diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..d03c1b619119d3c27ba5559c0d3b5b5bc3b82ffa --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c92b3c5f59e252fdaa2ad9c836218de4b4b4cf17fcc046a186fa52c55f68e91c +size 36533348 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3f5eb083dfd6b2e7283b81a31687c1b9ea8a3dc3 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2441ae34fc7d12946dba7b63379063e856ffc7c3e11ba5f7533efb1450562ca6 +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..0a32df6173b94e0faa31dcc1c960e260bfc9adfc --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 36, + "Matmul" : 12, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 6, + "Add" : 13, + "LayerNorm" : 13, + "Mul" : 12, + "Transpose" : 25, + "Gelu" : 8, + "Reshape" : 24 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..1a556286ea11b368906c6d7a5613793051843109 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/model.mil @@ -0,0 +1,393 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_20 = const()[name = tensor("op_20"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor([1])]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245888)))]; + tensor cast_187 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_32_cast = conv(bias = bias_3_to_fp16, dilations = var_30, groups = var_20, pad = var_32_pad_0, pad_type = var_32_pad_type_0, strides = var_28, weight = weight_3_to_fp16, x = cast_187); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_32_cast); + tensor var_36 = const()[name = tensor("op_36"), val = tensor(1)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor([2])]; + tensor var_47 = const()[name = tensor("op_47"), val = tensor([1])]; + tensor var_49_pad_type_0 = const()[name = tensor("op_49_pad_type_0"), val = tensor("custom")]; + tensor var_49_pad_0 = const()[name = tensor("op_49_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246976)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1819904)))]; + tensor var_49_cast = conv(bias = bias_7_to_fp16, dilations = var_47, groups = var_36, pad = var_49_pad_0, pad_type = var_49_pad_type_0, strides = var_45, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_49_cast); + tensor var_54 = const()[name = tensor("op_54"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1820992)))]; + tensor transpose_48 = transpose(perm = var_54, x = x_3_cast); + tensor var_57_cast = add(x = transpose_48, y = positional_embedding_to_fp16); + tensor var_70 = const()[name = tensor("op_70"), val = tensor(-1)]; + tensor var_87_axes_0 = const()[name = tensor("op_87_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3357056)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3358144)))]; + tensor var_76_to_fp16 = const()[name = tensor("op_76_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_87_cast = layer_norm(axes = var_87_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_76_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_57_cast); + tensor var_98_to_fp16 = const()[name = tensor("op_98_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3359232)))]; + tensor var_99_to_fp16 = const()[name = tensor("op_99_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3883584)))]; + tensor q_1_cast = linear(bias = var_99_to_fp16, weight = var_98_to_fp16, x = var_87_cast); + tensor var_102_to_fp16 = const()[name = tensor("op_102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3884672)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4409024)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_102_to_fp16, x = var_87_cast); + tensor var_106_to_fp16 = const()[name = tensor("op_106_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4410112)))]; + tensor var_107_to_fp16 = const()[name = tensor("op_107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4934464)))]; + tensor v_1_cast = linear(bias = var_107_to_fp16, weight = var_106_to_fp16, x = var_87_cast); + tensor var_115 = const()[name = tensor("op_115"), val = tensor([1, 1500, 8, -1])]; + tensor var_116_cast = reshape(shape = var_115, x = q_1_cast); + tensor const_42_to_fp16 = const()[name = tensor("const_42_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_116_cast, y = const_42_to_fp16); + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1500, 8, -1])]; + tensor var_123_cast = reshape(shape = var_122, x = k_1_cast); + tensor const_43_to_fp16 = const()[name = tensor("const_43_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_123_cast, y = const_43_to_fp16); + tensor var_129 = const()[name = tensor("op_129"), val = tensor([1, 1500, 8, -1])]; + tensor var_130_cast = reshape(shape = var_129, x = v_1_cast); + tensor var_131 = const()[name = tensor("op_131"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_45 = transpose(perm = transpose_13_perm_0, x = k_3_cast); + tensor transpose_46 = transpose(perm = transpose_12_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_46, y = transpose_45); + tensor var_135_cast = softmax(axis = var_70, x = qk_1_cast); + tensor var_137_transpose_x_0 = const()[name = tensor("op_137_transpose_x_0"), val = tensor(false)]; + tensor var_137_transpose_y_0 = const()[name = tensor("op_137_transpose_y_0"), val = tensor(false)]; + tensor transpose_47 = transpose(perm = var_131, x = var_130_cast); + tensor var_137_cast = matmul(transpose_x = var_137_transpose_x_0, transpose_y = var_137_transpose_y_0, x = var_135_cast, y = transpose_47); + tensor var_138 = const()[name = tensor("op_138"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 512])]; + tensor transpose_44 = transpose(perm = var_138, x = var_137_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_44); + tensor var_143_to_fp16 = const()[name = tensor("op_143_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4935552)))]; + tensor var_144_to_fp16 = const()[name = tensor("op_144_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5459904)))]; + tensor var_145_cast = linear(bias = var_144_to_fp16, weight = var_143_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_57_cast, y = var_145_cast); + tensor var_151_axes_0 = const()[name = tensor("op_151_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5460992)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5462080)))]; + tensor var_151_cast = layer_norm(axes = var_151_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_76_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_160_to_fp16 = const()[name = tensor("op_160_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5463168)))]; + tensor var_161_to_fp16 = const()[name = tensor("op_161_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7560384)))]; + tensor input_9_cast = linear(bias = var_161_to_fp16, weight = var_160_to_fp16, x = var_151_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_166_to_fp16 = const()[name = tensor("op_166_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7564544)))]; + tensor var_167_to_fp16 = const()[name = tensor("op_167_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9661760)))]; + tensor var_168_cast = linear(bias = var_167_to_fp16, weight = var_166_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_168_cast); + tensor var_177 = const()[name = tensor("op_177"), val = tensor(-1)]; + tensor var_194_axes_0 = const()[name = tensor("op_194_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9662848)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9663936)))]; + tensor var_183_to_fp16 = const()[name = tensor("op_183_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_194_cast = layer_norm(axes = var_194_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_183_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_205_to_fp16 = const()[name = tensor("op_205_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9665024)))]; + tensor var_206_to_fp16 = const()[name = tensor("op_206_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10189376)))]; + tensor q_5_cast = linear(bias = var_206_to_fp16, weight = var_205_to_fp16, x = var_194_cast); + tensor var_209_to_fp16 = const()[name = tensor("op_209_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10190464)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10714816)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_209_to_fp16, x = var_194_cast); + tensor var_213_to_fp16 = const()[name = tensor("op_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10715904)))]; + tensor var_214_to_fp16 = const()[name = tensor("op_214_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11240256)))]; + tensor v_5_cast = linear(bias = var_214_to_fp16, weight = var_213_to_fp16, x = var_194_cast); + tensor var_222 = const()[name = tensor("op_222"), val = tensor([1, 1500, 8, -1])]; + tensor var_223_cast = reshape(shape = var_222, x = q_5_cast); + tensor const_44_to_fp16 = const()[name = tensor("const_44_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_223_cast, y = const_44_to_fp16); + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 1500, 8, -1])]; + tensor var_230_cast = reshape(shape = var_229, x = k_5_cast); + tensor const_45_to_fp16 = const()[name = tensor("const_45_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_230_cast, y = const_45_to_fp16); + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 1500, 8, -1])]; + tensor var_237_cast = reshape(shape = var_236, x = v_5_cast); + tensor var_238 = const()[name = tensor("op_238"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_41 = transpose(perm = transpose_15_perm_0, x = k_7_cast); + tensor transpose_42 = transpose(perm = transpose_14_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_42, y = transpose_41); + tensor var_242_cast = softmax(axis = var_177, x = qk_3_cast); + tensor var_244_transpose_x_0 = const()[name = tensor("op_244_transpose_x_0"), val = tensor(false)]; + tensor var_244_transpose_y_0 = const()[name = tensor("op_244_transpose_y_0"), val = tensor(false)]; + tensor transpose_43 = transpose(perm = var_238, x = var_237_cast); + tensor var_244_cast = matmul(transpose_x = var_244_transpose_x_0, transpose_y = var_244_transpose_y_0, x = var_242_cast, y = transpose_43); + tensor var_245 = const()[name = tensor("op_245"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 512])]; + tensor transpose_40 = transpose(perm = var_245, x = var_244_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_40); + tensor var_250_to_fp16 = const()[name = tensor("op_250_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11241344)))]; + tensor var_251_to_fp16 = const()[name = tensor("op_251_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11765696)))]; + tensor var_252_cast = linear(bias = var_251_to_fp16, weight = var_250_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_252_cast); + tensor var_258_axes_0 = const()[name = tensor("op_258_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11766784)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11767872)))]; + tensor var_258_cast = layer_norm(axes = var_258_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_183_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11768960)))]; + tensor var_268_to_fp16 = const()[name = tensor("op_268_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13866176)))]; + tensor input_17_cast = linear(bias = var_268_to_fp16, weight = var_267_to_fp16, x = var_258_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_273_to_fp16 = const()[name = tensor("op_273_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13870336)))]; + tensor var_274_to_fp16 = const()[name = tensor("op_274_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15967552)))]; + tensor var_275_cast = linear(bias = var_274_to_fp16, weight = var_273_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_275_cast); + tensor var_284 = const()[name = tensor("op_284"), val = tensor(-1)]; + tensor var_301_axes_0 = const()[name = tensor("op_301_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15968640)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15969728)))]; + tensor var_290_to_fp16 = const()[name = tensor("op_290_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_301_cast = layer_norm(axes = var_301_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_290_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_312_to_fp16 = const()[name = tensor("op_312_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15970816)))]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16495168)))]; + tensor q_9_cast = linear(bias = var_313_to_fp16, weight = var_312_to_fp16, x = var_301_cast); + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16496256)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17020608)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_316_to_fp16, x = var_301_cast); + tensor var_320_to_fp16 = const()[name = tensor("op_320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17021696)))]; + tensor var_321_to_fp16 = const()[name = tensor("op_321_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17546048)))]; + tensor v_9_cast = linear(bias = var_321_to_fp16, weight = var_320_to_fp16, x = var_301_cast); + tensor var_329 = const()[name = tensor("op_329"), val = tensor([1, 1500, 8, -1])]; + tensor var_330_cast = reshape(shape = var_329, x = q_9_cast); + tensor const_46_to_fp16 = const()[name = tensor("const_46_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_330_cast, y = const_46_to_fp16); + tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1500, 8, -1])]; + tensor var_337_cast = reshape(shape = var_336, x = k_9_cast); + tensor const_47_to_fp16 = const()[name = tensor("const_47_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_337_cast, y = const_47_to_fp16); + tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 1500, 8, -1])]; + tensor var_344_cast = reshape(shape = var_343, x = v_9_cast); + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_37 = transpose(perm = transpose_17_perm_0, x = k_11_cast); + tensor transpose_38 = transpose(perm = transpose_16_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_38, y = transpose_37); + tensor var_349_cast = softmax(axis = var_284, x = qk_5_cast); + tensor var_351_transpose_x_0 = const()[name = tensor("op_351_transpose_x_0"), val = tensor(false)]; + tensor var_351_transpose_y_0 = const()[name = tensor("op_351_transpose_y_0"), val = tensor(false)]; + tensor transpose_39 = transpose(perm = var_345, x = var_344_cast); + tensor var_351_cast = matmul(transpose_x = var_351_transpose_x_0, transpose_y = var_351_transpose_y_0, x = var_349_cast, y = transpose_39); + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 512])]; + tensor transpose_36 = transpose(perm = var_352, x = var_351_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_36); + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17547136)))]; + tensor var_358_to_fp16 = const()[name = tensor("op_358_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18071488)))]; + tensor var_359_cast = linear(bias = var_358_to_fp16, weight = var_357_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_359_cast); + tensor var_365_axes_0 = const()[name = tensor("op_365_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18072576)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18073664)))]; + tensor var_365_cast = layer_norm(axes = var_365_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_290_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_374_to_fp16 = const()[name = tensor("op_374_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18074752)))]; + tensor var_375_to_fp16 = const()[name = tensor("op_375_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20171968)))]; + tensor input_25_cast = linear(bias = var_375_to_fp16, weight = var_374_to_fp16, x = var_365_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_380_to_fp16 = const()[name = tensor("op_380_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20176128)))]; + tensor var_381_to_fp16 = const()[name = tensor("op_381_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22273344)))]; + tensor var_382_cast = linear(bias = var_381_to_fp16, weight = var_380_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_382_cast); + tensor var_391 = const()[name = tensor("op_391"), val = tensor(-1)]; + tensor var_408_axes_0 = const()[name = tensor("op_408_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22274432)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22275520)))]; + tensor var_397_to_fp16 = const()[name = tensor("op_397_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_408_cast = layer_norm(axes = var_408_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_397_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_419_to_fp16 = const()[name = tensor("op_419_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22276608)))]; + tensor var_420_to_fp16 = const()[name = tensor("op_420_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22800960)))]; + tensor q_13_cast = linear(bias = var_420_to_fp16, weight = var_419_to_fp16, x = var_408_cast); + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22802048)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23326400)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_423_to_fp16, x = var_408_cast); + tensor var_427_to_fp16 = const()[name = tensor("op_427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23327488)))]; + tensor var_428_to_fp16 = const()[name = tensor("op_428_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23851840)))]; + tensor v_13_cast = linear(bias = var_428_to_fp16, weight = var_427_to_fp16, x = var_408_cast); + tensor var_436 = const()[name = tensor("op_436"), val = tensor([1, 1500, 8, -1])]; + tensor var_437_cast = reshape(shape = var_436, x = q_13_cast); + tensor const_48_to_fp16 = const()[name = tensor("const_48_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_437_cast, y = const_48_to_fp16); + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1500, 8, -1])]; + tensor var_444_cast = reshape(shape = var_443, x = k_13_cast); + tensor const_49_to_fp16 = const()[name = tensor("const_49_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_444_cast, y = const_49_to_fp16); + tensor var_450 = const()[name = tensor("op_450"), val = tensor([1, 1500, 8, -1])]; + tensor var_451_cast = reshape(shape = var_450, x = v_13_cast); + tensor var_452 = const()[name = tensor("op_452"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_33 = transpose(perm = transpose_19_perm_0, x = k_15_cast); + tensor transpose_34 = transpose(perm = transpose_18_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_34, y = transpose_33); + tensor var_456_cast = softmax(axis = var_391, x = qk_7_cast); + tensor var_458_transpose_x_0 = const()[name = tensor("op_458_transpose_x_0"), val = tensor(false)]; + tensor var_458_transpose_y_0 = const()[name = tensor("op_458_transpose_y_0"), val = tensor(false)]; + tensor transpose_35 = transpose(perm = var_452, x = var_451_cast); + tensor var_458_cast = matmul(transpose_x = var_458_transpose_x_0, transpose_y = var_458_transpose_y_0, x = var_456_cast, y = transpose_35); + tensor var_459 = const()[name = tensor("op_459"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 512])]; + tensor transpose_32 = transpose(perm = var_459, x = var_458_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_32); + tensor var_464_to_fp16 = const()[name = tensor("op_464_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23852928)))]; + tensor var_465_to_fp16 = const()[name = tensor("op_465_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24377280)))]; + tensor var_466_cast = linear(bias = var_465_to_fp16, weight = var_464_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_466_cast); + tensor var_472_axes_0 = const()[name = tensor("op_472_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24378368)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24379456)))]; + tensor var_472_cast = layer_norm(axes = var_472_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_397_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_481_to_fp16 = const()[name = tensor("op_481_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24380544)))]; + tensor var_482_to_fp16 = const()[name = tensor("op_482_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26477760)))]; + tensor input_33_cast = linear(bias = var_482_to_fp16, weight = var_481_to_fp16, x = var_472_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_487_to_fp16 = const()[name = tensor("op_487_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26481920)))]; + tensor var_488_to_fp16 = const()[name = tensor("op_488_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28579136)))]; + tensor var_489_cast = linear(bias = var_488_to_fp16, weight = var_487_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_489_cast); + tensor var_498 = const()[name = tensor("op_498"), val = tensor(-1)]; + tensor var_515_axes_0 = const()[name = tensor("op_515_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28580224)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28581312)))]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_515_cast = layer_norm(axes = var_515_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_504_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28582400)))]; + tensor var_527_to_fp16 = const()[name = tensor("op_527_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29106752)))]; + tensor q_17_cast = linear(bias = var_527_to_fp16, weight = var_526_to_fp16, x = var_515_cast); + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29107840)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29632192)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_530_to_fp16, x = var_515_cast); + tensor var_534_to_fp16 = const()[name = tensor("op_534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29633280)))]; + tensor var_535_to_fp16 = const()[name = tensor("op_535_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30157632)))]; + tensor v_17_cast = linear(bias = var_535_to_fp16, weight = var_534_to_fp16, x = var_515_cast); + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1500, 8, -1])]; + tensor var_544_cast = reshape(shape = var_543, x = q_17_cast); + tensor const_50_to_fp16 = const()[name = tensor("const_50_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_544_cast, y = const_50_to_fp16); + tensor var_550 = const()[name = tensor("op_550"), val = tensor([1, 1500, 8, -1])]; + tensor var_551_cast = reshape(shape = var_550, x = k_17_cast); + tensor const_51_to_fp16 = const()[name = tensor("const_51_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_551_cast, y = const_51_to_fp16); + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1500, 8, -1])]; + tensor var_558_cast = reshape(shape = var_557, x = v_17_cast); + tensor var_559 = const()[name = tensor("op_559"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_29 = transpose(perm = transpose_21_perm_0, x = k_19_cast); + tensor transpose_30 = transpose(perm = transpose_20_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_30, y = transpose_29); + tensor var_563_cast = softmax(axis = var_498, x = qk_9_cast); + tensor var_565_transpose_x_0 = const()[name = tensor("op_565_transpose_x_0"), val = tensor(false)]; + tensor var_565_transpose_y_0 = const()[name = tensor("op_565_transpose_y_0"), val = tensor(false)]; + tensor transpose_31 = transpose(perm = var_559, x = var_558_cast); + tensor var_565_cast = matmul(transpose_x = var_565_transpose_x_0, transpose_y = var_565_transpose_y_0, x = var_563_cast, y = transpose_31); + tensor var_566 = const()[name = tensor("op_566"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 512])]; + tensor transpose_28 = transpose(perm = var_566, x = var_565_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_28); + tensor var_571_to_fp16 = const()[name = tensor("op_571_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30158720)))]; + tensor var_572_to_fp16 = const()[name = tensor("op_572_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30683072)))]; + tensor var_573_cast = linear(bias = var_572_to_fp16, weight = var_571_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_573_cast); + tensor var_579_axes_0 = const()[name = tensor("op_579_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30684160)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30685248)))]; + tensor var_579_cast = layer_norm(axes = var_579_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_504_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_588_to_fp16 = const()[name = tensor("op_588_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30686336)))]; + tensor var_589_to_fp16 = const()[name = tensor("op_589_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32783552)))]; + tensor input_41_cast = linear(bias = var_589_to_fp16, weight = var_588_to_fp16, x = var_579_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_594_to_fp16 = const()[name = tensor("op_594_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32787712)))]; + tensor var_595_to_fp16 = const()[name = tensor("op_595_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34884928)))]; + tensor var_596_cast = linear(bias = var_595_to_fp16, weight = var_594_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_596_cast); + tensor var_605 = const()[name = tensor("op_605"), val = tensor(-1)]; + tensor var_622_axes_0 = const()[name = tensor("op_622_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34886016)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34887104)))]; + tensor var_611_to_fp16 = const()[name = tensor("op_611_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_622_cast = layer_norm(axes = var_622_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_611_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_633_to_fp16 = const()[name = tensor("op_633_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34888192)))]; + tensor var_634_to_fp16 = const()[name = tensor("op_634_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35412544)))]; + tensor q_21_cast = linear(bias = var_634_to_fp16, weight = var_633_to_fp16, x = var_622_cast); + tensor var_637_to_fp16 = const()[name = tensor("op_637_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35413632)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35937984)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_637_to_fp16, x = var_622_cast); + tensor var_641_to_fp16 = const()[name = tensor("op_641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35939072)))]; + tensor var_642_to_fp16 = const()[name = tensor("op_642_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36463424)))]; + tensor v_21_cast = linear(bias = var_642_to_fp16, weight = var_641_to_fp16, x = var_622_cast); + tensor var_650 = const()[name = tensor("op_650"), val = tensor([1, 1500, 8, -1])]; + tensor var_651_cast = reshape(shape = var_650, x = q_21_cast); + tensor const_52_to_fp16 = const()[name = tensor("const_52_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_651_cast, y = const_52_to_fp16); + tensor var_657 = const()[name = tensor("op_657"), val = tensor([1, 1500, 8, -1])]; + tensor var_658_cast = reshape(shape = var_657, x = k_21_cast); + tensor const_53_to_fp16 = const()[name = tensor("const_53_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_658_cast, y = const_53_to_fp16); + tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 1500, 8, -1])]; + tensor var_665_cast = reshape(shape = var_664, x = v_21_cast); + tensor var_666 = const()[name = tensor("op_666"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_25 = transpose(perm = transpose_23_perm_0, x = k_cast); + tensor transpose_26 = transpose(perm = transpose_22_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_26, y = transpose_25); + tensor var_670_cast = softmax(axis = var_605, x = qk_cast); + tensor var_672_transpose_x_0 = const()[name = tensor("op_672_transpose_x_0"), val = tensor(false)]; + tensor var_672_transpose_y_0 = const()[name = tensor("op_672_transpose_y_0"), val = tensor(false)]; + tensor transpose_27 = transpose(perm = var_666, x = var_665_cast); + tensor var_672_cast = matmul(transpose_x = var_672_transpose_x_0, transpose_y = var_672_transpose_y_0, x = var_670_cast, y = transpose_27); + tensor var_673 = const()[name = tensor("op_673"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 512])]; + tensor transpose_24 = transpose(perm = var_673, x = var_672_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_24); + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36464512)))]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36988864)))]; + tensor var_680_cast = linear(bias = var_679_to_fp16, weight = var_678_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_680_cast); + tensor var_686_axes_0 = const()[name = tensor("op_686_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36989952)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36991040)))]; + tensor var_686_cast = layer_norm(axes = var_686_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_611_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_695_to_fp16 = const()[name = tensor("op_695_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36992128)))]; + tensor var_696_to_fp16 = const()[name = tensor("op_696_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39089344)))]; + tensor input_49_cast = linear(bias = var_696_to_fp16, weight = var_695_to_fp16, x = var_686_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39093504)))]; + tensor var_702_to_fp16 = const()[name = tensor("op_702_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190720)))]; + tensor var_703_cast = linear(bias = var_702_to_fp16, weight = var_701_to_fp16, x = x_77_cast); + tensor x_cast = add(x = x_73_cast, y = var_703_cast); + tensor var_716_axes_0 = const()[name = tensor("op_716_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41191808)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41192896)))]; + tensor var_707_to_fp16 = const()[name = tensor("op_707_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_716_cast = layer_norm(axes = var_716_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_707_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_716_cast_to_fp32_dtype_0 = const()[name = tensor("op_716_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_716_cast_to_fp32_dtype_0, x = var_716_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..34fb6f861394395ee81e8b21a6d9df8fa5b9a46d --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa38c9d90f01effad1990d1f17e67cd8382c6486b8e5a4dbacf44439cd838a38 +size 41193984 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3f5eb083dfd6b2e7283b81a31687c1b9ea8a3dc3 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2441ae34fc7d12946dba7b63379063e856ffc7c3e11ba5f7533efb1450562ca6 +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..d08d52976d7269d52e5f1283ff74c505ab57d50f --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 36, + "Matmul" : 12, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 6, + "Add" : 13, + "LayerNorm" : 13, + "Mul" : 12, + "Transpose" : 25, + "Gelu" : 8, + "Reshape" : 24 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..1a556286ea11b368906c6d7a5613793051843109 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/model.mil @@ -0,0 +1,393 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_20 = const()[name = tensor("op_20"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor([1])]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245888)))]; + tensor cast_187 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_32_cast = conv(bias = bias_3_to_fp16, dilations = var_30, groups = var_20, pad = var_32_pad_0, pad_type = var_32_pad_type_0, strides = var_28, weight = weight_3_to_fp16, x = cast_187); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_32_cast); + tensor var_36 = const()[name = tensor("op_36"), val = tensor(1)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor([2])]; + tensor var_47 = const()[name = tensor("op_47"), val = tensor([1])]; + tensor var_49_pad_type_0 = const()[name = tensor("op_49_pad_type_0"), val = tensor("custom")]; + tensor var_49_pad_0 = const()[name = tensor("op_49_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246976)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1819904)))]; + tensor var_49_cast = conv(bias = bias_7_to_fp16, dilations = var_47, groups = var_36, pad = var_49_pad_0, pad_type = var_49_pad_type_0, strides = var_45, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_49_cast); + tensor var_54 = const()[name = tensor("op_54"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1820992)))]; + tensor transpose_48 = transpose(perm = var_54, x = x_3_cast); + tensor var_57_cast = add(x = transpose_48, y = positional_embedding_to_fp16); + tensor var_70 = const()[name = tensor("op_70"), val = tensor(-1)]; + tensor var_87_axes_0 = const()[name = tensor("op_87_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3357056)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3358144)))]; + tensor var_76_to_fp16 = const()[name = tensor("op_76_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_87_cast = layer_norm(axes = var_87_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_76_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_57_cast); + tensor var_98_to_fp16 = const()[name = tensor("op_98_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3359232)))]; + tensor var_99_to_fp16 = const()[name = tensor("op_99_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3883584)))]; + tensor q_1_cast = linear(bias = var_99_to_fp16, weight = var_98_to_fp16, x = var_87_cast); + tensor var_102_to_fp16 = const()[name = tensor("op_102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3884672)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4409024)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_102_to_fp16, x = var_87_cast); + tensor var_106_to_fp16 = const()[name = tensor("op_106_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4410112)))]; + tensor var_107_to_fp16 = const()[name = tensor("op_107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4934464)))]; + tensor v_1_cast = linear(bias = var_107_to_fp16, weight = var_106_to_fp16, x = var_87_cast); + tensor var_115 = const()[name = tensor("op_115"), val = tensor([1, 1500, 8, -1])]; + tensor var_116_cast = reshape(shape = var_115, x = q_1_cast); + tensor const_42_to_fp16 = const()[name = tensor("const_42_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_116_cast, y = const_42_to_fp16); + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1500, 8, -1])]; + tensor var_123_cast = reshape(shape = var_122, x = k_1_cast); + tensor const_43_to_fp16 = const()[name = tensor("const_43_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_123_cast, y = const_43_to_fp16); + tensor var_129 = const()[name = tensor("op_129"), val = tensor([1, 1500, 8, -1])]; + tensor var_130_cast = reshape(shape = var_129, x = v_1_cast); + tensor var_131 = const()[name = tensor("op_131"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_45 = transpose(perm = transpose_13_perm_0, x = k_3_cast); + tensor transpose_46 = transpose(perm = transpose_12_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_46, y = transpose_45); + tensor var_135_cast = softmax(axis = var_70, x = qk_1_cast); + tensor var_137_transpose_x_0 = const()[name = tensor("op_137_transpose_x_0"), val = tensor(false)]; + tensor var_137_transpose_y_0 = const()[name = tensor("op_137_transpose_y_0"), val = tensor(false)]; + tensor transpose_47 = transpose(perm = var_131, x = var_130_cast); + tensor var_137_cast = matmul(transpose_x = var_137_transpose_x_0, transpose_y = var_137_transpose_y_0, x = var_135_cast, y = transpose_47); + tensor var_138 = const()[name = tensor("op_138"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 512])]; + tensor transpose_44 = transpose(perm = var_138, x = var_137_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_44); + tensor var_143_to_fp16 = const()[name = tensor("op_143_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4935552)))]; + tensor var_144_to_fp16 = const()[name = tensor("op_144_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5459904)))]; + tensor var_145_cast = linear(bias = var_144_to_fp16, weight = var_143_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_57_cast, y = var_145_cast); + tensor var_151_axes_0 = const()[name = tensor("op_151_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5460992)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5462080)))]; + tensor var_151_cast = layer_norm(axes = var_151_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_76_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_160_to_fp16 = const()[name = tensor("op_160_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5463168)))]; + tensor var_161_to_fp16 = const()[name = tensor("op_161_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7560384)))]; + tensor input_9_cast = linear(bias = var_161_to_fp16, weight = var_160_to_fp16, x = var_151_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_166_to_fp16 = const()[name = tensor("op_166_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7564544)))]; + tensor var_167_to_fp16 = const()[name = tensor("op_167_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9661760)))]; + tensor var_168_cast = linear(bias = var_167_to_fp16, weight = var_166_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_168_cast); + tensor var_177 = const()[name = tensor("op_177"), val = tensor(-1)]; + tensor var_194_axes_0 = const()[name = tensor("op_194_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9662848)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9663936)))]; + tensor var_183_to_fp16 = const()[name = tensor("op_183_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_194_cast = layer_norm(axes = var_194_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_183_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_205_to_fp16 = const()[name = tensor("op_205_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9665024)))]; + tensor var_206_to_fp16 = const()[name = tensor("op_206_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10189376)))]; + tensor q_5_cast = linear(bias = var_206_to_fp16, weight = var_205_to_fp16, x = var_194_cast); + tensor var_209_to_fp16 = const()[name = tensor("op_209_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10190464)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10714816)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_209_to_fp16, x = var_194_cast); + tensor var_213_to_fp16 = const()[name = tensor("op_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10715904)))]; + tensor var_214_to_fp16 = const()[name = tensor("op_214_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11240256)))]; + tensor v_5_cast = linear(bias = var_214_to_fp16, weight = var_213_to_fp16, x = var_194_cast); + tensor var_222 = const()[name = tensor("op_222"), val = tensor([1, 1500, 8, -1])]; + tensor var_223_cast = reshape(shape = var_222, x = q_5_cast); + tensor const_44_to_fp16 = const()[name = tensor("const_44_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_223_cast, y = const_44_to_fp16); + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 1500, 8, -1])]; + tensor var_230_cast = reshape(shape = var_229, x = k_5_cast); + tensor const_45_to_fp16 = const()[name = tensor("const_45_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_230_cast, y = const_45_to_fp16); + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 1500, 8, -1])]; + tensor var_237_cast = reshape(shape = var_236, x = v_5_cast); + tensor var_238 = const()[name = tensor("op_238"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_41 = transpose(perm = transpose_15_perm_0, x = k_7_cast); + tensor transpose_42 = transpose(perm = transpose_14_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_42, y = transpose_41); + tensor var_242_cast = softmax(axis = var_177, x = qk_3_cast); + tensor var_244_transpose_x_0 = const()[name = tensor("op_244_transpose_x_0"), val = tensor(false)]; + tensor var_244_transpose_y_0 = const()[name = tensor("op_244_transpose_y_0"), val = tensor(false)]; + tensor transpose_43 = transpose(perm = var_238, x = var_237_cast); + tensor var_244_cast = matmul(transpose_x = var_244_transpose_x_0, transpose_y = var_244_transpose_y_0, x = var_242_cast, y = transpose_43); + tensor var_245 = const()[name = tensor("op_245"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 512])]; + tensor transpose_40 = transpose(perm = var_245, x = var_244_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_40); + tensor var_250_to_fp16 = const()[name = tensor("op_250_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11241344)))]; + tensor var_251_to_fp16 = const()[name = tensor("op_251_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11765696)))]; + tensor var_252_cast = linear(bias = var_251_to_fp16, weight = var_250_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_252_cast); + tensor var_258_axes_0 = const()[name = tensor("op_258_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11766784)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11767872)))]; + tensor var_258_cast = layer_norm(axes = var_258_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_183_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11768960)))]; + tensor var_268_to_fp16 = const()[name = tensor("op_268_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13866176)))]; + tensor input_17_cast = linear(bias = var_268_to_fp16, weight = var_267_to_fp16, x = var_258_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_273_to_fp16 = const()[name = tensor("op_273_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13870336)))]; + tensor var_274_to_fp16 = const()[name = tensor("op_274_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15967552)))]; + tensor var_275_cast = linear(bias = var_274_to_fp16, weight = var_273_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_275_cast); + tensor var_284 = const()[name = tensor("op_284"), val = tensor(-1)]; + tensor