292_cast_fp16_5 = split(axis = var_292_axis_0, split_sizes = tile_5, x = var_274_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_285_cast_fp16_0, var_277_cast_fp16_0))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_285_cast_fp16_1, var_277_cast_fp16_1))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_285_cast_fp16_2, var_277_cast_fp16_2))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_285_cast_fp16_3, var_277_cast_fp16_3))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_285_cast_fp16_4, var_277_cast_fp16_4))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_285_cast_fp16_5, var_277_cast_fp16_5))[name = tensor("aw_23_cast_fp16")]; + tensor var_311_cast_fp16 = softmax(axis = var_225, x = aw_13_cast_fp16)[name = tensor("op_311_cast_fp16")]; + tensor var_312_cast_fp16 = softmax(axis = var_225, x = aw_15_cast_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_cast_fp16 = softmax(axis = var_225, x = aw_17_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_cast_fp16 = softmax(axis = var_225, x = aw_19_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor var_315_cast_fp16 = softmax(axis = var_225, x = aw_21_cast_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = softmax(axis = var_225, x = aw_23_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor var_318_equation_0 = const()[name = tensor("op_318_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_318_cast_fp16 = einsum(equation = var_318_equation_0, values = (var_292_cast_fp16_0, var_311_cast_fp16))[name = tensor("op_318_cast_fp16")]; + tensor var_320_equation_0 = const()[name = tensor("op_320_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_320_cast_fp16 = einsum(equation = var_320_equation_0, values = (var_292_cast_fp16_1, var_312_cast_fp16))[name = tensor("op_320_cast_fp16")]; + tensor var_322_equation_0 = const()[name = tensor("op_322_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_322_cast_fp16 = einsum(equation = var_322_equation_0, values = (var_292_cast_fp16_2, var_313_cast_fp16))[name = tensor("op_322_cast_fp16")]; + tensor var_324_equation_0 = const()[name = tensor("op_324_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_324_cast_fp16 = einsum(equation = var_324_equation_0, values = (var_292_cast_fp16_3, var_314_cast_fp16))[name = tensor("op_324_cast_fp16")]; + tensor var_326_equation_0 = const()[name = tensor("op_326_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_326_cast_fp16 = einsum(equation = var_326_equation_0, values = (var_292_cast_fp16_4, var_315_cast_fp16))[name = tensor("op_326_cast_fp16")]; + tensor var_328_equation_0 = const()[name = tensor("op_328_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_328_cast_fp16 = einsum(equation = var_328_equation_0, values = (var_292_cast_fp16_5, var_316_cast_fp16))[name = tensor("op_328_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_225, interleave = input_15_interleave_0, values = (var_318_cast_fp16, var_320_cast_fp16, var_322_cast_fp16, var_324_cast_fp16, var_326_cast_fp16, var_328_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_337_pad_type_0 = const()[name = tensor("op_337_pad_type_0"), val = tensor("valid")]; + tensor var_337_strides_0 = const()[name = tensor("op_337_strides_0"), val = tensor([1, 1])]; + tensor var_337_pad_0 = const()[name = tensor("op_337_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_337_dilations_0 = const()[name = tensor("op_337_dilations_0"), val = tensor([1, 1])]; + tensor var_337_groups_0 = const()[name = tensor("op_337_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6660352)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6955328)))]; + tensor var_337_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_337_dilations_0, groups = var_337_groups_0, pad = var_337_pad_0, pad_type = var_337_pad_type_0, strides = var_337_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_337_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956160)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956992)))]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_347_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6957824)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8137536)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_373_pad_type_0 = const()[name = tensor("op_373_pad_type_0"), val = tensor("valid")]; + tensor var_373_strides_0 = const()[name = tensor("op_373_strides_0"), val = tensor([1, 1])]; + tensor var_373_pad_0 = const()[name = tensor("op_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_373_dilations_0 = const()[name = tensor("op_373_dilations_0"), val = tensor([1, 1])]; + tensor var_373_groups_0 = const()[name = tensor("op_373_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8140672)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9320384)))]; + tensor var_373_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_373_dilations_0, groups = var_373_groups_0, pad = var_373_pad_0, pad_type = var_373_pad_type_0, strides = var_373_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_373_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9321216)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322048)))]; + tensor var_398_to_fp16 = const()[name = tensor("op_398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_398_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_433_weight_0_to_fp16 = const()[name = tensor("op_433_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322880)))]; + tensor var_433_bias_0_to_fp16 = const()[name = tensor("op_433_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9617856)))]; + tensor var_433_cast_fp16 = conv(bias = var_433_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_433_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_433_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9618688)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_431_pad_type_0 = const()[name = tensor("op_431_pad_type_0"), val = tensor("valid")]; + tensor var_431_strides_0 = const()[name = tensor("op_431_strides_0"), val = tensor([1, 1])]; + tensor var_431_pad_0 = const()[name = tensor("op_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_431_dilations_0 = const()[name = tensor("op_431_dilations_0"), val = tensor([1, 1])]; + tensor var_431_groups_0 = const()[name = tensor("op_431_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9913664)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10208640)))]; + tensor var_431_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_431_dilations_0, groups = var_431_groups_0, pad = var_431_pad_0, pad_type = var_431_pad_type_0, strides = var_431_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_431_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_434_axis_0 = const()[name = tensor("op_434_axis_0"), val = tensor(1)]; + tensor var_434_cast_fp16_0, tensor var_434_cast_fp16_1, tensor var_434_cast_fp16_2, tensor var_434_cast_fp16_3, tensor var_434_cast_fp16_4, tensor var_434_cast_fp16_5 = split(axis = var_434_axis_0, split_sizes = tile_6, x = var_433_cast_fp16)[name = tensor("op_434_cast_fp16")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(3)]; + tensor var_441_cast_fp16 = transpose(perm = var_441_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_2")]; + tensor var_442_cast_fp16_0, tensor var_442_cast_fp16_1, tensor var_442_cast_fp16_2, tensor var_442_cast_fp16_3, tensor var_442_cast_fp16_4, tensor var_442_cast_fp16_5 = split(axis = var_442_axis_0, split_sizes = tile_7, x = var_441_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_449_axis_0 = const()[name = tensor("op_449_axis_0"), val = tensor(1)]; + tensor var_449_cast_fp16_0, tensor var_449_cast_fp16_1, tensor var_449_cast_fp16_2, tensor var_449_cast_fp16_3, tensor var_449_cast_fp16_4, tensor var_449_cast_fp16_5 = split(axis = var_449_axis_0, split_sizes = tile_8, x = var_431_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_442_cast_fp16_0, var_434_cast_fp16_0))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_442_cast_fp16_1, var_434_cast_fp16_1))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_442_cast_fp16_2, var_434_cast_fp16_2))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_442_cast_fp16_3, var_434_cast_fp16_3))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_442_cast_fp16_4, var_434_cast_fp16_4))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_442_cast_fp16_5, var_434_cast_fp16_5))[name = tensor("aw_35_cast_fp16")]; + tensor var_468_cast_fp16 = softmax(axis = var_382, x = aw_25_cast_fp16)[name = tensor("op_468_cast_fp16")]; + tensor var_469_cast_fp16 = softmax(axis = var_382, x = aw_27_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = softmax(axis = var_382, x = aw_29_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = softmax(axis = var_382, x = aw_31_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor var_472_cast_fp16 = softmax(axis = var_382, x = aw_33_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor var_473_cast_fp16 = softmax(axis = var_382, x = aw_35_cast_fp16)[name = tensor("op_473_cast_fp16")]; + tensor var_475_equation_0 = const()[name = tensor("op_475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_475_cast_fp16 = einsum(equation = var_475_equation_0, values = (var_449_cast_fp16_0, var_468_cast_fp16))[name = tensor("op_475_cast_fp16")]; + tensor var_477_equation_0 = const()[name = tensor("op_477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_477_cast_fp16 = einsum(equation = var_477_equation_0, values = (var_449_cast_fp16_1, var_469_cast_fp16))[name = tensor("op_477_cast_fp16")]; + tensor var_479_equation_0 = const()[name = tensor("op_479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_479_cast_fp16 = einsum(equation = var_479_equation_0, values = (var_449_cast_fp16_2, var_470_cast_fp16))[name = tensor("op_479_cast_fp16")]; + tensor var_481_equation_0 = const()[name = tensor("op_481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_481_cast_fp16 = einsum(equation = var_481_equation_0, values = (var_449_cast_fp16_3, var_471_cast_fp16))[name = tensor("op_481_cast_fp16")]; + tensor var_483_equation_0 = const()[name = tensor("op_483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_483_cast_fp16 = einsum(equation = var_483_equation_0, values = (var_449_cast_fp16_4, var_472_cast_fp16))[name = tensor("op_483_cast_fp16")]; + tensor var_485_equation_0 = const()[name = tensor("op_485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_485_cast_fp16 = einsum(equation = var_485_equation_0, values = (var_449_cast_fp16_5, var_473_cast_fp16))[name = tensor("op_485_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_382, interleave = input_25_interleave_0, values = (var_475_cast_fp16, var_477_cast_fp16, var_479_cast_fp16, var_481_cast_fp16, var_483_cast_fp16, var_485_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_494_pad_type_0 = const()[name = tensor("op_494_pad_type_0"), val = tensor("valid")]; + tensor var_494_strides_0 = const()[name = tensor("op_494_strides_0"), val = tensor([1, 1])]; + tensor var_494_pad_0 = const()[name = tensor("op_494_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_494_dilations_0 = const()[name = tensor("op_494_dilations_0"), val = tensor([1, 1])]; + tensor var_494_groups_0 = const()[name = tensor("op_494_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10209472)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10504448)))]; + tensor var_494_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_494_dilations_0, groups = var_494_groups_0, pad = var_494_pad_0, pad_type = var_494_pad_type_0, strides = var_494_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_494_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10505280)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506112)))]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_504_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506944)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11686656)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_530_pad_type_0 = const()[name = tensor("op_530_pad_type_0"), val = tensor("valid")]; + tensor var_530_strides_0 = const()[name = tensor("op_530_strides_0"), val = tensor([1, 1])]; + tensor var_530_pad_0 = const()[name = tensor("op_530_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_530_dilations_0 = const()[name = tensor("op_530_dilations_0"), val = tensor([1, 1])]; + tensor var_530_groups_0 = const()[name = tensor("op_530_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11689792)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12869504)))]; + tensor var_530_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_530_dilations_0, groups = var_530_groups_0, pad = var_530_pad_0, pad_type = var_530_pad_type_0, strides = var_530_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_530_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12870336)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12871168)))]; + tensor var_555_to_fp16 = const()[name = tensor("op_555_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_555_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_590_weight_0_to_fp16 = const()[name = tensor("op_590_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12872000)))]; + tensor var_590_bias_0_to_fp16 = const()[name = tensor("op_590_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13166976)))]; + tensor var_590_cast_fp16 = conv(bias = var_590_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_590_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13167808)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_588_pad_type_0 = const()[name = tensor("op_588_pad_type_0"), val = tensor("valid")]; + tensor var_588_strides_0 = const()[name = tensor("op_588_strides_0"), val = tensor([1, 1])]; + tensor var_588_pad_0 = const()[name = tensor("op_588_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_588_dilations_0 = const()[name = tensor("op_588_dilations_0"), val = tensor([1, 1])]; + tensor var_588_groups_0 = const()[name = tensor("op_588_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13462784)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13757760)))]; + tensor var_588_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_588_dilations_0, groups = var_588_groups_0, pad = var_588_pad_0, pad_type = var_588_pad_type_0, strides = var_588_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_588_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_591_axis_0 = const()[name = tensor("op_591_axis_0"), val = tensor(1)]; + tensor var_591_cast_fp16_0, tensor var_591_cast_fp16_1, tensor var_591_cast_fp16_2, tensor var_591_cast_fp16_3, tensor var_591_cast_fp16_4, tensor var_591_cast_fp16_5 = split(axis = var_591_axis_0, split_sizes = tile_9, x = var_590_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_598_perm_0 = const()[name = tensor("op_598_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_599_axis_0 = const()[name = tensor("op_599_axis_0"), val = tensor(3)]; + tensor var_598_cast_fp16 = transpose(perm = var_598_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_599_cast_fp16_0, tensor var_599_cast_fp16_1, tensor var_599_cast_fp16_2, tensor var_599_cast_fp16_3, tensor var_599_cast_fp16_4, tensor var_599_cast_fp16_5 = split(axis = var_599_axis_0, split_sizes = tile_10, x = var_598_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_606_axis_0 = const()[name = tensor("op_606_axis_0"), val = tensor(1)]; + tensor var_606_cast_fp16_0, tensor var_606_cast_fp16_1, tensor var_606_cast_fp16_2, tensor var_606_cast_fp16_3, tensor var_606_cast_fp16_4, tensor var_606_cast_fp16_5 = split(axis = var_606_axis_0, split_sizes = tile_11, x = var_588_cast_fp16)[name = tensor("op_606_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_599_cast_fp16_0, var_591_cast_fp16_0))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_599_cast_fp16_1, var_591_cast_fp16_1))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_599_cast_fp16_2, var_591_cast_fp16_2))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_599_cast_fp16_3, var_591_cast_fp16_3))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_599_cast_fp16_4, var_591_cast_fp16_4))[name = tensor("aw_45_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_599_cast_fp16_5, var_591_cast_fp16_5))[name = tensor("aw_cast_fp16")]; + tensor var_625_cast_fp16 = softmax(axis = var_539, x = aw_37_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor var_626_cast_fp16 = softmax(axis = var_539, x = aw_39_cast_fp16)[name = tensor("op_626_cast_fp16")]; + tensor var_627_cast_fp16 = softmax(axis = var_539, x = aw_41_cast_fp16)[name = tensor("op_627_cast_fp16")]; + tensor var_628_cast_fp16 = softmax(axis = var_539, x = aw_43_cast_fp16)[name = tensor("op_628_cast_fp16")]; + tensor var_629_cast_fp16 = softmax(axis = var_539, x = aw_45_cast_fp16)[name = tensor("op_629_cast_fp16")]; + tensor var_630_cast_fp16 = softmax(axis = var_539, x = aw_cast_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_632_equation_0 = const()[name = tensor("op_632_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_632_cast_fp16 = einsum(equation = var_632_equation_0, values = (var_606_cast_fp16_0, var_625_cast_fp16))[name = tensor("op_632_cast_fp16")]; + tensor var_634_equation_0 = const()[name = tensor("op_634_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_634_cast_fp16 = einsum(equation = var_634_equation_0, values = (var_606_cast_fp16_1, var_626_cast_fp16))[name = tensor("op_634_cast_fp16")]; + tensor var_636_equation_0 = const()[name = tensor("op_636_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_636_cast_fp16 = einsum(equation = var_636_equation_0, values = (var_606_cast_fp16_2, var_627_cast_fp16))[name = tensor("op_636_cast_fp16")]; + tensor var_638_equation_0 = const()[name = tensor("op_638_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_638_cast_fp16 = einsum(equation = var_638_equation_0, values = (var_606_cast_fp16_3, var_628_cast_fp16))[name = tensor("op_638_cast_fp16")]; + tensor var_640_equation_0 = const()[name = tensor("op_640_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_640_cast_fp16 = einsum(equation = var_640_equation_0, values = (var_606_cast_fp16_4, var_629_cast_fp16))[name = tensor("op_640_cast_fp16")]; + tensor var_642_equation_0 = const()[name = tensor("op_642_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_642_cast_fp16 = einsum(equation = var_642_equation_0, values = (var_606_cast_fp16_5, var_630_cast_fp16))[name = tensor("op_642_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_539, interleave = input_35_interleave_0, values = (var_632_cast_fp16, var_634_cast_fp16, var_636_cast_fp16, var_638_cast_fp16, var_640_cast_fp16, var_642_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_651_pad_type_0 = const()[name = tensor("op_651_pad_type_0"), val = tensor("valid")]; + tensor var_651_strides_0 = const()[name = tensor("op_651_strides_0"), val = tensor([1, 1])]; + tensor var_651_pad_0 = const()[name = tensor("op_651_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_651_dilations_0 = const()[name = tensor("op_651_dilations_0"), val = tensor([1, 1])]; + tensor var_651_groups_0 = const()[name = tensor("op_651_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13758592)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14053568)))]; + tensor var_651_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_651_dilations_0, groups = var_651_groups_0, pad = var_651_pad_0, pad_type = var_651_pad_type_0, strides = var_651_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_651_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_651_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14054400)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14055232)))]; + tensor var_661_to_fp16 = const()[name = tensor("op_661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_661_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14056064)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15235776)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_39_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_687_pad_type_0 = const()[name = tensor("op_687_pad_type_0"), val = tensor("valid")]; + tensor var_687_strides_0 = const()[name = tensor("op_687_strides_0"), val = tensor([1, 1])]; + tensor var_687_pad_0 = const()[name = tensor("op_687_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_687_dilations_0 = const()[name = tensor("op_687_dilations_0"), val = tensor([1, 1])]; + tensor var_687_groups_0 = const()[name = tensor("op_687_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15238912)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16418624)))]; + tensor var_687_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_687_dilations_0, groups = var_687_groups_0, pad = var_687_pad_0, pad_type = var_687_pad_type_0, strides = var_687_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_687_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16419456)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16420288)))]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_701_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_712_axes_0 = const()[name = tensor("op_712_axes_0"), val = tensor([2])]; + tensor var_712_cast_fp16 = squeeze(axes = var_712_axes_0, x = x_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_715_perm_0 = const()[name = tensor("op_715_perm_0"), val = tensor([0, 2, 1])]; + tensor var_715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_715_cast_fp16 = transpose(perm = var_715_perm_0, x = var_712_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_715_cast_fp16_to_fp32_dtype_0, x = var_715_cast_fp16)[name = tensor("cast_19")]; + } -> (output); +} \ No newline at end of file diff --git a/tiny.en/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin b/tiny.en/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d89d6272f407b442a55ccec48b193d3414930b6b --- /dev/null +++ b/tiny.en/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:040cc1dc03624b30f9f01e567d71b651729da26d98de36c72ab3266c85f68fab +size 16421120 diff --git a/tiny.en/ggml-tiny.en.bin b/tiny.en/ggml-tiny.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..17ad750438d1d42162fe06ab4b21aef2389d2137 --- /dev/null +++ b/tiny.en/ggml-tiny.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:921e4cf8686fdd993dcd081a5da5b6c365bfde1162e72b08d75ac75289920b1f +size 77704715 diff --git a/tiny/.DS_Store b/tiny/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..776b14f7da0c1e89472bb0067a6c0955b0fae242 Binary files /dev/null and b/tiny/.DS_Store differ diff --git a/tiny/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin b/tiny/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fb65fe66c88d9f34cbfc6d72c42eee5a375a471f --- /dev/null +++ b/tiny/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c811ab131cbe7fef6230b32b61cb04cc99fb5990e5fb70ab5d7ec907a4a124b2 +size 243 diff --git a/tiny/ggml-tiny-encoder.mlmodelc/coremldata.bin b/tiny/ggml-tiny-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..94d88431c84a536676db3d979d8407f9b298e689 --- /dev/null +++ b/tiny/ggml-tiny-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9df7e387da981d548738ae3570db8a375e2e55d16ae8e1374da84809fdd0c5a +size 320 diff --git a/tiny/ggml-tiny-encoder.mlmodelc/metadata.json b/tiny/ggml-tiny-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..c7f4d8063a97e36f6929a5717606155f752b1fcc --- /dev/null +++ b/tiny/ggml-tiny-encoder.mlmodelc/metadata.json @@ -0,0 +1,71 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1500 × 384)", + "shortDescription" : "", + "shape" : "[1, 1500, 384]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Concat" : 4, + "Gelu" : 6, + "LayerNorm" : 9, + "Transpose" : 5, + "Softmax" : 24, + "Squeeze" : 1, + "Cast" : 2, + "Add" : 9, + "Einsum" : 48, + "ExpandDims" : 1, + "Split" : 12, + "Conv" : 26 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "visionOS" : "1.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.source_dialect" : "TorchScript", + "com.github.apple.coremltools.version" : "8.3.0", + "com.github.apple.coremltools.source" : "torch==2.2.2" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 80 × 3000)", + "shortDescription" : "", + "shape" : "[1, 80, 3000]", + "name" : "logmel_data", + "type" : "MultiArray" + } + ], + "generatedClassName" : "coreml_encoder_tiny", + "method" : "predict" + } +] \ No newline at end of file diff --git a/tiny/ggml-tiny-encoder.mlmodelc/model.mil b/tiny/ggml-tiny-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..705ff46e2edf3b81a1e15e29babaefdf0ed49ab9 --- /dev/null +++ b/tiny/ggml-tiny-encoder.mlmodelc/model.mil @@ -0,0 +1,463 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor logmel_data) { + tensor var_28_pad_type_0 = const()[name = tensor("op_28_pad_type_0"), val = tensor("custom")]; + tensor var_28_pad_0 = const()[name = tensor("op_28_pad_0"), val = tensor([1, 1])]; + tensor var_28_strides_0 = const()[name = tensor("op_28_strides_0"), val = tensor([1])]; + tensor var_28_dilations_0 = const()[name = tensor("op_28_dilations_0"), val = tensor([1])]; + tensor var_28_groups_0 = const()[name = tensor("op_28_groups_0"), val = tensor(1)]; + tensor logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184448)))]; + tensor logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_20")]; + tensor var_28_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_28_dilations_0, groups = var_28_groups_0, pad = var_28_pad_0, pad_type = var_28_pad_type_0, strides = var_28_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor("op_28_cast_fp16")]; + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_28_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_46_pad_type_0 = const()[name = tensor("op_46_pad_type_0"), val = tensor("custom")]; + tensor var_46_pad_0 = const()[name = tensor("op_46_pad_0"), val = tensor([1, 1])]; + tensor var_46_strides_0 = const()[name = tensor("op_46_strides_0"), val = tensor([2])]; + tensor var_46_dilations_0 = const()[name = tensor("op_46_dilations_0"), val = tensor([1])]; + tensor var_46_groups_0 = const()[name = tensor("op_46_groups_0"), val = tensor(1)]; + tensor const_2_to_fp16 = const()[name = tensor("const_2_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185280)))]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070080)))]; + tensor var_46_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor("op_46_cast_fp16")]; + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_46_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_51_to_fp16 = const()[name = tensor("op_51_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070912)))]; + tensor var_53_cast_fp16 = add(x = x_3_cast_fp16, y = var_51_to_fp16)[name = tensor("op_53_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([2])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_53_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(1)]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([1])]; + tensor input_3_gamma_0_to_fp16 = const()[name = tensor("input_3_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2222976)))]; + tensor input_3_beta_0_to_fp16 = const()[name = tensor("input_3_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2223808)))]; + tensor var_84_to_fp16 = const()[name = tensor("op_84_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_84_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor q_1_pad_type_0 = const()[name = tensor("q_1_pad_type_0"), val = tensor("valid")]; + tensor q_1_strides_0 = const()[name = tensor("q_1_strides_0"), val = tensor([1, 1])]; + tensor q_1_pad_0 = const()[name = tensor("q_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_dilations_0 = const()[name = tensor("q_1_dilations_0"), val = tensor([1, 1])]; + tensor q_1_groups_0 = const()[name = tensor("q_1_groups_0"), val = tensor(1)]; + tensor var_119_weight_0_to_fp16 = const()[name = tensor("op_119_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2224640)))]; + tensor var_119_bias_0_to_fp16 = const()[name = tensor("op_119_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2519616)))]; + tensor var_119_cast_fp16 = conv(bias = var_119_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_119_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("op_119_cast_fp16")]; + tensor k_1_pad_type_0 = const()[name = tensor("k_1_pad_type_0"), val = tensor("valid")]; + tensor k_1_strides_0 = const()[name = tensor("k_1_strides_0"), val = tensor([1, 1])]; + tensor k_1_pad_0 = const()[name = tensor("k_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_1_dilations_0 = const()[name = tensor("k_1_dilations_0"), val = tensor([1, 1])]; + tensor k_1_groups_0 = const()[name = tensor("k_1_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_key_weight_to_fp16 = const()[name = tensor("blocks_0_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2520448)))]; + tensor k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor var_117_pad_type_0 = const()[name = tensor("op_117_pad_type_0"), val = tensor("valid")]; + tensor var_117_strides_0 = const()[name = tensor("op_117_strides_0"), val = tensor([1, 1])]; + tensor var_117_pad_0 = const()[name = tensor("op_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_117_dilations_0 = const()[name = tensor("op_117_dilations_0"), val = tensor([1, 1])]; + tensor var_117_groups_0 = const()[name = tensor("op_117_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_value_weight_to_fp16 = const()[name = tensor("blocks_0_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2815424)))]; + tensor blocks_0_attn_value_bias_to_fp16 = const()[name = tensor("blocks_0_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3110400)))]; + tensor var_117_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_117_dilations_0, groups = var_117_groups_0, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_117_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("op_117_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_120_axis_0 = const()[name = tensor("op_120_axis_0"), val = tensor(1)]; + tensor var_120_cast_fp16_0, tensor var_120_cast_fp16_1, tensor var_120_cast_fp16_2, tensor var_120_cast_fp16_3, tensor var_120_cast_fp16_4, tensor var_120_cast_fp16_5 = split(axis = var_120_axis_0, split_sizes = tile_0, x = var_119_cast_fp16)[name = tensor("op_120_cast_fp16")]; + tensor var_127_perm_0 = const()[name = tensor("op_127_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_128_axis_0 = const()[name = tensor("op_128_axis_0"), val = tensor(3)]; + tensor var_127_cast_fp16 = transpose(perm = var_127_perm_0, x = k_1_cast_fp16)[name = tensor("transpose_4")]; + tensor var_128_cast_fp16_0, tensor var_128_cast_fp16_1, tensor var_128_cast_fp16_2, tensor var_128_cast_fp16_3, tensor var_128_cast_fp16_4, tensor var_128_cast_fp16_5 = split(axis = var_128_axis_0, split_sizes = tile_1, x = var_127_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor tile_2 = const()[name = tensor("tile_2"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_135_axis_0 = const()[name = tensor("op_135_axis_0"), val = tensor(1)]; + tensor var_135_cast_fp16_0, tensor var_135_cast_fp16_1, tensor var_135_cast_fp16_2, tensor var_135_cast_fp16_3, tensor var_135_cast_fp16_4, tensor var_135_cast_fp16_5 = split(axis = var_135_axis_0, split_sizes = tile_2, x = var_117_cast_fp16)[name = tensor("op_135_cast_fp16")]; + tensor aw_1_equation_0 = const()[name = tensor("aw_1_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_128_cast_fp16_0, var_120_cast_fp16_0))[name = tensor("aw_1_cast_fp16")]; + tensor aw_3_equation_0 = const()[name = tensor("aw_3_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_128_cast_fp16_1, var_120_cast_fp16_1))[name = tensor("aw_3_cast_fp16")]; + tensor aw_5_equation_0 = const()[name = tensor("aw_5_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_128_cast_fp16_2, var_120_cast_fp16_2))[name = tensor("aw_5_cast_fp16")]; + tensor aw_7_equation_0 = const()[name = tensor("aw_7_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_128_cast_fp16_3, var_120_cast_fp16_3))[name = tensor("aw_7_cast_fp16")]; + tensor aw_9_equation_0 = const()[name = tensor("aw_9_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_128_cast_fp16_4, var_120_cast_fp16_4))[name = tensor("aw_9_cast_fp16")]; + tensor aw_11_equation_0 = const()[name = tensor("aw_11_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_128_cast_fp16_5, var_120_cast_fp16_5))[name = tensor("aw_11_cast_fp16")]; + tensor var_154_cast_fp16 = softmax(axis = var_68, x = aw_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_155_cast_fp16 = softmax(axis = var_68, x = aw_3_cast_fp16)[name = tensor("op_155_cast_fp16")]; + tensor var_156_cast_fp16 = softmax(axis = var_68, x = aw_5_cast_fp16)[name = tensor("op_156_cast_fp16")]; + tensor var_157_cast_fp16 = softmax(axis = var_68, x = aw_7_cast_fp16)[name = tensor("op_157_cast_fp16")]; + tensor var_158_cast_fp16 = softmax(axis = var_68, x = aw_9_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_159_cast_fp16 = softmax(axis = var_68, x = aw_11_cast_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_161_equation_0 = const()[name = tensor("op_161_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_161_cast_fp16 = einsum(equation = var_161_equation_0, values = (var_135_cast_fp16_0, var_154_cast_fp16))[name = tensor("op_161_cast_fp16")]; + tensor var_163_equation_0 = const()[name = tensor("op_163_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_163_cast_fp16 = einsum(equation = var_163_equation_0, values = (var_135_cast_fp16_1, var_155_cast_fp16))[name = tensor("op_163_cast_fp16")]; + tensor var_165_equation_0 = const()[name = tensor("op_165_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_165_cast_fp16 = einsum(equation = var_165_equation_0, values = (var_135_cast_fp16_2, var_156_cast_fp16))[name = tensor("op_165_cast_fp16")]; + tensor var_167_equation_0 = const()[name = tensor("op_167_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_167_cast_fp16 = einsum(equation = var_167_equation_0, values = (var_135_cast_fp16_3, var_157_cast_fp16))[name = tensor("op_167_cast_fp16")]; + tensor var_169_equation_0 = const()[name = tensor("op_169_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_169_cast_fp16 = einsum(equation = var_169_equation_0, values = (var_135_cast_fp16_4, var_158_cast_fp16))[name = tensor("op_169_cast_fp16")]; + tensor var_171_equation_0 = const()[name = tensor("op_171_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_171_cast_fp16 = einsum(equation = var_171_equation_0, values = (var_135_cast_fp16_5, var_159_cast_fp16))[name = tensor("op_171_cast_fp16")]; + tensor input_5_interleave_0 = const()[name = tensor("input_5_interleave_0"), val = tensor(false)]; + tensor input_5_cast_fp16 = concat(axis = var_68, interleave = input_5_interleave_0, values = (var_161_cast_fp16, var_163_cast_fp16, var_165_cast_fp16, var_167_cast_fp16, var_169_cast_fp16, var_171_cast_fp16))[name = tensor("input_5_cast_fp16")]; + tensor var_180_pad_type_0 = const()[name = tensor("op_180_pad_type_0"), val = tensor("valid")]; + tensor var_180_strides_0 = const()[name = tensor("op_180_strides_0"), val = tensor([1, 1])]; + tensor var_180_pad_0 = const()[name = tensor("op_180_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_180_dilations_0 = const()[name = tensor("op_180_dilations_0"), val = tensor([1, 1])]; + tensor var_180_groups_0 = const()[name = tensor("op_180_groups_0"), val = tensor(1)]; + tensor blocks_0_attn_out_weight_to_fp16 = const()[name = tensor("blocks_0_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3111232)))]; + tensor blocks_0_attn_out_bias_to_fp16 = const()[name = tensor("blocks_0_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3406208)))]; + tensor var_180_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_180_dilations_0, groups = var_180_groups_0, pad = var_180_pad_0, pad_type = var_180_pad_type_0, strides = var_180_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_180_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([1])]; + tensor input_7_gamma_0_to_fp16 = const()[name = tensor("input_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407040)))]; + tensor input_7_beta_0_to_fp16 = const()[name = tensor("input_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407872)))]; + tensor var_190_to_fp16 = const()[name = tensor("op_190_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_190_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; + tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; + tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3408704)))]; + tensor blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4588416)))]; + tensor input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_216_pad_type_0 = const()[name = tensor("op_216_pad_type_0"), val = tensor("valid")]; + tensor var_216_strides_0 = const()[name = tensor("op_216_strides_0"), val = tensor([1, 1])]; + tensor var_216_pad_0 = const()[name = tensor("op_216_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_216_dilations_0 = const()[name = tensor("op_216_dilations_0"), val = tensor([1, 1])]; + tensor var_216_groups_0 = const()[name = tensor("op_216_groups_0"), val = tensor(1)]; + tensor blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4591552)))]; + tensor blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5771264)))]; + tensor var_216_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_216_dilations_0, groups = var_216_groups_0, pad = var_216_pad_0, pad_type = var_216_pad_type_0, strides = var_216_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("op_216_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_216_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(1)]; + tensor input_13_axes_0 = const()[name = tensor("input_13_axes_0"), val = tensor([1])]; + tensor input_13_gamma_0_to_fp16 = const()[name = tensor("input_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772096)))]; + tensor input_13_beta_0_to_fp16 = const()[name = tensor("input_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772928)))]; + tensor var_241_to_fp16 = const()[name = tensor("op_241_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_241_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor q_3_pad_type_0 = const()[name = tensor("q_3_pad_type_0"), val = tensor("valid")]; + tensor q_3_strides_0 = const()[name = tensor("q_3_strides_0"), val = tensor([1, 1])]; + tensor q_3_pad_0 = const()[name = tensor("q_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_3_dilations_0 = const()[name = tensor("q_3_dilations_0"), val = tensor([1, 1])]; + tensor q_3_groups_0 = const()[name = tensor("q_3_groups_0"), val = tensor(1)]; + tensor var_276_weight_0_to_fp16 = const()[name = tensor("op_276_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5773760)))]; + tensor var_276_bias_0_to_fp16 = const()[name = tensor("op_276_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6068736)))]; + tensor var_276_cast_fp16 = conv(bias = var_276_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_276_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("op_276_cast_fp16")]; + tensor k_3_pad_type_0 = const()[name = tensor("k_3_pad_type_0"), val = tensor("valid")]; + tensor k_3_strides_0 = const()[name = tensor("k_3_strides_0"), val = tensor([1, 1])]; + tensor k_3_pad_0 = const()[name = tensor("k_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_3_dilations_0 = const()[name = tensor("k_3_dilations_0"), val = tensor([1, 1])]; + tensor k_3_groups_0 = const()[name = tensor("k_3_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_key_weight_to_fp16 = const()[name = tensor("blocks_1_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6069568)))]; + tensor k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_274_pad_type_0 = const()[name = tensor("op_274_pad_type_0"), val = tensor("valid")]; + tensor var_274_strides_0 = const()[name = tensor("op_274_strides_0"), val = tensor([1, 1])]; + tensor var_274_pad_0 = const()[name = tensor("op_274_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_274_dilations_0 = const()[name = tensor("op_274_dilations_0"), val = tensor([1, 1])]; + tensor var_274_groups_0 = const()[name = tensor("op_274_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_value_weight_to_fp16 = const()[name = tensor("blocks_1_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6364544)))]; + tensor blocks_1_attn_value_bias_to_fp16 = const()[name = tensor("blocks_1_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6659520)))]; + tensor var_274_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_274_dilations_0, groups = var_274_groups_0, pad = var_274_pad_0, pad_type = var_274_pad_type_0, strides = var_274_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("op_274_cast_fp16")]; + tensor tile_3 = const()[name = tensor("tile_3"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_277_axis_0 = const()[name = tensor("op_277_axis_0"), val = tensor(1)]; + tensor var_277_cast_fp16_0, tensor var_277_cast_fp16_1, tensor var_277_cast_fp16_2, tensor var_277_cast_fp16_3, tensor var_277_cast_fp16_4, tensor var_277_cast_fp16_5 = split(axis = var_277_axis_0, split_sizes = tile_3, x = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_4 = const()[name = tensor("tile_4"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_285_axis_0 = const()[name = tensor("op_285_axis_0"), val = tensor(3)]; + tensor var_284_cast_fp16 = transpose(perm = var_284_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_3")]; + tensor var_285_cast_fp16_0, tensor var_285_cast_fp16_1, tensor var_285_cast_fp16_2, tensor var_285_cast_fp16_3, tensor var_285_cast_fp16_4, tensor var_285_cast_fp16_5 = split(axis = var_285_axis_0, split_sizes = tile_4, x = var_284_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor tile_5 = const()[name = tensor("tile_5"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_292_axis_0 = const()[name = tensor("op_292_axis_0"), val = tensor(1)]; + tensor var_292_cast_fp16_0, tensor var_292_cast_fp16_1, tensor var_292_cast_fp16_2, tensor var_292_cast_fp16_3, tensor var_292_cast_fp16_4, tensor var_292_cast_fp16_5 = split(axis = var_292_axis_0, split_sizes = tile_5, x = var_274_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor aw_13_equation_0 = const()[name = tensor("aw_13_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_285_cast_fp16_0, var_277_cast_fp16_0))[name = tensor("aw_13_cast_fp16")]; + tensor aw_15_equation_0 = const()[name = tensor("aw_15_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_285_cast_fp16_1, var_277_cast_fp16_1))[name = tensor("aw_15_cast_fp16")]; + tensor aw_17_equation_0 = const()[name = tensor("aw_17_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_285_cast_fp16_2, var_277_cast_fp16_2))[name = tensor("aw_17_cast_fp16")]; + tensor aw_19_equation_0 = const()[name = tensor("aw_19_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_285_cast_fp16_3, var_277_cast_fp16_3))[name = tensor("aw_19_cast_fp16")]; + tensor aw_21_equation_0 = const()[name = tensor("aw_21_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_285_cast_fp16_4, var_277_cast_fp16_4))[name = tensor("aw_21_cast_fp16")]; + tensor aw_23_equation_0 = const()[name = tensor("aw_23_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_285_cast_fp16_5, var_277_cast_fp16_5))[name = tensor("aw_23_cast_fp16")]; + tensor var_311_cast_fp16 = softmax(axis = var_225, x = aw_13_cast_fp16)[name = tensor("op_311_cast_fp16")]; + tensor var_312_cast_fp16 = softmax(axis = var_225, x = aw_15_cast_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_cast_fp16 = softmax(axis = var_225, x = aw_17_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_cast_fp16 = softmax(axis = var_225, x = aw_19_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor var_315_cast_fp16 = softmax(axis = var_225, x = aw_21_cast_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = softmax(axis = var_225, x = aw_23_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor var_318_equation_0 = const()[name = tensor("op_318_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_318_cast_fp16 = einsum(equation = var_318_equation_0, values = (var_292_cast_fp16_0, var_311_cast_fp16))[name = tensor("op_318_cast_fp16")]; + tensor var_320_equation_0 = const()[name = tensor("op_320_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_320_cast_fp16 = einsum(equation = var_320_equation_0, values = (var_292_cast_fp16_1, var_312_cast_fp16))[name = tensor("op_320_cast_fp16")]; + tensor var_322_equation_0 = const()[name = tensor("op_322_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_322_cast_fp16 = einsum(equation = var_322_equation_0, values = (var_292_cast_fp16_2, var_313_cast_fp16))[name = tensor("op_322_cast_fp16")]; + tensor var_324_equation_0 = const()[name = tensor("op_324_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_324_cast_fp16 = einsum(equation = var_324_equation_0, values = (var_292_cast_fp16_3, var_314_cast_fp16))[name = tensor("op_324_cast_fp16")]; + tensor var_326_equation_0 = const()[name = tensor("op_326_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_326_cast_fp16 = einsum(equation = var_326_equation_0, values = (var_292_cast_fp16_4, var_315_cast_fp16))[name = tensor("op_326_cast_fp16")]; + tensor var_328_equation_0 = const()[name = tensor("op_328_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_328_cast_fp16 = einsum(equation = var_328_equation_0, values = (var_292_cast_fp16_5, var_316_cast_fp16))[name = tensor("op_328_cast_fp16")]; + tensor input_15_interleave_0 = const()[name = tensor("input_15_interleave_0"), val = tensor(false)]; + tensor input_15_cast_fp16 = concat(axis = var_225, interleave = input_15_interleave_0, values = (var_318_cast_fp16, var_320_cast_fp16, var_322_cast_fp16, var_324_cast_fp16, var_326_cast_fp16, var_328_cast_fp16))[name = tensor("input_15_cast_fp16")]; + tensor var_337_pad_type_0 = const()[name = tensor("op_337_pad_type_0"), val = tensor("valid")]; + tensor var_337_strides_0 = const()[name = tensor("op_337_strides_0"), val = tensor([1, 1])]; + tensor var_337_pad_0 = const()[name = tensor("op_337_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_337_dilations_0 = const()[name = tensor("op_337_dilations_0"), val = tensor([1, 1])]; + tensor var_337_groups_0 = const()[name = tensor("op_337_groups_0"), val = tensor(1)]; + tensor blocks_1_attn_out_weight_to_fp16 = const()[name = tensor("blocks_1_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6660352)))]; + tensor blocks_1_attn_out_bias_to_fp16 = const()[name = tensor("blocks_1_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6955328)))]; + tensor var_337_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_337_dilations_0, groups = var_337_groups_0, pad = var_337_pad_0, pad_type = var_337_pad_type_0, strides = var_337_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_337_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([1])]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956160)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956992)))]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_347_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6957824)))]; + tensor blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8137536)))]; + tensor input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_373_pad_type_0 = const()[name = tensor("op_373_pad_type_0"), val = tensor("valid")]; + tensor var_373_strides_0 = const()[name = tensor("op_373_strides_0"), val = tensor([1, 1])]; + tensor var_373_pad_0 = const()[name = tensor("op_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_373_dilations_0 = const()[name = tensor("op_373_dilations_0"), val = tensor([1, 1])]; + tensor var_373_groups_0 = const()[name = tensor("op_373_groups_0"), val = tensor(1)]; + tensor blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8140672)))]; + tensor blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9320384)))]; + tensor var_373_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_373_dilations_0, groups = var_373_groups_0, pad = var_373_pad_0, pad_type = var_373_pad_type_0, strides = var_373_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_373_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor(1)]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([1])]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9321216)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322048)))]; + tensor var_398_to_fp16 = const()[name = tensor("op_398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_398_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor q_5_pad_type_0 = const()[name = tensor("q_5_pad_type_0"), val = tensor("valid")]; + tensor q_5_strides_0 = const()[name = tensor("q_5_strides_0"), val = tensor([1, 1])]; + tensor q_5_pad_0 = const()[name = tensor("q_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_5_dilations_0 = const()[name = tensor("q_5_dilations_0"), val = tensor([1, 1])]; + tensor q_5_groups_0 = const()[name = tensor("q_5_groups_0"), val = tensor(1)]; + tensor var_433_weight_0_to_fp16 = const()[name = tensor("op_433_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322880)))]; + tensor var_433_bias_0_to_fp16 = const()[name = tensor("op_433_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9617856)))]; + tensor var_433_cast_fp16 = conv(bias = var_433_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_433_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("op_433_cast_fp16")]; + tensor k_5_pad_type_0 = const()[name = tensor("k_5_pad_type_0"), val = tensor("valid")]; + tensor k_5_strides_0 = const()[name = tensor("k_5_strides_0"), val = tensor([1, 1])]; + tensor k_5_pad_0 = const()[name = tensor("k_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_5_dilations_0 = const()[name = tensor("k_5_dilations_0"), val = tensor([1, 1])]; + tensor k_5_groups_0 = const()[name = tensor("k_5_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_key_weight_to_fp16 = const()[name = tensor("blocks_2_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9618688)))]; + tensor k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_431_pad_type_0 = const()[name = tensor("op_431_pad_type_0"), val = tensor("valid")]; + tensor var_431_strides_0 = const()[name = tensor("op_431_strides_0"), val = tensor([1, 1])]; + tensor var_431_pad_0 = const()[name = tensor("op_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_431_dilations_0 = const()[name = tensor("op_431_dilations_0"), val = tensor([1, 1])]; + tensor var_431_groups_0 = const()[name = tensor("op_431_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_value_weight_to_fp16 = const()[name = tensor("blocks_2_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9913664)))]; + tensor blocks_2_attn_value_bias_to_fp16 = const()[name = tensor("blocks_2_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10208640)))]; + tensor var_431_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_431_dilations_0, groups = var_431_groups_0, pad = var_431_pad_0, pad_type = var_431_pad_type_0, strides = var_431_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("op_431_cast_fp16")]; + tensor tile_6 = const()[name = tensor("tile_6"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_434_axis_0 = const()[name = tensor("op_434_axis_0"), val = tensor(1)]; + tensor var_434_cast_fp16_0, tensor var_434_cast_fp16_1, tensor var_434_cast_fp16_2, tensor var_434_cast_fp16_3, tensor var_434_cast_fp16_4, tensor var_434_cast_fp16_5 = split(axis = var_434_axis_0, split_sizes = tile_6, x = var_433_cast_fp16)[name = tensor("op_434_cast_fp16")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_7 = const()[name = tensor("tile_7"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(3)]; + tensor var_441_cast_fp16 = transpose(perm = var_441_perm_0, x = k_5_cast_fp16)[name = tensor("transpose_2")]; + tensor var_442_cast_fp16_0, tensor var_442_cast_fp16_1, tensor var_442_cast_fp16_2, tensor var_442_cast_fp16_3, tensor var_442_cast_fp16_4, tensor var_442_cast_fp16_5 = split(axis = var_442_axis_0, split_sizes = tile_7, x = var_441_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor tile_8 = const()[name = tensor("tile_8"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_449_axis_0 = const()[name = tensor("op_449_axis_0"), val = tensor(1)]; + tensor var_449_cast_fp16_0, tensor var_449_cast_fp16_1, tensor var_449_cast_fp16_2, tensor var_449_cast_fp16_3, tensor var_449_cast_fp16_4, tensor var_449_cast_fp16_5 = split(axis = var_449_axis_0, split_sizes = tile_8, x = var_431_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor aw_25_equation_0 = const()[name = tensor("aw_25_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_442_cast_fp16_0, var_434_cast_fp16_0))[name = tensor("aw_25_cast_fp16")]; + tensor aw_27_equation_0 = const()[name = tensor("aw_27_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_442_cast_fp16_1, var_434_cast_fp16_1))[name = tensor("aw_27_cast_fp16")]; + tensor aw_29_equation_0 = const()[name = tensor("aw_29_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_442_cast_fp16_2, var_434_cast_fp16_2))[name = tensor("aw_29_cast_fp16")]; + tensor aw_31_equation_0 = const()[name = tensor("aw_31_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_442_cast_fp16_3, var_434_cast_fp16_3))[name = tensor("aw_31_cast_fp16")]; + tensor aw_33_equation_0 = const()[name = tensor("aw_33_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_442_cast_fp16_4, var_434_cast_fp16_4))[name = tensor("aw_33_cast_fp16")]; + tensor aw_35_equation_0 = const()[name = tensor("aw_35_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_442_cast_fp16_5, var_434_cast_fp16_5))[name = tensor("aw_35_cast_fp16")]; + tensor var_468_cast_fp16 = softmax(axis = var_382, x = aw_25_cast_fp16)[name = tensor("op_468_cast_fp16")]; + tensor var_469_cast_fp16 = softmax(axis = var_382, x = aw_27_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = softmax(axis = var_382, x = aw_29_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = softmax(axis = var_382, x = aw_31_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor var_472_cast_fp16 = softmax(axis = var_382, x = aw_33_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor var_473_cast_fp16 = softmax(axis = var_382, x = aw_35_cast_fp16)[name = tensor("op_473_cast_fp16")]; + tensor var_475_equation_0 = const()[name = tensor("op_475_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_475_cast_fp16 = einsum(equation = var_475_equation_0, values = (var_449_cast_fp16_0, var_468_cast_fp16))[name = tensor("op_475_cast_fp16")]; + tensor var_477_equation_0 = const()[name = tensor("op_477_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_477_cast_fp16 = einsum(equation = var_477_equation_0, values = (var_449_cast_fp16_1, var_469_cast_fp16))[name = tensor("op_477_cast_fp16")]; + tensor var_479_equation_0 = const()[name = tensor("op_479_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_479_cast_fp16 = einsum(equation = var_479_equation_0, values = (var_449_cast_fp16_2, var_470_cast_fp16))[name = tensor("op_479_cast_fp16")]; + tensor var_481_equation_0 = const()[name = tensor("op_481_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_481_cast_fp16 = einsum(equation = var_481_equation_0, values = (var_449_cast_fp16_3, var_471_cast_fp16))[name = tensor("op_481_cast_fp16")]; + tensor var_483_equation_0 = const()[name = tensor("op_483_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_483_cast_fp16 = einsum(equation = var_483_equation_0, values = (var_449_cast_fp16_4, var_472_cast_fp16))[name = tensor("op_483_cast_fp16")]; + tensor var_485_equation_0 = const()[name = tensor("op_485_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_485_cast_fp16 = einsum(equation = var_485_equation_0, values = (var_449_cast_fp16_5, var_473_cast_fp16))[name = tensor("op_485_cast_fp16")]; + tensor input_25_interleave_0 = const()[name = tensor("input_25_interleave_0"), val = tensor(false)]; + tensor input_25_cast_fp16 = concat(axis = var_382, interleave = input_25_interleave_0, values = (var_475_cast_fp16, var_477_cast_fp16, var_479_cast_fp16, var_481_cast_fp16, var_483_cast_fp16, var_485_cast_fp16))[name = tensor("input_25_cast_fp16")]; + tensor var_494_pad_type_0 = const()[name = tensor("op_494_pad_type_0"), val = tensor("valid")]; + tensor var_494_strides_0 = const()[name = tensor("op_494_strides_0"), val = tensor([1, 1])]; + tensor var_494_pad_0 = const()[name = tensor("op_494_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_494_dilations_0 = const()[name = tensor("op_494_dilations_0"), val = tensor([1, 1])]; + tensor var_494_groups_0 = const()[name = tensor("op_494_groups_0"), val = tensor(1)]; + tensor blocks_2_attn_out_weight_to_fp16 = const()[name = tensor("blocks_2_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10209472)))]; + tensor blocks_2_attn_out_bias_to_fp16 = const()[name = tensor("blocks_2_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10504448)))]; + tensor var_494_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_494_dilations_0, groups = var_494_groups_0, pad = var_494_pad_0, pad_type = var_494_pad_type_0, strides = var_494_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_494_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([1])]; + tensor input_27_gamma_0_to_fp16 = const()[name = tensor("input_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10505280)))]; + tensor input_27_beta_0_to_fp16 = const()[name = tensor("input_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506112)))]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_504_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506944)))]; + tensor blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11686656)))]; + tensor input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_530_pad_type_0 = const()[name = tensor("op_530_pad_type_0"), val = tensor("valid")]; + tensor var_530_strides_0 = const()[name = tensor("op_530_strides_0"), val = tensor([1, 1])]; + tensor var_530_pad_0 = const()[name = tensor("op_530_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_530_dilations_0 = const()[name = tensor("op_530_dilations_0"), val = tensor([1, 1])]; + tensor var_530_groups_0 = const()[name = tensor("op_530_groups_0"), val = tensor(1)]; + tensor blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11689792)))]; + tensor blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12869504)))]; + tensor var_530_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_530_dilations_0, groups = var_530_groups_0, pad = var_530_pad_0, pad_type = var_530_pad_type_0, strides = var_530_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_530_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor(1)]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([1])]; + tensor input_33_gamma_0_to_fp16 = const()[name = tensor("input_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12870336)))]; + tensor input_33_beta_0_to_fp16 = const()[name = tensor("input_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12871168)))]; + tensor var_555_to_fp16 = const()[name = tensor("op_555_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_555_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor q_pad_type_0 = const()[name = tensor("q_pad_type_0"), val = tensor("valid")]; + tensor q_strides_0 = const()[name = tensor("q_strides_0"), val = tensor([1, 1])]; + tensor q_pad_0 = const()[name = tensor("q_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor q_dilations_0 = const()[name = tensor("q_dilations_0"), val = tensor([1, 1])]; + tensor q_groups_0 = const()[name = tensor("q_groups_0"), val = tensor(1)]; + tensor var_590_weight_0_to_fp16 = const()[name = tensor("op_590_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12872000)))]; + tensor var_590_bias_0_to_fp16 = const()[name = tensor("op_590_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13166976)))]; + tensor var_590_cast_fp16 = conv(bias = var_590_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_590_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor k_pad_type_0 = const()[name = tensor("k_pad_type_0"), val = tensor("valid")]; + tensor k_strides_0 = const()[name = tensor("k_strides_0"), val = tensor([1, 1])]; + tensor k_pad_0 = const()[name = tensor("k_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor k_dilations_0 = const()[name = tensor("k_dilations_0"), val = tensor([1, 1])]; + tensor k_groups_0 = const()[name = tensor("k_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_key_weight_to_fp16 = const()[name = tensor("blocks_3_attn_key_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13167808)))]; + tensor k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_588_pad_type_0 = const()[name = tensor("op_588_pad_type_0"), val = tensor("valid")]; + tensor var_588_strides_0 = const()[name = tensor("op_588_strides_0"), val = tensor([1, 1])]; + tensor var_588_pad_0 = const()[name = tensor("op_588_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_588_dilations_0 = const()[name = tensor("op_588_dilations_0"), val = tensor([1, 1])]; + tensor var_588_groups_0 = const()[name = tensor("op_588_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_value_weight_to_fp16 = const()[name = tensor("blocks_3_attn_value_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13462784)))]; + tensor blocks_3_attn_value_bias_to_fp16 = const()[name = tensor("blocks_3_attn_value_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13757760)))]; + tensor var_588_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_588_dilations_0, groups = var_588_groups_0, pad = var_588_pad_0, pad_type = var_588_pad_type_0, strides = var_588_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("op_588_cast_fp16")]; + tensor tile_9 = const()[name = tensor("tile_9"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_591_axis_0 = const()[name = tensor("op_591_axis_0"), val = tensor(1)]; + tensor var_591_cast_fp16_0, tensor var_591_cast_fp16_1, tensor var_591_cast_fp16_2, tensor var_591_cast_fp16_3, tensor var_591_cast_fp16_4, tensor var_591_cast_fp16_5 = split(axis = var_591_axis_0, split_sizes = tile_9, x = var_590_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_598_perm_0 = const()[name = tensor("op_598_perm_0"), val = tensor([0, 3, 2, 1])]; + tensor tile_10 = const()[name = tensor("tile_10"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_599_axis_0 = const()[name = tensor("op_599_axis_0"), val = tensor(3)]; + tensor var_598_cast_fp16 = transpose(perm = var_598_perm_0, x = k_cast_fp16)[name = tensor("transpose_1")]; + tensor var_599_cast_fp16_0, tensor var_599_cast_fp16_1, tensor var_599_cast_fp16_2, tensor var_599_cast_fp16_3, tensor var_599_cast_fp16_4, tensor var_599_cast_fp16_5 = split(axis = var_599_axis_0, split_sizes = tile_10, x = var_598_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor tile_11 = const()[name = tensor("tile_11"), val = tensor([64, 64, 64, 64, 64, 64])]; + tensor var_606_axis_0 = const()[name = tensor("op_606_axis_0"), val = tensor(1)]; + tensor var_606_cast_fp16_0, tensor var_606_cast_fp16_1, tensor var_606_cast_fp16_2, tensor var_606_cast_fp16_3, tensor var_606_cast_fp16_4, tensor var_606_cast_fp16_5 = split(axis = var_606_axis_0, split_sizes = tile_11, x = var_588_cast_fp16)[name = tensor("op_606_cast_fp16")]; + tensor aw_37_equation_0 = const()[name = tensor("aw_37_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_599_cast_fp16_0, var_591_cast_fp16_0))[name = tensor("aw_37_cast_fp16")]; + tensor aw_39_equation_0 = const()[name = tensor("aw_39_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_599_cast_fp16_1, var_591_cast_fp16_1))[name = tensor("aw_39_cast_fp16")]; + tensor aw_41_equation_0 = const()[name = tensor("aw_41_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_599_cast_fp16_2, var_591_cast_fp16_2))[name = tensor("aw_41_cast_fp16")]; + tensor aw_43_equation_0 = const()[name = tensor("aw_43_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_599_cast_fp16_3, var_591_cast_fp16_3))[name = tensor("aw_43_cast_fp16")]; + tensor aw_45_equation_0 = const()[name = tensor("aw_45_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_599_cast_fp16_4, var_591_cast_fp16_4))[name = tensor("aw_45_cast_fp16")]; + tensor aw_equation_0 = const()[name = tensor("aw_equation_0"), val = tensor("bkhc,bchq->bkhq")]; + tensor aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_599_cast_fp16_5, var_591_cast_fp16_5))[name = tensor("aw_cast_fp16")]; + tensor var_625_cast_fp16 = softmax(axis = var_539, x = aw_37_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor var_626_cast_fp16 = softmax(axis = var_539, x = aw_39_cast_fp16)[name = tensor("op_626_cast_fp16")]; + tensor var_627_cast_fp16 = softmax(axis = var_539, x = aw_41_cast_fp16)[name = tensor("op_627_cast_fp16")]; + tensor var_628_cast_fp16 = softmax(axis = var_539, x = aw_43_cast_fp16)[name = tensor("op_628_cast_fp16")]; + tensor var_629_cast_fp16 = softmax(axis = var_539, x = aw_45_cast_fp16)[name = tensor("op_629_cast_fp16")]; + tensor var_630_cast_fp16 = softmax(axis = var_539, x = aw_cast_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_632_equation_0 = const()[name = tensor("op_632_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_632_cast_fp16 = einsum(equation = var_632_equation_0, values = (var_606_cast_fp16_0, var_625_cast_fp16))[name = tensor("op_632_cast_fp16")]; + tensor var_634_equation_0 = const()[name = tensor("op_634_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_634_cast_fp16 = einsum(equation = var_634_equation_0, values = (var_606_cast_fp16_1, var_626_cast_fp16))[name = tensor("op_634_cast_fp16")]; + tensor var_636_equation_0 = const()[name = tensor("op_636_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_636_cast_fp16 = einsum(equation = var_636_equation_0, values = (var_606_cast_fp16_2, var_627_cast_fp16))[name = tensor("op_636_cast_fp16")]; + tensor var_638_equation_0 = const()[name = tensor("op_638_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_638_cast_fp16 = einsum(equation = var_638_equation_0, values = (var_606_cast_fp16_3, var_628_cast_fp16))[name = tensor("op_638_cast_fp16")]; + tensor var_640_equation_0 = const()[name = tensor("op_640_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_640_cast_fp16 = einsum(equation = var_640_equation_0, values = (var_606_cast_fp16_4, var_629_cast_fp16))[name = tensor("op_640_cast_fp16")]; + tensor var_642_equation_0 = const()[name = tensor("op_642_equation_0"), val = tensor("bchk,bkhq->bchq")]; + tensor var_642_cast_fp16 = einsum(equation = var_642_equation_0, values = (var_606_cast_fp16_5, var_630_cast_fp16))[name = tensor("op_642_cast_fp16")]; + tensor input_35_interleave_0 = const()[name = tensor("input_35_interleave_0"), val = tensor(false)]; + tensor input_35_cast_fp16 = concat(axis = var_539, interleave = input_35_interleave_0, values = (var_632_cast_fp16, var_634_cast_fp16, var_636_cast_fp16, var_638_cast_fp16, var_640_cast_fp16, var_642_cast_fp16))[name = tensor("input_35_cast_fp16")]; + tensor var_651_pad_type_0 = const()[name = tensor("op_651_pad_type_0"), val = tensor("valid")]; + tensor var_651_strides_0 = const()[name = tensor("op_651_strides_0"), val = tensor([1, 1])]; + tensor var_651_pad_0 = const()[name = tensor("op_651_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_651_dilations_0 = const()[name = tensor("op_651_dilations_0"), val = tensor([1, 1])]; + tensor var_651_groups_0 = const()[name = tensor("op_651_groups_0"), val = tensor(1)]; + tensor blocks_3_attn_out_weight_to_fp16 = const()[name = tensor("blocks_3_attn_out_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13758592)))]; + tensor blocks_3_attn_out_bias_to_fp16 = const()[name = tensor("blocks_3_attn_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14053568)))]; + tensor var_651_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_651_dilations_0, groups = var_651_groups_0, pad = var_651_pad_0, pad_type = var_651_pad_type_0, strides = var_651_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("op_651_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_651_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor input_37_axes_0 = const()[name = tensor("input_37_axes_0"), val = tensor([1])]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14054400)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14055232)))]; + tensor var_661_to_fp16 = const()[name = tensor("op_661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_661_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14056064)))]; + tensor blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15235776)))]; + tensor input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_39_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_687_pad_type_0 = const()[name = tensor("op_687_pad_type_0"), val = tensor("valid")]; + tensor var_687_strides_0 = const()[name = tensor("op_687_strides_0"), val = tensor([1, 1])]; + tensor var_687_pad_0 = const()[name = tensor("op_687_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_687_dilations_0 = const()[name = tensor("op_687_dilations_0"), val = tensor([1, 1])]; + tensor var_687_groups_0 = const()[name = tensor("op_687_groups_0"), val = tensor(1)]; + tensor blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15238912)))]; + tensor blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16418624)))]; + tensor var_687_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_687_dilations_0, groups = var_687_groups_0, pad = var_687_pad_0, pad_type = var_687_pad_type_0, strides = var_687_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_687_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([1])]; + tensor x_gamma_0_to_fp16 = const()[name = tensor("x_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16419456)))]; + tensor x_beta_0_to_fp16 = const()[name = tensor("x_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16420288)))]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_701_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_712_axes_0 = const()[name = tensor("op_712_axes_0"), val = tensor([2])]; + tensor var_712_cast_fp16 = squeeze(axes = var_712_axes_0, x = x_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_715_perm_0 = const()[name = tensor("op_715_perm_0"), val = tensor([0, 2, 1])]; + tensor var_715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_715_cast_fp16 = transpose(perm = var_715_perm_0, x = var_712_cast_fp16)[name = tensor("transpose_0")]; + tensor output = cast(dtype = var_715_cast_fp16_to_fp32_dtype_0, x = var_715_cast_fp16)[name = tensor("cast_19")]; + } -> (output); +} \ No newline at end of file diff --git a/tiny/ggml-tiny-encoder.mlmodelc/weights/weight.bin b/tiny/ggml-tiny-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..59aa32c2ef3eb1490176f23728b6407163bd8b1b --- /dev/null +++ b/tiny/ggml-tiny-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17aa929eb61d433fa68217d73a8aec7125af1b4afe39b5c8f27d61d80e2c1a80 +size 16421120 diff --git a/tiny/ggml-tiny.bin b/tiny/ggml-tiny.bin new file mode 100644 index 0000000000000000000000000000000000000000..d144f735b005ae8cbfa04a49e22fe40faa24dbec --- /dev/null +++ b/tiny/ggml-tiny.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be07e048e1e599ad46341c8d2a135645097a538221678b7acdd1b1919c6e1b21 +size 77691713