var_301_axes_0 = const()[name = tensor("op_301_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15968640)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15969728)))]; + tensor var_290_to_fp16 = const()[name = tensor("op_290_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_301_cast = layer_norm(axes = var_301_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_290_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_312_to_fp16 = const()[name = tensor("op_312_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15970816)))]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16495168)))]; + tensor q_9_cast = linear(bias = var_313_to_fp16, weight = var_312_to_fp16, x = var_301_cast); + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16496256)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17020608)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_316_to_fp16, x = var_301_cast); + tensor var_320_to_fp16 = const()[name = tensor("op_320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17021696)))]; + tensor var_321_to_fp16 = const()[name = tensor("op_321_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17546048)))]; + tensor v_9_cast = linear(bias = var_321_to_fp16, weight = var_320_to_fp16, x = var_301_cast); + tensor var_329 = const()[name = tensor("op_329"), val = tensor([1, 1500, 8, -1])]; + tensor var_330_cast = reshape(shape = var_329, x = q_9_cast); + tensor const_46_to_fp16 = const()[name = tensor("const_46_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_330_cast, y = const_46_to_fp16); + tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1500, 8, -1])]; + tensor var_337_cast = reshape(shape = var_336, x = k_9_cast); + tensor const_47_to_fp16 = const()[name = tensor("const_47_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_337_cast, y = const_47_to_fp16); + tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 1500, 8, -1])]; + tensor var_344_cast = reshape(shape = var_343, x = v_9_cast); + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_37 = transpose(perm = transpose_17_perm_0, x = k_11_cast); + tensor transpose_38 = transpose(perm = transpose_16_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_38, y = transpose_37); + tensor var_349_cast = softmax(axis = var_284, x = qk_5_cast); + tensor var_351_transpose_x_0 = const()[name = tensor("op_351_transpose_x_0"), val = tensor(false)]; + tensor var_351_transpose_y_0 = const()[name = tensor("op_351_transpose_y_0"), val = tensor(false)]; + tensor transpose_39 = transpose(perm = var_345, x = var_344_cast); + tensor var_351_cast = matmul(transpose_x = var_351_transpose_x_0, transpose_y = var_351_transpose_y_0, x = var_349_cast, y = transpose_39); + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 512])]; + tensor transpose_36 = transpose(perm = var_352, x = var_351_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_36); + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17547136)))]; + tensor var_358_to_fp16 = const()[name = tensor("op_358_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18071488)))]; + tensor var_359_cast = linear(bias = var_358_to_fp16, weight = var_357_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_359_cast); + tensor var_365_axes_0 = const()[name = tensor("op_365_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18072576)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18073664)))]; + tensor var_365_cast = layer_norm(axes = var_365_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_290_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_374_to_fp16 = const()[name = tensor("op_374_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18074752)))]; + tensor var_375_to_fp16 = const()[name = tensor("op_375_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20171968)))]; + tensor input_25_cast = linear(bias = var_375_to_fp16, weight = var_374_to_fp16, x = var_365_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_380_to_fp16 = const()[name = tensor("op_380_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20176128)))]; + tensor var_381_to_fp16 = const()[name = tensor("op_381_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22273344)))]; + tensor var_382_cast = linear(bias = var_381_to_fp16, weight = var_380_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_382_cast); + tensor var_391 = const()[name = tensor("op_391"), val = tensor(-1)]; + tensor var_408_axes_0 = const()[name = tensor("op_408_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22274432)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22275520)))]; + tensor var_397_to_fp16 = const()[name = tensor("op_397_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_408_cast = layer_norm(axes = var_408_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_397_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_419_to_fp16 = const()[name = tensor("op_419_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22276608)))]; + tensor var_420_to_fp16 = const()[name = tensor("op_420_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22800960)))]; + tensor q_13_cast = linear(bias = var_420_to_fp16, weight = var_419_to_fp16, x = var_408_cast); + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22802048)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23326400)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_423_to_fp16, x = var_408_cast); + tensor var_427_to_fp16 = const()[name = tensor("op_427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23327488)))]; + tensor var_428_to_fp16 = const()[name = tensor("op_428_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23851840)))]; + tensor v_13_cast = linear(bias = var_428_to_fp16, weight = var_427_to_fp16, x = var_408_cast); + tensor var_436 = const()[name = tensor("op_436"), val = tensor([1, 1500, 8, -1])]; + tensor var_437_cast = reshape(shape = var_436, x = q_13_cast); + tensor const_48_to_fp16 = const()[name = tensor("const_48_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_437_cast, y = const_48_to_fp16); + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1500, 8, -1])]; + tensor var_444_cast = reshape(shape = var_443, x = k_13_cast); + tensor const_49_to_fp16 = const()[name = tensor("const_49_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_444_cast, y = const_49_to_fp16); + tensor var_450 = const()[name = tensor("op_450"), val = tensor([1, 1500, 8, -1])]; + tensor var_451_cast = reshape(shape = var_450, x = v_13_cast); + tensor var_452 = const()[name = tensor("op_452"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_33 = transpose(perm = transpose_19_perm_0, x = k_15_cast); + tensor transpose_34 = transpose(perm = transpose_18_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_34, y = transpose_33); + tensor var_456_cast = softmax(axis = var_391, x = qk_7_cast); + tensor var_458_transpose_x_0 = const()[name = tensor("op_458_transpose_x_0"), val = tensor(false)]; + tensor var_458_transpose_y_0 = const()[name = tensor("op_458_transpose_y_0"), val = tensor(false)]; + tensor transpose_35 = transpose(perm = var_452, x = var_451_cast); + tensor var_458_cast = matmul(transpose_x = var_458_transpose_x_0, transpose_y = var_458_transpose_y_0, x = var_456_cast, y = transpose_35); + tensor var_459 = const()[name = tensor("op_459"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 512])]; + tensor transpose_32 = transpose(perm = var_459, x = var_458_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_32); + tensor var_464_to_fp16 = const()[name = tensor("op_464_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23852928)))]; + tensor var_465_to_fp16 = const()[name = tensor("op_465_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24377280)))]; + tensor var_466_cast = linear(bias = var_465_to_fp16, weight = var_464_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_466_cast); + tensor var_472_axes_0 = const()[name = tensor("op_472_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24378368)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24379456)))]; + tensor var_472_cast = layer_norm(axes = var_472_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_397_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_481_to_fp16 = const()[name = tensor("op_481_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24380544)))]; + tensor var_482_to_fp16 = const()[name = tensor("op_482_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26477760)))]; + tensor input_33_cast = linear(bias = var_482_to_fp16, weight = var_481_to_fp16, x = var_472_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_487_to_fp16 = const()[name = tensor("op_487_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26481920)))]; + tensor var_488_to_fp16 = const()[name = tensor("op_488_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28579136)))]; + tensor var_489_cast = linear(bias = var_488_to_fp16, weight = var_487_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_489_cast); + tensor var_498 = const()[name = tensor("op_498"), val = tensor(-1)]; + tensor var_515_axes_0 = const()[name = tensor("op_515_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28580224)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28581312)))]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_515_cast = layer_norm(axes = var_515_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_504_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28582400)))]; + tensor var_527_to_fp16 = const()[name = tensor("op_527_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29106752)))]; + tensor q_17_cast = linear(bias = var_527_to_fp16, weight = var_526_to_fp16, x = var_515_cast); + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29107840)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29632192)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_530_to_fp16, x = var_515_cast); + tensor var_534_to_fp16 = const()[name = tensor("op_534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29633280)))]; + tensor var_535_to_fp16 = const()[name = tensor("op_535_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30157632)))]; + tensor v_17_cast = linear(bias = var_535_to_fp16, weight = var_534_to_fp16, x = var_515_cast); + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1500, 8, -1])]; + tensor var_544_cast = reshape(shape = var_543, x = q_17_cast); + tensor const_50_to_fp16 = const()[name = tensor("const_50_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_544_cast, y = const_50_to_fp16); + tensor var_550 = const()[name = tensor("op_550"), val = tensor([1, 1500, 8, -1])]; + tensor var_551_cast = reshape(shape = var_550, x = k_17_cast); + tensor const_51_to_fp16 = const()[name = tensor("const_51_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_551_cast, y = const_51_to_fp16); + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1500, 8, -1])]; + tensor var_558_cast = reshape(shape = var_557, x = v_17_cast); + tensor var_559 = const()[name = tensor("op_559"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_29 = transpose(perm = transpose_21_perm_0, x = k_19_cast); + tensor transpose_30 = transpose(perm = transpose_20_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_30, y = transpose_29); + tensor var_563_cast = softmax(axis = var_498, x = qk_9_cast); + tensor var_565_transpose_x_0 = const()[name = tensor("op_565_transpose_x_0"), val = tensor(false)]; + tensor var_565_transpose_y_0 = const()[name = tensor("op_565_transpose_y_0"), val = tensor(false)]; + tensor transpose_31 = transpose(perm = var_559, x = var_558_cast); + tensor var_565_cast = matmul(transpose_x = var_565_transpose_x_0, transpose_y = var_565_transpose_y_0, x = var_563_cast, y = transpose_31); + tensor var_566 = const()[name = tensor("op_566"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 512])]; + tensor transpose_28 = transpose(perm = var_566, x = var_565_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_28); + tensor var_571_to_fp16 = const()[name = tensor("op_571_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30158720)))]; + tensor var_572_to_fp16 = const()[name = tensor("op_572_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30683072)))]; + tensor var_573_cast = linear(bias = var_572_to_fp16, weight = var_571_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_573_cast); + tensor var_579_axes_0 = const()[name = tensor("op_579_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30684160)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30685248)))]; + tensor var_579_cast = layer_norm(axes = var_579_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_504_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_588_to_fp16 = const()[name = tensor("op_588_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30686336)))]; + tensor var_589_to_fp16 = const()[name = tensor("op_589_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32783552)))]; + tensor input_41_cast = linear(bias = var_589_to_fp16, weight = var_588_to_fp16, x = var_579_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_594_to_fp16 = const()[name = tensor("op_594_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32787712)))]; + tensor var_595_to_fp16 = const()[name = tensor("op_595_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34884928)))]; + tensor var_596_cast = linear(bias = var_595_to_fp16, weight = var_594_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_596_cast); + tensor var_605 = const()[name = tensor("op_605"), val = tensor(-1)]; + tensor var_622_axes_0 = const()[name = tensor("op_622_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34886016)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34887104)))]; + tensor var_611_to_fp16 = const()[name = tensor("op_611_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_622_cast = layer_norm(axes = var_622_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_611_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_633_to_fp16 = const()[name = tensor("op_633_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34888192)))]; + tensor var_634_to_fp16 = const()[name = tensor("op_634_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35412544)))]; + tensor q_21_cast = linear(bias = var_634_to_fp16, weight = var_633_to_fp16, x = var_622_cast); + tensor var_637_to_fp16 = const()[name = tensor("op_637_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35413632)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35937984)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_637_to_fp16, x = var_622_cast); + tensor var_641_to_fp16 = const()[name = tensor("op_641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35939072)))]; + tensor var_642_to_fp16 = const()[name = tensor("op_642_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36463424)))]; + tensor v_21_cast = linear(bias = var_642_to_fp16, weight = var_641_to_fp16, x = var_622_cast); + tensor var_650 = const()[name = tensor("op_650"), val = tensor([1, 1500, 8, -1])]; + tensor var_651_cast = reshape(shape = var_650, x = q_21_cast); + tensor const_52_to_fp16 = const()[name = tensor("const_52_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_651_cast, y = const_52_to_fp16); + tensor var_657 = const()[name = tensor("op_657"), val = tensor([1, 1500, 8, -1])]; + tensor var_658_cast = reshape(shape = var_657, x = k_21_cast); + tensor const_53_to_fp16 = const()[name = tensor("const_53_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_658_cast, y = const_53_to_fp16); + tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 1500, 8, -1])]; + tensor var_665_cast = reshape(shape = var_664, x = v_21_cast); + tensor var_666 = const()[name = tensor("op_666"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_25 = transpose(perm = transpose_23_perm_0, x = k_cast); + tensor transpose_26 = transpose(perm = transpose_22_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_26, y = transpose_25); + tensor var_670_cast = softmax(axis = var_605, x = qk_cast); + tensor var_672_transpose_x_0 = const()[name = tensor("op_672_transpose_x_0"), val = tensor(false)]; + tensor var_672_transpose_y_0 = const()[name = tensor("op_672_transpose_y_0"), val = tensor(false)]; + tensor transpose_27 = transpose(perm = var_666, x = var_665_cast); + tensor var_672_cast = matmul(transpose_x = var_672_transpose_x_0, transpose_y = var_672_transpose_y_0, x = var_670_cast, y = transpose_27); + tensor var_673 = const()[name = tensor("op_673"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 512])]; + tensor transpose_24 = transpose(perm = var_673, x = var_672_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_24); + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36464512)))]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36988864)))]; + tensor var_680_cast = linear(bias = var_679_to_fp16, weight = var_678_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_680_cast); + tensor var_686_axes_0 = const()[name = tensor("op_686_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36989952)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36991040)))]; + tensor var_686_cast = layer_norm(axes = var_686_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_611_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_695_to_fp16 = const()[name = tensor("op_695_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36992128)))]; + tensor var_696_to_fp16 = const()[name = tensor("op_696_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39089344)))]; + tensor input_49_cast = linear(bias = var_696_to_fp16, weight = var_695_to_fp16, x = var_686_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39093504)))]; + tensor var_702_to_fp16 = const()[name = tensor("op_702_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190720)))]; + tensor var_703_cast = linear(bias = var_702_to_fp16, weight = var_701_to_fp16, x = x_77_cast); + tensor x_cast = add(x = x_73_cast, y = var_703_cast); + tensor var_716_axes_0 = const()[name = tensor("op_716_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41191808)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41192896)))]; + tensor var_707_to_fp16 = const()[name = tensor("op_707_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_716_cast = layer_norm(axes = var_716_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_707_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_716_cast_to_fp32_dtype_0 = const()[name = tensor("op_716_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_716_cast_to_fp32_dtype_0, x = var_716_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..5a4e6dc2f9e2a8593c315ff37e6c4c2c548dcdcc --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-base.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1578c2b8925c0ff867a8bb96dd56c0eb75df3b909ca9507dc2aa1e437c296dc9 +size 41193984 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..d87922edf7affa104ef6e9e094c17ade37bcacc4 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f44eecece0d8b55dbd8c82e5d7815af613273c24ed7600e7f8fc78a1b0265ea4 +size 1145487954 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a868bd81bfcf766ee3c1023614854d6b682b9351 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5aec358d66255e1eecb43c5df3acc31815bed1b5d23ec4caa89cb7dfe1b89cc4 +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..fb5c7d72f8a8bfd9f03d6e62a52303c9c41e1ed7 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 192, + "Matmul" : 64, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 32, + "Add" : 65, + "LayerNorm" : 65, + "Mul" : 64, + "Transpose" : 129, + "Gelu" : 34, + "Reshape" : 128 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..66fd37b448b28df30396f27bdd155e6975fc1fb9 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/model.mil @@ -0,0 +1,1927 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor([1])]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614528)))]; + tensor cast_967 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_84_cast = conv(bias = bias_3_to_fp16, dilations = var_82, groups = var_72, pad = var_84_pad_0, pad_type = var_84_pad_type_0, strides = var_80, weight = weight_3_to_fp16, x = cast_967); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_84_cast); + tensor var_88 = const()[name = tensor("op_88"), val = tensor(1)]; + tensor var_97 = const()[name = tensor("op_97"), val = tensor([2])]; + tensor var_99 = const()[name = tensor("op_99"), val = tensor([1])]; + tensor var_101_pad_type_0 = const()[name = tensor("op_101_pad_type_0"), val = tensor("custom")]; + tensor var_101_pad_0 = const()[name = tensor("op_101_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617152)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10447616)))]; + tensor var_101_cast = conv(bias = bias_7_to_fp16, dilations = var_99, groups = var_88, pad = var_101_pad_0, pad_type = var_101_pad_type_0, strides = var_97, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_101_cast); + tensor var_106 = const()[name = tensor("op_106"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10450240)))]; + tensor transpose_256 = transpose(perm = var_106, x = x_3_cast); + tensor var_109_cast = add(x = transpose_256, y = positional_embedding_to_fp16); + tensor var_122 = const()[name = tensor("op_122"), val = tensor(-1)]; + tensor var_139_axes_0 = const()[name = tensor("op_139_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14290304)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14292928)))]; + tensor var_128_to_fp16 = const()[name = tensor("op_128_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_139_cast = layer_norm(axes = var_139_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_128_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_109_cast); + tensor var_150_to_fp16 = const()[name = tensor("op_150_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14295552)))]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17572416)))]; + tensor q_1_cast = linear(bias = var_151_to_fp16, weight = var_150_to_fp16, x = var_139_cast); + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17575040)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20851904)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_154_to_fp16, x = var_139_cast); + tensor var_158_to_fp16 = const()[name = tensor("op_158_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20854528)))]; + tensor var_159_to_fp16 = const()[name = tensor("op_159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24131392)))]; + tensor v_1_cast = linear(bias = var_159_to_fp16, weight = var_158_to_fp16, x = var_139_cast); + tensor var_167 = const()[name = tensor("op_167"), val = tensor([1, 1500, 20, -1])]; + tensor var_168_cast = reshape(shape = var_167, x = q_1_cast); + tensor const_224_to_fp16 = const()[name = tensor("const_224_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_168_cast, y = const_224_to_fp16); + tensor var_174 = const()[name = tensor("op_174"), val = tensor([1, 1500, 20, -1])]; + tensor var_175_cast = reshape(shape = var_174, x = k_1_cast); + tensor const_225_to_fp16 = const()[name = tensor("const_225_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_175_cast, y = const_225_to_fp16); + tensor var_181 = const()[name = tensor("op_181"), val = tensor([1, 1500, 20, -1])]; + tensor var_182_cast = reshape(shape = var_181, x = v_1_cast); + tensor var_183 = const()[name = tensor("op_183"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_64_perm_0 = const()[name = tensor("transpose_64_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_65_perm_0 = const()[name = tensor("transpose_65_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_253 = transpose(perm = transpose_65_perm_0, x = k_3_cast); + tensor transpose_254 = transpose(perm = transpose_64_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_254, y = transpose_253); + tensor var_187_cast = softmax(axis = var_122, x = qk_1_cast); + tensor var_189_transpose_x_0 = const()[name = tensor("op_189_transpose_x_0"), val = tensor(false)]; + tensor var_189_transpose_y_0 = const()[name = tensor("op_189_transpose_y_0"), val = tensor(false)]; + tensor transpose_255 = transpose(perm = var_183, x = var_182_cast); + tensor var_189_cast = matmul(transpose_x = var_189_transpose_x_0, transpose_y = var_189_transpose_y_0, x = var_187_cast, y = transpose_255); + tensor var_190 = const()[name = tensor("op_190"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 1280])]; + tensor transpose_252 = transpose(perm = var_190, x = var_189_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_252); + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24134016)))]; + tensor var_196_to_fp16 = const()[name = tensor("op_196_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27410880)))]; + tensor var_197_cast = linear(bias = var_196_to_fp16, weight = var_195_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_109_cast, y = var_197_cast); + tensor var_203_axes_0 = const()[name = tensor("op_203_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27413504)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27416128)))]; + tensor var_203_cast = layer_norm(axes = var_203_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_128_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_212_to_fp16 = const()[name = tensor("op_212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27418752)))]; + tensor var_213_to_fp16 = const()[name = tensor("op_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40526016)))]; + tensor input_9_cast = linear(bias = var_213_to_fp16, weight = var_212_to_fp16, x = var_203_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_218_to_fp16 = const()[name = tensor("op_218_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40536320)))]; + tensor var_219_to_fp16 = const()[name = tensor("op_219_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53643584)))]; + tensor var_220_cast = linear(bias = var_219_to_fp16, weight = var_218_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_220_cast); + tensor var_229 = const()[name = tensor("op_229"), val = tensor(-1)]; + tensor var_246_axes_0 = const()[name = tensor("op_246_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53646208)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53648832)))]; + tensor var_235_to_fp16 = const()[name = tensor("op_235_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_246_cast = layer_norm(axes = var_246_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_235_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_257_to_fp16 = const()[name = tensor("op_257_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53651456)))]; + tensor var_258_to_fp16 = const()[name = tensor("op_258_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56928320)))]; + tensor q_5_cast = linear(bias = var_258_to_fp16, weight = var_257_to_fp16, x = var_246_cast); + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56930944)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60207808)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_261_to_fp16, x = var_246_cast); + tensor var_265_to_fp16 = const()[name = tensor("op_265_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60210432)))]; + tensor var_266_to_fp16 = const()[name = tensor("op_266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63487296)))]; + tensor v_5_cast = linear(bias = var_266_to_fp16, weight = var_265_to_fp16, x = var_246_cast); + tensor var_274 = const()[name = tensor("op_274"), val = tensor([1, 1500, 20, -1])]; + tensor var_275_cast = reshape(shape = var_274, x = q_5_cast); + tensor const_226_to_fp16 = const()[name = tensor("const_226_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_275_cast, y = const_226_to_fp16); + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1500, 20, -1])]; + tensor var_282_cast = reshape(shape = var_281, x = k_5_cast); + tensor const_227_to_fp16 = const()[name = tensor("const_227_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_282_cast, y = const_227_to_fp16); + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1500, 20, -1])]; + tensor var_289_cast = reshape(shape = var_288, x = v_5_cast); + tensor var_290 = const()[name = tensor("op_290"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_66_perm_0 = const()[name = tensor("transpose_66_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_67_perm_0 = const()[name = tensor("transpose_67_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_249 = transpose(perm = transpose_67_perm_0, x = k_7_cast); + tensor transpose_250 = transpose(perm = transpose_66_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_250, y = transpose_249); + tensor var_294_cast = softmax(axis = var_229, x = qk_3_cast); + tensor var_296_transpose_x_0 = const()[name = tensor("op_296_transpose_x_0"), val = tensor(false)]; + tensor var_296_transpose_y_0 = const()[name = tensor("op_296_transpose_y_0"), val = tensor(false)]; + tensor transpose_251 = transpose(perm = var_290, x = var_289_cast); + tensor var_296_cast = matmul(transpose_x = var_296_transpose_x_0, transpose_y = var_296_transpose_y_0, x = var_294_cast, y = transpose_251); + tensor var_297 = const()[name = tensor("op_297"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 1280])]; + tensor transpose_248 = transpose(perm = var_297, x = var_296_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_248); + tensor var_302_to_fp16 = const()[name = tensor("op_302_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63489920)))]; + tensor var_303_to_fp16 = const()[name = tensor("op_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66766784)))]; + tensor var_304_cast = linear(bias = var_303_to_fp16, weight = var_302_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_304_cast); + tensor var_310_axes_0 = const()[name = tensor("op_310_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66769408)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66772032)))]; + tensor var_310_cast = layer_norm(axes = var_310_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_235_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66774656)))]; + tensor var_320_to_fp16 = const()[name = tensor("op_320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79881920)))]; + tensor input_17_cast = linear(bias = var_320_to_fp16, weight = var_319_to_fp16, x = var_310_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_325_to_fp16 = const()[name = tensor("op_325_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79892224)))]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92999488)))]; + tensor var_327_cast = linear(bias = var_326_to_fp16, weight = var_325_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_327_cast); + tensor var_336 = const()[name = tensor("op_336"), val = tensor(-1)]; + tensor var_353_axes_0 = const()[name = tensor("op_353_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93002112)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93004736)))]; + tensor var_342_to_fp16 = const()[name = tensor("op_342_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_353_cast = layer_norm(axes = var_353_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_342_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_364_to_fp16 = const()[name = tensor("op_364_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93007360)))]; + tensor var_365_to_fp16 = const()[name = tensor("op_365_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96284224)))]; + tensor q_9_cast = linear(bias = var_365_to_fp16, weight = var_364_to_fp16, x = var_353_cast); + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96286848)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99563712)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_368_to_fp16, x = var_353_cast); + tensor var_372_to_fp16 = const()[name = tensor("op_372_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99566336)))]; + tensor var_373_to_fp16 = const()[name = tensor("op_373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102843200)))]; + tensor v_9_cast = linear(bias = var_373_to_fp16, weight = var_372_to_fp16, x = var_353_cast); + tensor var_381 = const()[name = tensor("op_381"), val = tensor([1, 1500, 20, -1])]; + tensor var_382_cast = reshape(shape = var_381, x = q_9_cast); + tensor const_228_to_fp16 = const()[name = tensor("const_228_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_382_cast, y = const_228_to_fp16); + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1500, 20, -1])]; + tensor var_389_cast = reshape(shape = var_388, x = k_9_cast); + tensor const_229_to_fp16 = const()[name = tensor("const_229_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_389_cast, y = const_229_to_fp16); + tensor var_395 = const()[name = tensor("op_395"), val = tensor([1, 1500, 20, -1])]; + tensor var_396_cast = reshape(shape = var_395, x = v_9_cast); + tensor var_397 = const()[name = tensor("op_397"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_68_perm_0 = const()[name = tensor("transpose_68_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_245 = transpose(perm = transpose_69_perm_0, x = k_11_cast); + tensor transpose_246 = transpose(perm = transpose_68_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_246, y = transpose_245); + tensor var_401_cast = softmax(axis = var_336, x = qk_5_cast); + tensor var_403_transpose_x_0 = const()[name = tensor("op_403_transpose_x_0"), val = tensor(false)]; + tensor var_403_transpose_y_0 = const()[name = tensor("op_403_transpose_y_0"), val = tensor(false)]; + tensor transpose_247 = transpose(perm = var_397, x = var_396_cast); + tensor var_403_cast = matmul(transpose_x = var_403_transpose_x_0, transpose_y = var_403_transpose_y_0, x = var_401_cast, y = transpose_247); + tensor var_404 = const()[name = tensor("op_404"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 1280])]; + tensor transpose_244 = transpose(perm = var_404, x = var_403_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_244); + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102845824)))]; + tensor var_410_to_fp16 = const()[name = tensor("op_410_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106122688)))]; + tensor var_411_cast = linear(bias = var_410_to_fp16, weight = var_409_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_411_cast); + tensor var_417_axes_0 = const()[name = tensor("op_417_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106125312)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106127936)))]; + tensor var_417_cast = layer_norm(axes = var_417_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_342_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106130560)))]; + tensor var_427_to_fp16 = const()[name = tensor("op_427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119237824)))]; + tensor input_25_cast = linear(bias = var_427_to_fp16, weight = var_426_to_fp16, x = var_417_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_432_to_fp16 = const()[name = tensor("op_432_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119248128)))]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132355392)))]; + tensor var_434_cast = linear(bias = var_433_to_fp16, weight = var_432_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_434_cast); + tensor var_443 = const()[name = tensor("op_443"), val = tensor(-1)]; + tensor var_460_axes_0 = const()[name = tensor("op_460_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132358016)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132360640)))]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_460_cast = layer_norm(axes = var_460_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_449_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132363264)))]; + tensor var_472_to_fp16 = const()[name = tensor("op_472_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135640128)))]; + tensor q_13_cast = linear(bias = var_472_to_fp16, weight = var_471_to_fp16, x = var_460_cast); + tensor var_475_to_fp16 = const()[name = tensor("op_475_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135642752)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138919616)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_475_to_fp16, x = var_460_cast); + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138922240)))]; + tensor var_480_to_fp16 = const()[name = tensor("op_480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142199104)))]; + tensor v_13_cast = linear(bias = var_480_to_fp16, weight = var_479_to_fp16, x = var_460_cast); + tensor var_488 = const()[name = tensor("op_488"), val = tensor([1, 1500, 20, -1])]; + tensor var_489_cast = reshape(shape = var_488, x = q_13_cast); + tensor const_230_to_fp16 = const()[name = tensor("const_230_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_489_cast, y = const_230_to_fp16); + tensor var_495 = const()[name = tensor("op_495"), val = tensor([1, 1500, 20, -1])]; + tensor var_496_cast = reshape(shape = var_495, x = k_13_cast); + tensor const_231_to_fp16 = const()[name = tensor("const_231_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_496_cast, y = const_231_to_fp16); + tensor var_502 = const()[name = tensor("op_502"), val = tensor([1, 1500, 20, -1])]; + tensor var_503_cast = reshape(shape = var_502, x = v_13_cast); + tensor var_504 = const()[name = tensor("op_504"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_241 = transpose(perm = transpose_71_perm_0, x = k_15_cast); + tensor transpose_242 = transpose(perm = transpose_70_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_242, y = transpose_241); + tensor var_508_cast = softmax(axis = var_443, x = qk_7_cast); + tensor var_510_transpose_x_0 = const()[name = tensor("op_510_transpose_x_0"), val = tensor(false)]; + tensor var_510_transpose_y_0 = const()[name = tensor("op_510_transpose_y_0"), val = tensor(false)]; + tensor transpose_243 = transpose(perm = var_504, x = var_503_cast); + tensor var_510_cast = matmul(transpose_x = var_510_transpose_x_0, transpose_y = var_510_transpose_y_0, x = var_508_cast, y = transpose_243); + tensor var_511 = const()[name = tensor("op_511"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 1280])]; + tensor transpose_240 = transpose(perm = var_511, x = var_510_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_240); + tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142201728)))]; + tensor var_517_to_fp16 = const()[name = tensor("op_517_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145478592)))]; + tensor var_518_cast = linear(bias = var_517_to_fp16, weight = var_516_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_518_cast); + tensor var_524_axes_0 = const()[name = tensor("op_524_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145481216)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145483840)))]; + tensor var_524_cast = layer_norm(axes = var_524_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_449_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_533_to_fp16 = const()[name = tensor("op_533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145486464)))]; + tensor var_534_to_fp16 = const()[name = tensor("op_534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158593728)))]; + tensor input_33_cast = linear(bias = var_534_to_fp16, weight = var_533_to_fp16, x = var_524_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_539_to_fp16 = const()[name = tensor("op_539_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158604032)))]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171711296)))]; + tensor var_541_cast = linear(bias = var_540_to_fp16, weight = var_539_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_541_cast); + tensor var_550 = const()[name = tensor("op_550"), val = tensor(-1)]; + tensor var_567_axes_0 = const()[name = tensor("op_567_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171713920)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171716544)))]; + tensor var_556_to_fp16 = const()[name = tensor("op_556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_567_cast = layer_norm(axes = var_567_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_556_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_578_to_fp16 = const()[name = tensor("op_578_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171719168)))]; + tensor var_579_to_fp16 = const()[name = tensor("op_579_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174996032)))]; + tensor q_17_cast = linear(bias = var_579_to_fp16, weight = var_578_to_fp16, x = var_567_cast); + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174998656)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178275520)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_582_to_fp16, x = var_567_cast); + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178278144)))]; + tensor var_587_to_fp16 = const()[name = tensor("op_587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181555008)))]; + tensor v_17_cast = linear(bias = var_587_to_fp16, weight = var_586_to_fp16, x = var_567_cast); + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1500, 20, -1])]; + tensor var_596_cast = reshape(shape = var_595, x = q_17_cast); + tensor const_232_to_fp16 = const()[name = tensor("const_232_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_596_cast, y = const_232_to_fp16); + tensor var_602 = const()[name = tensor("op_602"), val = tensor([1, 1500, 20, -1])]; + tensor var_603_cast = reshape(shape = var_602, x = k_17_cast); + tensor const_233_to_fp16 = const()[name = tensor("const_233_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_603_cast, y = const_233_to_fp16); + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1500, 20, -1])]; + tensor var_610_cast = reshape(shape = var_609, x = v_17_cast); + tensor var_611 = const()[name = tensor("op_611"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_237 = transpose(perm = transpose_73_perm_0, x = k_19_cast); + tensor transpose_238 = transpose(perm = transpose_72_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_238, y = transpose_237); + tensor var_615_cast = softmax(axis = var_550, x = qk_9_cast); + tensor var_617_transpose_x_0 = const()[name = tensor("op_617_transpose_x_0"), val = tensor(false)]; + tensor var_617_transpose_y_0 = const()[name = tensor("op_617_transpose_y_0"), val = tensor(false)]; + tensor transpose_239 = transpose(perm = var_611, x = var_610_cast); + tensor var_617_cast = matmul(transpose_x = var_617_transpose_x_0, transpose_y = var_617_transpose_y_0, x = var_615_cast, y = transpose_239); + tensor var_618 = const()[name = tensor("op_618"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 1280])]; + tensor transpose_236 = transpose(perm = var_618, x = var_617_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_236); + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181557632)))]; + tensor var_624_to_fp16 = const()[name = tensor("op_624_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184834496)))]; + tensor var_625_cast = linear(bias = var_624_to_fp16, weight = var_623_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_625_cast); + tensor var_631_axes_0 = const()[name = tensor("op_631_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184837120)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839744)))]; + tensor var_631_cast = layer_norm(axes = var_631_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_556_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_640_to_fp16 = const()[name = tensor("op_640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184842368)))]; + tensor var_641_to_fp16 = const()[name = tensor("op_641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197949632)))]; + tensor input_41_cast = linear(bias = var_641_to_fp16, weight = var_640_to_fp16, x = var_631_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_646_to_fp16 = const()[name = tensor("op_646_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197959936)))]; + tensor var_647_to_fp16 = const()[name = tensor("op_647_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211067200)))]; + tensor var_648_cast = linear(bias = var_647_to_fp16, weight = var_646_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_648_cast); + tensor var_657 = const()[name = tensor("op_657"), val = tensor(-1)]; + tensor var_674_axes_0 = const()[name = tensor("op_674_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211069824)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211072448)))]; + tensor var_663_to_fp16 = const()[name = tensor("op_663_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_674_cast = layer_norm(axes = var_674_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_663_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_685_to_fp16 = const()[name = tensor("op_685_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211075072)))]; + tensor var_686_to_fp16 = const()[name = tensor("op_686_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214351936)))]; + tensor q_21_cast = linear(bias = var_686_to_fp16, weight = var_685_to_fp16, x = var_674_cast); + tensor var_689_to_fp16 = const()[name = tensor("op_689_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214354560)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217631424)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_689_to_fp16, x = var_674_cast); + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217634048)))]; + tensor var_694_to_fp16 = const()[name = tensor("op_694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220910912)))]; + tensor v_21_cast = linear(bias = var_694_to_fp16, weight = var_693_to_fp16, x = var_674_cast); + tensor var_702 = const()[name = tensor("op_702"), val = tensor([1, 1500, 20, -1])]; + tensor var_703_cast = reshape(shape = var_702, x = q_21_cast); + tensor const_234_to_fp16 = const()[name = tensor("const_234_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast = mul(x = var_703_cast, y = const_234_to_fp16); + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1500, 20, -1])]; + tensor var_710_cast = reshape(shape = var_709, x = k_21_cast); + tensor const_235_to_fp16 = const()[name = tensor("const_235_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast = mul(x = var_710_cast, y = const_235_to_fp16); + tensor var_716 = const()[name = tensor("op_716"), val = tensor([1, 1500, 20, -1])]; + tensor var_717_cast = reshape(shape = var_716, x = v_21_cast); + tensor var_718 = const()[name = tensor("op_718"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_233 = transpose(perm = transpose_75_perm_0, x = k_23_cast); + tensor transpose_234 = transpose(perm = transpose_74_perm_0, x = q_23_cast); + tensor qk_11_cast = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_234, y = transpose_233); + tensor var_722_cast = softmax(axis = var_657, x = qk_11_cast); + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_235 = transpose(perm = var_718, x = var_717_cast); + tensor var_724_cast = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = var_722_cast, y = transpose_235); + tensor var_725 = const()[name = tensor("op_725"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 1280])]; + tensor transpose_232 = transpose(perm = var_725, x = var_724_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_232); + tensor var_730_to_fp16 = const()[name = tensor("op_730_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220913536)))]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224190400)))]; + tensor var_732_cast = linear(bias = var_731_to_fp16, weight = var_730_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_732_cast); + tensor var_738_axes_0 = const()[name = tensor("op_738_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224193024)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224195648)))]; + tensor var_738_cast = layer_norm(axes = var_738_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_663_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_747_to_fp16 = const()[name = tensor("op_747_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224198272)))]; + tensor var_748_to_fp16 = const()[name = tensor("op_748_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237305536)))]; + tensor input_49_cast = linear(bias = var_748_to_fp16, weight = var_747_to_fp16, x = var_738_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_753_to_fp16 = const()[name = tensor("op_753_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237315840)))]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250423104)))]; + tensor var_755_cast = linear(bias = var_754_to_fp16, weight = var_753_to_fp16, x = x_77_cast); + tensor x_79_cast = add(x = x_73_cast, y = var_755_cast); + tensor var_764 = const()[name = tensor("op_764"), val = tensor(-1)]; + tensor var_781_axes_0 = const()[name = tensor("op_781_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250425728)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250428352)))]; + tensor var_770_to_fp16 = const()[name = tensor("op_770_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_781_cast = layer_norm(axes = var_781_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_770_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast); + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250430976)))]; + tensor var_793_to_fp16 = const()[name = tensor("op_793_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253707840)))]; + tensor q_25_cast = linear(bias = var_793_to_fp16, weight = var_792_to_fp16, x = var_781_cast); + tensor var_796_to_fp16 = const()[name = tensor("op_796_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253710464)))]; + tensor k_25_bias_0_to_fp16 = const()[name = tensor("k_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256987328)))]; + tensor k_25_cast = linear(bias = k_25_bias_0_to_fp16, weight = var_796_to_fp16, x = var_781_cast); + tensor var_800_to_fp16 = const()[name = tensor("op_800_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256989952)))]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260266816)))]; + tensor v_25_cast = linear(bias = var_801_to_fp16, weight = var_800_to_fp16, x = var_781_cast); + tensor var_809 = const()[name = tensor("op_809"), val = tensor([1, 1500, 20, -1])]; + tensor var_810_cast = reshape(shape = var_809, x = q_25_cast); + tensor const_236_to_fp16 = const()[name = tensor("const_236_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast = mul(x = var_810_cast, y = const_236_to_fp16); + tensor var_816 = const()[name = tensor("op_816"), val = tensor([1, 1500, 20, -1])]; + tensor var_817_cast = reshape(shape = var_816, x = k_25_cast); + tensor const_237_to_fp16 = const()[name = tensor("const_237_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast = mul(x = var_817_cast, y = const_237_to_fp16); + tensor var_823 = const()[name = tensor("op_823"), val = tensor([1, 1500, 20, -1])]; + tensor var_824_cast = reshape(shape = var_823, x = v_25_cast); + tensor var_825 = const()[name = tensor("op_825"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_229 = transpose(perm = transpose_77_perm_0, x = k_27_cast); + tensor transpose_230 = transpose(perm = transpose_76_perm_0, x = q_27_cast); + tensor qk_13_cast = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_230, y = transpose_229); + tensor var_829_cast = softmax(axis = var_764, x = qk_13_cast); + tensor var_831_transpose_x_0 = const()[name = tensor("op_831_transpose_x_0"), val = tensor(false)]; + tensor var_831_transpose_y_0 = const()[name = tensor("op_831_transpose_y_0"), val = tensor(false)]; + tensor transpose_231 = transpose(perm = var_825, x = var_824_cast); + tensor var_831_cast = matmul(transpose_x = var_831_transpose_x_0, transpose_y = var_831_transpose_y_0, x = var_829_cast, y = transpose_231); + tensor var_832 = const()[name = tensor("op_832"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 1280])]; + tensor transpose_228 = transpose(perm = var_832, x = var_831_cast); + tensor x_83_cast = reshape(shape = concat_6, x = transpose_228); + tensor var_837_to_fp16 = const()[name = tensor("op_837_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260269440)))]; + tensor var_838_to_fp16 = const()[name = tensor("op_838_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263546304)))]; + tensor var_839_cast = linear(bias = var_838_to_fp16, weight = var_837_to_fp16, x = x_83_cast); + tensor x_85_cast = add(x = x_79_cast, y = var_839_cast); + tensor var_845_axes_0 = const()[name = tensor("op_845_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263548928)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263551552)))]; + tensor var_845_cast = layer_norm(axes = var_845_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_770_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast); + tensor var_854_to_fp16 = const()[name = tensor("op_854_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263554176)))]; + tensor var_855_to_fp16 = const()[name = tensor("op_855_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276661440)))]; + tensor input_57_cast = linear(bias = var_855_to_fp16, weight = var_854_to_fp16, x = var_845_cast); + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast = gelu(mode = x_89_mode_0, x = input_57_cast); + tensor var_860_to_fp16 = const()[name = tensor("op_860_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276671744)))]; + tensor var_861_to_fp16 = const()[name = tensor("op_861_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289779008)))]; + tensor var_862_cast = linear(bias = var_861_to_fp16, weight = var_860_to_fp16, x = x_89_cast); + tensor x_91_cast = add(x = x_85_cast, y = var_862_cast); + tensor var_871 = const()[name = tensor("op_871"), val = tensor(-1)]; + tensor var_888_axes_0 = const()[name = tensor("op_888_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289781632)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289784256)))]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_888_cast = layer_norm(axes = var_888_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_877_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast); + tensor var_899_to_fp16 = const()[name = tensor("op_899_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289786880)))]; + tensor var_900_to_fp16 = const()[name = tensor("op_900_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293063744)))]; + tensor q_29_cast = linear(bias = var_900_to_fp16, weight = var_899_to_fp16, x = var_888_cast); + tensor var_903_to_fp16 = const()[name = tensor("op_903_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293066368)))]; + tensor k_29_bias_0_to_fp16 = const()[name = tensor("k_29_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296343232)))]; + tensor k_29_cast = linear(bias = k_29_bias_0_to_fp16, weight = var_903_to_fp16, x = var_888_cast); + tensor var_907_to_fp16 = const()[name = tensor("op_907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296345856)))]; + tensor var_908_to_fp16 = const()[name = tensor("op_908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299622720)))]; + tensor v_29_cast = linear(bias = var_908_to_fp16, weight = var_907_to_fp16, x = var_888_cast); + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1500, 20, -1])]; + tensor var_917_cast = reshape(shape = var_916, x = q_29_cast); + tensor const_238_to_fp16 = const()[name = tensor("const_238_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast = mul(x = var_917_cast, y = const_238_to_fp16); + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1500, 20, -1])]; + tensor var_924_cast = reshape(shape = var_923, x = k_29_cast); + tensor const_239_to_fp16 = const()[name = tensor("const_239_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast = mul(x = var_924_cast, y = const_239_to_fp16); + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 1500, 20, -1])]; + tensor var_931_cast = reshape(shape = var_930, x = v_29_cast); + tensor var_932 = const()[name = tensor("op_932"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_225 = transpose(perm = transpose_79_perm_0, x = k_31_cast); + tensor transpose_226 = transpose(perm = transpose_78_perm_0, x = q_31_cast); + tensor qk_15_cast = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_226, y = transpose_225); + tensor var_936_cast = softmax(axis = var_871, x = qk_15_cast); + tensor var_938_transpose_x_0 = const()[name = tensor("op_938_transpose_x_0"), val = tensor(false)]; + tensor var_938_transpose_y_0 = const()[name = tensor("op_938_transpose_y_0"), val = tensor(false)]; + tensor transpose_227 = transpose(perm = var_932, x = var_931_cast); + tensor var_938_cast = matmul(transpose_x = var_938_transpose_x_0, transpose_y = var_938_transpose_y_0, x = var_936_cast, y = transpose_227); + tensor var_939 = const()[name = tensor("op_939"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 1280])]; + tensor transpose_224 = transpose(perm = var_939, x = var_938_cast); + tensor x_95_cast = reshape(shape = concat_7, x = transpose_224); + tensor var_944_to_fp16 = const()[name = tensor("op_944_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299625344)))]; + tensor var_945_to_fp16 = const()[name = tensor("op_945_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302902208)))]; + tensor var_946_cast = linear(bias = var_945_to_fp16, weight = var_944_to_fp16, x = x_95_cast); + tensor x_97_cast = add(x = x_91_cast, y = var_946_cast); + tensor var_952_axes_0 = const()[name = tensor("op_952_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302904832)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302907456)))]; + tensor var_952_cast = layer_norm(axes = var_952_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_877_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast); + tensor var_961_to_fp16 = const()[name = tensor("op_961_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302910080)))]; + tensor var_962_to_fp16 = const()[name = tensor("op_962_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316017344)))]; + tensor input_65_cast = linear(bias = var_962_to_fp16, weight = var_961_to_fp16, x = var_952_cast); + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast = gelu(mode = x_101_mode_0, x = input_65_cast); + tensor var_967_to_fp16 = const()[name = tensor("op_967_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316027648)))]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329134912)))]; + tensor var_969_cast = linear(bias = var_968_to_fp16, weight = var_967_to_fp16, x = x_101_cast); + tensor x_103_cast = add(x = x_97_cast, y = var_969_cast); + tensor var_978 = const()[name = tensor("op_978"), val = tensor(-1)]; + tensor var_995_axes_0 = const()[name = tensor("op_995_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329137536)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329140160)))]; + tensor var_984_to_fp16 = const()[name = tensor("op_984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_995_cast = layer_norm(axes = var_995_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_984_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast); + tensor var_1006_to_fp16 = const()[name = tensor("op_1006_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329142784)))]; + tensor var_1007_to_fp16 = const()[name = tensor("op_1007_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332419648)))]; + tensor q_33_cast = linear(bias = var_1007_to_fp16, weight = var_1006_to_fp16, x = var_995_cast); + tensor var_1010_to_fp16 = const()[name = tensor("op_1010_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332422272)))]; + tensor k_33_bias_0_to_fp16 = const()[name = tensor("k_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335699136)))]; + tensor k_33_cast = linear(bias = k_33_bias_0_to_fp16, weight = var_1010_to_fp16, x = var_995_cast); + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335701760)))]; + tensor var_1015_to_fp16 = const()[name = tensor("op_1015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338978624)))]; + tensor v_33_cast = linear(bias = var_1015_to_fp16, weight = var_1014_to_fp16, x = var_995_cast); + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 1500, 20, -1])]; + tensor var_1024_cast = reshape(shape = var_1023, x = q_33_cast); + tensor const_240_to_fp16 = const()[name = tensor("const_240_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast = mul(x = var_1024_cast, y = const_240_to_fp16); + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([1, 1500, 20, -1])]; + tensor var_1031_cast = reshape(shape = var_1030, x = k_33_cast); + tensor const_241_to_fp16 = const()[name = tensor("const_241_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast = mul(x = var_1031_cast, y = const_241_to_fp16); + tensor var_1037 = const()[name = tensor("op_1037"), val = tensor([1, 1500, 20, -1])]; + tensor var_1038_cast = reshape(shape = var_1037, x = v_33_cast); + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_221 = transpose(perm = transpose_81_perm_0, x = k_35_cast); + tensor transpose_222 = transpose(perm = transpose_80_perm_0, x = q_35_cast); + tensor qk_17_cast = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_222, y = transpose_221); + tensor var_1043_cast = softmax(axis = var_978, x = qk_17_cast); + tensor var_1045_transpose_x_0 = const()[name = tensor("op_1045_transpose_x_0"), val = tensor(false)]; + tensor var_1045_transpose_y_0 = const()[name = tensor("op_1045_transpose_y_0"), val = tensor(false)]; + tensor transpose_223 = transpose(perm = var_1039, x = var_1038_cast); + tensor var_1045_cast = matmul(transpose_x = var_1045_transpose_x_0, transpose_y = var_1045_transpose_y_0, x = var_1043_cast, y = transpose_223); + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 1280])]; + tensor transpose_220 = transpose(perm = var_1046, x = var_1045_cast); + tensor x_107_cast = reshape(shape = concat_8, x = transpose_220); + tensor var_1051_to_fp16 = const()[name = tensor("op_1051_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338981248)))]; + tensor var_1052_to_fp16 = const()[name = tensor("op_1052_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342258112)))]; + tensor var_1053_cast = linear(bias = var_1052_to_fp16, weight = var_1051_to_fp16, x = x_107_cast); + tensor x_109_cast = add(x = x_103_cast, y = var_1053_cast); + tensor var_1059_axes_0 = const()[name = tensor("op_1059_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342260736)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342263360)))]; + tensor var_1059_cast = layer_norm(axes = var_1059_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_984_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast); + tensor var_1068_to_fp16 = const()[name = tensor("op_1068_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342265984)))]; + tensor var_1069_to_fp16 = const()[name = tensor("op_1069_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355373248)))]; + tensor input_73_cast = linear(bias = var_1069_to_fp16, weight = var_1068_to_fp16, x = var_1059_cast); + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast = gelu(mode = x_113_mode_0, x = input_73_cast); + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355383552)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368490816)))]; + tensor var_1076_cast = linear(bias = var_1075_to_fp16, weight = var_1074_to_fp16, x = x_113_cast); + tensor x_115_cast = add(x = x_109_cast, y = var_1076_cast); + tensor var_1085 = const()[name = tensor("op_1085"), val = tensor(-1)]; + tensor var_1102_axes_0 = const()[name = tensor("op_1102_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368493440)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368496064)))]; + tensor var_1091_to_fp16 = const()[name = tensor("op_1091_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1102_cast = layer_norm(axes = var_1102_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1091_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast); + tensor var_1113_to_fp16 = const()[name = tensor("op_1113_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368498688)))]; + tensor var_1114_to_fp16 = const()[name = tensor("op_1114_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371775552)))]; + tensor q_37_cast = linear(bias = var_1114_to_fp16, weight = var_1113_to_fp16, x = var_1102_cast); + tensor var_1117_to_fp16 = const()[name = tensor("op_1117_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371778176)))]; + tensor k_37_bias_0_to_fp16 = const()[name = tensor("k_37_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375055040)))]; + tensor k_37_cast = linear(bias = k_37_bias_0_to_fp16, weight = var_1117_to_fp16, x = var_1102_cast); + tensor var_1121_to_fp16 = const()[name = tensor("op_1121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375057664)))]; + tensor var_1122_to_fp16 = const()[name = tensor("op_1122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378334528)))]; + tensor v_37_cast = linear(bias = var_1122_to_fp16, weight = var_1121_to_fp16, x = var_1102_cast); + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1500, 20, -1])]; + tensor var_1131_cast = reshape(shape = var_1130, x = q_37_cast); + tensor const_242_to_fp16 = const()[name = tensor("const_242_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast = mul(x = var_1131_cast, y = const_242_to_fp16); + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([1, 1500, 20, -1])]; + tensor var_1138_cast = reshape(shape = var_1137, x = k_37_cast); + tensor const_243_to_fp16 = const()[name = tensor("const_243_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast = mul(x = var_1138_cast, y = const_243_to_fp16); + tensor var_1144 = const()[name = tensor("op_1144"), val = tensor([1, 1500, 20, -1])]; + tensor var_1145_cast = reshape(shape = var_1144, x = v_37_cast); + tensor var_1146 = const()[name = tensor("op_1146"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_217 = transpose(perm = transpose_83_perm_0, x = k_39_cast); + tensor transpose_218 = transpose(perm = transpose_82_perm_0, x = q_39_cast); + tensor qk_19_cast = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_218, y = transpose_217); + tensor var_1150_cast = softmax(axis = var_1085, x = qk_19_cast); + tensor var_1152_transpose_x_0 = const()[name = tensor("op_1152_transpose_x_0"), val = tensor(false)]; + tensor var_1152_transpose_y_0 = const()[name = tensor("op_1152_transpose_y_0"), val = tensor(false)]; + tensor transpose_219 = transpose(perm = var_1146, x = var_1145_cast); + tensor var_1152_cast = matmul(transpose_x = var_1152_transpose_x_0, transpose_y = var_1152_transpose_y_0, x = var_1150_cast, y = transpose_219); + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 1280])]; + tensor transpose_216 = transpose(perm = var_1153, x = var_1152_cast); + tensor x_119_cast = reshape(shape = concat_9, x = transpose_216); + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378337152)))]; + tensor var_1159_to_fp16 = const()[name = tensor("op_1159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381614016)))]; + tensor var_1160_cast = linear(bias = var_1159_to_fp16, weight = var_1158_to_fp16, x = x_119_cast); + tensor x_121_cast = add(x = x_115_cast, y = var_1160_cast); + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381616640)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381619264)))]; + tensor var_1166_cast = layer_norm(axes = var_1166_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1091_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast); + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381621888)))]; + tensor var_1176_to_fp16 = const()[name = tensor("op_1176_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394729152)))]; + tensor input_81_cast = linear(bias = var_1176_to_fp16, weight = var_1175_to_fp16, x = var_1166_cast); + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast = gelu(mode = x_125_mode_0, x = input_81_cast); + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394739456)))]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407846720)))]; + tensor var_1183_cast = linear(bias = var_1182_to_fp16, weight = var_1181_to_fp16, x = x_125_cast); + tensor x_127_cast = add(x = x_121_cast, y = var_1183_cast); + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor(-1)]; + tensor var_1209_axes_0 = const()[name = tensor("op_1209_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407849344)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407851968)))]; + tensor var_1198_to_fp16 = const()[name = tensor("op_1198_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1209_cast = layer_norm(axes = var_1209_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast); + tensor var_1220_to_fp16 = const()[name = tensor("op_1220_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407854592)))]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411131456)))]; + tensor q_41_cast = linear(bias = var_1221_to_fp16, weight = var_1220_to_fp16, x = var_1209_cast); + tensor var_1224_to_fp16 = const()[name = tensor("op_1224_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411134080)))]; + tensor k_41_bias_0_to_fp16 = const()[name = tensor("k_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414410944)))]; + tensor k_41_cast = linear(bias = k_41_bias_0_to_fp16, weight = var_1224_to_fp16, x = var_1209_cast); + tensor var_1228_to_fp16 = const()[name = tensor("op_1228_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414413568)))]; + tensor var_1229_to_fp16 = const()[name = tensor("op_1229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417690432)))]; + tensor v_41_cast = linear(bias = var_1229_to_fp16, weight = var_1228_to_fp16, x = var_1209_cast); + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1500, 20, -1])]; + tensor var_1238_cast = reshape(shape = var_1237, x = q_41_cast); + tensor const_244_to_fp16 = const()[name = tensor("const_244_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast = mul(x = var_1238_cast, y = const_244_to_fp16); + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([1, 1500, 20, -1])]; + tensor var_1245_cast = reshape(shape = var_1244, x = k_41_cast); + tensor const_245_to_fp16 = const()[name = tensor("const_245_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast = mul(x = var_1245_cast, y = const_245_to_fp16); + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 1500, 20, -1])]; + tensor var_1252_cast = reshape(shape = var_1251, x = v_41_cast); + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_213 = transpose(perm = transpose_85_perm_0, x = k_43_cast); + tensor transpose_214 = transpose(perm = transpose_84_perm_0, x = q_43_cast); + tensor qk_21_cast = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_214, y = transpose_213); + tensor var_1257_cast = softmax(axis = var_1192, x = qk_21_cast); + tensor var_1259_transpose_x_0 = const()[name = tensor("op_1259_transpose_x_0"), val = tensor(false)]; + tensor var_1259_transpose_y_0 = const()[name = tensor("op_1259_transpose_y_0"), val = tensor(false)]; + tensor transpose_215 = transpose(perm = var_1253, x = var_1252_cast); + tensor var_1259_cast = matmul(transpose_x = var_1259_transpose_x_0, transpose_y = var_1259_transpose_y_0, x = var_1257_cast, y = transpose_215); + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 1280])]; + tensor transpose_212 = transpose(perm = var_1260, x = var_1259_cast); + tensor x_131_cast = reshape(shape = concat_10, x = transpose_212); + tensor var_1265_to_fp16 = const()[name = tensor("op_1265_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417693056)))]; + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420969920)))]; + tensor var_1267_cast = linear(bias = var_1266_to_fp16, weight = var_1265_to_fp16, x = x_131_cast); + tensor x_133_cast = add(x = x_127_cast, y = var_1267_cast); + tensor var_1273_axes_0 = const()[name = tensor("op_1273_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420972544)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420975168)))]; + tensor var_1273_cast = layer_norm(axes = var_1273_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast); + tensor var_1282_to_fp16 = const()[name = tensor("op_1282_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420977792)))]; + tensor var_1283_to_fp16 = const()[name = tensor("op_1283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434085056)))]; + tensor input_89_cast = linear(bias = var_1283_to_fp16, weight = var_1282_to_fp16, x = var_1273_cast); + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast = gelu(mode = x_137_mode_0, x = input_89_cast); + tensor var_1288_to_fp16 = const()[name = tensor("op_1288_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434095360)))]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447202624)))]; + tensor var_1290_cast = linear(bias = var_1289_to_fp16, weight = var_1288_to_fp16, x = x_137_cast); + tensor x_139_cast = add(x = x_133_cast, y = var_1290_cast); + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor(-1)]; + tensor var_1316_axes_0 = const()[name = tensor("op_1316_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447205248)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447207872)))]; + tensor var_1305_to_fp16 = const()[name = tensor("op_1305_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1316_cast = layer_norm(axes = var_1316_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1305_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast); + tensor var_1327_to_fp16 = const()[name = tensor("op_1327_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447210496)))]; + tensor var_1328_to_fp16 = const()[name = tensor("op_1328_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450487360)))]; + tensor q_45_cast = linear(bias = var_1328_to_fp16, weight = var_1327_to_fp16, x = var_1316_cast); + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450489984)))]; + tensor k_45_bias_0_to_fp16 = const()[name = tensor("k_45_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453766848)))]; + tensor k_45_cast = linear(bias = k_45_bias_0_to_fp16, weight = var_1331_to_fp16, x = var_1316_cast); + tensor var_1335_to_fp16 = const()[name = tensor("op_1335_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453769472)))]; + tensor var_1336_to_fp16 = const()[name = tensor("op_1336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457046336)))]; + tensor v_45_cast = linear(bias = var_1336_to_fp16, weight = var_1335_to_fp16, x = var_1316_cast); + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([1, 1500, 20, -1])]; + tensor var_1345_cast = reshape(shape = var_1344, x = q_45_cast); + tensor const_246_to_fp16 = const()[name = tensor("const_246_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_47_cast = mul(x = var_1345_cast, y = const_246_to_fp16); + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([1, 1500, 20, -1])]; + tensor var_1352_cast = reshape(shape = var_1351, x = k_45_cast); + tensor const_247_to_fp16 = const()[name = tensor("const_247_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_47_cast = mul(x = var_1352_cast, y = const_247_to_fp16); + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 1500, 20, -1])]; + tensor var_1359_cast = reshape(shape = var_1358, x = v_45_cast); + tensor var_1360 = const()[name = tensor("op_1360"), val = tensor([0, 2, 1, 3])]; + tensor qk_23_transpose_x_0 = const()[name = tensor("qk_23_transpose_x_0"), val = tensor(false)]; + tensor qk_23_transpose_y_0 = const()[name = tensor("qk_23_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_209 = transpose(perm = transpose_87_perm_0, x = k_47_cast); + tensor transpose_210 = transpose(perm = transpose_86_perm_0, x = q_47_cast); + tensor qk_23_cast = matmul(transpose_x = qk_23_transpose_x_0, transpose_y = qk_23_transpose_y_0, x = transpose_210, y = transpose_209); + tensor var_1364_cast = softmax(axis = var_1299, x = qk_23_cast); + tensor var_1366_transpose_x_0 = const()[name = tensor("op_1366_transpose_x_0"), val = tensor(false)]; + tensor var_1366_transpose_y_0 = const()[name = tensor("op_1366_transpose_y_0"), val = tensor(false)]; + tensor transpose_211 = transpose(perm = var_1360, x = var_1359_cast); + tensor var_1366_cast = matmul(transpose_x = var_1366_transpose_x_0, transpose_y = var_1366_transpose_y_0, x = var_1364_cast, y = transpose_211); + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 1280])]; + tensor transpose_208 = transpose(perm = var_1367, x = var_1366_cast); + tensor x_143_cast = reshape(shape = concat_11, x = transpose_208); + tensor var_1372_to_fp16 = const()[name = tensor("op_1372_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457048960)))]; + tensor var_1373_to_fp16 = const()[name = tensor("op_1373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460325824)))]; + tensor var_1374_cast = linear(bias = var_1373_to_fp16, weight = var_1372_to_fp16, x = x_143_cast); + tensor x_145_cast = add(x = x_139_cast, y = var_1374_cast); + tensor var_1380_axes_0 = const()[name = tensor("op_1380_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460328448)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460331072)))]; + tensor var_1380_cast = layer_norm(axes = var_1380_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1305_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast); + tensor var_1389_to_fp16 = const()[name = tensor("op_1389_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460333696)))]; + tensor var_1390_to_fp16 = const()[name = tensor("op_1390_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473440960)))]; + tensor input_97_cast = linear(bias = var_1390_to_fp16, weight = var_1389_to_fp16, x = var_1380_cast); + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast = gelu(mode = x_149_mode_0, x = input_97_cast); + tensor var_1395_to_fp16 = const()[name = tensor("op_1395_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473451264)))]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486558528)))]; + tensor var_1397_cast = linear(bias = var_1396_to_fp16, weight = var_1395_to_fp16, x = x_149_cast); + tensor x_151_cast = add(x = x_145_cast, y = var_1397_cast); + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor var_1423_axes_0 = const()[name = tensor("op_1423_axes_0"), val = tensor([-1])]; + tensor blocks_12_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_12_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486561152)))]; + tensor blocks_12_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_12_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486563776)))]; + tensor var_1412_to_fp16 = const()[name = tensor("op_1412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1423_cast = layer_norm(axes = var_1423_axes_0, beta = blocks_12_attn_ln_bias_to_fp16, epsilon = var_1412_to_fp16, gamma = blocks_12_attn_ln_weight_to_fp16, x = x_151_cast); + tensor var_1434_to_fp16 = const()[name = tensor("op_1434_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486566400)))]; + tensor var_1435_to_fp16 = const()[name = tensor("op_1435_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489843264)))]; + tensor q_49_cast = linear(bias = var_1435_to_fp16, weight = var_1434_to_fp16, x = var_1423_cast); + tensor var_1438_to_fp16 = const()[name = tensor("op_1438_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489845888)))]; + tensor k_49_bias_0_to_fp16 = const()[name = tensor("k_49_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493122752)))]; + tensor k_49_cast = linear(bias = k_49_bias_0_to_fp16, weight = var_1438_to_fp16, x = var_1423_cast); + tensor var_1442_to_fp16 = const()[name = tensor("op_1442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493125376)))]; + tensor var_1443_to_fp16 = const()[name = tensor("op_1443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496402240)))]; + tensor v_49_cast = linear(bias = var_1443_to_fp16, weight = var_1442_to_fp16, x = var_1423_cast); + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([1, 1500, 20, -1])]; + tensor var_1452_cast = reshape(shape = var_1451, x = q_49_cast); + tensor const_248_to_fp16 = const()[name = tensor("const_248_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_51_cast = mul(x = var_1452_cast, y = const_248_to_fp16); + tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([1, 1500, 20, -1])]; + tensor var_1459_cast = reshape(shape = var_1458, x = k_49_cast); + tensor const_249_to_fp16 = const()[name = tensor("const_249_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_51_cast = mul(x = var_1459_cast, y = const_249_to_fp16); + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1500, 20, -1])]; + tensor var_1466_cast = reshape(shape = var_1465, x = v_49_cast); + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([0, 2, 1, 3])]; + tensor qk_25_transpose_x_0 = const()[name = tensor("qk_25_transpose_x_0"), val = tensor(false)]; + tensor qk_25_transpose_y_0 = const()[name = tensor("qk_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_205 = transpose(perm = transpose_89_perm_0, x = k_51_cast); + tensor transpose_206 = transpose(perm = transpose_88_perm_0, x = q_51_cast); + tensor qk_25_cast = matmul(transpose_x = qk_25_transpose_x_0, transpose_y = qk_25_transpose_y_0, x = transpose_206, y = transpose_205); + tensor var_1471_cast = softmax(axis = var_1406, x = qk_25_cast); + tensor var_1473_transpose_x_0 = const()[name = tensor("op_1473_transpose_x_0"), val = tensor(false)]; + tensor var_1473_transpose_y_0 = const()[name = tensor("op_1473_transpose_y_0"), val = tensor(false)]; + tensor transpose_207 = transpose(perm = var_1467, x = var_1466_cast); + tensor var_1473_cast = matmul(transpose_x = var_1473_transpose_x_0, transpose_y = var_1473_transpose_y_0, x = var_1471_cast, y = transpose_207); + tensor var_1474 = const()[name = tensor("op_1474"), val = tensor([0, 2, 1, 3])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 1500, 1280])]; + tensor transpose_204 = transpose(perm = var_1474, x = var_1473_cast); + tensor x_155_cast = reshape(shape = concat_12, x = transpose_204); + tensor var_1479_to_fp16 = const()[name = tensor("op_1479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496404864)))]; + tensor var_1480_to_fp16 = const()[name = tensor("op_1480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499681728)))]; + tensor var_1481_cast = linear(bias = var_1480_to_fp16, weight = var_1479_to_fp16, x = x_155_cast); + tensor x_157_cast = add(x = x_151_cast, y = var_1481_cast); + tensor var_1487_axes_0 = const()[name = tensor("op_1487_axes_0"), val = tensor([-1])]; + tensor blocks_12_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499684352)))]; + tensor blocks_12_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499686976)))]; + tensor var_1487_cast = layer_norm(axes = var_1487_axes_0, beta = blocks_12_mlp_ln_bias_to_fp16, epsilon = var_1412_to_fp16, gamma = blocks_12_mlp_ln_weight_to_fp16, x = x_157_cast); + tensor var_1496_to_fp16 = const()[name = tensor("op_1496_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499689600)))]; + tensor var_1497_to_fp16 = const()[name = tensor("op_1497_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512796864)))]; + tensor input_105_cast = linear(bias = var_1497_to_fp16, weight = var_1496_to_fp16, x = var_1487_cast); + tensor x_161_mode_0 = const()[name = tensor("x_161_mode_0"), val = tensor("EXACT")]; + tensor x_161_cast = gelu(mode = x_161_mode_0, x = input_105_cast); + tensor var_1502_to_fp16 = const()[name = tensor("op_1502_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512807168)))]; + tensor var_1503_to_fp16 = const()[name = tensor("op_1503_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525914432)))]; + tensor var_1504_cast = linear(bias = var_1503_to_fp16, weight = var_1502_to_fp16, x = x_161_cast); + tensor x_163_cast = add(x = x_157_cast, y = var_1504_cast); + tensor var_1513 = const()[name = tensor("op_1513"), val = tensor(-1)]; + tensor var_1530_axes_0 = const()[name = tensor("op_1530_axes_0"), val = tensor([-1])]; + tensor blocks_13_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_13_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525917056)))]; + tensor blocks_13_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_13_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525919680)))]; + tensor var_1519_to_fp16 = const()[name = tensor("op_1519_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1530_cast = layer_norm(axes = var_1530_axes_0, beta = blocks_13_attn_ln_bias_to_fp16, epsilon = var_1519_to_fp16, gamma = blocks_13_attn_ln_weight_to_fp16, x = x_163_cast); + tensor var_1541_to_fp16 = const()[name = tensor("op_1541_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525922304)))]; + tensor var_1542_to_fp16 = const()[name = tensor("op_1542_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529199168)))]; + tensor q_53_cast = linear(bias = var_1542_to_fp16, weight = var_1541_to_fp16, x = var_1530_cast); + tensor var_1545_to_fp16 = const()[name = tensor("op_1545_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529201792)))]; + tensor k_53_bias_0_to_fp16 = const()[name = tensor("k_53_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532478656)))]; + tensor k_53_cast = linear(bias = k_53_bias_0_to_fp16, weight = var_1545_to_fp16, x = var_1530_cast); + tensor var_1549_to_fp16 = const()[name = tensor("op_1549_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532481280)))]; + tensor var_1550_to_fp16 = const()[name = tensor("op_1550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535758144)))]; + tensor v_53_cast = linear(bias = var_1550_to_fp16, weight = var_1549_to_fp16, x = var_1530_cast); + tensor var_1558 = const()[name = tensor("op_1558"), val = tensor([1, 1500, 20, -1])]; + tensor var_1559_cast = reshape(shape = var_1558, x = q_53_cast); + tensor const_250_to_fp16 = const()[name = tensor("const_250_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_55_cast = mul(x = var_1559_cast, y = const_250_to_fp16); + tensor var_1565 = const()[name = tensor("op_1565"), val = tensor([1, 1500, 20, -1])]; + tensor var_1566_cast = reshape(shape = var_1565, x = k_53_cast); + tensor const_251_to_fp16 = const()[name = tensor("const_251_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_55_cast = mul(x = var_1566_cast, y = const_251_to_fp16); + tensor var_1572 = const()[name = tensor("op_1572"), val = tensor([1, 1500, 20, -1])]; + tensor var_1573_cast = reshape(shape = var_1572, x = v_53_cast); + tensor var_1574 = const()[name = tensor("op_1574"), val = tensor([0, 2, 1, 3])]; + tensor qk_27_transpose_x_0 = const()[name = tensor("qk_27_transpose_x_0"), val = tensor(false)]; + tensor qk_27_transpose_y_0 = const()[name = tensor("qk_27_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_201 = transpose(perm = transpose_91_perm_0, x = k_55_cast); + tensor transpose_202 = transpose(perm = transpose_90_perm_0, x = q_55_cast); + tensor qk_27_cast = matmul(transpose_x = qk_27_transpose_x_0, transpose_y = qk_27_transpose_y_0, x = transpose_202, y = transpose_201); + tensor var_1578_cast = softmax(axis = var_1513, x = qk_27_cast); + tensor var_1580_transpose_x_0 = const()[name = tensor("op_1580_transpose_x_0"), val = tensor(false)]; + tensor var_1580_transpose_y_0 = const()[name = tensor("op_1580_transpose_y_0"), val = tensor(false)]; + tensor transpose_203 = transpose(perm = var_1574, x = var_1573_cast); + tensor var_1580_cast = matmul(transpose_x = var_1580_transpose_x_0, transpose_y = var_1580_transpose_y_0, x = var_1578_cast, y = transpose_203); + tensor var_1581 = const()[name = tensor("op_1581"), val = tensor([0, 2, 1, 3])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([1, 1500, 1280])]; + tensor transpose_200 = transpose(perm = var_1581, x = var_1580_cast); + tensor x_167_cast = reshape(shape = concat_13, x = transpose_200); + tensor var_1586_to_fp16 = const()[name = tensor("op_1586_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535760768)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539037632)))]; + tensor var_1588_cast = linear(bias = var_1587_to_fp16, weight = var_1586_to_fp16, x = x_167_cast); + tensor x_169_cast = add(x = x_163_cast, y = var_1588_cast); + tensor var_1594_axes_0 = const()[name = tensor("op_1594_axes_0"), val = tensor([-1])]; + tensor blocks_13_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539040256)))]; + tensor blocks_13_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539042880)))]; + tensor var_1594_cast = layer_norm(axes = var_1594_axes_0, beta = blocks_13_mlp_ln_bias_to_fp16, epsilon = var_1519_to_fp16, gamma = blocks_13_mlp_ln_weight_to_fp16, x = x_169_cast); + tensor var_1603_to_fp16 = const()[name = tensor("op_1603_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539045504)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552152768)))]; + tensor input_113_cast = linear(bias = var_1604_to_fp16, weight = var_1603_to_fp16, x = var_1594_cast); + tensor x_173_mode_0 = const()[name = tensor("x_173_mode_0"), val = tensor("EXACT")]; + tensor x_173_cast = gelu(mode = x_173_mode_0, x = input_113_cast); + tensor var_1609_to_fp16 = const()[name = tensor("op_1609_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552163072)))]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565270336)))]; + tensor var_1611_cast = linear(bias = var_1610_to_fp16, weight = var_1609_to_fp16, x = x_173_cast); + tensor x_175_cast = add(x = x_169_cast, y = var_1611_cast); + tensor var_1620 = const()[name = tensor("op_1620"), val = tensor(-1)]; + tensor var_1637_axes_0 = const()[name = tensor("op_1637_axes_0"), val = tensor([-1])]; + tensor blocks_14_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_14_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565272960)))]; + tensor blocks_14_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_14_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565275584)))]; + tensor var_1626_to_fp16 = const()[name = tensor("op_1626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1637_cast = layer_norm(axes = var_1637_axes_0, beta = blocks_14_attn_ln_bias_to_fp16, epsilon = var_1626_to_fp16, gamma = blocks_14_attn_ln_weight_to_fp16, x = x_175_cast); + tensor var_1648_to_fp16 = const()[name = tensor("op_1648_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565278208)))]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568555072)))]; + tensor q_57_cast = linear(bias = var_1649_to_fp16, weight = var_1648_to_fp16, x = var_1637_cast); + tensor var_1652_to_fp16 = const()[name = tensor("op_1652_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568557696)))]; + tensor k_57_bias_0_to_fp16 = const()[name = tensor("k_57_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571834560)))]; + tensor k_57_cast = linear(bias = k_57_bias_0_to_fp16, weight = var_1652_to_fp16, x = var_1637_cast); + tensor var_1656_to_fp16 = const()[name = tensor("op_1656_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571837184)))]; + tensor var_1657_to_fp16 = const()[name = tensor("op_1657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575114048)))]; + tensor v_57_cast = linear(bias = var_1657_to_fp16, weight = var_1656_to_fp16, x = var_1637_cast); + tensor var_1665 = const()[name = tensor("op_1665"), val = tensor([1, 1500, 20, -1])]; + tensor var_1666_cast = reshape(shape = var_1665, x = q_57_cast); + tensor const_252_to_fp16 = const()[name = tensor("const_252_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_59_cast = mul(x = var_1666_cast, y = const_252_to_fp16); + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 1500, 20, -1])]; + tensor var_1673_cast = reshape(shape = var_1672, x = k_57_cast); + tensor const_253_to_fp16 = const()[name = tensor("const_253_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_59_cast = mul(x = var_1673_cast, y = const_253_to_fp16); + tensor var_1679 = const()[name = tensor("op_1679"), val = tensor([1, 1500, 20, -1])]; + tensor var_1680_cast = reshape(shape = var_1679, x = v_57_cast); + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([0, 2, 1, 3])]; + tensor qk_29_transpose_x_0 = const()[name = tensor("qk_29_transpose_x_0"), val = tensor(false)]; + tensor qk_29_transpose_y_0 = const()[name = tensor("qk_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_197 = transpose(perm = transpose_93_perm_0, x = k_59_cast); + tensor transpose_198 = transpose(perm = transpose_92_perm_0, x = q_59_cast); + tensor qk_29_cast = matmul(transpose_x = qk_29_transpose_x_0, transpose_y = qk_29_transpose_y_0, x = transpose_198, y = transpose_197); + tensor var_1685_cast = softmax(axis = var_1620, x = qk_29_cast); + tensor var_1687_transpose_x_0 = const()[name = tensor("op_1687_transpose_x_0"), val = tensor(false)]; + tensor var_1687_transpose_y_0 = const()[name = tensor("op_1687_transpose_y_0"), val = tensor(false)]; + tensor transpose_199 = transpose(perm = var_1681, x = var_1680_cast); + tensor var_1687_cast = matmul(transpose_x = var_1687_transpose_x_0, transpose_y = var_1687_transpose_y_0, x = var_1685_cast, y = transpose_199); + tensor var_1688 = const()[name = tensor("op_1688"), val = tensor([0, 2, 1, 3])]; + tensor concat_14 = const()[name = tensor("concat_14"), val = tensor([1, 1500, 1280])]; + tensor transpose_196 = transpose(perm = var_1688, x = var_1687_cast); + tensor x_179_cast = reshape(shape = concat_14, x = transpose_196); + tensor var_1693_to_fp16 = const()[name = tensor("op_1693_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575116672)))]; + tensor var_1694_to_fp16 = const()[name = tensor("op_1694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578393536)))]; + tensor var_1695_cast = linear(bias = var_1694_to_fp16, weight = var_1693_to_fp16, x = x_179_cast); + tensor x_181_cast = add(x = x_175_cast, y = var_1695_cast); + tensor var_1701_axes_0 = const()[name = tensor("op_1701_axes_0"), val = tensor([-1])]; + tensor blocks_14_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578396160)))]; + tensor blocks_14_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578398784)))]; + tensor var_1701_cast = layer_norm(axes = var_1701_axes_0, beta = blocks_14_mlp_ln_bias_to_fp16, epsilon = var_1626_to_fp16, gamma = blocks_14_mlp_ln_weight_to_fp16, x = x_181_cast); + tensor var_1710_to_fp16 = const()[name = tensor("op_1710_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578401408)))]; + tensor var_1711_to_fp16 = const()[name = tensor("op_1711_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591508672)))]; + tensor input_121_cast = linear(bias = var_1711_to_fp16, weight = var_1710_to_fp16, x = var_1701_cast); + tensor x_185_mode_0 = const()[name = tensor("x_185_mode_0"), val = tensor("EXACT")]; + tensor x_185_cast = gelu(mode = x_185_mode_0, x = input_121_cast); + tensor var_1716_to_fp16 = const()[name = tensor("op_1716_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591518976)))]; + tensor var_1717_to_fp16 = const()[name = tensor("op_1717_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604626240)))]; + tensor var_1718_cast = linear(bias = var_1717_to_fp16, weight = var_1716_to_fp16, x = x_185_cast); + tensor x_187_cast = add(x = x_181_cast, y = var_1718_cast); + tensor var_1727 = const()[name = tensor("op_1727"), val = tensor(-1)]; + tensor var_1744_axes_0 = const()[name = tensor("op_1744_axes_0"), val = tensor([-1])]; + tensor blocks_15_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_15_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604628864)))]; + tensor blocks_15_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_15_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604631488)))]; + tensor var_1733_to_fp16 = const()[name = tensor("op_1733_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1744_cast = layer_norm(axes = var_1744_axes_0, beta = blocks_15_attn_ln_bias_to_fp16, epsilon = var_1733_to_fp16, gamma = blocks_15_attn_ln_weight_to_fp16, x = x_187_cast); + tensor var_1755_to_fp16 = const()[name = tensor("op_1755_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604634112)))]; + tensor var_1756_to_fp16 = const()[name = tensor("op_1756_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607910976)))]; + tensor q_61_cast = linear(bias = var_1756_to_fp16, weight = var_1755_to_fp16, x = var_1744_cast); + tensor var_1759_to_fp16 = const()[name = tensor("op_1759_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607913600)))]; + tensor k_61_bias_0_to_fp16 = const()[name = tensor("k_61_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611190464)))]; + tensor k_61_cast = linear(bias = k_61_bias_0_to_fp16, weight = var_1759_to_fp16, x = var_1744_cast); + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611193088)))]; + tensor var_1764_to_fp16 = const()[name = tensor("op_1764_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614469952)))]; + tensor v_61_cast = linear(bias = var_1764_to_fp16, weight = var_1763_to_fp16, x = var_1744_cast); + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1500, 20, -1])]; + tensor var_1773_cast = reshape(shape = var_1772, x = q_61_cast); + tensor const_254_to_fp16 = const()[name = tensor("const_254_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_63_cast = mul(x = var_1773_cast, y = const_254_to_fp16); + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 1500, 20, -1])]; + tensor var_1780_cast = reshape(shape = var_1779, x = k_61_cast); + tensor const_255_to_fp16 = const()[name = tensor("const_255_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_63_cast = mul(x = var_1780_cast, y = const_255_to_fp16); + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 1500, 20, -1])]; + tensor var_1787_cast = reshape(shape = var_1786, x = v_61_cast); + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([0, 2, 1, 3])]; + tensor qk_31_transpose_x_0 = const()[name = tensor("qk_31_transpose_x_0"), val = tensor(false)]; + tensor qk_31_transpose_y_0 = const()[name = tensor("qk_31_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_193 = transpose(perm = transpose_95_perm_0, x = k_63_cast); + tensor transpose_194 = transpose(perm = transpose_94_perm_0, x = q_63_cast); + tensor qk_31_cast = matmul(transpose_x = qk_31_transpose_x_0, transpose_y = qk_31_transpose_y_0, x = transpose_194, y = transpose_193); + tensor var_1792_cast = softmax(axis = var_1727, x = qk_31_cast); + tensor var_1794_transpose_x_0 = const()[name = tensor("op_1794_transpose_x_0"), val = tensor(false)]; + tensor var_1794_transpose_y_0 = const()[name = tensor("op_1794_transpose_y_0"), val = tensor(false)]; + tensor transpose_195 = transpose(perm = var_1788, x = var_1787_cast); + tensor var_1794_cast = matmul(transpose_x = var_1794_transpose_x_0, transpose_y = var_1794_transpose_y_0, x = var_1792_cast, y = transpose_195); + tensor var_1795 = const()[name = tensor("op_1795"), val = tensor([0, 2, 1, 3])]; + tensor concat_15 = const()[name = tensor("concat_15"), val = tensor([1, 1500, 1280])]; + tensor transpose_192 = transpose(perm = var_1795, x = var_1794_cast); + tensor x_191_cast = reshape(shape = concat_15, x = transpose_192); + tensor var_1800_to_fp16 = const()[name = tensor("op_1800_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614472576)))]; + tensor var_1801_to_fp16 = const()[name = tensor("op_1801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617749440)))]; + tensor var_1802_cast = linear(bias = var_1801_to_fp16, weight = var_1800_to_fp16, x = x_191_cast); + tensor x_193_cast = add(x = x_187_cast, y = var_1802_cast); + tensor var_1808_axes_0 = const()[name = tensor("op_1808_axes_0"), val = tensor([-1])]; + tensor blocks_15_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617752064)))]; + tensor blocks_15_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617754688)))]; + tensor var_1808_cast = layer_norm(axes = var_1808_axes_0, beta = blocks_15_mlp_ln_bias_to_fp16, epsilon = var_1733_to_fp16, gamma = blocks_15_mlp_ln_weight_to_fp16, x = x_193_cast); + tensor var_1817_to_fp16 = const()[name = tensor("op_1817_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617757312)))]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630864576)))]; + tensor input_129_cast = linear(bias = var_1818_to_fp16, weight = var_1817_to_fp16, x = var_1808_cast); + tensor x_197_mode_0 = const()[name = tensor("x_197_mode_0"), val = tensor("EXACT")]; + tensor x_197_cast = gelu(mode = x_197_mode_0, x = input_129_cast); + tensor var_1823_to_fp16 = const()[name = tensor("op_1823_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630874880)))]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643982144)))]; + tensor var_1825_cast = linear(bias = var_1824_to_fp16, weight = var_1823_to_fp16, x = x_197_cast); + tensor x_199_cast = add(x = x_193_cast, y = var_1825_cast); + tensor var_1834 = const()[name = tensor("op_1834"), val = tensor(-1)]; + tensor var_1851_axes_0 = const()[name = tensor("op_1851_axes_0"), val = tensor([-1])]; + tensor blocks_16_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_16_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643984768)))]; + tensor blocks_16_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_16_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643987392)))]; + tensor var_1840_to_fp16 = const()[name = tensor("op_1840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1851_cast = layer_norm(axes = var_1851_axes_0, beta = blocks_16_attn_ln_bias_to_fp16, epsilon = var_1840_to_fp16, gamma = blocks_16_attn_ln_weight_to_fp16, x = x_199_cast); + tensor var_1862_to_fp16 = const()[name = tensor("op_1862_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643990016)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647266880)))]; + tensor q_65_cast = linear(bias = var_1863_to_fp16, weight = var_1862_to_fp16, x = var_1851_cast); + tensor var_1866_to_fp16 = const()[name = tensor("op_1866_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647269504)))]; + tensor k_65_bias_0_to_fp16 = const()[name = tensor("k_65_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650546368)))]; + tensor k_65_cast = linear(bias = k_65_bias_0_to_fp16, weight = var_1866_to_fp16, x = var_1851_cast); + tensor var_1870_to_fp16 = const()[name = tensor("op_1870_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650548992)))]; + tensor var_1871_to_fp16 = const()[name = tensor("op_1871_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653825856)))]; + tensor v_65_cast = linear(bias = var_1871_to_fp16, weight = var_1870_to_fp16, x = var_1851_cast); + tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([1, 1500, 20, -1])]; + tensor var_1880_cast = reshape(shape = var_1879, x = q_65_cast); + tensor const_256_to_fp16 = const()[name = tensor("const_256_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_67_cast = mul(x = var_1880_cast, y = const_256_to_fp16); + tensor var_1886 = const()[name = tensor("op_1886"), val = tensor([1, 1500, 20, -1])]; + tensor var_1887_cast = reshape(shape = var_1886, x = k_65_cast); + tensor const_257_to_fp16 = const()[name = tensor("const_257_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_67_cast = mul(x = var_1887_cast, y = const_257_to_fp16); + tensor var_1893 = const()[name = tensor("op_1893"), val = tensor([1, 1500, 20, -1])]; + tensor var_1894_cast = reshape(shape = var_1893, x = v_65_cast); + tensor var_1895 = const()[name = tensor("op_1895"), val = tensor([0, 2, 1, 3])]; + tensor qk_33_transpose_x_0 = const()[name = tensor("qk_33_transpose_x_0"), val = tensor(false)]; + tensor qk_33_transpose_y_0 = const()[name = tensor("qk_33_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_189 = transpose(perm = transpose_97_perm_0, x = k_67_cast); + tensor transpose_190 = transpose(perm = transpose_96_perm_0, x = q_67_cast); + tensor qk_33_cast = matmul(transpose_x = qk_33_transpose_x_0, transpose_y = qk_33_transpose_y_0, x = transpose_190, y = transpose_189); + tensor var_1899_cast = softmax(axis = var_1834, x = qk_33_cast); + tensor var_1901_transpose_x_0 = const()[name = tensor("op_1901_transpose_x_0"), val = tensor(false)]; + tensor var_1901_transpose_y_0 = const()[name = tensor("op_1901_transpose_y_0"), val = tensor(false)]; + tensor transpose_191 = transpose(perm = var_1895, x = var_1894_cast); + tensor var_1901_cast = matmul(transpose_x = var_1901_transpose_x_0, transpose_y = var_1901_transpose_y_0, x = var_1899_cast, y = transpose_191); + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([1, 1500, 1280])]; + tensor transpose_188 = transpose(perm = var_1902, x = var_1901_cast); + tensor x_203_cast = reshape(shape = concat_16, x = transpose_188); + tensor var_1907_to_fp16 = const()[name = tensor("op_1907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653828480)))]; + tensor var_1908_to_fp16 = const()[name = tensor("op_1908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657105344)))]; + tensor var_1909_cast = linear(bias = var_1908_to_fp16, weight = var_1907_to_fp16, x = x_203_cast); + tensor x_205_cast = add(x = x_199_cast, y = var_1909_cast); + tensor var_1915_axes_0 = const()[name = tensor("op_1915_axes_0"), val = tensor([-1])]; + tensor blocks_16_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657107968)))]; + tensor blocks_16_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657110592)))]; + tensor var_1915_cast = layer_norm(axes = var_1915_axes_0, beta = blocks_16_mlp_ln_bias_to_fp16, epsilon = var_1840_to_fp16, gamma = blocks_16_mlp_ln_weight_to_fp16, x = x_205_cast); + tensor var_1924_to_fp16 = const()[name = tensor("op_1924_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657113216)))]; + tensor var_1925_to_fp16 = const()[name = tensor("op_1925_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670220480)))]; + tensor input_137_cast = linear(bias = var_1925_to_fp16, weight = var_1924_to_fp16, x = var_1915_cast); + tensor x_209_mode_0 = const()[name = tensor("x_209_mode_0"), val = tensor("EXACT")]; + tensor x_209_cast = gelu(mode = x_209_mode_0, x = input_137_cast); + tensor var_1930_to_fp16 = const()[name = tensor("op_1930_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670230784)))]; + tensor var_1931_to_fp16 = const()[name = tensor("op_1931_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683338048)))]; + tensor var_1932_cast = linear(bias = var_1931_to_fp16, weight = var_1930_to_fp16, x = x_209_cast); + tensor x_211_cast = add(x = x_205_cast, y = var_1932_cast); + tensor var_1941 = const()[name = tensor("op_1941"), val = tensor(-1)]; + tensor var_1958_axes_0 = const()[name = tensor("op_1958_axes_0"), val = tensor([-1])]; + tensor blocks_17_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_17_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683340672)))]; + tensor blocks_17_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_17_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683343296)))]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1958_cast = layer_norm(axes = var_1958_axes_0, beta = blocks_17_attn_ln_bias_to_fp16, epsilon = var_1947_to_fp16, gamma = blocks_17_attn_ln_weight_to_fp16, x = x_211_cast); + tensor var_1969_to_fp16 = const()[name = tensor("op_1969_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683345920)))]; + tensor var_1970_to_fp16 = const()[name = tensor("op_1970_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686622784)))]; + tensor q_69_cast = linear(bias = var_1970_to_fp16, weight = var_1969_to_fp16, x = var_1958_cast); + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686625408)))]; + tensor k_69_bias_0_to_fp16 = const()[name = tensor("k_69_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689902272)))]; + tensor k_69_cast = linear(bias = k_69_bias_0_to_fp16, weight = var_1973_to_fp16, x = var_1958_cast); + tensor var_1977_to_fp16 = const()[name = tensor("op_1977_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689904896)))]; + tensor var_1978_to_fp16 = const()[name = tensor("op_1978_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693181760)))]; + tensor v_69_cast = linear(bias = var_1978_to_fp16, weight = var_1977_to_fp16, x = var_1958_cast); + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 1500, 20, -1])]; + tensor var_1987_cast = reshape(shape = var_1986, x = q_69_cast); + tensor const_258_to_fp16 = const()[name = tensor("const_258_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_71_cast = mul(x = var_1987_cast, y = const_258_to_fp16); + tensor var_1993 = const()[name = tensor("op_1993"), val = tensor([1, 1500, 20, -1])]; + tensor var_1994_cast = reshape(shape = var_1993, x = k_69_cast); + tensor const_259_to_fp16 = const()[name = tensor("const_259_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_71_cast = mul(x = var_1994_cast, y = const_259_to_fp16); + tensor var_2000 = const()[name = tensor("op_2000"), val = tensor([1, 1500, 20, -1])]; + tensor var_2001_cast = reshape(shape = var_2000, x = v_69_cast); + tensor var_2002 = const()[name = tensor("op_2002"), val = tensor([0, 2, 1, 3])]; + tensor qk_35_transpose_x_0 = const()[name = tensor("qk_35_transpose_x_0"), val = tensor(false)]; + tensor qk_35_transpose_y_0 = const()[name = tensor("qk_35_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_185 = transpose(perm = transpose_99_perm_0, x = k_71_cast); + tensor transpose_186 = transpose(perm = transpose_98_perm_0, x = q_71_cast); + tensor qk_35_cast = matmul(transpose_x = qk_35_transpose_x_0, transpose_y = qk_35_transpose_y_0, x = transpose_186, y = transpose_185); + tensor var_2006_cast = softmax(axis = var_1941, x = qk_35_cast); + tensor var_2008_transpose_x_0 = const()[name = tensor("op_2008_transpose_x_0"), val = tensor(false)]; + tensor var_2008_transpose_y_0 = const()[name = tensor("op_2008_transpose_y_0"), val = tensor(false)]; + tensor transpose_187 = transpose(perm = var_2002, x = var_2001_cast); + tensor var_2008_cast = matmul(transpose_x = var_2008_transpose_x_0, transpose_y = var_2008_transpose_y_0, x = var_2006_cast, y = transpose_187); + tensor var_2009 = const()[name = tensor("op_2009"), val = tensor([0, 2, 1, 3])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([1, 1500, 1280])]; + tensor transpose_184 = transpose(perm = var_2009, x = var_2008_cast); + tensor x_215_cast = reshape(shape = concat_17, x = transpose_184); + tensor var_2014_to_fp16 = const()[name = tensor("op_2014_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693184384)))]; + tensor var_2015_to_fp16 = const()[name = tensor("op_2015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696461248)))]; + tensor var_2016_cast = linear(bias = var_2015_to_fp16, weight = var_2014_to_fp16, x = x_215_cast); + tensor x_217_cast = add(x = x_211_cast, y = var_2016_cast); + tensor var_2022_axes_0 = const()[name = tensor("op_2022_axes_0"), val = tensor([-1])]; + tensor blocks_17_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696463872)))]; + tensor blocks_17_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696466496)))]; + tensor var_2022_cast = layer_norm(axes = var_2022_axes_0, beta = blocks_17_mlp_ln_bias_to_fp16, epsilon = var_1947_to_fp16, gamma = blocks_17_mlp_ln_weight_to_fp16, x = x_217_cast); + tensor var_2031_to_fp16 = const()[name = tensor("op_2031_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696469120)))]; + tensor var_2032_to_fp16 = const()[name = tensor("op_2032_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709576384)))]; + tensor input_145_cast = linear(bias = var_2032_to_fp16, weight = var_2031_to_fp16, x = var_2022_cast); + tensor x_221_mode_0 = const()[name = tensor("x_221_mode_0"), val = tensor("EXACT")]; + tensor x_221_cast = gelu(mode = x_221_mode_0, x = input_145_cast); + tensor var_2037_to_fp16 = const()[name = tensor("op_2037_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709586688)))]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722693952)))]; + tensor var_2039_cast = linear(bias = var_2038_to_fp16, weight = var_2037_to_fp16, x = x_221_cast); + tensor x_223_cast = add(x = x_217_cast, y = var_2039_cast); + tensor var_2048 = const()[name = tensor("op_2048"), val = tensor(-1)]; + tensor var_2065_axes_0 = const()[name = tensor("op_2065_axes_0"), val = tensor([-1])]; + tensor blocks_18_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_18_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722696576)))]; + tensor blocks_18_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_18_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722699200)))]; + tensor var_2054_to_fp16 = const()[name = tensor("op_2054_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2065_cast = layer_norm(axes = var_2065_axes_0, beta = blocks_18_attn_ln_bias_to_fp16, epsilon = var_2054_to_fp16, gamma = blocks_18_attn_ln_weight_to_fp16, x = x_223_cast); + tensor var_2076_to_fp16 = const()[name = tensor("op_2076_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722701824)))]; + tensor var_2077_to_fp16 = const()[name = tensor("op_2077_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725978688)))]; + tensor q_73_cast = linear(bias = var_2077_to_fp16, weight = var_2076_to_fp16, x = var_2065_cast); + tensor var_2080_to_fp16 = const()[name = tensor("op_2080_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725981312)))]; + tensor k_73_bias_0_to_fp16 = const()[name = tensor("k_73_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729258176)))]; + tensor k_73_cast = linear(bias = k_73_bias_0_to_fp16, weight = var_2080_to_fp16, x = var_2065_cast); + tensor var_2084_to_fp16 = const()[name = tensor("op_2084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729260800)))]; + tensor var_2085_to_fp16 = const()[name = tensor("op_2085_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732537664)))]; + tensor v_73_cast = linear(bias = var_2085_to_fp16, weight = var_2084_to_fp16, x = var_2065_cast); + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([1, 1500, 20, -1])]; + tensor var_2094_cast = reshape(shape = var_2093, x = q_73_cast); + tensor const_260_to_fp16 = const()[name = tensor("const_260_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_75_cast = mul(x = var_2094_cast, y = const_260_to_fp16); + tensor var_2100 = const()[name = tensor("op_2100"), val = tensor([1, 1500, 20, -1])]; + tensor var_2101_cast = reshape(shape = var_2100, x = k_73_cast); + tensor const_261_to_fp16 = const()[name = tensor("const_261_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_75_cast = mul(x = var_2101_cast, y = const_261_to_fp16); + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 1500, 20, -1])]; + tensor var_2108_cast = reshape(shape = var_2107, x = v_73_cast); + tensor var_2109 = const()[name = tensor("op_2109"), val = tensor([0, 2, 1, 3])]; + tensor qk_37_transpose_x_0 = const()[name = tensor("qk_37_transpose_x_0"), val = tensor(false)]; + tensor qk_37_transpose_y_0 = const()[name = tensor("qk_37_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_181 = transpose(perm = transpose_101_perm_0, x = k_75_cast); + tensor transpose_182 = transpose(perm = transpose_100_perm_0, x = q_75_cast); + tensor qk_37_cast = matmul(transpose_x = qk_37_transpose_x_0, transpose_y = qk_37_transpose_y_0, x = transpose_182, y = transpose_181); + tensor var_2113_cast = softmax(axis = var_2048, x = qk_37_cast); + tensor var_2115_transpose_x_0 = const()[name = tensor("op_2115_transpose_x_0"), val = tensor(false)]; + tensor var_2115_transpose_y_0 = const()[name = tensor("op_2115_transpose_y_0"), val = tensor(false)]; + tensor transpose_183 = transpose(perm = var_2109, x = var_2108_cast); + tensor var_2115_cast = matmul(transpose_x = var_2115_transpose_x_0, transpose_y = var_2115_transpose_y_0, x = var_2113_cast, y = transpose_183); + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor([0, 2, 1, 3])]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 1500, 1280])]; + tensor transpose_180 = transpose(perm = var_2116, x = var_2115_cast); + tensor x_227_cast = reshape(shape = concat_18, x = transpose_180); + tensor var_2121_to_fp16 = const()[name = tensor("op_2121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732540288)))]; + tensor var_2122_to_fp16 = const()[name = tensor("op_2122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735817152)))]; + tensor var_2123_cast = linear(bias = var_2122_to_fp16, weight = var_2121_to_fp16, x = x_227_cast); + tensor x_229_cast = add(x = x_223_cast, y = var_2123_cast); + tensor var_2129_axes_0 = const()[name = tensor("op_2129_axes_0"), val = tensor([-1])]; + tensor blocks_18_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735819776)))]; + tensor blocks_18_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735822400)))]; + tensor var_2129_cast = layer_norm(axes = var_2129_axes_0, beta = blocks_18_mlp_ln_bias_to_fp16, epsilon = var_2054_to_fp16, gamma = blocks_18_mlp_ln_weight_to_fp16, x = x_229_cast); + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735825024)))]; + tensor var_2139_to_fp16 = const()[name = tensor("op_2139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748932288)))]; + tensor input_153_cast = linear(bias = var_2139_to_fp16, weight = var_2138_to_fp16, x = var_2129_cast); + tensor x_233_mode_0 = const()[name = tensor("x_233_mode_0"), val = tensor("EXACT")]; + tensor x_233_cast = gelu(mode = x_233_mode_0, x = input_153_cast); + tensor var_2144_to_fp16 = const()[name = tensor("op_2144_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748942592)))]; + tensor var_2145_to_fp16 = const()[name = tensor("op_2145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762049856)))]; + tensor var_2146_cast = linear(bias = var_2145_to_fp16, weight = var_2144_to_fp16, x = x_233_cast); + tensor x_235_cast = add(x = x_229_cast, y = var_2146_cast); + tensor var_2155 = const()[name = tensor("op_2155"), val = tensor(-1)]; + tensor var_2172_axes_0 = const()[name = tensor("op_2172_axes_0"), val = tensor([-1])]; + tensor blocks_19_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_19_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762052480)))]; + tensor blocks_19_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_19_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762055104)))]; + tensor var_2161_to_fp16 = const()[name = tensor("op_2161_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2172_cast = layer_norm(axes = var_2172_axes_0, beta = blocks_19_attn_ln_bias_to_fp16, epsilon = var_2161_to_fp16, gamma = blocks_19_attn_ln_weight_to_fp16, x = x_235_cast); + tensor var_2183_to_fp16 = const()[name = tensor("op_2183_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762057728)))]; + tensor var_2184_to_fp16 = const()[name = tensor("op_2184_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765334592)))]; + tensor q_77_cast = linear(bias = var_2184_to_fp16, weight = var_2183_to_fp16, x = var_2172_cast); + tensor var_2187_to_fp16 = const()[name = tensor("op_2187_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765337216)))]; + tensor k_77_bias_0_to_fp16 = const()[name = tensor("k_77_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768614080)))]; + tensor k_77_cast = linear(bias = k_77_bias_0_to_fp16, weight = var_2187_to_fp16, x = var_2172_cast); + tensor var_2191_to_fp16 = const()[name = tensor("op_2191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768616704)))]; + tensor var_2192_to_fp16 = const()[name = tensor("op_2192_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771893568)))]; + tensor v_77_cast = linear(bias = var_2192_to_fp16, weight = var_2191_to_fp16, x = var_2172_cast); + tensor var_2200 = const()[name = tensor("op_2200"), val = tensor([1, 1500, 20, -1])]; + tensor var_2201_cast = reshape(shape = var_2200, x = q_77_cast); + tensor const_262_to_fp16 = const()[name = tensor("const_262_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_79_cast = mul(x = var_2201_cast, y = const_262_to_fp16); + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([1, 1500, 20, -1])]; + tensor var_2208_cast = reshape(shape = var_2207, x = k_77_cast); + tensor const_263_to_fp16 = const()[name = tensor("const_263_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_79_cast = mul(x = var_2208_cast, y = const_263_to_fp16); + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1500, 20, -1])]; + tensor var_2215_cast = reshape(shape = var_2214, x = v_77_cast); + tensor var_2216 = const()[name = tensor("op_2216"), val = tensor([0, 2, 1, 3])]; + tensor qk_39_transpose_x_0 = const()[name = tensor("qk_39_transpose_x_0"), val = tensor(false)]; + tensor qk_39_transpose_y_0 = const()[name = tensor("qk_39_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_177 = transpose(perm = transpose_103_perm_0, x = k_79_cast); + tensor transpose_178 = transpose(perm = transpose_102_perm_0, x = q_79_cast); + tensor qk_39_cast = matmul(transpose_x = qk_39_transpose_x_0, transpose_y = qk_39_transpose_y_0, x = transpose_178, y = transpose_177); + tensor var_2220_cast = softmax(axis = var_2155, x = qk_39_cast); + tensor var_2222_transpose_x_0 = const()[name = tensor("op_2222_transpose_x_0"), val = tensor(false)]; + tensor var_2222_transpose_y_0 = const()[name = tensor("op_2222_transpose_y_0"), val = tensor(false)]; + tensor transpose_179 = transpose(perm = var_2216, x = var_2215_cast); + tensor var_2222_cast = matmul(transpose_x = var_2222_transpose_x_0, transpose_y = var_2222_transpose_y_0, x = var_2220_cast, y = transpose_179); + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([0, 2, 1, 3])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([1, 1500, 1280])]; + tensor transpose_176 = transpose(perm = var_2223, x = var_2222_cast); + tensor x_239_cast = reshape(shape = concat_19, x = transpose_176); + tensor var_2228_to_fp16 = const()[name = tensor("op_2228_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771896192)))]; + tensor var_2229_to_fp16 = const()[name = tensor("op_2229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775173056)))]; + tensor var_2230_cast = linear(bias = var_2229_to_fp16, weight = var_2228_to_fp16, x = x_239_cast); + tensor x_241_cast = add(x = x_235_cast, y = var_2230_cast); + tensor var_2236_axes_0 = const()[name = tensor("op_2236_axes_0"), val = tensor([-1])]; + tensor blocks_19_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775175680)))]; + tensor blocks_19_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775178304)))]; + tensor var_2236_cast = layer_norm(axes = var_2236_axes_0, beta = blocks_19_mlp_ln_bias_to_fp16, epsilon = var_2161_to_fp16, gamma = blocks_19_mlp_ln_weight_to_fp16, x = x_241_cast); + tensor var_2245_to_fp16 = const()[name = tensor("op_2245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775180928)))]; + tensor var_2246_to_fp16 = const()[name = tensor("op_2246_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788288192)))]; + tensor input_161_cast = linear(bias = var_2246_to_fp16, weight = var_2245_to_fp16, x = var_2236_cast); + tensor x_245_mode_0 = const()[name = tensor("x_245_mode_0"), val = tensor("EXACT")]; + tensor x_245_cast = gelu(mode = x_245_mode_0, x = input_161_cast); + tensor var_2251_to_fp16 = const()[name = tensor("op_2251_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788298496)))]; + tensor var_2252_to_fp16 = const()[name = tensor("op_2252_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801405760)))]; + tensor var_2253_cast = linear(bias = var_2252_to_fp16, weight = var_2251_to_fp16, x = x_245_cast); + tensor x_247_cast = add(x = x_241_cast, y = var_2253_cast); + tensor var_2262 = const()[name = tensor("op_2262"), val = tensor(-1)]; + tensor var_2279_axes_0 = const()[name = tensor("op_2279_axes_0"), val = tensor([-1])]; + tensor blocks_20_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_20_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801408384)))]; + tensor blocks_20_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_20_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801411008)))]; + tensor var_2268_to_fp16 = const()[name = tensor("op_2268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2279_cast = layer_norm(axes = var_2279_axes_0, beta = blocks_20_attn_ln_bias_to_fp16, epsilon = var_2268_to_fp16, gamma = blocks_20_attn_ln_weight_to_fp16, x = x_247_cast); + tensor var_2290_to_fp16 = const()[name = tensor("op_2290_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801413632)))]; + tensor var_2291_to_fp16 = const()[name = tensor("op_2291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804690496)))]; + tensor q_81_cast = linear(bias = var_2291_to_fp16, weight = var_2290_to_fp16, x = var_2279_cast); + tensor var_2294_to_fp16 = const()[name = tensor("op_2294_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804693120)))]; + tensor k_81_bias_0_to_fp16 = const()[name = tensor("k_81_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807969984)))]; + tensor k_81_cast = linear(bias = k_81_bias_0_to_fp16, weight = var_2294_to_fp16, x = var_2279_cast); + tensor var_2298_to_fp16 = const()[name = tensor("op_2298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807972608)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811249472)))]; + tensor v_81_cast = linear(bias = var_2299_to_fp16, weight = var_2298_to_fp16, x = var_2279_cast); + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 1500, 20, -1])]; + tensor var_2308_cast = reshape(shape = var_2307, x = q_81_cast); + tensor const_264_to_fp16 = const()[name = tensor("const_264_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_83_cast = mul(x = var_2308_cast, y = const_264_to_fp16); + tensor var_2314 = const()[name = tensor("op_2314"), val = tensor([1, 1500, 20, -1])]; + tensor var_2315_cast = reshape(shape = var_2314, x = k_81_cast); + tensor const_265_to_fp16 = const()[name = tensor("const_265_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_83_cast = mul(x = var_2315_cast, y = const_265_to_fp16); + tensor var_2321 = const()[name = tensor("op_2321"), val = tensor([1, 1500, 20, -1])]; + tensor var_2322_cast = reshape(shape = var_2321, x = v_81_cast); + tensor var_2323 = const()[name = tensor("op_2323"), val = tensor([0, 2, 1, 3])]; + tensor qk_41_transpose_x_0 = const()[name = tensor("qk_41_transpose_x_0"), val = tensor(false)]; + tensor qk_41_transpose_y_0 = const()[name = tensor("qk_41_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_173 = transpose(perm = transpose_105_perm_0, x = k_83_cast); + tensor transpose_174 = transpose(perm = transpose_104_perm_0, x = q_83_cast); + tensor qk_41_cast = matmul(transpose_x = qk_41_transpose_x_0, transpose_y = qk_41_transpose_y_0, x = transpose_174, y = transpose_173); + tensor var_2327_cast = softmax(axis = var_2262, x = qk_41_cast); + tensor var_2329_transpose_x_0 = const()[name = tensor("op_2329_transpose_x_0"), val = tensor(false)]; + tensor var_2329_transpose_y_0 = const()[name = tensor("op_2329_transpose_y_0"), val = tensor(false)]; + tensor transpose_175 = transpose(perm = var_2323, x = var_2322_cast); + tensor var_2329_cast = matmul(transpose_x = var_2329_transpose_x_0, transpose_y = var_2329_transpose_y_0, x = var_2327_cast, y = transpose_175); + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 1500, 1280])]; + tensor transpose_172 = transpose(perm = var_2330, x = var_2329_cast); + tensor x_251_cast = reshape(shape = concat_20, x = transpose_172); + tensor var_2335_to_fp16 = const()[name = tensor("op_2335_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811252096)))]; + tensor var_2336_to_fp16 = const()[name = tensor("op_2336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814528960)))]; + tensor var_2337_cast = linear(bias = var_2336_to_fp16, weight = var_2335_to_fp16, x = x_251_cast); + tensor x_253_cast = add(x = x_247_cast, y = var_2337_cast); + tensor var_2343_axes_0 = const()[name = tensor("op_2343_axes_0"), val = tensor([-1])]; + tensor blocks_20_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814531584)))]; + tensor blocks_20_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814534208)))]; + tensor var_2343_cast = layer_norm(axes = var_2343_axes_0, beta = blocks_20_mlp_ln_bias_to_fp16, epsilon = var_2268_to_fp16, gamma = blocks_20_mlp_ln_weight_to_fp16, x = x_253_cast); + tensor var_2352_to_fp16 = const()[name = tensor("op_2352_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814536832)))]; + tensor var_2353_to_fp16 = const()[name = tensor("op_2353_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827644096)))]; + tensor input_169_cast = linear(bias = var_2353_to_fp16, weight = var_2352_to_fp16, x = var_2343_cast); + tensor x_257_mode_0 = const()[name = tensor("x_257_mode_0"), val = tensor("EXACT")]; + tensor x_257_cast = gelu(mode = x_257_mode_0, x = input_169_cast); + tensor var_2358_to_fp16 = const()[name = tensor("op_2358_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827654400)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840761664)))]; + tensor var_2360_cast = linear(bias = var_2359_to_fp16, weight = var_2358_to_fp16, x = x_257_cast); + tensor x_259_cast = add(x = x_253_cast, y = var_2360_cast); + tensor var_2369 = const()[name = tensor("op_2369"), val = tensor(-1)]; + tensor var_2386_axes_0 = const()[name = tensor("op_2386_axes_0"), val = tensor([-1])]; + tensor blocks_21_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_21_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840764288)))]; + tensor blocks_21_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_21_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840766912)))]; + tensor var_2375_to_fp16 = const()[name = tensor("op_2375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2386_cast = layer_norm(axes = var_2386_axes_0, beta = blocks_21_attn_ln_bias_to_fp16, epsilon = var_2375_to_fp16, gamma = blocks_21_attn_ln_weight_to_fp16, x = x_259_cast); + tensor var_2397_to_fp16 = const()[name = tensor("op_2397_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840769536)))]; + tensor var_2398_to_fp16 = const()[name = tensor("op_2398_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844046400)))]; + tensor q_85_cast = linear(bias = var_2398_to_fp16, weight = var_2397_to_fp16, x = var_2386_cast); + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844049024)))]; + tensor k_85_bias_0_to_fp16 = const()[name = tensor("k_85_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847325888)))]; + tensor k_85_cast = linear(bias = k_85_bias_0_to_fp16, weight = var_2401_to_fp16, x = var_2386_cast); + tensor var_2405_to_fp16 = const()[name = tensor("op_2405_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847328512)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850605376)))]; + tensor v_85_cast = linear(bias = var_2406_to_fp16, weight = var_2405_to_fp16, x = var_2386_cast); + tensor var_2414 = const()[name = tensor("op_2414"), val = tensor([1, 1500, 20, -1])]; + tensor var_2415_cast = reshape(shape = var_2414, x = q_85_cast); + tensor const_266_to_fp16 = const()[name = tensor("const_266_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_87_cast = mul(x = var_2415_cast, y = const_266_to_fp16); + tensor var_2421 = const()[name = tensor("op_2421"), val = tensor([1, 1500, 20, -1])]; + tensor var_2422_cast = reshape(shape = var_2421, x = k_85_cast); + tensor const_267_to_fp16 = const()[name = tensor("const_267_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_87_cast = mul(x = var_2422_cast, y = const_267_to_fp16); + tensor var_2428 = const()[name = tensor("op_2428"), val = tensor([1, 1500, 20, -1])]; + tensor var_2429_cast = reshape(shape = var_2428, x = v_85_cast); + tensor var_2430 = const()[name = tensor("op_2430"), val = tensor([0, 2, 1, 3])]; + tensor qk_43_transpose_x_0 = const()[name = tensor("qk_43_transpose_x_0"), val = tensor(false)]; + tensor qk_43_transpose_y_0 = const()[name = tensor("qk_43_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_169 = transpose(perm = transpose_107_perm_0, x = k_87_cast); + tensor transpose_170 = transpose(perm = transpose_106_perm_0, x = q_87_cast); + tensor qk_43_cast = matmul(transpose_x = qk_43_transpose_x_0, transpose_y = qk_43_transpose_y_0, x = transpose_170, y = transpose_169); + tensor var_2434_cast = softmax(axis = var_2369, x = qk_43_cast); + tensor var_2436_transpose_x_0 = const()[name = tensor("op_2436_transpose_x_0"), val = tensor(false)]; + tensor var_2436_transpose_y_0 = const()[name = tensor("op_2436_transpose_y_0"), val = tensor(false)]; + tensor transpose_171 = transpose(perm = var_2430, x = var_2429_cast); + tensor var_2436_cast = matmul(transpose_x = var_2436_transpose_x_0, transpose_y = var_2436_transpose_y_0, x = var_2434_cast, y = transpose_171); + tensor var_2437 = const()[name = tensor("op_2437"), val = tensor([0, 2, 1, 3])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([1, 1500, 1280])]; + tensor transpose_168 = transpose(perm = var_2437, x = var_2436_cast); + tensor x_263_cast = reshape(shape = concat_21, x = transpose_168); + tensor var_2442_to_fp16 = const()[name = tensor("op_2442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850608000)))]; + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853884864)))]; + tensor var_2444_cast = linear(bias = var_2443_to_fp16, weight = var_2442_to_fp16, x = x_263_cast); + tensor x_265_cast = add(x = x_259_cast, y = var_2444_cast); + tensor var_2450_axes_0 = const()[name = tensor("op_2450_axes_0"), val = tensor([-1])]; + tensor blocks_21_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853887488)))]; + tensor blocks_21_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853890112)))]; + tensor var_2450_cast = layer_norm(axes = var_2450_axes_0, beta = blocks_21_mlp_ln_bias_to_fp16, epsilon = var_2375_to_fp16, gamma = blocks_21_mlp_ln_weight_to_fp16, x = x_265_cast); + tensor var_2459_to_fp16 = const()[name = tensor("op_2459_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853892736)))]; + tensor var_2460_to_fp16 = const()[name = tensor("op_2460_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867000000)))]; + tensor input_177_cast = linear(bias = var_2460_to_fp16, weight = var_2459_to_fp16, x = var_2450_cast); + tensor x_269_mode_0 = const()[name = tensor("x_269_mode_0"), val = tensor("EXACT")]; + tensor x_269_cast = gelu(mode = x_269_mode_0, x = input_177_cast); + tensor var_2465_to_fp16 = const()[name = tensor("op_2465_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867010304)))]; + tensor var_2466_to_fp16 = const()[name = tensor("op_2466_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880117568)))]; + tensor var_2467_cast = linear(bias = var_2466_to_fp16, weight = var_2465_to_fp16, x = x_269_cast); + tensor x_271_cast = add(x = x_265_cast, y = var_2467_cast); + tensor var_2476 = const()[name = tensor("op_2476"), val = tensor(-1)]; + tensor var_2493_axes_0 = const()[name = tensor("op_2493_axes_0"), val = tensor([-1])]; + tensor blocks_22_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_22_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880120192)))]; + tensor blocks_22_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_22_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880122816)))]; + tensor var_2482_to_fp16 = const()[name = tensor("op_2482_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2493_cast = layer_norm(axes = var_2493_axes_0, beta = blocks_22_attn_ln_bias_to_fp16, epsilon = var_2482_to_fp16, gamma = blocks_22_attn_ln_weight_to_fp16, x = x_271_cast); + tensor var_2504_to_fp16 = const()[name = tensor("op_2504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880125440)))]; + tensor var_2505_to_fp16 = const()[name = tensor("op_2505_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883402304)))]; + tensor q_89_cast = linear(bias = var_2505_to_fp16, weight = var_2504_to_fp16, x = var_2493_cast); + tensor var_2508_to_fp16 = const()[name = tensor("op_2508_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883404928)))]; + tensor k_89_bias_0_to_fp16 = const()[name = tensor("k_89_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886681792)))]; + tensor k_89_cast = linear(bias = k_89_bias_0_to_fp16, weight = var_2508_to_fp16, x = var_2493_cast); + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886684416)))]; + tensor var_2513_to_fp16 = const()[name = tensor("op_2513_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889961280)))]; + tensor v_89_cast = linear(bias = var_2513_to_fp16, weight = var_2512_to_fp16, x = var_2493_cast); + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([1, 1500, 20, -1])]; + tensor var_2522_cast = reshape(shape = var_2521, x = q_89_cast); + tensor const_268_to_fp16 = const()[name = tensor("const_268_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_91_cast = mul(x = var_2522_cast, y = const_268_to_fp16); + tensor var_2528 = const()[name = tensor("op_2528"), val = tensor([1, 1500, 20, -1])]; + tensor var_2529_cast = reshape(shape = var_2528, x = k_89_cast); + tensor const_269_to_fp16 = const()[name = tensor("const_269_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_91_cast = mul(x = var_2529_cast, y = const_269_to_fp16); + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor([1, 1500, 20, -1])]; + tensor var_2536_cast = reshape(shape = var_2535, x = v_89_cast); + tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([0, 2, 1, 3])]; + tensor qk_45_transpose_x_0 = const()[name = tensor("qk_45_transpose_x_0"), val = tensor(false)]; + tensor qk_45_transpose_y_0 = const()[name = tensor("qk_45_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_165 = transpose(perm = transpose_109_perm_0, x = k_91_cast); + tensor transpose_166 = transpose(perm = transpose_108_perm_0, x = q_91_cast); + tensor qk_45_cast = matmul(transpose_x = qk_45_transpose_x_0, transpose_y = qk_45_transpose_y_0, x = transpose_166, y = transpose_165); + tensor var_2541_cast = softmax(axis = var_2476, x = qk_45_cast); + tensor var_2543_transpose_x_0 = const()[name = tensor("op_2543_transpose_x_0"), val = tensor(false)]; + tensor var_2543_transpose_y_0 = const()[name = tensor("op_2543_transpose_y_0"), val = tensor(false)]; + tensor transpose_167 = transpose(perm = var_2537, x = var_2536_cast); + tensor var_2543_cast = matmul(transpose_x = var_2543_transpose_x_0, transpose_y = var_2543_transpose_y_0, x = var_2541_cast, y = transpose_167); + tensor var_2544 = const()[name = tensor("op_2544"), val = tensor([0, 2, 1, 3])]; + tensor concat_22 = const()[name = tensor("concat_22"), val = tensor([1, 1500, 1280])]; + tensor transpose_164 = transpose(perm = var_2544, x = var_2543_cast); + tensor x_275_cast = reshape(shape = concat_22, x = transpose_164); + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889963904)))]; + tensor var_2550_to_fp16 = const()[name = tensor("op_2550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893240768)))]; + tensor var_2551_cast = linear(bias = var_2550_to_fp16, weight = var_2549_to_fp16, x = x_275_cast); + tensor x_277_cast = add(x = x_271_cast, y = var_2551_cast); + tensor var_2557_axes_0 = const()[name = tensor("op_2557_axes_0"), val = tensor([-1])]; + tensor blocks_22_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893243392)))]; + tensor blocks_22_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893246016)))]; + tensor var_2557_cast = layer_norm(axes = var_2557_axes_0, beta = blocks_22_mlp_ln_bias_to_fp16, epsilon = var_2482_to_fp16, gamma = blocks_22_mlp_ln_weight_to_fp16, x = x_277_cast); + tensor var_2566_to_fp16 = const()[name = tensor("op_2566_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893248640)))]; + tensor var_2567_to_fp16 = const()[name = tensor("op_2567_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906355904)))]; + tensor input_185_cast = linear(bias = var_2567_to_fp16, weight = var_2566_to_fp16, x = var_2557_cast); + tensor x_281_mode_0 = const()[name = tensor("x_281_mode_0"), val = tensor("EXACT")]; + tensor x_281_cast = gelu(mode = x_281_mode_0, x = input_185_cast); + tensor var_2572_to_fp16 = const()[name = tensor("op_2572_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906366208)))]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919473472)))]; + tensor var_2574_cast = linear(bias = var_2573_to_fp16, weight = var_2572_to_fp16, x = x_281_cast); + tensor x_283_cast = add(x = x_277_cast, y = var_2574_cast); + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor var_2600_axes_0 = const()[name = tensor("op_2600_axes_0"), val = tensor([-1])]; + tensor blocks_23_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_23_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919476096)))]; + tensor blocks_23_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_23_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919478720)))]; + tensor var_2589_to_fp16 = const()[name = tensor("op_2589_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2600_cast = layer_norm(axes = var_2600_axes_0, beta = blocks_23_attn_ln_bias_to_fp16, epsilon = var_2589_to_fp16, gamma = blocks_23_attn_ln_weight_to_fp16, x = x_283_cast); + tensor var_2611_to_fp16 = const()[name = tensor("op_2611_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919481344)))]; + tensor var_2612_to_fp16 = const()[name = tensor("op_2612_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922758208)))]; + tensor q_93_cast = linear(bias = var_2612_to_fp16, weight = var_2611_to_fp16, x = var_2600_cast); + tensor var_2615_to_fp16 = const()[name = tensor("op_2615_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922760832)))]; + tensor k_93_bias_0_to_fp16 = const()[name = tensor("k_93_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926037696)))]; + tensor k_93_cast = linear(bias = k_93_bias_0_to_fp16, weight = var_2615_to_fp16, x = var_2600_cast); + tensor var_2619_to_fp16 = const()[name = tensor("op_2619_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926040320)))]; + tensor var_2620_to_fp16 = const()[name = tensor("op_2620_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929317184)))]; + tensor v_93_cast = linear(bias = var_2620_to_fp16, weight = var_2619_to_fp16, x = var_2600_cast); + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 1500, 20, -1])]; + tensor var_2629_cast = reshape(shape = var_2628, x = q_93_cast); + tensor const_270_to_fp16 = const()[name = tensor("const_270_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_95_cast = mul(x = var_2629_cast, y = const_270_to_fp16); + tensor var_2635 = const()[name = tensor("op_2635"), val = tensor([1, 1500, 20, -1])]; + tensor var_2636_cast = reshape(shape = var_2635, x = k_93_cast); + tensor const_271_to_fp16 = const()[name = tensor("const_271_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_95_cast = mul(x = var_2636_cast, y = const_271_to_fp16); + tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 1500, 20, -1])]; + tensor var_2643_cast = reshape(shape = var_2642, x = v_93_cast); + tensor var_2644 = const()[name = tensor("op_2644"), val = tensor([0, 2, 1, 3])]; + tensor qk_47_transpose_x_0 = const()[name = tensor("qk_47_transpose_x_0"), val = tensor(false)]; + tensor qk_47_transpose_y_0 = const()[name = tensor("qk_47_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_161 = transpose(perm = transpose_111_perm_0, x = k_95_cast); + tensor transpose_162 = transpose(perm = transpose_110_perm_0, x = q_95_cast); + tensor qk_47_cast = matmul(transpose_x = qk_47_transpose_x_0, transpose_y = qk_47_transpose_y_0, x = transpose_162, y = transpose_161); + tensor var_2648_cast = softmax(axis = var_2583, x = qk_47_cast); + tensor var_2650_transpose_x_0 = const()[name = tensor("op_2650_transpose_x_0"), val = tensor(false)]; + tensor var_2650_transpose_y_0 = const()[name = tensor("op_2650_transpose_y_0"), val = tensor(false)]; + tensor transpose_163 = transpose(perm = var_2644, x = var_2643_cast); + tensor var_2650_cast = matmul(transpose_x = var_2650_transpose_x_0, transpose_y = var_2650_transpose_y_0, x = var_2648_cast, y = transpose_163); + tensor var_2651 = const()[name = tensor("op_2651"), val = tensor([0, 2, 1, 3])]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([1, 1500, 1280])]; + tensor transpose_160 = transpose(perm = var_2651, x = var_2650_cast); + tensor x_287_cast = reshape(shape = concat_23, x = transpose_160); + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929319808)))]; + tensor var_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932596672)))]; + tensor var_2658_cast = linear(bias = var_2657_to_fp16, weight = var_2656_to_fp16, x = x_287_cast); + tensor x_289_cast = add(x = x_283_cast, y = var_2658_cast); + tensor var_2664_axes_0 = const()[name = tensor("op_2664_axes_0"), val = tensor([-1])]; + tensor blocks_23_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932599296)))]; + tensor blocks_23_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932601920)))]; + tensor var_2664_cast = layer_norm(axes = var_2664_axes_0, beta = blocks_23_mlp_ln_bias_to_fp16, epsilon = var_2589_to_fp16, gamma = blocks_23_mlp_ln_weight_to_fp16, x = x_289_cast); + tensor var_2673_to_fp16 = const()[name = tensor("op_2673_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932604544)))]; + tensor var_2674_to_fp16 = const()[name = tensor("op_2674_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945711808)))]; + tensor input_193_cast = linear(bias = var_2674_to_fp16, weight = var_2673_to_fp16, x = var_2664_cast); + tensor x_293_mode_0 = const()[name = tensor("x_293_mode_0"), val = tensor("EXACT")]; + tensor x_293_cast = gelu(mode = x_293_mode_0, x = input_193_cast); + tensor var_2679_to_fp16 = const()[name = tensor("op_2679_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945722112)))]; + tensor var_2680_to_fp16 = const()[name = tensor("op_2680_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958829376)))]; + tensor var_2681_cast = linear(bias = var_2680_to_fp16, weight = var_2679_to_fp16, x = x_293_cast); + tensor x_295_cast = add(x = x_289_cast, y = var_2681_cast); + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor(-1)]; + tensor var_2707_axes_0 = const()[name = tensor("op_2707_axes_0"), val = tensor([-1])]; + tensor blocks_24_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_24_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958832000)))]; + tensor blocks_24_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_24_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958834624)))]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2707_cast = layer_norm(axes = var_2707_axes_0, beta = blocks_24_attn_ln_bias_to_fp16, epsilon = var_2696_to_fp16, gamma = blocks_24_attn_ln_weight_to_fp16, x = x_295_cast); + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958837248)))]; + tensor var_2719_to_fp16 = const()[name = tensor("op_2719_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962114112)))]; + tensor q_97_cast = linear(bias = var_2719_to_fp16, weight = var_2718_to_fp16, x = var_2707_cast); + tensor var_2722_to_fp16 = const()[name = tensor("op_2722_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962116736)))]; + tensor k_97_bias_0_to_fp16 = const()[name = tensor("k_97_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965393600)))]; + tensor k_97_cast = linear(bias = k_97_bias_0_to_fp16, weight = var_2722_to_fp16, x = var_2707_cast); + tensor var_2726_to_fp16 = const()[name = tensor("op_2726_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965396224)))]; + tensor var_2727_to_fp16 = const()[name = tensor("op_2727_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968673088)))]; + tensor v_97_cast = linear(bias = var_2727_to_fp16, weight = var_2726_to_fp16, x = var_2707_cast); + tensor var_2735 = const()[name = tensor("op_2735"), val = tensor([1, 1500, 20, -1])]; + tensor var_2736_cast = reshape(shape = var_2735, x = q_97_cast); + tensor const_272_to_fp16 = const()[name = tensor("const_272_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_99_cast = mul(x = var_2736_cast, y = const_272_to_fp16); + tensor var_2742 = const()[name = tensor("op_2742"), val = tensor([1, 1500, 20, -1])]; + tensor var_2743_cast = reshape(shape = var_2742, x = k_97_cast); + tensor const_273_to_fp16 = const()[name = tensor("const_273_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_99_cast = mul(x = var_2743_cast, y = const_273_to_fp16); + tensor var_2749 = const()[name = tensor("op_2749"), val = tensor([1, 1500, 20, -1])]; + tensor var_2750_cast = reshape(shape = var_2749, x = v_97_cast); + tensor var_2751 = const()[name = tensor("op_2751"), val = tensor([0, 2, 1, 3])]; + tensor qk_49_transpose_x_0 = const()[name = tensor("qk_49_transpose_x_0"), val = tensor(false)]; + tensor qk_49_transpose_y_0 = const()[name = tensor("qk_49_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_157 = transpose(perm = transpose_113_perm_0, x = k_99_cast); + tensor transpose_158 = transpose(perm = transpose_112_perm_0, x = q_99_cast); + tensor qk_49_cast = matmul(transpose_x = qk_49_transpose_x_0, transpose_y = qk_49_transpose_y_0, x = transpose_158, y = transpose_157); + tensor var_2755_cast = softmax(axis = var_2690, x = qk_49_cast); + tensor var_2757_transpose_x_0 = const()[name = tensor("op_2757_transpose_x_0"), val = tensor(false)]; + tensor var_2757_transpose_y_0 = const()[name = tensor("op_2757_transpose_y_0"), val = tensor(false)]; + tensor transpose_159 = transpose(perm = var_2751, x = var_2750_cast); + tensor var_2757_cast = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = var_2755_cast, y = transpose_159); + tensor var_2758 = const()[name = tensor("op_2758"), val = tensor([0, 2, 1, 3])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([1, 1500, 1280])]; + tensor transpose_156 = transpose(perm = var_2758, x = var_2757_cast); + tensor x_299_cast = reshape(shape = concat_24, x = transpose_156); + tensor var_2763_to_fp16 = const()[name = tensor("op_2763_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968675712)))]; + tensor var_2764_to_fp16 = const()[name = tensor("op_2764_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971952576)))]; + tensor var_2765_cast = linear(bias = var_2764_to_fp16, weight = var_2763_to_fp16, x = x_299_cast); + tensor x_301_cast = add(x = x_295_cast, y = var_2765_cast); + tensor var_2771_axes_0 = const()[name = tensor("op_2771_axes_0"), val = tensor([-1])]; + tensor blocks_24_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971955200)))]; + tensor blocks_24_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971957824)))]; + tensor var_2771_cast = layer_norm(axes = var_2771_axes_0, beta = blocks_24_mlp_ln_bias_to_fp16, epsilon = var_2696_to_fp16, gamma = blocks_24_mlp_ln_weight_to_fp16, x = x_301_cast); + tensor var_2780_to_fp16 = const()[name = tensor("op_2780_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971960448)))]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985067712)))]; + tensor input_201_cast = linear(bias = var_2781_to_fp16, weight = var_2780_to_fp16, x = var_2771_cast); + tensor x_305_mode_0 = const()[name = tensor("x_305_mode_0"), val = tensor("EXACT")]; + tensor x_305_cast = gelu(mode = x_305_mode_0, x = input_201_cast); + tensor var_2786_to_fp16 = const()[name = tensor("op_2786_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985078016)))]; + tensor var_2787_to_fp16 = const()[name = tensor("op_2787_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998185280)))]; + tensor var_2788_cast = linear(bias = var_2787_to_fp16, weight = var_2786_to_fp16, x = x_305_cast); + tensor x_307_cast = add(x = x_301_cast, y = var_2788_cast); + tensor var_2797 = const()[name = tensor("op_2797"), val = tensor(-1)]; + tensor var_2814_axes_0 = const()[name = tensor("op_2814_axes_0"), val = tensor([-1])]; + tensor blocks_25_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_25_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998187904)))]; + tensor blocks_25_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_25_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998190528)))]; + tensor var_2803_to_fp16 = const()[name = tensor("op_2803_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2814_cast = layer_norm(axes = var_2814_axes_0, beta = blocks_25_attn_ln_bias_to_fp16, epsilon = var_2803_to_fp16, gamma = blocks_25_attn_ln_weight_to_fp16, x = x_307_cast); + tensor var_2825_to_fp16 = const()[name = tensor("op_2825_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998193152)))]; + tensor var_2826_to_fp16 = const()[name = tensor("op_2826_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001470016)))]; + tensor q_101_cast = linear(bias = var_2826_to_fp16, weight = var_2825_to_fp16, x = var_2814_cast); + tensor var_2829_to_fp16 = const()[name = tensor("op_2829_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001472640)))]; + tensor k_101_bias_0_to_fp16 = const()[name = tensor("k_101_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1004749504)))]; + tensor k_101_cast = linear(bias = k_101_bias_0_to_fp16, weight = var_2829_to_fp16, x = var_2814_cast); + tensor var_2833_to_fp16 = const()[name = tensor("op_2833_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1004752128)))]; + tensor var_2834_to_fp16 = const()[name = tensor("op_2834_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008028992)))]; + tensor v_101_cast = linear(bias = var_2834_to_fp16, weight = var_2833_to_fp16, x = var_2814_cast); + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 1500, 20, -1])]; + tensor var_2843_cast = reshape(shape = var_2842, x = q_101_cast); + tensor const_274_to_fp16 = const()[name = tensor("const_274_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_103_cast = mul(x = var_2843_cast, y = const_274_to_fp16); + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1500, 20, -1])]; + tensor var_2850_cast = reshape(shape = var_2849, x = k_101_cast); + tensor const_275_to_fp16 = const()[name = tensor("const_275_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_103_cast = mul(x = var_2850_cast, y = const_275_to_fp16); + tensor var_2856 = const()[name = tensor("op_2856"), val = tensor([1, 1500, 20, -1])]; + tensor var_2857_cast = reshape(shape = var_2856, x = v_101_cast); + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([0, 2, 1, 3])]; + tensor qk_51_transpose_x_0 = const()[name = tensor("qk_51_transpose_x_0"), val = tensor(false)]; + tensor qk_51_transpose_y_0 = const()[name = tensor("qk_51_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_153 = transpose(perm = transpose_115_perm_0, x = k_103_cast); + tensor transpose_154 = transpose(perm = transpose_114_perm_0, x = q_103_cast); + tensor qk_51_cast = matmul(transpose_x = qk_51_transpose_x_0, transpose_y = qk_51_transpose_y_0, x = transpose_154, y = transpose_153); + tensor var_2862_cast = softmax(axis = var_2797, x = qk_51_cast); + tensor var_2864_transpose_x_0 = const()[name = tensor("op_2864_transpose_x_0"), val = tensor(false)]; + tensor var_2864_transpose_y_0 = const()[name = tensor("op_2864_transpose_y_0"), val = tensor(false)]; + tensor transpose_155 = transpose(perm = var_2858, x = var_2857_cast); + tensor var_2864_cast = matmul(transpose_x = var_2864_transpose_x_0, transpose_y = var_2864_transpose_y_0, x = var_2862_cast, y = transpose_155); + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor([0, 2, 1, 3])]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 1500, 1280])]; + tensor transpose_152 = transpose(perm = var_2865, x = var_2864_cast); + tensor x_311_cast = reshape(shape = concat_25, x = transpose_152); + tensor var_2870_to_fp16 = const()[name = tensor("op_2870_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008031616)))]; + tensor var_2871_to_fp16 = const()[name = tensor("op_2871_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011308480)))]; + tensor var_2872_cast = linear(bias = var_2871_to_fp16, weight = var_2870_to_fp16, x = x_311_cast); + tensor x_313_cast = add(x = x_307_cast, y = var_2872_cast); + tensor var_2878_axes_0 = const()[name = tensor("op_2878_axes_0"), val = tensor([-1])]; + tensor blocks_25_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011311104)))]; + tensor blocks_25_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011313728)))]; + tensor var_2878_cast = layer_norm(axes = var_2878_axes_0, beta = blocks_25_mlp_ln_bias_to_fp16, epsilon = var_2803_to_fp16, gamma = blocks_25_mlp_ln_weight_to_fp16, x = x_313_cast); + tensor var_2887_to_fp16 = const()[name = tensor("op_2887_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011316352)))]; + tensor var_2888_to_fp16 = const()[name = tensor("op_2888_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024423616)))]; + tensor input_209_cast = linear(bias = var_2888_to_fp16, weight = var_2887_to_fp16, x = var_2878_cast); + tensor x_317_mode_0 = const()[name = tensor("x_317_mode_0"), val = tensor("EXACT")]; + tensor x_317_cast = gelu(mode = x_317_mode_0, x = input_209_cast); + tensor var_2893_to_fp16 = const()[name = tensor("op_2893_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024433920)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037541184)))]; + tensor var_2895_cast = linear(bias = var_2894_to_fp16, weight = var_2893_to_fp16, x = x_317_cast); + tensor x_319_cast = add(x = x_313_cast, y = var_2895_cast); + tensor var_2904 = const()[name = tensor("op_2904"), val = tensor(-1)]; + tensor var_2921_axes_0 = const()[name = tensor("op_2921_axes_0"), val = tensor([-1])]; + tensor blocks_26_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_26_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037543808)))]; + tensor blocks_26_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_26_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037546432)))]; + tensor var_2910_to_fp16 = const()[name = tensor("op_2910_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2921_cast = layer_norm(axes = var_2921_axes_0, beta = blocks_26_attn_ln_bias_to_fp16, epsilon = var_2910_to_fp16, gamma = blocks_26_attn_ln_weight_to_fp16, x = x_319_cast); + tensor var_2932_to_fp16 = const()[name = tensor("op_2932_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037549056)))]; + tensor var_2933_to_fp16 = const()[name = tensor("op_2933_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040825920)))]; + tensor q_105_cast = linear(bias = var_2933_to_fp16, weight = var_2932_to_fp16, x = var_2921_cast); + tensor var_2936_to_fp16 = const()[name = tensor("op_2936_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040828544)))]; + tensor k_105_bias_0_to_fp16 = const()[name = tensor("k_105_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044105408)))]; + tensor k_105_cast = linear(bias = k_105_bias_0_to_fp16, weight = var_2936_to_fp16, x = var_2921_cast); + tensor var_2940_to_fp16 = const()[name = tensor("op_2940_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044108032)))]; + tensor var_2941_to_fp16 = const()[name = tensor("op_2941_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047384896)))]; + tensor v_105_cast = linear(bias = var_2941_to_fp16, weight = var_2940_to_fp16, x = var_2921_cast); + tensor var_2949 = const()[name = tensor("op_2949"), val = tensor([1, 1500, 20, -1])]; + tensor var_2950_cast = reshape(shape = var_2949, x = q_105_cast); + tensor const_276_to_fp16 = const()[name = tensor("const_276_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_107_cast = mul(x = var_2950_cast, y = const_276_to_fp16); + tensor var_2956 = const()[name = tensor("op_2956"), val = tensor([1, 1500, 20, -1])]; + tensor var_2957_cast = reshape(shape = var_2956, x = k_105_cast); + tensor const_277_to_fp16 = const()[name = tensor("const_277_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_107_cast = mul(x = var_2957_cast, y = const_277_to_fp16); + tensor var_2963 = const()[name = tensor("op_2963"), val = tensor([1, 1500, 20, -1])]; + tensor var_2964_cast = reshape(shape = var_2963, x = v_105_cast); + tensor var_2965 = const()[name = tensor("op_2965"), val = tensor([0, 2, 1, 3])]; + tensor qk_53_transpose_x_0 = const()[name = tensor("qk_53_transpose_x_0"), val = tensor(false)]; + tensor qk_53_transpose_y_0 = const()[name = tensor("qk_53_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_149 = transpose(perm = transpose_117_perm_0, x = k_107_cast); + tensor transpose_150 = transpose(perm = transpose_116_perm_0, x = q_107_cast); + tensor qk_53_cast = matmul(transpose_x = qk_53_transpose_x_0, transpose_y = qk_53_transpose_y_0, x = transpose_150, y = transpose_149); + tensor var_2969_cast = softmax(axis = var_2904, x = qk_53_cast); + tensor var_2971_transpose_x_0 = const()[name = tensor("op_2971_transpose_x_0"), val = tensor(false)]; + tensor var_2971_transpose_y_0 = const()[name = tensor("op_2971_transpose_y_0"), val = tensor(false)]; + tensor transpose_151 = transpose(perm = var_2965, x = var_2964_cast); + tensor var_2971_cast = matmul(transpose_x = var_2971_transpose_x_0, transpose_y = var_2971_transpose_y_0, x = var_2969_cast, y = transpose_151); + tensor var_2972 = const()[name = tensor("op_2972"), val = tensor([0, 2, 1, 3])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([1, 1500, 1280])]; + tensor transpose_148 = transpose(perm = var_2972, x = var_2971_cast); + tensor x_323_cast = reshape(shape = concat_26, x = transpose_148); + tensor var_2977_to_fp16 = const()[name = tensor("op_2977_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047387520)))]; + tensor var_2978_to_fp16 = const()[name = tensor("op_2978_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050664384)))]; + tensor var_2979_cast = linear(bias = var_2978_to_fp16, weight = var_2977_to_fp16, x = x_323_cast); + tensor x_325_cast = add(x = x_319_cast, y = var_2979_cast); + tensor var_2985_axes_0 = const()[name = tensor("op_2985_axes_0"), val = tensor([-1])]; + tensor blocks_26_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050667008)))]; + tensor blocks_26_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050669632)))]; + tensor var_2985_cast = layer_norm(axes = var_2985_axes_0, beta = blocks_26_mlp_ln_bias_to_fp16, epsilon = var_2910_to_fp16, gamma = blocks_26_mlp_ln_weight_to_fp16, x = x_325_cast); + tensor var_2994_to_fp16 = const()[name = tensor("op_2994_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050672256)))]; + tensor var_2995_to_fp16 = const()[name = tensor("op_2995_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063779520)))]; + tensor input_217_cast = linear(bias = var_2995_to_fp16, weight = var_2994_to_fp16, x = var_2985_cast); + tensor x_329_mode_0 = const()[name = tensor("x_329_mode_0"), val = tensor("EXACT")]; + tensor x_329_cast = gelu(mode = x_329_mode_0, x = input_217_cast); + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063789824)))]; + tensor var_3001_to_fp16 = const()[name = tensor("op_3001_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076897088)))]; + tensor var_3002_cast = linear(bias = var_3001_to_fp16, weight = var_3000_to_fp16, x = x_329_cast); + tensor x_331_cast = add(x = x_325_cast, y = var_3002_cast); + tensor var_3011 = const()[name = tensor("op_3011"), val = tensor(-1)]; + tensor var_3028_axes_0 = const()[name = tensor("op_3028_axes_0"), val = tensor([-1])]; + tensor blocks_27_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_27_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076899712)))]; + tensor blocks_27_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_27_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076902336)))]; + tensor var_3017_to_fp16 = const()[name = tensor("op_3017_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3028_cast = layer_norm(axes = var_3028_axes_0, beta = blocks_27_attn_ln_bias_to_fp16, epsilon = var_3017_to_fp16, gamma = blocks_27_attn_ln_weight_to_fp16, x = x_331_cast); + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076904960)))]; + tensor var_3040_to_fp16 = const()[name = tensor("op_3040_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080181824)))]; + tensor q_109_cast = linear(bias = var_3040_to_fp16, weight = var_3039_to_fp16, x = var_3028_cast); + tensor var_3043_to_fp16 = const()[name = tensor("op_3043_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080184448)))]; + tensor k_109_bias_0_to_fp16 = const()[name = tensor("k_109_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083461312)))]; + tensor k_109_cast = linear(bias = k_109_bias_0_to_fp16, weight = var_3043_to_fp16, x = var_3028_cast); + tensor var_3047_to_fp16 = const()[name = tensor("op_3047_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083463936)))]; + tensor var_3048_to_fp16 = const()[name = tensor("op_3048_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086740800)))]; + tensor v_109_cast = linear(bias = var_3048_to_fp16, weight = var_3047_to_fp16, x = var_3028_cast); + tensor var_3056 = const()[name = tensor("op_3056"), val = tensor([1, 1500, 20, -1])]; + tensor var_3057_cast = reshape(shape = var_3056, x = q_109_cast); + tensor const_278_to_fp16 = const()[name = tensor("const_278_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_111_cast = mul(x = var_3057_cast, y = const_278_to_fp16); + tensor var_3063 = const()[name = tensor("op_3063"), val = tensor([1, 1500, 20, -1])]; + tensor var_3064_cast = reshape(shape = var_3063, x = k_109_cast); + tensor const_279_to_fp16 = const()[name = tensor("const_279_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_111_cast = mul(x = var_3064_cast, y = const_279_to_fp16); + tensor var_3070 = const()[name = tensor("op_3070"), val = tensor([1, 1500, 20, -1])]; + tensor var_3071_cast = reshape(shape = var_3070, x = v_109_cast); + tensor var_3072 = const()[name = tensor("op_3072"), val = tensor([0, 2, 1, 3])]; + tensor qk_55_transpose_x_0 = const()[name = tensor("qk_55_transpose_x_0"), val = tensor(false)]; + tensor qk_55_transpose_y_0 = const()[name = tensor("qk_55_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_145 = transpose(perm = transpose_119_perm_0, x = k_111_cast); + tensor transpose_146 = transpose(perm = transpose_118_perm_0, x = q_111_cast); + tensor qk_55_cast = matmul(transpose_x = qk_55_transpose_x_0, transpose_y = qk_55_transpose_y_0, x = transpose_146, y = transpose_145); + tensor var_3076_cast = softmax(axis = var_3011, x = qk_55_cast); + tensor var_3078_transpose_x_0 = const()[name = tensor("op_3078_transpose_x_0"), val = tensor(false)]; + tensor var_3078_transpose_y_0 = const()[name = tensor("op_3078_transpose_y_0"), val = tensor(false)]; + tensor transpose_147 = transpose(perm = var_3072, x = var_3071_cast); + tensor var_3078_cast = matmul(transpose_x = var_3078_transpose_x_0, transpose_y = var_3078_transpose_y_0, x = var_3076_cast, y = transpose_147); + tensor var_3079 = const()[name = tensor("op_3079"), val = tensor([0, 2, 1, 3])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([1, 1500, 1280])]; + tensor transpose_144 = transpose(perm = var_3079, x = var_3078_cast); + tensor x_335_cast = reshape(shape = concat_27, x = transpose_144); + tensor var_3084_to_fp16 = const()[name = tensor("op_3084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086743424)))]; + tensor var_3085_to_fp16 = const()[name = tensor("op_3085_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090020288)))]; + tensor var_3086_cast = linear(bias = var_3085_to_fp16, weight = var_3084_to_fp16, x = x_335_cast); + tensor x_337_cast = add(x = x_331_cast, y = var_3086_cast); + tensor var_3092_axes_0 = const()[name = tensor("op_3092_axes_0"), val = tensor([-1])]; + tensor blocks_27_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090022912)))]; + tensor blocks_27_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090025536)))]; + tensor var_3092_cast = layer_norm(axes = var_3092_axes_0, beta = blocks_27_mlp_ln_bias_to_fp16, epsilon = var_3017_to_fp16, gamma = blocks_27_mlp_ln_weight_to_fp16, x = x_337_cast); + tensor var_3101_to_fp16 = const()[name = tensor("op_3101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090028160)))]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103135424)))]; + tensor input_225_cast = linear(bias = var_3102_to_fp16, weight = var_3101_to_fp16, x = var_3092_cast); + tensor x_341_mode_0 = const()[name = tensor("x_341_mode_0"), val = tensor("EXACT")]; + tensor x_341_cast = gelu(mode = x_341_mode_0, x = input_225_cast); + tensor var_3107_to_fp16 = const()[name = tensor("op_3107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103145728)))]; + tensor var_3108_to_fp16 = const()[name = tensor("op_3108_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116252992)))]; + tensor var_3109_cast = linear(bias = var_3108_to_fp16, weight = var_3107_to_fp16, x = x_341_cast); + tensor x_343_cast = add(x = x_337_cast, y = var_3109_cast); + tensor var_3118 = const()[name = tensor("op_3118"), val = tensor(-1)]; + tensor var_3135_axes_0 = const()[name = tensor("op_3135_axes_0"), val = tensor([-1])]; + tensor blocks_28_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_28_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116255616)))]; + tensor blocks_28_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_28_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116258240)))]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3135_cast = layer_norm(axes = var_3135_axes_0, beta = blocks_28_attn_ln_bias_to_fp16, epsilon = var_3124_to_fp16, gamma = blocks_28_attn_ln_weight_to_fp16, x = x_343_cast); + tensor var_3146_to_fp16 = const()[name = tensor("op_3146_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116260864)))]; + tensor var_3147_to_fp16 = const()[name = tensor("op_3147_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119537728)))]; + tensor q_113_cast = linear(bias = var_3147_to_fp16, weight = var_3146_to_fp16, x = var_3135_cast); + tensor var_3150_to_fp16 = const()[name = tensor("op_3150_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119540352)))]; + tensor k_113_bias_0_to_fp16 = const()[name = tensor("k_113_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1122817216)))]; + tensor k_113_cast = linear(bias = k_113_bias_0_to_fp16, weight = var_3150_to_fp16, x = var_3135_cast); + tensor var_3154_to_fp16 = const()[name = tensor("op_3154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1122819840)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126096704)))]; + tensor v_113_cast = linear(bias = var_3155_to_fp16, weight = var_3154_to_fp16, x = var_3135_cast); + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 1500, 20, -1])]; + tensor var_3164_cast = reshape(shape = var_3163, x = q_113_cast); + tensor const_280_to_fp16 = const()[name = tensor("const_280_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_115_cast = mul(x = var_3164_cast, y = const_280_to_fp16); + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 1500, 20, -1])]; + tensor var_3171_cast = reshape(shape = var_3170, x = k_113_cast); + tensor const_281_to_fp16 = const()[name = tensor("const_281_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_115_cast = mul(x = var_3171_cast, y = const_281_to_fp16); + tensor var_3177 = const()[name = tensor("op_3177"), val = tensor([1, 1500, 20, -1])]; + tensor var_3178_cast = reshape(shape = var_3177, x = v_113_cast); + tensor var_3179 = const()[name = tensor("op_3179"), val = tensor([0, 2, 1, 3])]; + tensor qk_57_transpose_x_0 = const()[name = tensor("qk_57_transpose_x_0"), val = tensor(false)]; + tensor qk_57_transpose_y_0 = const()[name = tensor("qk_57_transpose_y_0"), val = tensor(false)]; + tensor transpose_120_perm_0 = const()[name = tensor("transpose_120_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_121_perm_0 = const()[name = tensor("transpose_121_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_141 = transpose(perm = transpose_121_perm_0, x = k_115_cast); + tensor transpose_142 = transpose(perm = transpose_120_perm_0, x = q_115_cast); + tensor qk_57_cast = matmul(transpose_x = qk_57_transpose_x_0, transpose_y = qk_57_transpose_y_0, x = transpose_142, y = transpose_141); + tensor var_3183_cast = softmax(axis = var_3118, x = qk_57_cast); + tensor var_3185_transpose_x_0 = const()[name = tensor("op_3185_transpose_x_0"), val = tensor(false)]; + tensor var_3185_transpose_y_0 = const()[name = tensor("op_3185_transpose_y_0"), val = tensor(false)]; + tensor transpose_143 = transpose(perm = var_3179, x = var_3178_cast); + tensor var_3185_cast = matmul(transpose_x = var_3185_transpose_x_0, transpose_y = var_3185_transpose_y_0, x = var_3183_cast, y = transpose_143); + tensor var_3186 = const()[name = tensor("op_3186"), val = tensor([0, 2, 1, 3])]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 1500, 1280])]; + tensor transpose_140 = transpose(perm = var_3186, x = var_3185_cast); + tensor x_347_cast = reshape(shape = concat_28, x = transpose_140); + tensor var_3191_to_fp16 = const()[name = tensor("op_3191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126099328)))]; + tensor var_3192_to_fp16 = const()[name = tensor("op_3192_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129376192)))]; + tensor var_3193_cast = linear(bias = var_3192_to_fp16, weight = var_3191_to_fp16, x = x_347_cast); + tensor x_349_cast = add(x = x_343_cast, y = var_3193_cast); + tensor var_3199_axes_0 = const()[name = tensor("op_3199_axes_0"), val = tensor([-1])]; + tensor blocks_28_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129378816)))]; + tensor blocks_28_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129381440)))]; + tensor var_3199_cast = layer_norm(axes = var_3199_axes_0, beta = blocks_28_mlp_ln_bias_to_fp16, epsilon = var_3124_to_fp16, gamma = blocks_28_mlp_ln_weight_to_fp16, x = x_349_cast); + tensor var_3208_to_fp16 = const()[name = tensor("op_3208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129384064)))]; + tensor var_3209_to_fp16 = const()[name = tensor("op_3209_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142491328)))]; + tensor input_233_cast = linear(bias = var_3209_to_fp16, weight = var_3208_to_fp16, x = var_3199_cast); + tensor x_353_mode_0 = const()[name = tensor("x_353_mode_0"), val = tensor("EXACT")]; + tensor x_353_cast = gelu(mode = x_353_mode_0, x = input_233_cast); + tensor var_3214_to_fp16 = const()[name = tensor("op_3214_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142501632)))]; + tensor var_3215_to_fp16 = const()[name = tensor("op_3215_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155608896)))]; + tensor var_3216_cast = linear(bias = var_3215_to_fp16, weight = var_3214_to_fp16, x = x_353_cast); + tensor x_355_cast = add(x = x_349_cast, y = var_3216_cast); + tensor var_3225 = const()[name = tensor("op_3225"), val = tensor(-1)]; + tensor var_3242_axes_0 = const()[name = tensor("op_3242_axes_0"), val = tensor([-1])]; + tensor blocks_29_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_29_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155611520)))]; + tensor blocks_29_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_29_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155614144)))]; + tensor var_3231_to_fp16 = const()[name = tensor("op_3231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3242_cast = layer_norm(axes = var_3242_axes_0, beta = blocks_29_attn_ln_bias_to_fp16, epsilon = var_3231_to_fp16, gamma = blocks_29_attn_ln_weight_to_fp16, x = x_355_cast); + tensor var_3253_to_fp16 = const()[name = tensor("op_3253_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155616768)))]; + tensor var_3254_to_fp16 = const()[name = tensor("op_3254_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158893632)))]; + tensor q_117_cast = linear(bias = var_3254_to_fp16, weight = var_3253_to_fp16, x = var_3242_cast); + tensor var_3257_to_fp16 = const()[name = tensor("op_3257_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158896256)))]; + tensor k_117_bias_0_to_fp16 = const()[name = tensor("k_117_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162173120)))]; + tensor k_117_cast = linear(bias = k_117_bias_0_to_fp16, weight = var_3257_to_fp16, x = var_3242_cast); + tensor var_3261_to_fp16 = const()[name = tensor("op_3261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162175744)))]; + tensor var_3262_to_fp16 = const()[name = tensor("op_3262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165452608)))]; + tensor v_117_cast = linear(bias = var_3262_to_fp16, weight = var_3261_to_fp16, x = var_3242_cast); + tensor var_3270 = const()[name = tensor("op_3270"), val = tensor([1, 1500, 20, -1])]; + tensor var_3271_cast = reshape(shape = var_3270, x = q_117_cast); + tensor const_282_to_fp16 = const()[name = tensor("const_282_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_119_cast = mul(x = var_3271_cast, y = const_282_to_fp16); + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1500, 20, -1])]; + tensor var_3278_cast = reshape(shape = var_3277, x = k_117_cast); + tensor const_283_to_fp16 = const()[name = tensor("const_283_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_119_cast = mul(x = var_3278_cast, y = const_283_to_fp16); + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([1, 1500, 20, -1])]; + tensor var_3285_cast = reshape(shape = var_3284, x = v_117_cast); + tensor var_3286 = const()[name = tensor("op_3286"), val = tensor([0, 2, 1, 3])]; + tensor qk_59_transpose_x_0 = const()[name = tensor("qk_59_transpose_x_0"), val = tensor(false)]; + tensor qk_59_transpose_y_0 = const()[name = tensor("qk_59_transpose_y_0"), val = tensor(false)]; + tensor transpose_122_perm_0 = const()[name = tensor("transpose_122_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_123_perm_0 = const()[name = tensor("transpose_123_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_137 = transpose(perm = transpose_123_perm_0, x = k_119_cast); + tensor transpose_138 = transpose(perm = transpose_122_perm_0, x = q_119_cast); + tensor qk_59_cast = matmul(transpose_x = qk_59_transpose_x_0, transpose_y = qk_59_transpose_y_0, x = transpose_138, y = transpose_137); + tensor var_3290_cast = softmax(axis = var_3225, x = qk_59_cast); + tensor var_3292_transpose_x_0 = const()[name = tensor("op_3292_transpose_x_0"), val = tensor(false)]; + tensor var_3292_transpose_y_0 = const()[name = tensor("op_3292_transpose_y_0"), val = tensor(false)]; + tensor transpose_139 = transpose(perm = var_3286, x = var_3285_cast); + tensor var_3292_cast = matmul(transpose_x = var_3292_transpose_x_0, transpose_y = var_3292_transpose_y_0, x = var_3290_cast, y = transpose_139); + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([0, 2, 1, 3])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([1, 1500, 1280])]; + tensor transpose_136 = transpose(perm = var_3293, x = var_3292_cast); + tensor x_359_cast = reshape(shape = concat_29, x = transpose_136); + tensor var_3298_to_fp16 = const()[name = tensor("op_3298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165455232)))]; + tensor var_3299_to_fp16 = const()[name = tensor("op_3299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168732096)))]; + tensor var_3300_cast = linear(bias = var_3299_to_fp16, weight = var_3298_to_fp16, x = x_359_cast); + tensor x_361_cast = add(x = x_355_cast, y = var_3300_cast); + tensor var_3306_axes_0 = const()[name = tensor("op_3306_axes_0"), val = tensor([-1])]; + tensor blocks_29_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168734720)))]; + tensor blocks_29_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168737344)))]; + tensor var_3306_cast = layer_norm(axes = var_3306_axes_0, beta = blocks_29_mlp_ln_bias_to_fp16, epsilon = var_3231_to_fp16, gamma = blocks_29_mlp_ln_weight_to_fp16, x = x_361_cast); + tensor var_3315_to_fp16 = const()[name = tensor("op_3315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168739968)))]; + tensor var_3316_to_fp16 = const()[name = tensor("op_3316_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181847232)))]; + tensor input_241_cast = linear(bias = var_3316_to_fp16, weight = var_3315_to_fp16, x = var_3306_cast); + tensor x_365_mode_0 = const()[name = tensor("x_365_mode_0"), val = tensor("EXACT")]; + tensor x_365_cast = gelu(mode = x_365_mode_0, x = input_241_cast); + tensor var_3321_to_fp16 = const()[name = tensor("op_3321_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181857536)))]; + tensor var_3322_to_fp16 = const()[name = tensor("op_3322_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194964800)))]; + tensor var_3323_cast = linear(bias = var_3322_to_fp16, weight = var_3321_to_fp16, x = x_365_cast); + tensor x_367_cast = add(x = x_361_cast, y = var_3323_cast); + tensor var_3332 = const()[name = tensor("op_3332"), val = tensor(-1)]; + tensor var_3349_axes_0 = const()[name = tensor("op_3349_axes_0"), val = tensor([-1])]; + tensor blocks_30_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_30_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194967424)))]; + tensor blocks_30_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_30_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194970048)))]; + tensor var_3338_to_fp16 = const()[name = tensor("op_3338_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3349_cast = layer_norm(axes = var_3349_axes_0, beta = blocks_30_attn_ln_bias_to_fp16, epsilon = var_3338_to_fp16, gamma = blocks_30_attn_ln_weight_to_fp16, x = x_367_cast); + tensor var_3360_to_fp16 = const()[name = tensor("op_3360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194972672)))]; + tensor var_3361_to_fp16 = const()[name = tensor("op_3361_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198249536)))]; + tensor q_121_cast = linear(bias = var_3361_to_fp16, weight = var_3360_to_fp16, x = var_3349_cast); + tensor var_3364_to_fp16 = const()[name = tensor("op_3364_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198252160)))]; + tensor k_121_bias_0_to_fp16 = const()[name = tensor("k_121_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201529024)))]; + tensor k_121_cast = linear(bias = k_121_bias_0_to_fp16, weight = var_3364_to_fp16, x = var_3349_cast); + tensor var_3368_to_fp16 = const()[name = tensor("op_3368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201531648)))]; + tensor var_3369_to_fp16 = const()[name = tensor("op_3369_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204808512)))]; + tensor v_121_cast = linear(bias = var_3369_to_fp16, weight = var_3368_to_fp16, x = var_3349_cast); + tensor var_3377 = const()[name = tensor("op_3377"), val = tensor([1, 1500, 20, -1])]; + tensor var_3378_cast = reshape(shape = var_3377, x = q_121_cast); + tensor const_284_to_fp16 = const()[name = tensor("const_284_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_123_cast = mul(x = var_3378_cast, y = const_284_to_fp16); + tensor var_3384 = const()[name = tensor("op_3384"), val = tensor([1, 1500, 20, -1])]; + tensor var_3385_cast = reshape(shape = var_3384, x = k_121_cast); + tensor const_285_to_fp16 = const()[name = tensor("const_285_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_123_cast = mul(x = var_3385_cast, y = const_285_to_fp16); + tensor var_3391 = const()[name = tensor("op_3391"), val = tensor([1, 1500, 20, -1])]; + tensor var_3392_cast = reshape(shape = var_3391, x = v_121_cast); + tensor var_3393 = const()[name = tensor("op_3393"), val = tensor([0, 2, 1, 3])]; + tensor qk_61_transpose_x_0 = const()[name = tensor("qk_61_transpose_x_0"), val = tensor(false)]; + tensor qk_61_transpose_y_0 = const()[name = tensor("qk_61_transpose_y_0"), val = tensor(false)]; + tensor transpose_124_perm_0 = const()[name = tensor("transpose_124_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_125_perm_0 = const()[name = tensor("transpose_125_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_133 = transpose(perm = transpose_125_perm_0, x = k_123_cast); + tensor transpose_134 = transpose(perm = transpose_124_perm_0, x = q_123_cast); + tensor qk_61_cast = matmul(transpose_x = qk_61_transpose_x_0, transpose_y = qk_61_transpose_y_0, x = transpose_134, y = transpose_133); + tensor var_3397_cast = softmax(axis = var_3332, x = qk_61_cast); + tensor var_3399_transpose_x_0 = const()[name = tensor("op_3399_transpose_x_0"), val = tensor(false)]; + tensor var_3399_transpose_y_0 = const()[name = tensor("op_3399_transpose_y_0"), val = tensor(false)]; + tensor transpose_135 = transpose(perm = var_3393, x = var_3392_cast); + tensor var_3399_cast = matmul(transpose_x = var_3399_transpose_x_0, transpose_y = var_3399_transpose_y_0, x = var_3397_cast, y = transpose_135); + tensor var_3400 = const()[name = tensor("op_3400"), val = tensor([0, 2, 1, 3])]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([1, 1500, 1280])]; + tensor transpose_132 = transpose(perm = var_3400, x = var_3399_cast); + tensor x_371_cast = reshape(shape = concat_30, x = transpose_132); + tensor var_3405_to_fp16 = const()[name = tensor("op_3405_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204811136)))]; + tensor var_3406_to_fp16 = const()[name = tensor("op_3406_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208088000)))]; + tensor var_3407_cast = linear(bias = var_3406_to_fp16, weight = var_3405_to_fp16, x = x_371_cast); + tensor x_373_cast = add(x = x_367_cast, y = var_3407_cast); + tensor var_3413_axes_0 = const()[name = tensor("op_3413_axes_0"), val = tensor([-1])]; + tensor blocks_30_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208090624)))]; + tensor blocks_30_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208093248)))]; + tensor var_3413_cast = layer_norm(axes = var_3413_axes_0, beta = blocks_30_mlp_ln_bias_to_fp16, epsilon = var_3338_to_fp16, gamma = blocks_30_mlp_ln_weight_to_fp16, x = x_373_cast); + tensor var_3422_to_fp16 = const()[name = tensor("op_3422_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208095872)))]; + tensor var_3423_to_fp16 = const()[name = tensor("op_3423_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221203136)))]; + tensor input_249_cast = linear(bias = var_3423_to_fp16, weight = var_3422_to_fp16, x = var_3413_cast); + tensor x_377_mode_0 = const()[name = tensor("x_377_mode_0"), val = tensor("EXACT")]; + tensor x_377_cast = gelu(mode = x_377_mode_0, x = input_249_cast); + tensor var_3428_to_fp16 = const()[name = tensor("op_3428_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221213440)))]; + tensor var_3429_to_fp16 = const()[name = tensor("op_3429_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234320704)))]; + tensor var_3430_cast = linear(bias = var_3429_to_fp16, weight = var_3428_to_fp16, x = x_377_cast); + tensor x_379_cast = add(x = x_373_cast, y = var_3430_cast); + tensor var_3439 = const()[name = tensor("op_3439"), val = tensor(-1)]; + tensor var_3456_axes_0 = const()[name = tensor("op_3456_axes_0"), val = tensor([-1])]; + tensor blocks_31_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_31_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234323328)))]; + tensor blocks_31_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_31_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234325952)))]; + tensor var_3445_to_fp16 = const()[name = tensor("op_3445_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3456_cast = layer_norm(axes = var_3456_axes_0, beta = blocks_31_attn_ln_bias_to_fp16, epsilon = var_3445_to_fp16, gamma = blocks_31_attn_ln_weight_to_fp16, x = x_379_cast); + tensor var_3467_to_fp16 = const()[name = tensor("op_3467_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234328576)))]; + tensor var_3468_to_fp16 = const()[name = tensor("op_3468_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237605440)))]; + tensor q_125_cast = linear(bias = var_3468_to_fp16, weight = var_3467_to_fp16, x = var_3456_cast); + tensor var_3471_to_fp16 = const()[name = tensor("op_3471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237608064)))]; + tensor k_125_bias_0_to_fp16 = const()[name = tensor("k_125_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1240884928)))]; + tensor k_125_cast = linear(bias = k_125_bias_0_to_fp16, weight = var_3471_to_fp16, x = var_3456_cast); + tensor var_3475_to_fp16 = const()[name = tensor("op_3475_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1240887552)))]; + tensor var_3476_to_fp16 = const()[name = tensor("op_3476_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244164416)))]; + tensor v_125_cast = linear(bias = var_3476_to_fp16, weight = var_3475_to_fp16, x = var_3456_cast); + tensor var_3484 = const()[name = tensor("op_3484"), val = tensor([1, 1500, 20, -1])]; + tensor var_3485_cast = reshape(shape = var_3484, x = q_125_cast); + tensor const_286_to_fp16 = const()[name = tensor("const_286_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_3485_cast, y = const_286_to_fp16); + tensor var_3491 = const()[name = tensor("op_3491"), val = tensor([1, 1500, 20, -1])]; + tensor var_3492_cast = reshape(shape = var_3491, x = k_125_cast); + tensor const_287_to_fp16 = const()[name = tensor("const_287_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_3492_cast, y = const_287_to_fp16); + tensor var_3498 = const()[name = tensor("op_3498"), val = tensor([1, 1500, 20, -1])]; + tensor var_3499_cast = reshape(shape = var_3498, x = v_125_cast); + tensor var_3500 = const()[name = tensor("op_3500"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_126_perm_0 = const()[name = tensor("transpose_126_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_127_perm_0 = const()[name = tensor("transpose_127_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_129 = transpose(perm = transpose_127_perm_0, x = k_cast); + tensor transpose_130 = transpose(perm = transpose_126_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_130, y = transpose_129); + tensor var_3504_cast = softmax(axis = var_3439, x = qk_cast); + tensor var_3506_transpose_x_0 = const()[name = tensor("op_3506_transpose_x_0"), val = tensor(false)]; + tensor var_3506_transpose_y_0 = const()[name = tensor("op_3506_transpose_y_0"), val = tensor(false)]; + tensor transpose_131 = transpose(perm = var_3500, x = var_3499_cast); + tensor var_3506_cast = matmul(transpose_x = var_3506_transpose_x_0, transpose_y = var_3506_transpose_y_0, x = var_3504_cast, y = transpose_131); + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor([0, 2, 1, 3])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([1, 1500, 1280])]; + tensor transpose_128 = transpose(perm = var_3507, x = var_3506_cast); + tensor x_383_cast = reshape(shape = concat_31, x = transpose_128); + tensor var_3512_to_fp16 = const()[name = tensor("op_3512_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244167040)))]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247443904)))]; + tensor var_3514_cast = linear(bias = var_3513_to_fp16, weight = var_3512_to_fp16, x = x_383_cast); + tensor x_385_cast = add(x = x_379_cast, y = var_3514_cast); + tensor var_3520_axes_0 = const()[name = tensor("op_3520_axes_0"), val = tensor([-1])]; + tensor blocks_31_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247446528)))]; + tensor blocks_31_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247449152)))]; + tensor var_3520_cast = layer_norm(axes = var_3520_axes_0, beta = blocks_31_mlp_ln_bias_to_fp16, epsilon = var_3445_to_fp16, gamma = blocks_31_mlp_ln_weight_to_fp16, x = x_385_cast); + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247451776)))]; + tensor var_3530_to_fp16 = const()[name = tensor("op_3530_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260559040)))]; + tensor input_257_cast = linear(bias = var_3530_to_fp16, weight = var_3529_to_fp16, x = var_3520_cast); + tensor x_389_mode_0 = const()[name = tensor("x_389_mode_0"), val = tensor("EXACT")]; + tensor x_389_cast = gelu(mode = x_389_mode_0, x = input_257_cast); + tensor var_3535_to_fp16 = const()[name = tensor("op_3535_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260569344)))]; + tensor var_3536_to_fp16 = const()[name = tensor("op_3536_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273676608)))]; + tensor var_3537_cast = linear(bias = var_3536_to_fp16, weight = var_3535_to_fp16, x = x_389_cast); + tensor x_cast = add(x = x_385_cast, y = var_3537_cast); + tensor var_3550_axes_0 = const()[name = tensor("op_3550_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273679232)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273681856)))]; + tensor var_3541_to_fp16 = const()[name = tensor("op_3541_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3550_cast = layer_norm(axes = var_3550_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_3541_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_3550_cast_to_fp32_dtype_0 = const()[name = tensor("op_3550_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_3550_cast_to_fp32_dtype_0, x = var_3550_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f49a02ba6b98952be5932e993e097e28fb75746e --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v1-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19c3a5cd99b4004e00cc8f4cb18cacc4df290b734fff984350c698302b8da1cb +size 1273684480 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..4c966d6067a1d7bed07810f153202d0921f01b16 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21ffccc363739b0e4d703d36e393cb1003a54543946eb8cf51cfd61b7e483c10 +size 1144200907 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a868bd81bfcf766ee3c1023614854d6b682b9351 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5aec358d66255e1eecb43c5df3acc31815bed1b5d23ec4caa89cb7dfe1b89cc4 +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..8557aa58375c203cf681bfbdaebcee23ed5d4f2a --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 192, + "Matmul" : 64, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 32, + "Add" : 65, + "LayerNorm" : 65, + "Mul" : 64, + "Transpose" : 129, + "Gelu" : 34, + "Reshape" : 128 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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", + "method" : "predict" + } +] \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..66fd37b448b28df30396f27bdd155e6975fc1fb9 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/model.mil @@ -0,0 +1,1927 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor([1])]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614528)))]; + tensor cast_967 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_84_cast = conv(bias = bias_3_to_fp16, dilations = var_82, groups = var_72, pad = var_84_pad_0, pad_type = var_84_pad_type_0, strides = var_80, weight = weight_3_to_fp16, x = cast_967); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_84_cast); + tensor var_88 = const()[name = tensor("op_88"), val = tensor(1)]; + tensor var_97 = const()[name = tensor("op_97"), val = tensor([2])]; + tensor var_99 = const()[name = tensor("op_99"), val = tensor([1])]; + tensor var_101_pad_type_0 = const()[name = tensor("op_101_pad_type_0"), val = tensor("custom")]; + tensor var_101_pad_0 = const()[name = tensor("op_101_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617152)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10447616)))]; + tensor var_101_cast = conv(bias = bias_7_to_fp16, dilations = var_99, groups = var_88, pad = var_101_pad_0, pad_type = var_101_pad_type_0, strides = var_97, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_101_cast); + tensor var_106 = const()[name = tensor("op_106"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10450240)))]; + tensor transpose_256 = transpose(perm = var_106, x = x_3_cast); + tensor var_109_cast = add(x = transpose_256, y = positional_embedding_to_fp16); + tensor var_122 = const()[name = tensor("op_122"), val = tensor(-1)]; + tensor var_139_axes_0 = const()[name = tensor("op_139_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14290304)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14292928)))]; + tensor var_128_to_fp16 = const()[name = tensor("op_128_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_139_cast = layer_norm(axes = var_139_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_128_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_109_cast); + tensor var_150_to_fp16 = const()[name = tensor("op_150_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14295552)))]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17572416)))]; + tensor q_1_cast = linear(bias = var_151_to_fp16, weight = var_150_to_fp16, x = var_139_cast); + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17575040)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20851904)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_154_to_fp16, x = var_139_cast); + tensor var_158_to_fp16 = const()[name = tensor("op_158_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20854528)))]; + tensor var_159_to_fp16 = const()[name = tensor("op_159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24131392)))]; + tensor v_1_cast = linear(bias = var_159_to_fp16, weight = var_158_to_fp16, x = var_139_cast); + tensor var_167 = const()[name = tensor("op_167"), val = tensor([1, 1500, 20, -1])]; + tensor var_168_cast = reshape(shape = var_167, x = q_1_cast); + tensor const_224_to_fp16 = const()[name = tensor("const_224_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_168_cast, y = const_224_to_fp16); + tensor var_174 = const()[name = tensor("op_174"), val = tensor([1, 1500, 20, -1])]; + tensor var_175_cast = reshape(shape = var_174, x = k_1_cast); + tensor const_225_to_fp16 = const()[name = tensor("const_225_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_175_cast, y = const_225_to_fp16); + tensor var_181 = const()[name = tensor("op_181"), val = tensor([1, 1500, 20, -1])]; + tensor var_182_cast = reshape(shape = var_181, x = v_1_cast); + tensor var_183 = const()[name = tensor("op_183"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_64_perm_0 = const()[name = tensor("transpose_64_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_65_perm_0 = const()[name = tensor("transpose_65_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_253 = transpose(perm = transpose_65_perm_0, x = k_3_cast); + tensor transpose_254 = transpose(perm = transpose_64_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_254, y = transpose_253); + tensor var_187_cast = softmax(axis = var_122, x = qk_1_cast); + tensor var_189_transpose_x_0 = const()[name = tensor("op_189_transpose_x_0"), val = tensor(false)]; + tensor var_189_transpose_y_0 = const()[name = tensor("op_189_transpose_y_0"), val = tensor(false)]; + tensor transpose_255 = transpose(perm = var_183, x = var_182_cast); + tensor var_189_cast = matmul(transpose_x = var_189_transpose_x_0, transpose_y = var_189_transpose_y_0, x = var_187_cast, y = transpose_255); + tensor var_190 = const()[name = tensor("op_190"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 1280])]; + tensor transpose_252 = transpose(perm = var_190, x = var_189_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_252); + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24134016)))]; + tensor var_196_to_fp16 = const()[name = tensor("op_196_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27410880)))]; + tensor var_197_cast = linear(bias = var_196_to_fp16, weight = var_195_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_109_cast, y = var_197_cast); + tensor var_203_axes_0 = const()[name = tensor("op_203_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27413504)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27416128)))]; + tensor var_203_cast = layer_norm(axes = var_203_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_128_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_212_to_fp16 = const()[name = tensor("op_212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27418752)))]; + tensor var_213_to_fp16 = const()[name = tensor("op_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40526016)))]; + tensor input_9_cast = linear(bias = var_213_to_fp16, weight = var_212_to_fp16, x = var_203_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_218_to_fp16 = const()[name = tensor("op_218_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40536320)))]; + tensor var_219_to_fp16 = const()[name = tensor("op_219_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53643584)))]; + tensor var_220_cast = linear(bias = var_219_to_fp16, weight = var_218_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_220_cast); + tensor var_229 = const()[name = tensor("op_229"), val = tensor(-1)]; + tensor var_246_axes_0 = const()[name = tensor("op_246_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53646208)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53648832)))]; + tensor var_235_to_fp16 = const()[name = tensor("op_235_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_246_cast = layer_norm(axes = var_246_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_235_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_257_to_fp16 = const()[name = tensor("op_257_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53651456)))]; + tensor var_258_to_fp16 = const()[name = tensor("op_258_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56928320)))]; + tensor q_5_cast = linear(bias = var_258_to_fp16, weight = var_257_to_fp16, x = var_246_cast); + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56930944)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60207808)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_261_to_fp16, x = var_246_cast); + tensor var_265_to_fp16 = const()[name = tensor("op_265_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60210432)))]; + tensor var_266_to_fp16 = const()[name = tensor("op_266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63487296)))]; + tensor v_5_cast = linear(bias = var_266_to_fp16, weight = var_265_to_fp16, x = var_246_cast); + tensor var_274 = const()[name = tensor("op_274"), val = tensor([1, 1500, 20, -1])]; + tensor var_275_cast = reshape(shape = var_274, x = q_5_cast); + tensor const_226_to_fp16 = const()[name = tensor("const_226_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_275_cast, y = const_226_to_fp16); + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1500, 20, -1])]; + tensor var_282_cast = reshape(shape = var_281, x = k_5_cast); + tensor const_227_to_fp16 = const()[name = tensor("const_227_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_282_cast, y = const_227_to_fp16); + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1500, 20, -1])]; + tensor var_289_cast = reshape(shape = var_288, x = v_5_cast); + tensor var_290 = const()[name = tensor("op_290"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_66_perm_0 = const()[name = tensor("transpose_66_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_67_perm_0 = const()[name = tensor("transpose_67_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_249 = transpose(perm = transpose_67_perm_0, x = k_7_cast); + tensor transpose_250 = transpose(perm = transpose_66_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_250, y = transpose_249); + tensor var_294_cast = softmax(axis = var_229, x = qk_3_cast); + tensor var_296_transpose_x_0 = const()[name = tensor("op_296_transpose_x_0"), val = tensor(false)]; + tensor var_296_transpose_y_0 = const()[name = tensor("op_296_transpose_y_0"), val = tensor(false)]; + tensor transpose_251 = transpose(perm = var_290, x = var_289_cast); + tensor var_296_cast = matmul(transpose_x = var_296_transpose_x_0, transpose_y = var_296_transpose_y_0, x = var_294_cast, y = transpose_251); + tensor var_297 = const()[name = tensor("op_297"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 1280])]; + tensor transpose_248 = transpose(perm = var_297, x = var_296_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_248); + tensor var_302_to_fp16 = const()[name = tensor("op_302_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63489920)))]; + tensor var_303_to_fp16 = const()[name = tensor("op_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66766784)))]; + tensor var_304_cast = linear(bias = var_303_to_fp16, weight = var_302_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_304_cast); + tensor var_310_axes_0 = const()[name = tensor("op_310_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66769408)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66772032)))]; + tensor var_310_cast = layer_norm(axes = var_310_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_235_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66774656)))]; + tensor var_320_to_fp16 = const()[name = tensor("op_320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79881920)))]; + tensor input_17_cast = linear(bias = var_320_to_fp16, weight = var_319_to_fp16, x = var_310_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_325_to_fp16 = const()[name = tensor("op_325_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79892224)))]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92999488)))]; + tensor var_327_cast = linear(bias = var_326_to_fp16, weight = var_325_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_327_cast); + tensor var_336 = const()[name = tensor("op_336"), val = tensor(-1)]; + tensor var_353_axes_0 = const()[name = tensor("op_353_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93002112)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93004736)))]; + tensor var_342_to_fp16 = const()[name = tensor("op_342_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_353_cast = layer_norm(axes = var_353_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_342_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_364_to_fp16 = const()[name = tensor("op_364_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93007360)))]; + tensor var_365_to_fp16 = const()[name = tensor("op_365_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96284224)))]; + tensor q_9_cast = linear(bias = var_365_to_fp16, weight = var_364_to_fp16, x = var_353_cast); + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96286848)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99563712)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_368_to_fp16, x = var_353_cast); + tensor var_372_to_fp16 = const()[name = tensor("op_372_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99566336)))]; + tensor var_373_to_fp16 = const()[name = tensor("op_373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102843200)))]; + tensor v_9_cast = linear(bias = var_373_to_fp16, weight = var_372_to_fp16, x = var_353_cast); + tensor var_381 = const()[name = tensor("op_381"), val = tensor([1, 1500, 20, -1])]; + tensor var_382_cast = reshape(shape = var_381, x = q_9_cast); + tensor const_228_to_fp16 = const()[name = tensor("const_228_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_382_cast, y = const_228_to_fp16); + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1500, 20, -1])]; + tensor var_389_cast = reshape(shape = var_388, x = k_9_cast); + tensor const_229_to_fp16 = const()[name = tensor("const_229_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_389_cast, y = const_229_to_fp16); + tensor var_395 = const()[name = tensor("op_395"), val = tensor([1, 1500, 20, -1])]; + tensor var_396_cast = reshape(shape = var_395, x = v_9_cast); + tensor var_397 = const()[name = tensor("op_397"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_68_perm_0 = const()[name = tensor("transpose_68_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_245 = transpose(perm = transpose_69_perm_0, x = k_11_cast); + tensor transpose_246 = transpose(perm = transpose_68_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_246, y = transpose_245); + tensor var_401_cast = softmax(axis = var_336, x = qk_5_cast); + tensor var_403_transpose_x_0 = const()[name = tensor("op_403_transpose_x_0"), val = tensor(false)]; + tensor var_403_transpose_y_0 = const()[name = tensor("op_403_transpose_y_0"), val = tensor(false)]; + tensor transpose_247 = transpose(perm = var_397, x = var_396_cast); + tensor var_403_cast = matmul(transpose_x = var_403_transpose_x_0, transpose_y = var_403_transpose_y_0, x = var_401_cast, y = transpose_247); + tensor var_404 = const()[name = tensor("op_404"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 1280])]; + tensor transpose_244 = transpose(perm = var_404, x = var_403_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_244); + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102845824)))]; + tensor var_410_to_fp16 = const()[name = tensor("op_410_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106122688)))]; + tensor var_411_cast = linear(bias = var_410_to_fp16, weight = var_409_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_411_cast); + tensor var_417_axes_0 = const()[name = tensor("op_417_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106125312)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106127936)))]; + tensor var_417_cast = layer_norm(axes = var_417_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_342_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106130560)))]; + tensor var_427_to_fp16 = const()[name = tensor("op_427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119237824)))]; + tensor input_25_cast = linear(bias = var_427_to_fp16, weight = var_426_to_fp16, x = var_417_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_432_to_fp16 = const()[name = tensor("op_432_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119248128)))]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132355392)))]; + tensor var_434_cast = linear(bias = var_433_to_fp16, weight = var_432_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_434_cast); + tensor var_443 = const()[name = tensor("op_443"), val = tensor(-1)]; + tensor var_460_axes_0 = const()[name = tensor("op_460_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132358016)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132360640)))]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_460_cast = layer_norm(axes = var_460_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_449_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132363264)))]; + tensor var_472_to_fp16 = const()[name = tensor("op_472_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135640128)))]; + tensor q_13_cast = linear(bias = var_472_to_fp16, weight = var_471_to_fp16, x = var_460_cast); + tensor var_475_to_fp16 = const()[name = tensor("op_475_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135642752)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138919616)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_475_to_fp16, x = var_460_cast); + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138922240)))]; + tensor var_480_to_fp16 = const()[name = tensor("op_480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142199104)))]; + tensor v_13_cast = linear(bias = var_480_to_fp16, weight = var_479_to_fp16, x = var_460_cast); + tensor var_488 = const()[name = tensor("op_488"), val = tensor([1, 1500, 20, -1])]; + tensor var_489_cast = reshape(shape = var_488, x = q_13_cast); + tensor const_230_to_fp16 = const()[name = tensor("const_230_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_489_cast, y = const_230_to_fp16); + tensor var_495 = const()[name = tensor("op_495"), val = tensor([1, 1500, 20, -1])]; + tensor var_496_cast = reshape(shape = var_495, x = k_13_cast); + tensor const_231_to_fp16 = const()[name = tensor("const_231_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_496_cast, y = const_231_to_fp16); + tensor var_502 = const()[name = tensor("op_502"), val = tensor([1, 1500, 20, -1])]; + tensor var_503_cast = reshape(shape = var_502, x = v_13_cast); + tensor var_504 = const()[name = tensor("op_504"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_241 = transpose(perm = transpose_71_perm_0, x = k_15_cast); + tensor transpose_242 = transpose(perm = transpose_70_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_242, y = transpose_241); + tensor var_508_cast = softmax(axis = var_443, x = qk_7_cast); + tensor var_510_transpose_x_0 = const()[name = tensor("op_510_transpose_x_0"), val = tensor(false)]; + tensor var_510_transpose_y_0 = const()[name = tensor("op_510_transpose_y_0"), val = tensor(false)]; + tensor transpose_243 = transpose(perm = var_504, x = var_503_cast); + tensor var_510_cast = matmul(transpose_x = var_510_transpose_x_0, transpose_y = var_510_transpose_y_0, x = var_508_cast, y = transpose_243); + tensor var_511 = const()[name = tensor("op_511"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 1280])]; + tensor transpose_240 = transpose(perm = var_511, x = var_510_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_240); + tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142201728)))]; + tensor var_517_to_fp16 = const()[name = tensor("op_517_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145478592)))]; + tensor var_518_cast = linear(bias = var_517_to_fp16, weight = var_516_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_518_cast); + tensor var_524_axes_0 = const()[name = tensor("op_524_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145481216)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145483840)))]; + tensor var_524_cast = layer_norm(axes = var_524_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_449_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_533_to_fp16 = const()[name = tensor("op_533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145486464)))]; + tensor var_534_to_fp16 = const()[name = tensor("op_534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158593728)))]; + tensor input_33_cast = linear(bias = var_534_to_fp16, weight = var_533_to_fp16, x = var_524_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_539_to_fp16 = const()[name = tensor("op_539_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158604032)))]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171711296)))]; + tensor var_541_cast = linear(bias = var_540_to_fp16, weight = var_539_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_541_cast); + tensor var_550 = const()[name = tensor("op_550"), val = tensor(-1)]; + tensor var_567_axes_0 = const()[name = tensor("op_567_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171713920)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171716544)))]; + tensor var_556_to_fp16 = const()[name = tensor("op_556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_567_cast = layer_norm(axes = var_567_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_556_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_578_to_fp16 = const()[name = tensor("op_578_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171719168)))]; + tensor var_579_to_fp16 = const()[name = tensor("op_579_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174996032)))]; + tensor q_17_cast = linear(bias = var_579_to_fp16, weight = var_578_to_fp16, x = var_567_cast); + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174998656)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178275520)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_582_to_fp16, x = var_567_cast); + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178278144)))]; + tensor var_587_to_fp16 = const()[name = tensor("op_587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181555008)))]; + tensor v_17_cast = linear(bias = var_587_to_fp16, weight = var_586_to_fp16, x = var_567_cast); + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1500, 20, -1])]; + tensor var_596_cast = reshape(shape = var_595, x = q_17_cast); + tensor const_232_to_fp16 = const()[name = tensor("const_232_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_596_cast, y = const_232_to_fp16); + tensor var_602 = const()[name = tensor("op_602"), val = tensor([1, 1500, 20, -1])]; + tensor var_603_cast = reshape(shape = var_602, x = k_17_cast); + tensor const_233_to_fp16 = const()[name = tensor("const_233_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_603_cast, y = const_233_to_fp16); + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1500, 20, -1])]; + tensor var_610_cast = reshape(shape = var_609, x = v_17_cast); + tensor var_611 = const()[name = tensor("op_611"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_237 = transpose(perm = transpose_73_perm_0, x = k_19_cast); + tensor transpose_238 = transpose(perm = transpose_72_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_238, y = transpose_237); + tensor var_615_cast = softmax(axis = var_550, x = qk_9_cast); + tensor var_617_transpose_x_0 = const()[name = tensor("op_617_transpose_x_0"), val = tensor(false)]; + tensor var_617_transpose_y_0 = const()[name = tensor("op_617_transpose_y_0"), val = tensor(false)]; + tensor transpose_239 = transpose(perm = var_611, x = var_610_cast); + tensor var_617_cast = matmul(transpose_x = var_617_transpose_x_0, transpose_y = var_617_transpose_y_0, x = var_615_cast, y = transpose_239); + tensor var_618 = const()[name = tensor("op_618"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 1280])]; + tensor transpose_236 = transpose(perm = var_618, x = var_617_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_236); + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181557632)))]; + tensor var_624_to_fp16 = const()[name = tensor("op_624_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184834496)))]; + tensor var_625_cast = linear(bias = var_624_to_fp16, weight = var_623_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_625_cast); + tensor var_631_axes_0 = const()[name = tensor("op_631_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184837120)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839744)))]; + tensor var_631_cast = layer_norm(axes = var_631_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_556_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_640_to_fp16 = const()[name = tensor("op_640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184842368)))]; + tensor var_641_to_fp16 = const()[name = tensor("op_641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197949632)))]; + tensor input_41_cast = linear(bias = var_641_to_fp16, weight = var_640_to_fp16, x = var_631_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_646_to_fp16 = const()[name = tensor("op_646_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197959936)))]; + tensor var_647_to_fp16 = const()[name = tensor("op_647_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211067200)))]; + tensor var_648_cast = linear(bias = var_647_to_fp16, weight = var_646_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_648_cast); + tensor var_657 = const()[name = tensor("op_657"), val = tensor(-1)]; + tensor var_674_axes_0 = const()[name = tensor("op_674_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211069824)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211072448)))]; + tensor var_663_to_fp16 = const()[name = tensor("op_663_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_674_cast = layer_norm(axes = var_674_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_663_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_685_to_fp16 = const()[name = tensor("op_685_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211075072)))]; + tensor var_686_to_fp16 = const()[name = tensor("op_686_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214351936)))]; + tensor q_21_cast = linear(bias = var_686_to_fp16, weight = var_685_to_fp16, x = var_674_cast); + tensor var_689_to_fp16 = const()[name = tensor("op_689_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214354560)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217631424)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_689_to_fp16, x = var_674_cast); + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217634048)))]; + tensor var_694_to_fp16 = const()[name = tensor("op_694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220910912)))]; + tensor v_21_cast = linear(bias = var_694_to_fp16, weight = var_693_to_fp16, x = var_674_cast); + tensor var_702 = const()[name = tensor("op_702"), val = tensor([1, 1500, 20, -1])]; + tensor var_703_cast = reshape(shape = var_702, x = q_21_cast); + tensor const_234_to_fp16 = const()[name = tensor("const_234_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast = mul(x = var_703_cast, y = const_234_to_fp16); + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1500, 20, -1])]; + tensor var_710_cast = reshape(shape = var_709, x = k_21_cast); + tensor const_235_to_fp16 = const()[name = tensor("const_235_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast = mul(x = var_710_cast, y = const_235_to_fp16); + tensor var_716 = const()[name = tensor("op_716"), val = tensor([1, 1500, 20, -1])]; + tensor var_717_cast = reshape(shape = var_716, x = v_21_cast); + tensor var_718 = const()[name = tensor("op_718"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_233 = transpose(perm = transpose_75_perm_0, x = k_23_cast); + tensor transpose_234 = transpose(perm = transpose_74_perm_0, x = q_23_cast); + tensor qk_11_cast = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_234, y = transpose_233); + tensor var_722_cast = softmax(axis = var_657, x = qk_11_cast); + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_235 = transpose(perm = var_718, x = var_717_cast); + tensor var_724_cast = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = var_722_cast, y = transpose_235); + tensor var_725 = const()[name = tensor("op_725"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 1280])]; + tensor transpose_232 = transpose(perm = var_725, x = var_724_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_232); + tensor var_730_to_fp16 = const()[name = tensor("op_730_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220913536)))]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224190400)))]; + tensor var_732_cast = linear(bias = var_731_to_fp16, weight = var_730_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_732_cast); + tensor var_738_axes_0 = const()[name = tensor("op_738_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224193024)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224195648)))]; + tensor var_738_cast = layer_norm(axes = var_738_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_663_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_747_to_fp16 = const()[name = tensor("op_747_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224198272)))]; + tensor var_748_to_fp16 = const()[name = tensor("op_748_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237305536)))]; + tensor input_49_cast = linear(bias = var_748_to_fp16, weight = var_747_to_fp16, x = var_738_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_753_to_fp16 = const()[name = tensor("op_753_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237315840)))]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250423104)))]; + tensor var_755_cast = linear(bias = var_754_to_fp16, weight = var_753_to_fp16, x = x_77_cast); + tensor x_79_cast = add(x = x_73_cast, y = var_755_cast); + tensor var_764 = const()[name = tensor("op_764"), val = tensor(-1)]; + tensor var_781_axes_0 = const()[name = tensor("op_781_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250425728)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250428352)))]; + tensor var_770_to_fp16 = const()[name = tensor("op_770_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_781_cast = layer_norm(axes = var_781_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_770_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast); + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250430976)))]; + tensor var_793_to_fp16 = const()[name = tensor("op_793_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253707840)))]; + tensor q_25_cast = linear(bias = var_793_to_fp16, weight = var_792_to_fp16, x = var_781_cast); + tensor var_796_to_fp16 = const()[name = tensor("op_796_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253710464)))]; + tensor k_25_bias_0_to_fp16 = const()[name = tensor("k_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256987328)))]; + tensor k_25_cast = linear(bias = k_25_bias_0_to_fp16, weight = var_796_to_fp16, x = var_781_cast); + tensor var_800_to_fp16 = const()[name = tensor("op_800_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256989952)))]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260266816)))]; + tensor v_25_cast = linear(bias = var_801_to_fp16, weight = var_800_to_fp16, x = var_781_cast); + tensor var_809 = const()[name = tensor("op_809"), val = tensor([1, 1500, 20, -1])]; + tensor var_810_cast = reshape(shape = var_809, x = q_25_cast); + tensor const_236_to_fp16 = const()[name = tensor("const_236_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast = mul(x = var_810_cast, y = const_236_to_fp16); + tensor var_816 = const()[name = tensor("op_816"), val = tensor([1, 1500, 20, -1])]; + tensor var_817_cast = reshape(shape = var_816, x = k_25_cast); + tensor const_237_to_fp16 = const()[name = tensor("const_237_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast = mul(x = var_817_cast, y = const_237_to_fp16); + tensor var_823 = const()[name = tensor("op_823"), val = tensor([1, 1500, 20, -1])]; + tensor var_824_cast = reshape(shape = var_823, x = v_25_cast); + tensor var_825 = const()[name = tensor("op_825"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_229 = transpose(perm = transpose_77_perm_0, x = k_27_cast); + tensor transpose_230 = transpose(perm = transpose_76_perm_0, x = q_27_cast); + tensor qk_13_cast = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_230, y = transpose_229); + tensor var_829_cast = softmax(axis = var_764, x = qk_13_cast); + tensor var_831_transpose_x_0 = const()[name = tensor("op_831_transpose_x_0"), val = tensor(false)]; + tensor var_831_transpose_y_0 = const()[name = tensor("op_831_transpose_y_0"), val = tensor(false)]; + tensor transpose_231 = transpose(perm = var_825, x = var_824_cast); + tensor var_831_cast = matmul(transpose_x = var_831_transpose_x_0, transpose_y = var_831_transpose_y_0, x = var_829_cast, y = transpose_231); + tensor var_832 = const()[name = tensor("op_832"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 1280])]; + tensor transpose_228 = transpose(perm = var_832, x = var_831_cast); + tensor x_83_cast = reshape(shape = concat_6, x = transpose_228); + tensor var_837_to_fp16 = const()[name = tensor("op_837_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260269440)))]; + tensor var_838_to_fp16 = const()[name = tensor("op_838_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263546304)))]; + tensor var_839_cast = linear(bias = var_838_to_fp16, weight = var_837_to_fp16, x = x_83_cast); + tensor x_85_cast = add(x = x_79_cast, y = var_839_cast); + tensor var_845_axes_0 = const()[name = tensor("op_845_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263548928)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263551552)))]; + tensor var_845_cast = layer_norm(axes = var_845_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_770_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast); + tensor var_854_to_fp16 = const()[name = tensor("op_854_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263554176)))]; + tensor var_855_to_fp16 = const()[name = tensor("op_855_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276661440)))]; + tensor input_57_cast = linear(bias = var_855_to_fp16, weight = var_854_to_fp16, x = var_845_cast); + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast = gelu(mode = x_89_mode_0, x = input_57_cast); + tensor var_860_to_fp16 = const()[name = tensor("op_860_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276671744)))]; + tensor var_861_to_fp16 = const()[name = tensor("op_861_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289779008)))]; + tensor var_862_cast = linear(bias = var_861_to_fp16, weight = var_860_to_fp16, x = x_89_cast); + tensor x_91_cast = add(x = x_85_cast, y = var_862_cast); + tensor var_871 = const()[name = tensor("op_871"), val = tensor(-1)]; + tensor var_888_axes_0 = const()[name = tensor("op_888_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289781632)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289784256)))]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_888_cast = layer_norm(axes = var_888_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_877_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast); + tensor var_899_to_fp16 = const()[name = tensor("op_899_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289786880)))]; + tensor var_900_to_fp16 = const()[name = tensor("op_900_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293063744)))]; + tensor q_29_cast = linear(bias = var_900_to_fp16, weight = var_899_to_fp16, x = var_888_cast); + tensor var_903_to_fp16 = const()[name = tensor("op_903_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293066368)))]; + tensor k_29_bias_0_to_fp16 = const()[name = tensor("k_29_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296343232)))]; + tensor k_29_cast = linear(bias = k_29_bias_0_to_fp16, weight = var_903_to_fp16, x = var_888_cast); + tensor var_907_to_fp16 = const()[name = tensor("op_907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296345856)))]; + tensor var_908_to_fp16 = const()[name = tensor("op_908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299622720)))]; + tensor v_29_cast = linear(bias = var_908_to_fp16, weight = var_907_to_fp16, x = var_888_cast); + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1500, 20, -1])]; + tensor var_917_cast = reshape(shape = var_916, x = q_29_cast); + tensor const_238_to_fp16 = const()[name = tensor("const_238_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast = mul(x = var_917_cast, y = const_238_to_fp16); + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1500, 20, -1])]; + tensor var_924_cast = reshape(shape = var_923, x = k_29_cast); + tensor const_239_to_fp16 = const()[name = tensor("const_239_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast = mul(x = var_924_cast, y = const_239_to_fp16); + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 1500, 20, -1])]; + tensor var_931_cast = reshape(shape = var_930, x = v_29_cast); + tensor var_932 = const()[name = tensor("op_932"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_225 = transpose(perm = transpose_79_perm_0, x = k_31_cast); + tensor transpose_226 = transpose(perm = transpose_78_perm_0, x = q_31_cast); + tensor qk_15_cast = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_226, y = transpose_225); + tensor var_936_cast = softmax(axis = var_871, x = qk_15_cast); + tensor var_938_transpose_x_0 = const()[name = tensor("op_938_transpose_x_0"), val = tensor(false)]; + tensor var_938_transpose_y_0 = const()[name = tensor("op_938_transpose_y_0"), val = tensor(false)]; + tensor transpose_227 = transpose(perm = var_932, x = var_931_cast); + tensor var_938_cast = matmul(transpose_x = var_938_transpose_x_0, transpose_y = var_938_transpose_y_0, x = var_936_cast, y = transpose_227); + tensor var_939 = const()[name = tensor("op_939"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 1280])]; + tensor transpose_224 = transpose(perm = var_939, x = var_938_cast); + tensor x_95_cast = reshape(shape = concat_7, x = transpose_224); + tensor var_944_to_fp16 = const()[name = tensor("op_944_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299625344)))]; + tensor var_945_to_fp16 = const()[name = tensor("op_945_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302902208)))]; + tensor var_946_cast = linear(bias = var_945_to_fp16, weight = var_944_to_fp16, x = x_95_cast); + tensor x_97_cast = add(x = x_91_cast, y = var_946_cast); + tensor var_952_axes_0 = const()[name = tensor("op_952_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302904832)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302907456)))]; + tensor var_952_cast = layer_norm(axes = var_952_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_877_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast); + tensor var_961_to_fp16 = const()[name = tensor("op_961_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302910080)))]; + tensor var_962_to_fp16 = const()[name = tensor("op_962_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316017344)))]; + tensor input_65_cast = linear(bias = var_962_to_fp16, weight = var_961_to_fp16, x = var_952_cast); + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast = gelu(mode = x_101_mode_0, x = input_65_cast); + tensor var_967_to_fp16 = const()[name = tensor("op_967_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316027648)))]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329134912)))]; + tensor var_969_cast = linear(bias = var_968_to_fp16, weight = var_967_to_fp16, x = x_101_cast); + tensor x_103_cast = add(x = x_97_cast, y = var_969_cast); + tensor var_978 = const()[name = tensor("op_978"), val = tensor(-1)]; + tensor var_995_axes_0 = const()[name = tensor("op_995_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329137536)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329140160)))]; + tensor var_984_to_fp16 = const()[name = tensor("op_984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_995_cast = layer_norm(axes = var_995_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_984_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast); + tensor var_1006_to_fp16 = const()[name = tensor("op_1006_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329142784)))]; + tensor var_1007_to_fp16 = const()[name = tensor("op_1007_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332419648)))]; + tensor q_33_cast = linear(bias = var_1007_to_fp16, weight = var_1006_to_fp16, x = var_995_cast); + tensor var_1010_to_fp16 = const()[name = tensor("op_1010_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332422272)))]; + tensor k_33_bias_0_to_fp16 = const()[name = tensor("k_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335699136)))]; + tensor k_33_cast = linear(bias = k_33_bias_0_to_fp16, weight = var_1010_to_fp16, x = var_995_cast); + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335701760)))]; + tensor var_1015_to_fp16 = const()[name = tensor("op_1015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338978624)))]; + tensor v_33_cast = linear(bias = var_1015_to_fp16, weight = var_1014_to_fp16, x = var_995_cast); + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 1500, 20, -1])]; + tensor var_1024_cast = reshape(shape = var_1023, x = q_33_cast); + tensor const_240_to_fp16 = const()[name = tensor("const_240_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast = mul(x = var_1024_cast, y = const_240_to_fp16); + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([1, 1500, 20, -1])]; + tensor var_1031_cast = reshape(shape = var_1030, x = k_33_cast); + tensor const_241_to_fp16 = const()[name = tensor("const_241_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast = mul(x = var_1031_cast, y = const_241_to_fp16); + tensor var_1037 = const()[name = tensor("op_1037"), val = tensor([1, 1500, 20, -1])]; + tensor var_1038_cast = reshape(shape = var_1037, x = v_33_cast); + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_221 = transpose(perm = transpose_81_perm_0, x = k_35_cast); + tensor transpose_222 = transpose(perm = transpose_80_perm_0, x = q_35_cast); + tensor qk_17_cast = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_222, y = transpose_221); + tensor var_1043_cast = softmax(axis = var_978, x = qk_17_cast); + tensor var_1045_transpose_x_0 = const()[name = tensor("op_1045_transpose_x_0"), val = tensor(false)]; + tensor var_1045_transpose_y_0 = const()[name = tensor("op_1045_transpose_y_0"), val = tensor(false)]; + tensor transpose_223 = transpose(perm = var_1039, x = var_1038_cast); + tensor var_1045_cast = matmul(transpose_x = var_1045_transpose_x_0, transpose_y = var_1045_transpose_y_0, x = var_1043_cast, y = transpose_223); + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 1280])]; + tensor transpose_220 = transpose(perm = var_1046, x = var_1045_cast); + tensor x_107_cast = reshape(shape = concat_8, x = transpose_220); + tensor var_1051_to_fp16 = const()[name = tensor("op_1051_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338981248)))]; + tensor var_1052_to_fp16 = const()[name = tensor("op_1052_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342258112)))]; + tensor var_1053_cast = linear(bias = var_1052_to_fp16, weight = var_1051_to_fp16, x = x_107_cast); + tensor x_109_cast = add(x = x_103_cast, y = var_1053_cast); + tensor var_1059_axes_0 = const()[name = tensor("op_1059_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342260736)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342263360)))]; + tensor var_1059_cast = layer_norm(axes = var_1059_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_984_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast); + tensor var_1068_to_fp16 = const()[name = tensor("op_1068_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342265984)))]; + tensor var_1069_to_fp16 = const()[name = tensor("op_1069_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355373248)))]; + tensor input_73_cast = linear(bias = var_1069_to_fp16, weight = var_1068_to_fp16, x = var_1059_cast); + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast = gelu(mode = x_113_mode_0, x = input_73_cast); + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355383552)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368490816)))]; + tensor var_1076_cast = linear(bias = var_1075_to_fp16, weight = var_1074_to_fp16, x = x_113_cast); + tensor x_115_cast = add(x = x_109_cast, y = var_1076_cast); + tensor var_1085 = const()[name = tensor("op_1085"), val = tensor(-1)]; + tensor var_1102_axes_0 = const()[name = tensor("op_1102_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368493440)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368496064)))]; + tensor var_1091_to_fp16 = const()[name = tensor("op_1091_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1102_cast = layer_norm(axes = var_1102_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1091_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast); + tensor var_1113_to_fp16 = const()[name = tensor("op_1113_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368498688)))]; + tensor var_1114_to_fp16 = const()[name = tensor("op_1114_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371775552)))]; + tensor q_37_cast = linear(bias = var_1114_to_fp16, weight = var_1113_to_fp16, x = var_1102_cast); + tensor var_1117_to_fp16 = const()[name = tensor("op_1117_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371778176)))]; + tensor k_37_bias_0_to_fp16 = const()[name = tensor("k_37_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375055040)))]; + tensor k_37_cast = linear(bias = k_37_bias_0_to_fp16, weight = var_1117_to_fp16, x = var_1102_cast); + tensor var_1121_to_fp16 = const()[name = tensor("op_1121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375057664)))]; + tensor var_1122_to_fp16 = const()[name = tensor("op_1122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378334528)))]; + tensor v_37_cast = linear(bias = var_1122_to_fp16, weight = var_1121_to_fp16, x = var_1102_cast); + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1500, 20, -1])]; + tensor var_1131_cast = reshape(shape = var_1130, x = q_37_cast); + tensor const_242_to_fp16 = const()[name = tensor("const_242_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast = mul(x = var_1131_cast, y = const_242_to_fp16); + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([1, 1500, 20, -1])]; + tensor var_1138_cast = reshape(shape = var_1137, x = k_37_cast); + tensor const_243_to_fp16 = const()[name = tensor("const_243_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast = mul(x = var_1138_cast, y = const_243_to_fp16); + tensor var_1144 = const()[name = tensor("op_1144"), val = tensor([1, 1500, 20, -1])]; + tensor var_1145_cast = reshape(shape = var_1144, x = v_37_cast); + tensor var_1146 = const()[name = tensor("op_1146"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_217 = transpose(perm = transpose_83_perm_0, x = k_39_cast); + tensor transpose_218 = transpose(perm = transpose_82_perm_0, x = q_39_cast); + tensor qk_19_cast = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_218, y = transpose_217); + tensor var_1150_cast = softmax(axis = var_1085, x = qk_19_cast); + tensor var_1152_transpose_x_0 = const()[name = tensor("op_1152_transpose_x_0"), val = tensor(false)]; + tensor var_1152_transpose_y_0 = const()[name = tensor("op_1152_transpose_y_0"), val = tensor(false)]; + tensor transpose_219 = transpose(perm = var_1146, x = var_1145_cast); + tensor var_1152_cast = matmul(transpose_x = var_1152_transpose_x_0, transpose_y = var_1152_transpose_y_0, x = var_1150_cast, y = transpose_219); + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 1280])]; + tensor transpose_216 = transpose(perm = var_1153, x = var_1152_cast); + tensor x_119_cast = reshape(shape = concat_9, x = transpose_216); + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378337152)))]; + tensor var_1159_to_fp16 = const()[name = tensor("op_1159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381614016)))]; + tensor var_1160_cast = linear(bias = var_1159_to_fp16, weight = var_1158_to_fp16, x = x_119_cast); + tensor x_121_cast = add(x = x_115_cast, y = var_1160_cast); + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381616640)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381619264)))]; + tensor var_1166_cast = layer_norm(axes = var_1166_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1091_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast); + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381621888)))]; + tensor var_1176_to_fp16 = const()[name = tensor("op_1176_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394729152)))]; + tensor input_81_cast = linear(bias = var_1176_to_fp16, weight = var_1175_to_fp16, x = var_1166_cast); + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast = gelu(mode = x_125_mode_0, x = input_81_cast); + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394739456)))]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407846720)))]; + tensor var_1183_cast = linear(bias = var_1182_to_fp16, weight = var_1181_to_fp16, x = x_125_cast); + tensor x_127_cast = add(x = x_121_cast, y = var_1183_cast); + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor(-1)]; + tensor var_1209_axes_0 = const()[name = tensor("op_1209_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407849344)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407851968)))]; + tensor var_1198_to_fp16 = const()[name = tensor("op_1198_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1209_cast = layer_norm(axes = var_1209_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast); + tensor var_1220_to_fp16 = const()[name = tensor("op_1220_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407854592)))]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411131456)))]; + tensor q_41_cast = linear(bias = var_1221_to_fp16, weight = var_1220_to_fp16, x = var_1209_cast); + tensor var_1224_to_fp16 = const()[name = tensor("op_1224_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411134080)))]; + tensor k_41_bias_0_to_fp16 = const()[name = tensor("k_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414410944)))]; + tensor k_41_cast = linear(bias = k_41_bias_0_to_fp16, weight = var_1224_to_fp16, x = var_1209_cast); + tensor var_1228_to_fp16 = const()[name = tensor("op_1228_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414413568)))]; + tensor var_1229_to_fp16 = const()[name = tensor("op_1229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417690432)))]; + tensor v_41_cast = linear(bias = var_1229_to_fp16, weight = var_1228_to_fp16, x = var_1209_cast); + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1500, 20, -1])]; + tensor var_1238_cast = reshape(shape = var_1237, x = q_41_cast); + tensor const_244_to_fp16 = const()[name = tensor("const_244_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast = mul(x = var_1238_cast, y = const_244_to_fp16); + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([1, 1500, 20, -1])]; + tensor var_1245_cast = reshape(shape = var_1244, x = k_41_cast); + tensor const_245_to_fp16 = const()[name = tensor("const_245_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast = mul(x = var_1245_cast, y = const_245_to_fp16); + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 1500, 20, -1])]; + tensor var_1252_cast = reshape(shape = var_1251, x = v_41_cast); + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_213 = transpose(perm = transpose_85_perm_0, x = k_43_cast); + tensor transpose_214 = transpose(perm = transpose_84_perm_0, x = q_43_cast); + tensor qk_21_cast = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_214, y = transpose_213); + tensor var_1257_cast = softmax(axis = var_1192, x = qk_21_cast); + tensor var_1259_transpose_x_0 = const()[name = tensor("op_1259_transpose_x_0"), val = tensor(false)]; + tensor var_1259_transpose_y_0 = const()[name = tensor("op_1259_transpose_y_0"), val = tensor(false)]; + tensor transpose_215 = transpose(perm = var_1253, x = var_1252_cast); + tensor var_1259_cast = matmul(transpose_x = var_1259_transpose_x_0, transpose_y = var_1259_transpose_y_0, x = var_1257_cast, y = transpose_215); + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 1280])]; + tensor transpose_212 = transpose(perm = var_1260, x = var_1259_cast); + tensor x_131_cast = reshape(shape = concat_10, x = transpose_212); + tensor var_1265_to_fp16 = const()[name = tensor("op_1265_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417693056)))]; + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420969920)))]; + tensor var_1267_cast = linear(bias = var_1266_to_fp16, weight = var_1265_to_fp16, x = x_131_cast); + tensor x_133_cast = add(x = x_127_cast, y = var_1267_cast); + tensor var_1273_axes_0 = const()[name = tensor("op_1273_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420972544)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420975168)))]; + tensor var_1273_cast = layer_norm(axes = var_1273_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast); + tensor var_1282_to_fp16 = const()[name = tensor("op_1282_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420977792)))]; + tensor var_1283_to_fp16 = const()[name = tensor("op_1283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434085056)))]; + tensor input_89_cast = linear(bias = var_1283_to_fp16, weight = var_1282_to_fp16, x = var_1273_cast); + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast = gelu(mode = x_137_mode_0, x = input_89_cast); + tensor var_1288_to_fp16 = const()[name = tensor("op_1288_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434095360)))]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447202624)))]; + tensor var_1290_cast = linear(bias = var_1289_to_fp16, weight = var_1288_to_fp16, x = x_137_cast); + tensor x_139_cast = add(x = x_133_cast, y = var_1290_cast); + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor(-1)]; + tensor var_1316_axes_0 = const()[name = tensor("op_1316_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447205248)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447207872)))]; + tensor var_1305_to_fp16 = const()[name = tensor("op_1305_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1316_cast = layer_norm(axes = var_1316_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1305_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast); + tensor var_1327_to_fp16 = const()[name = tensor("op_1327_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447210496)))]; + tensor var_1328_to_fp16 = const()[name = tensor("op_1328_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450487360)))]; + tensor q_45_cast = linear(bias = var_1328_to_fp16, weight = var_1327_to_fp16, x = var_1316_cast); + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450489984)))]; + tensor k_45_bias_0_to_fp16 = const()[name = tensor("k_45_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453766848)))]; + tensor k_45_cast = linear(bias = k_45_bias_0_to_fp16, weight = var_1331_to_fp16, x = var_1316_cast); + tensor var_1335_to_fp16 = const()[name = tensor("op_1335_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453769472)))]; + tensor var_1336_to_fp16 = const()[name = tensor("op_1336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457046336)))]; + tensor v_45_cast = linear(bias = var_1336_to_fp16, weight = var_1335_to_fp16, x = var_1316_cast); + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([1, 1500, 20, -1])]; + tensor var_1345_cast = reshape(shape = var_1344, x = q_45_cast); + tensor const_246_to_fp16 = const()[name = tensor("const_246_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_47_cast = mul(x = var_1345_cast, y = const_246_to_fp16); + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([1, 1500, 20, -1])]; + tensor var_1352_cast = reshape(shape = var_1351, x = k_45_cast); + tensor const_247_to_fp16 = const()[name = tensor("const_247_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_47_cast = mul(x = var_1352_cast, y = const_247_to_fp16); + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 1500, 20, -1])]; + tensor var_1359_cast = reshape(shape = var_1358, x = v_45_cast); + tensor var_1360 = const()[name = tensor("op_1360"), val = tensor([0, 2, 1, 3])]; + tensor qk_23_transpose_x_0 = const()[name = tensor("qk_23_transpose_x_0"), val = tensor(false)]; + tensor qk_23_transpose_y_0 = const()[name = tensor("qk_23_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_209 = transpose(perm = transpose_87_perm_0, x = k_47_cast); + tensor transpose_210 = transpose(perm = transpose_86_perm_0, x = q_47_cast); + tensor qk_23_cast = matmul(transpose_x = qk_23_transpose_x_0, transpose_y = qk_23_transpose_y_0, x = transpose_210, y = transpose_209); + tensor var_1364_cast = softmax(axis = var_1299, x = qk_23_cast); + tensor var_1366_transpose_x_0 = const()[name = tensor("op_1366_transpose_x_0"), val = tensor(false)]; + tensor var_1366_transpose_y_0 = const()[name = tensor("op_1366_transpose_y_0"), val = tensor(false)]; + tensor transpose_211 = transpose(perm = var_1360, x = var_1359_cast); + tensor var_1366_cast = matmul(transpose_x = var_1366_transpose_x_0, transpose_y = var_1366_transpose_y_0, x = var_1364_cast, y = transpose_211); + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 1280])]; + tensor transpose_208 = transpose(perm = var_1367, x = var_1366_cast); + tensor x_143_cast = reshape(shape = concat_11, x = transpose_208); + tensor var_1372_to_fp16 = const()[name = tensor("op_1372_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457048960)))]; + tensor var_1373_to_fp16 = const()[name = tensor("op_1373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460325824)))]; + tensor var_1374_cast = linear(bias = var_1373_to_fp16, weight = var_1372_to_fp16, x = x_143_cast); + tensor x_145_cast = add(x = x_139_cast, y = var_1374_cast); + tensor var_1380_axes_0 = const()[name = tensor("op_1380_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460328448)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460331072)))]; + tensor var_1380_cast = layer_norm(axes = var_1380_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1305_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast); + tensor var_1389_to_fp16 = const()[name = tensor("op_1389_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460333696)))]; + tensor var_1390_to_fp16 = const()[name = tensor("op_1390_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473440960)))]; + tensor input_97_cast = linear(bias = var_1390_to_fp16, weight = var_1389_to_fp16, x = var_1380_cast); + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast = gelu(mode = x_149_mode_0, x = input_97_cast); + tensor var_1395_to_fp16 = const()[name = tensor("op_1395_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473451264)))]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486558528)))]; + tensor var_1397_cast = linear(bias = var_1396_to_fp16, weight = var_1395_to_fp16, x = x_149_cast); + tensor x_151_cast = add(x = x_145_cast, y = var_1397_cast); + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor var_1423_axes_0 = const()[name = tensor("op_1423_axes_0"), val = tensor([-1])]; + tensor blocks_12_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_12_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486561152)))]; + tensor blocks_12_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_12_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486563776)))]; + tensor var_1412_to_fp16 = const()[name = tensor("op_1412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1423_cast = layer_norm(axes = var_1423_axes_0, beta = blocks_12_attn_ln_bias_to_fp16, epsilon = var_1412_to_fp16, gamma = blocks_12_attn_ln_weight_to_fp16, x = x_151_cast); + tensor var_1434_to_fp16 = const()[name = tensor("op_1434_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486566400)))]; + tensor var_1435_to_fp16 = const()[name = tensor("op_1435_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489843264)))]; + tensor q_49_cast = linear(bias = var_1435_to_fp16, weight = var_1434_to_fp16, x = var_1423_cast); + tensor var_1438_to_fp16 = const()[name = tensor("op_1438_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489845888)))]; + tensor k_49_bias_0_to_fp16 = const()[name = tensor("k_49_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493122752)))]; + tensor k_49_cast = linear(bias = k_49_bias_0_to_fp16, weight = var_1438_to_fp16, x = var_1423_cast); + tensor var_1442_to_fp16 = const()[name = tensor("op_1442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493125376)))]; + tensor var_1443_to_fp16 = const()[name = tensor("op_1443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496402240)))]; + tensor v_49_cast = linear(bias = var_1443_to_fp16, weight = var_1442_to_fp16, x = var_1423_cast); + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([1, 1500, 20, -1])]; + tensor var_1452_cast = reshape(shape = var_1451, x = q_49_cast); + tensor const_248_to_fp16 = const()[name = tensor("const_248_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_51_cast = mul(x = var_1452_cast, y = const_248_to_fp16); + tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([1, 1500, 20, -1])]; + tensor var_1459_cast = reshape(shape = var_1458, x = k_49_cast); + tensor const_249_to_fp16 = const()[name = tensor("const_249_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_51_cast = mul(x = var_1459_cast, y = const_249_to_fp16); + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1500, 20, -1])]; + tensor var_1466_cast = reshape(shape = var_1465, x = v_49_cast); + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([0, 2, 1, 3])]; + tensor qk_25_transpose_x_0 = const()[name = tensor("qk_25_transpose_x_0"), val = tensor(false)]; + tensor qk_25_transpose_y_0 = const()[name = tensor("qk_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_205 = transpose(perm = transpose_89_perm_0, x = k_51_cast); + tensor transpose_206 = transpose(perm = transpose_88_perm_0, x = q_51_cast); + tensor qk_25_cast = matmul(transpose_x = qk_25_transpose_x_0, transpose_y = qk_25_transpose_y_0, x = transpose_206, y = transpose_205); + tensor var_1471_cast = softmax(axis = var_1406, x = qk_25_cast); + tensor var_1473_transpose_x_0 = const()[name = tensor("op_1473_transpose_x_0"), val = tensor(false)]; + tensor var_1473_transpose_y_0 = const()[name = tensor("op_1473_transpose_y_0"), val = tensor(false)]; + tensor transpose_207 = transpose(perm = var_1467, x = var_1466_cast); + tensor var_1473_cast = matmul(transpose_x = var_1473_transpose_x_0, transpose_y = var_1473_transpose_y_0, x = var_1471_cast, y = transpose_207); + tensor var_1474 = const()[name = tensor("op_1474"), val = tensor([0, 2, 1, 3])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 1500, 1280])]; + tensor transpose_204 = transpose(perm = var_1474, x = var_1473_cast); + tensor x_155_cast = reshape(shape = concat_12, x = transpose_204); + tensor var_1479_to_fp16 = const()[name = tensor("op_1479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496404864)))]; + tensor var_1480_to_fp16 = const()[name = tensor("op_1480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499681728)))]; + tensor var_1481_cast = linear(bias = var_1480_to_fp16, weight = var_1479_to_fp16, x = x_155_cast); + tensor x_157_cast = add(x = x_151_cast, y = var_1481_cast); + tensor var_1487_axes_0 = const()[name = tensor("op_1487_axes_0"), val = tensor([-1])]; + tensor blocks_12_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499684352)))]; + tensor blocks_12_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499686976)))]; + tensor var_1487_cast = layer_norm(axes = var_1487_axes_0, beta = blocks_12_mlp_ln_bias_to_fp16, epsilon = var_1412_to_fp16, gamma = blocks_12_mlp_ln_weight_to_fp16, x = x_157_cast); + tensor var_1496_to_fp16 = const()[name = tensor("op_1496_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499689600)))]; + tensor var_1497_to_fp16 = const()[name = tensor("op_1497_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512796864)))]; + tensor input_105_cast = linear(bias = var_1497_to_fp16, weight = var_1496_to_fp16, x = var_1487_cast); + tensor x_161_mode_0 = const()[name = tensor("x_161_mode_0"), val = tensor("EXACT")]; + tensor x_161_cast = gelu(mode = x_161_mode_0, x = input_105_cast); + tensor var_1502_to_fp16 = const()[name = tensor("op_1502_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512807168)))]; + tensor var_1503_to_fp16 = const()[name = tensor("op_1503_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525914432)))]; + tensor var_1504_cast = linear(bias = var_1503_to_fp16, weight = var_1502_to_fp16, x = x_161_cast); + tensor x_163_cast = add(x = x_157_cast, y = var_1504_cast); + tensor var_1513 = const()[name = tensor("op_1513"), val = tensor(-1)]; + tensor var_1530_axes_0 = const()[name = tensor("op_1530_axes_0"), val = tensor([-1])]; + tensor blocks_13_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_13_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525917056)))]; + tensor blocks_13_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_13_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525919680)))]; + tensor var_1519_to_fp16 = const()[name = tensor("op_1519_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1530_cast = layer_norm(axes = var_1530_axes_0, beta = blocks_13_attn_ln_bias_to_fp16, epsilon = var_1519_to_fp16, gamma = blocks_13_attn_ln_weight_to_fp16, x = x_163_cast); + tensor var_1541_to_fp16 = const()[name = tensor("op_1541_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525922304)))]; + tensor var_1542_to_fp16 = const()[name = tensor("op_1542_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529199168)))]; + tensor q_53_cast = linear(bias = var_1542_to_fp16, weight = var_1541_to_fp16, x = var_1530_cast); + tensor var_1545_to_fp16 = const()[name = tensor("op_1545_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529201792)))]; + tensor k_53_bias_0_to_fp16 = const()[name = tensor("k_53_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532478656)))]; + tensor k_53_cast = linear(bias = k_53_bias_0_to_fp16, weight = var_1545_to_fp16, x = var_1530_cast); + tensor var_1549_to_fp16 = const()[name = tensor("op_1549_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532481280)))]; + tensor var_1550_to_fp16 = const()[name = tensor("op_1550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535758144)))]; + tensor v_53_cast = linear(bias = var_1550_to_fp16, weight = var_1549_to_fp16, x = var_1530_cast); + tensor var_1558 = const()[name = tensor("op_1558"), val = tensor([1, 1500, 20, -1])]; + tensor var_1559_cast = reshape(shape = var_1558, x = q_53_cast); + tensor const_250_to_fp16 = const()[name = tensor("const_250_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_55_cast = mul(x = var_1559_cast, y = const_250_to_fp16); + tensor var_1565 = const()[name = tensor("op_1565"), val = tensor([1, 1500, 20, -1])]; + tensor var_1566_cast = reshape(shape = var_1565, x = k_53_cast); + tensor const_251_to_fp16 = const()[name = tensor("const_251_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_55_cast = mul(x = var_1566_cast, y = const_251_to_fp16); + tensor var_1572 = const()[name = tensor("op_1572"), val = tensor([1, 1500, 20, -1])]; + tensor var_1573_cast = reshape(shape = var_1572, x = v_53_cast); + tensor var_1574 = const()[name = tensor("op_1574"), val = tensor([0, 2, 1, 3])]; + tensor qk_27_transpose_x_0 = const()[name = tensor("qk_27_transpose_x_0"), val = tensor(false)]; + tensor qk_27_transpose_y_0 = const()[name = tensor("qk_27_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_201 = transpose(perm = transpose_91_perm_0, x = k_55_cast); + tensor transpose_202 = transpose(perm = transpose_90_perm_0, x = q_55_cast); + tensor qk_27_cast = matmul(transpose_x = qk_27_transpose_x_0, transpose_y = qk_27_transpose_y_0, x = transpose_202, y = transpose_201); + tensor var_1578_cast = softmax(axis = var_1513, x = qk_27_cast); + tensor var_1580_transpose_x_0 = const()[name = tensor("op_1580_transpose_x_0"), val = tensor(false)]; + tensor var_1580_transpose_y_0 = const()[name = tensor("op_1580_transpose_y_0"), val = tensor(false)]; + tensor transpose_203 = transpose(perm = var_1574, x = var_1573_cast); + tensor var_1580_cast = matmul(transpose_x = var_1580_transpose_x_0, transpose_y = var_1580_transpose_y_0, x = var_1578_cast, y = transpose_203); + tensor var_1581 = const()[name = tensor("op_1581"), val = tensor([0, 2, 1, 3])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([1, 1500, 1280])]; + tensor transpose_200 = transpose(perm = var_1581, x = var_1580_cast); + tensor x_167_cast = reshape(shape = concat_13, x = transpose_200); + tensor var_1586_to_fp16 = const()[name = tensor("op_1586_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535760768)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539037632)))]; + tensor var_1588_cast = linear(bias = var_1587_to_fp16, weight = var_1586_to_fp16, x = x_167_cast); + tensor x_169_cast = add(x = x_163_cast, y = var_1588_cast); + tensor var_1594_axes_0 = const()[name = tensor("op_1594_axes_0"), val = tensor([-1])]; + tensor blocks_13_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539040256)))]; + tensor blocks_13_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539042880)))]; + tensor var_1594_cast = layer_norm(axes = var_1594_axes_0, beta = blocks_13_mlp_ln_bias_to_fp16, epsilon = var_1519_to_fp16, gamma = blocks_13_mlp_ln_weight_to_fp16, x = x_169_cast); + tensor var_1603_to_fp16 = const()[name = tensor("op_1603_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539045504)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552152768)))]; + tensor input_113_cast = linear(bias = var_1604_to_fp16, weight = var_1603_to_fp16, x = var_1594_cast); + tensor x_173_mode_0 = const()[name = tensor("x_173_mode_0"), val = tensor("EXACT")]; + tensor x_173_cast = gelu(mode = x_173_mode_0, x = input_113_cast); + tensor var_1609_to_fp16 = const()[name = tensor("op_1609_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552163072)))]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565270336)))]; + tensor var_1611_cast = linear(bias = var_1610_to_fp16, weight = var_1609_to_fp16, x = x_173_cast); + tensor x_175_cast = add(x = x_169_cast, y = var_1611_cast); + tensor var_1620 = const()[name = tensor("op_1620"), val = tensor(-1)]; + tensor var_1637_axes_0 = const()[name = tensor("op_1637_axes_0"), val = tensor([-1])]; + tensor blocks_14_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_14_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565272960)))]; + tensor blocks_14_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_14_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565275584)))]; + tensor var_1626_to_fp16 = const()[name = tensor("op_1626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1637_cast = layer_norm(axes = var_1637_axes_0, beta = blocks_14_attn_ln_bias_to_fp16, epsilon = var_1626_to_fp16, gamma = blocks_14_attn_ln_weight_to_fp16, x = x_175_cast); + tensor var_1648_to_fp16 = const()[name = tensor("op_1648_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565278208)))]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568555072)))]; + tensor q_57_cast = linear(bias = var_1649_to_fp16, weight = var_1648_to_fp16, x = var_1637_cast); + tensor var_1652_to_fp16 = const()[name = tensor("op_1652_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568557696)))]; + tensor k_57_bias_0_to_fp16 = const()[name = tensor("k_57_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571834560)))]; + tensor k_57_cast = linear(bias = k_57_bias_0_to_fp16, weight = var_1652_to_fp16, x = var_1637_cast); + tensor var_1656_to_fp16 = const()[name = tensor("op_1656_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571837184)))]; + tensor var_1657_to_fp16 = const()[name = tensor("op_1657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575114048)))]; + tensor v_57_cast = linear(bias = var_1657_to_fp16, weight = var_1656_to_fp16, x = var_1637_cast); + tensor var_1665 = const()[name = tensor("op_1665"), val = tensor([1, 1500, 20, -1])]; + tensor var_1666_cast = reshape(shape = var_1665, x = q_57_cast); + tensor const_252_to_fp16 = const()[name = tensor("const_252_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_59_cast = mul(x = var_1666_cast, y = const_252_to_fp16); + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 1500, 20, -1])]; + tensor var_1673_cast = reshape(shape = var_1672, x = k_57_cast); + tensor const_253_to_fp16 = const()[name = tensor("const_253_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_59_cast = mul(x = var_1673_cast, y = const_253_to_fp16); + tensor var_1679 = const()[name = tensor("op_1679"), val = tensor([1, 1500, 20, -1])]; + tensor var_1680_cast = reshape(shape = var_1679, x = v_57_cast); + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([0, 2, 1, 3])]; + tensor qk_29_transpose_x_0 = const()[name = tensor("qk_29_transpose_x_0"), val = tensor(false)]; + tensor qk_29_transpose_y_0 = const()[name = tensor("qk_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_197 = transpose(perm = transpose_93_perm_0, x = k_59_cast); + tensor transpose_198 = transpose(perm = transpose_92_perm_0, x = q_59_cast); + tensor qk_29_cast = matmul(transpose_x = qk_29_transpose_x_0, transpose_y = qk_29_transpose_y_0, x = transpose_198, y = transpose_197); + tensor var_1685_cast = softmax(axis = var_1620, x = qk_29_cast); + tensor var_1687_transpose_x_0 = const()[name = tensor("op_1687_transpose_x_0"), val = tensor(false)]; + tensor var_1687_transpose_y_0 = const()[name = tensor("op_1687_transpose_y_0"), val = tensor(false)]; + tensor transpose_199 = transpose(perm = var_1681, x = var_1680_cast); + tensor var_1687_cast = matmul(transpose_x = var_1687_transpose_x_0, transpose_y = var_1687_transpose_y_0, x = var_1685_cast, y = transpose_199); + tensor var_1688 = const()[name = tensor("op_1688"), val = tensor([0, 2, 1, 3])]; + tensor concat_14 = const()[name = tensor("concat_14"), val = tensor([1, 1500, 1280])]; + tensor transpose_196 = transpose(perm = var_1688, x = var_1687_cast); + tensor x_179_cast = reshape(shape = concat_14, x = transpose_196); + tensor var_1693_to_fp16 = const()[name = tensor("op_1693_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575116672)))]; + tensor var_1694_to_fp16 = const()[name = tensor("op_1694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578393536)))]; + tensor var_1695_cast = linear(bias = var_1694_to_fp16, weight = var_1693_to_fp16, x = x_179_cast); + tensor x_181_cast = add(x = x_175_cast, y = var_1695_cast); + tensor var_1701_axes_0 = const()[name = tensor("op_1701_axes_0"), val = tensor([-1])]; + tensor blocks_14_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578396160)))]; + tensor blocks_14_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578398784)))]; + tensor var_1701_cast = layer_norm(axes = var_1701_axes_0, beta = blocks_14_mlp_ln_bias_to_fp16, epsilon = var_1626_to_fp16, gamma = blocks_14_mlp_ln_weight_to_fp16, x = x_181_cast); + tensor var_1710_to_fp16 = const()[name = tensor("op_1710_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578401408)))]; + tensor var_1711_to_fp16 = const()[name = tensor("op_1711_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591508672)))]; + tensor input_121_cast = linear(bias = var_1711_to_fp16, weight = var_1710_to_fp16, x = var_1701_cast); + tensor x_185_mode_0 = const()[name = tensor("x_185_mode_0"), val = tensor("EXACT")]; + tensor x_185_cast = gelu(mode = x_185_mode_0, x = input_121_cast); + tensor var_1716_to_fp16 = const()[name = tensor("op_1716_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591518976)))]; + tensor var_1717_to_fp16 = const()[name = tensor("op_1717_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604626240)))]; + tensor var_1718_cast = linear(bias = var_1717_to_fp16, weight = var_1716_to_fp16, x = x_185_cast); + tensor x_187_cast = add(x = x_181_cast, y = var_1718_cast); + tensor var_1727 = const()[name = tensor("op_1727"), val = tensor(-1)]; + tensor var_1744_axes_0 = const()[name = tensor("op_1744_axes_0"), val = tensor([-1])]; + tensor blocks_15_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_15_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604628864)))]; + tensor blocks_15_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_15_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604631488)))]; + tensor var_1733_to_fp16 = const()[name = tensor("op_1733_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1744_cast = layer_norm(axes = var_1744_axes_0, beta = blocks_15_attn_ln_bias_to_fp16, epsilon = var_1733_to_fp16, gamma = blocks_15_attn_ln_weight_to_fp16, x = x_187_cast); + tensor var_1755_to_fp16 = const()[name = tensor("op_1755_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604634112)))]; + tensor var_1756_to_fp16 = const()[name = tensor("op_1756_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607910976)))]; + tensor q_61_cast = linear(bias = var_1756_to_fp16, weight = var_1755_to_fp16, x = var_1744_cast); + tensor var_1759_to_fp16 = const()[name = tensor("op_1759_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(607913600)))]; + tensor k_61_bias_0_to_fp16 = const()[name = tensor("k_61_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611190464)))]; + tensor k_61_cast = linear(bias = k_61_bias_0_to_fp16, weight = var_1759_to_fp16, x = var_1744_cast); + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611193088)))]; + tensor var_1764_to_fp16 = const()[name = tensor("op_1764_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614469952)))]; + tensor v_61_cast = linear(bias = var_1764_to_fp16, weight = var_1763_to_fp16, x = var_1744_cast); + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1500, 20, -1])]; + tensor var_1773_cast = reshape(shape = var_1772, x = q_61_cast); + tensor const_254_to_fp16 = const()[name = tensor("const_254_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_63_cast = mul(x = var_1773_cast, y = const_254_to_fp16); + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 1500, 20, -1])]; + tensor var_1780_cast = reshape(shape = var_1779, x = k_61_cast); + tensor const_255_to_fp16 = const()[name = tensor("const_255_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_63_cast = mul(x = var_1780_cast, y = const_255_to_fp16); + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 1500, 20, -1])]; + tensor var_1787_cast = reshape(shape = var_1786, x = v_61_cast); + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([0, 2, 1, 3])]; + tensor qk_31_transpose_x_0 = const()[name = tensor("qk_31_transpose_x_0"), val = tensor(false)]; + tensor qk_31_transpose_y_0 = const()[name = tensor("qk_31_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_193 = transpose(perm = transpose_95_perm_0, x = k_63_cast); + tensor transpose_194 = transpose(perm = transpose_94_perm_0, x = q_63_cast); + tensor qk_31_cast = matmul(transpose_x = qk_31_transpose_x_0, transpose_y = qk_31_transpose_y_0, x = transpose_194, y = transpose_193); + tensor var_1792_cast = softmax(axis = var_1727, x = qk_31_cast); + tensor var_1794_transpose_x_0 = const()[name = tensor("op_1794_transpose_x_0"), val = tensor(false)]; + tensor var_1794_transpose_y_0 = const()[name = tensor("op_1794_transpose_y_0"), val = tensor(false)]; + tensor transpose_195 = transpose(perm = var_1788, x = var_1787_cast); + tensor var_1794_cast = matmul(transpose_x = var_1794_transpose_x_0, transpose_y = var_1794_transpose_y_0, x = var_1792_cast, y = transpose_195); + tensor var_1795 = const()[name = tensor("op_1795"), val = tensor([0, 2, 1, 3])]; + tensor concat_15 = const()[name = tensor("concat_15"), val = tensor([1, 1500, 1280])]; + tensor transpose_192 = transpose(perm = var_1795, x = var_1794_cast); + tensor x_191_cast = reshape(shape = concat_15, x = transpose_192); + tensor var_1800_to_fp16 = const()[name = tensor("op_1800_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614472576)))]; + tensor var_1801_to_fp16 = const()[name = tensor("op_1801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617749440)))]; + tensor var_1802_cast = linear(bias = var_1801_to_fp16, weight = var_1800_to_fp16, x = x_191_cast); + tensor x_193_cast = add(x = x_187_cast, y = var_1802_cast); + tensor var_1808_axes_0 = const()[name = tensor("op_1808_axes_0"), val = tensor([-1])]; + tensor blocks_15_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617752064)))]; + tensor blocks_15_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617754688)))]; + tensor var_1808_cast = layer_norm(axes = var_1808_axes_0, beta = blocks_15_mlp_ln_bias_to_fp16, epsilon = var_1733_to_fp16, gamma = blocks_15_mlp_ln_weight_to_fp16, x = x_193_cast); + tensor var_1817_to_fp16 = const()[name = tensor("op_1817_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617757312)))]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630864576)))]; + tensor input_129_cast = linear(bias = var_1818_to_fp16, weight = var_1817_to_fp16, x = var_1808_cast); + tensor x_197_mode_0 = const()[name = tensor("x_197_mode_0"), val = tensor("EXACT")]; + tensor x_197_cast = gelu(mode = x_197_mode_0, x = input_129_cast); + tensor var_1823_to_fp16 = const()[name = tensor("op_1823_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630874880)))]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643982144)))]; + tensor var_1825_cast = linear(bias = var_1824_to_fp16, weight = var_1823_to_fp16, x = x_197_cast); + tensor x_199_cast = add(x = x_193_cast, y = var_1825_cast); + tensor var_1834 = const()[name = tensor("op_1834"), val = tensor(-1)]; + tensor var_1851_axes_0 = const()[name = tensor("op_1851_axes_0"), val = tensor([-1])]; + tensor blocks_16_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_16_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643984768)))]; + tensor blocks_16_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_16_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643987392)))]; + tensor var_1840_to_fp16 = const()[name = tensor("op_1840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1851_cast = layer_norm(axes = var_1851_axes_0, beta = blocks_16_attn_ln_bias_to_fp16, epsilon = var_1840_to_fp16, gamma = blocks_16_attn_ln_weight_to_fp16, x = x_199_cast); + tensor var_1862_to_fp16 = const()[name = tensor("op_1862_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(643990016)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647266880)))]; + tensor q_65_cast = linear(bias = var_1863_to_fp16, weight = var_1862_to_fp16, x = var_1851_cast); + tensor var_1866_to_fp16 = const()[name = tensor("op_1866_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647269504)))]; + tensor k_65_bias_0_to_fp16 = const()[name = tensor("k_65_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650546368)))]; + tensor k_65_cast = linear(bias = k_65_bias_0_to_fp16, weight = var_1866_to_fp16, x = var_1851_cast); + tensor var_1870_to_fp16 = const()[name = tensor("op_1870_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650548992)))]; + tensor var_1871_to_fp16 = const()[name = tensor("op_1871_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653825856)))]; + tensor v_65_cast = linear(bias = var_1871_to_fp16, weight = var_1870_to_fp16, x = var_1851_cast); + tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([1, 1500, 20, -1])]; + tensor var_1880_cast = reshape(shape = var_1879, x = q_65_cast); + tensor const_256_to_fp16 = const()[name = tensor("const_256_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_67_cast = mul(x = var_1880_cast, y = const_256_to_fp16); + tensor var_1886 = const()[name = tensor("op_1886"), val = tensor([1, 1500, 20, -1])]; + tensor var_1887_cast = reshape(shape = var_1886, x = k_65_cast); + tensor const_257_to_fp16 = const()[name = tensor("const_257_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_67_cast = mul(x = var_1887_cast, y = const_257_to_fp16); + tensor var_1893 = const()[name = tensor("op_1893"), val = tensor([1, 1500, 20, -1])]; + tensor var_1894_cast = reshape(shape = var_1893, x = v_65_cast); + tensor var_1895 = const()[name = tensor("op_1895"), val = tensor([0, 2, 1, 3])]; + tensor qk_33_transpose_x_0 = const()[name = tensor("qk_33_transpose_x_0"), val = tensor(false)]; + tensor qk_33_transpose_y_0 = const()[name = tensor("qk_33_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_189 = transpose(perm = transpose_97_perm_0, x = k_67_cast); + tensor transpose_190 = transpose(perm = transpose_96_perm_0, x = q_67_cast); + tensor qk_33_cast = matmul(transpose_x = qk_33_transpose_x_0, transpose_y = qk_33_transpose_y_0, x = transpose_190, y = transpose_189); + tensor var_1899_cast = softmax(axis = var_1834, x = qk_33_cast); + tensor var_1901_transpose_x_0 = const()[name = tensor("op_1901_transpose_x_0"), val = tensor(false)]; + tensor var_1901_transpose_y_0 = const()[name = tensor("op_1901_transpose_y_0"), val = tensor(false)]; + tensor transpose_191 = transpose(perm = var_1895, x = var_1894_cast); + tensor var_1901_cast = matmul(transpose_x = var_1901_transpose_x_0, transpose_y = var_1901_transpose_y_0, x = var_1899_cast, y = transpose_191); + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([1, 1500, 1280])]; + tensor transpose_188 = transpose(perm = var_1902, x = var_1901_cast); + tensor x_203_cast = reshape(shape = concat_16, x = transpose_188); + tensor var_1907_to_fp16 = const()[name = tensor("op_1907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(653828480)))]; + tensor var_1908_to_fp16 = const()[name = tensor("op_1908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657105344)))]; + tensor var_1909_cast = linear(bias = var_1908_to_fp16, weight = var_1907_to_fp16, x = x_203_cast); + tensor x_205_cast = add(x = x_199_cast, y = var_1909_cast); + tensor var_1915_axes_0 = const()[name = tensor("op_1915_axes_0"), val = tensor([-1])]; + tensor blocks_16_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657107968)))]; + tensor blocks_16_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657110592)))]; + tensor var_1915_cast = layer_norm(axes = var_1915_axes_0, beta = blocks_16_mlp_ln_bias_to_fp16, epsilon = var_1840_to_fp16, gamma = blocks_16_mlp_ln_weight_to_fp16, x = x_205_cast); + tensor var_1924_to_fp16 = const()[name = tensor("op_1924_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657113216)))]; + tensor var_1925_to_fp16 = const()[name = tensor("op_1925_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670220480)))]; + tensor input_137_cast = linear(bias = var_1925_to_fp16, weight = var_1924_to_fp16, x = var_1915_cast); + tensor x_209_mode_0 = const()[name = tensor("x_209_mode_0"), val = tensor("EXACT")]; + tensor x_209_cast = gelu(mode = x_209_mode_0, x = input_137_cast); + tensor var_1930_to_fp16 = const()[name = tensor("op_1930_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670230784)))]; + tensor var_1931_to_fp16 = const()[name = tensor("op_1931_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683338048)))]; + tensor var_1932_cast = linear(bias = var_1931_to_fp16, weight = var_1930_to_fp16, x = x_209_cast); + tensor x_211_cast = add(x = x_205_cast, y = var_1932_cast); + tensor var_1941 = const()[name = tensor("op_1941"), val = tensor(-1)]; + tensor var_1958_axes_0 = const()[name = tensor("op_1958_axes_0"), val = tensor([-1])]; + tensor blocks_17_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_17_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683340672)))]; + tensor blocks_17_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_17_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683343296)))]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1958_cast = layer_norm(axes = var_1958_axes_0, beta = blocks_17_attn_ln_bias_to_fp16, epsilon = var_1947_to_fp16, gamma = blocks_17_attn_ln_weight_to_fp16, x = x_211_cast); + tensor var_1969_to_fp16 = const()[name = tensor("op_1969_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683345920)))]; + tensor var_1970_to_fp16 = const()[name = tensor("op_1970_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686622784)))]; + tensor q_69_cast = linear(bias = var_1970_to_fp16, weight = var_1969_to_fp16, x = var_1958_cast); + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686625408)))]; + tensor k_69_bias_0_to_fp16 = const()[name = tensor("k_69_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689902272)))]; + tensor k_69_cast = linear(bias = k_69_bias_0_to_fp16, weight = var_1973_to_fp16, x = var_1958_cast); + tensor var_1977_to_fp16 = const()[name = tensor("op_1977_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689904896)))]; + tensor var_1978_to_fp16 = const()[name = tensor("op_1978_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693181760)))]; + tensor v_69_cast = linear(bias = var_1978_to_fp16, weight = var_1977_to_fp16, x = var_1958_cast); + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 1500, 20, -1])]; + tensor var_1987_cast = reshape(shape = var_1986, x = q_69_cast); + tensor const_258_to_fp16 = const()[name = tensor("const_258_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_71_cast = mul(x = var_1987_cast, y = const_258_to_fp16); + tensor var_1993 = const()[name = tensor("op_1993"), val = tensor([1, 1500, 20, -1])]; + tensor var_1994_cast = reshape(shape = var_1993, x = k_69_cast); + tensor const_259_to_fp16 = const()[name = tensor("const_259_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_71_cast = mul(x = var_1994_cast, y = const_259_to_fp16); + tensor var_2000 = const()[name = tensor("op_2000"), val = tensor([1, 1500, 20, -1])]; + tensor var_2001_cast = reshape(shape = var_2000, x = v_69_cast); + tensor var_2002 = const()[name = tensor("op_2002"), val = tensor([0, 2, 1, 3])]; + tensor qk_35_transpose_x_0 = const()[name = tensor("qk_35_transpose_x_0"), val = tensor(false)]; + tensor qk_35_transpose_y_0 = const()[name = tensor("qk_35_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_185 = transpose(perm = transpose_99_perm_0, x = k_71_cast); + tensor transpose_186 = transpose(perm = transpose_98_perm_0, x = q_71_cast); + tensor qk_35_cast = matmul(transpose_x = qk_35_transpose_x_0, transpose_y = qk_35_transpose_y_0, x = transpose_186, y = transpose_185); + tensor var_2006_cast = softmax(axis = var_1941, x = qk_35_cast); + tensor var_2008_transpose_x_0 = const()[name = tensor("op_2008_transpose_x_0"), val = tensor(false)]; + tensor var_2008_transpose_y_0 = const()[name = tensor("op_2008_transpose_y_0"), val = tensor(false)]; + tensor transpose_187 = transpose(perm = var_2002, x = var_2001_cast); + tensor var_2008_cast = matmul(transpose_x = var_2008_transpose_x_0, transpose_y = var_2008_transpose_y_0, x = var_2006_cast, y = transpose_187); + tensor var_2009 = const()[name = tensor("op_2009"), val = tensor([0, 2, 1, 3])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([1, 1500, 1280])]; + tensor transpose_184 = transpose(perm = var_2009, x = var_2008_cast); + tensor x_215_cast = reshape(shape = concat_17, x = transpose_184); + tensor var_2014_to_fp16 = const()[name = tensor("op_2014_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693184384)))]; + tensor var_2015_to_fp16 = const()[name = tensor("op_2015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696461248)))]; + tensor var_2016_cast = linear(bias = var_2015_to_fp16, weight = var_2014_to_fp16, x = x_215_cast); + tensor x_217_cast = add(x = x_211_cast, y = var_2016_cast); + tensor var_2022_axes_0 = const()[name = tensor("op_2022_axes_0"), val = tensor([-1])]; + tensor blocks_17_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696463872)))]; + tensor blocks_17_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696466496)))]; + tensor var_2022_cast = layer_norm(axes = var_2022_axes_0, beta = blocks_17_mlp_ln_bias_to_fp16, epsilon = var_1947_to_fp16, gamma = blocks_17_mlp_ln_weight_to_fp16, x = x_217_cast); + tensor var_2031_to_fp16 = const()[name = tensor("op_2031_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696469120)))]; + tensor var_2032_to_fp16 = const()[name = tensor("op_2032_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709576384)))]; + tensor input_145_cast = linear(bias = var_2032_to_fp16, weight = var_2031_to_fp16, x = var_2022_cast); + tensor x_221_mode_0 = const()[name = tensor("x_221_mode_0"), val = tensor("EXACT")]; + tensor x_221_cast = gelu(mode = x_221_mode_0, x = input_145_cast); + tensor var_2037_to_fp16 = const()[name = tensor("op_2037_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709586688)))]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722693952)))]; + tensor var_2039_cast = linear(bias = var_2038_to_fp16, weight = var_2037_to_fp16, x = x_221_cast); + tensor x_223_cast = add(x = x_217_cast, y = var_2039_cast); + tensor var_2048 = const()[name = tensor("op_2048"), val = tensor(-1)]; + tensor var_2065_axes_0 = const()[name = tensor("op_2065_axes_0"), val = tensor([-1])]; + tensor blocks_18_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_18_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722696576)))]; + tensor blocks_18_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_18_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722699200)))]; + tensor var_2054_to_fp16 = const()[name = tensor("op_2054_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2065_cast = layer_norm(axes = var_2065_axes_0, beta = blocks_18_attn_ln_bias_to_fp16, epsilon = var_2054_to_fp16, gamma = blocks_18_attn_ln_weight_to_fp16, x = x_223_cast); + tensor var_2076_to_fp16 = const()[name = tensor("op_2076_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722701824)))]; + tensor var_2077_to_fp16 = const()[name = tensor("op_2077_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725978688)))]; + tensor q_73_cast = linear(bias = var_2077_to_fp16, weight = var_2076_to_fp16, x = var_2065_cast); + tensor var_2080_to_fp16 = const()[name = tensor("op_2080_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(725981312)))]; + tensor k_73_bias_0_to_fp16 = const()[name = tensor("k_73_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729258176)))]; + tensor k_73_cast = linear(bias = k_73_bias_0_to_fp16, weight = var_2080_to_fp16, x = var_2065_cast); + tensor var_2084_to_fp16 = const()[name = tensor("op_2084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729260800)))]; + tensor var_2085_to_fp16 = const()[name = tensor("op_2085_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732537664)))]; + tensor v_73_cast = linear(bias = var_2085_to_fp16, weight = var_2084_to_fp16, x = var_2065_cast); + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([1, 1500, 20, -1])]; + tensor var_2094_cast = reshape(shape = var_2093, x = q_73_cast); + tensor const_260_to_fp16 = const()[name = tensor("const_260_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_75_cast = mul(x = var_2094_cast, y = const_260_to_fp16); + tensor var_2100 = const()[name = tensor("op_2100"), val = tensor([1, 1500, 20, -1])]; + tensor var_2101_cast = reshape(shape = var_2100, x = k_73_cast); + tensor const_261_to_fp16 = const()[name = tensor("const_261_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_75_cast = mul(x = var_2101_cast, y = const_261_to_fp16); + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 1500, 20, -1])]; + tensor var_2108_cast = reshape(shape = var_2107, x = v_73_cast); + tensor var_2109 = const()[name = tensor("op_2109"), val = tensor([0, 2, 1, 3])]; + tensor qk_37_transpose_x_0 = const()[name = tensor("qk_37_transpose_x_0"), val = tensor(false)]; + tensor qk_37_transpose_y_0 = const()[name = tensor("qk_37_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_181 = transpose(perm = transpose_101_perm_0, x = k_75_cast); + tensor transpose_182 = transpose(perm = transpose_100_perm_0, x = q_75_cast); + tensor qk_37_cast = matmul(transpose_x = qk_37_transpose_x_0, transpose_y = qk_37_transpose_y_0, x = transpose_182, y = transpose_181); + tensor var_2113_cast = softmax(axis = var_2048, x = qk_37_cast); + tensor var_2115_transpose_x_0 = const()[name = tensor("op_2115_transpose_x_0"), val = tensor(false)]; + tensor var_2115_transpose_y_0 = const()[name = tensor("op_2115_transpose_y_0"), val = tensor(false)]; + tensor transpose_183 = transpose(perm = var_2109, x = var_2108_cast); + tensor var_2115_cast = matmul(transpose_x = var_2115_transpose_x_0, transpose_y = var_2115_transpose_y_0, x = var_2113_cast, y = transpose_183); + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor([0, 2, 1, 3])]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 1500, 1280])]; + tensor transpose_180 = transpose(perm = var_2116, x = var_2115_cast); + tensor x_227_cast = reshape(shape = concat_18, x = transpose_180); + tensor var_2121_to_fp16 = const()[name = tensor("op_2121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732540288)))]; + tensor var_2122_to_fp16 = const()[name = tensor("op_2122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735817152)))]; + tensor var_2123_cast = linear(bias = var_2122_to_fp16, weight = var_2121_to_fp16, x = x_227_cast); + tensor x_229_cast = add(x = x_223_cast, y = var_2123_cast); + tensor var_2129_axes_0 = const()[name = tensor("op_2129_axes_0"), val = tensor([-1])]; + tensor blocks_18_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735819776)))]; + tensor blocks_18_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735822400)))]; + tensor var_2129_cast = layer_norm(axes = var_2129_axes_0, beta = blocks_18_mlp_ln_bias_to_fp16, epsilon = var_2054_to_fp16, gamma = blocks_18_mlp_ln_weight_to_fp16, x = x_229_cast); + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735825024)))]; + tensor var_2139_to_fp16 = const()[name = tensor("op_2139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748932288)))]; + tensor input_153_cast = linear(bias = var_2139_to_fp16, weight = var_2138_to_fp16, x = var_2129_cast); + tensor x_233_mode_0 = const()[name = tensor("x_233_mode_0"), val = tensor("EXACT")]; + tensor x_233_cast = gelu(mode = x_233_mode_0, x = input_153_cast); + tensor var_2144_to_fp16 = const()[name = tensor("op_2144_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748942592)))]; + tensor var_2145_to_fp16 = const()[name = tensor("op_2145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762049856)))]; + tensor var_2146_cast = linear(bias = var_2145_to_fp16, weight = var_2144_to_fp16, x = x_233_cast); + tensor x_235_cast = add(x = x_229_cast, y = var_2146_cast); + tensor var_2155 = const()[name = tensor("op_2155"), val = tensor(-1)]; + tensor var_2172_axes_0 = const()[name = tensor("op_2172_axes_0"), val = tensor([-1])]; + tensor blocks_19_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_19_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762052480)))]; + tensor blocks_19_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_19_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762055104)))]; + tensor var_2161_to_fp16 = const()[name = tensor("op_2161_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2172_cast = layer_norm(axes = var_2172_axes_0, beta = blocks_19_attn_ln_bias_to_fp16, epsilon = var_2161_to_fp16, gamma = blocks_19_attn_ln_weight_to_fp16, x = x_235_cast); + tensor var_2183_to_fp16 = const()[name = tensor("op_2183_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762057728)))]; + tensor var_2184_to_fp16 = const()[name = tensor("op_2184_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765334592)))]; + tensor q_77_cast = linear(bias = var_2184_to_fp16, weight = var_2183_to_fp16, x = var_2172_cast); + tensor var_2187_to_fp16 = const()[name = tensor("op_2187_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765337216)))]; + tensor k_77_bias_0_to_fp16 = const()[name = tensor("k_77_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768614080)))]; + tensor k_77_cast = linear(bias = k_77_bias_0_to_fp16, weight = var_2187_to_fp16, x = var_2172_cast); + tensor var_2191_to_fp16 = const()[name = tensor("op_2191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768616704)))]; + tensor var_2192_to_fp16 = const()[name = tensor("op_2192_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771893568)))]; + tensor v_77_cast = linear(bias = var_2192_to_fp16, weight = var_2191_to_fp16, x = var_2172_cast); + tensor var_2200 = const()[name = tensor("op_2200"), val = tensor([1, 1500, 20, -1])]; + tensor var_2201_cast = reshape(shape = var_2200, x = q_77_cast); + tensor const_262_to_fp16 = const()[name = tensor("const_262_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_79_cast = mul(x = var_2201_cast, y = const_262_to_fp16); + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([1, 1500, 20, -1])]; + tensor var_2208_cast = reshape(shape = var_2207, x = k_77_cast); + tensor const_263_to_fp16 = const()[name = tensor("const_263_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_79_cast = mul(x = var_2208_cast, y = const_263_to_fp16); + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1500, 20, -1])]; + tensor var_2215_cast = reshape(shape = var_2214, x = v_77_cast); + tensor var_2216 = const()[name = tensor("op_2216"), val = tensor([0, 2, 1, 3])]; + tensor qk_39_transpose_x_0 = const()[name = tensor("qk_39_transpose_x_0"), val = tensor(false)]; + tensor qk_39_transpose_y_0 = const()[name = tensor("qk_39_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_177 = transpose(perm = transpose_103_perm_0, x = k_79_cast); + tensor transpose_178 = transpose(perm = transpose_102_perm_0, x = q_79_cast); + tensor qk_39_cast = matmul(transpose_x = qk_39_transpose_x_0, transpose_y = qk_39_transpose_y_0, x = transpose_178, y = transpose_177); + tensor var_2220_cast = softmax(axis = var_2155, x = qk_39_cast); + tensor var_2222_transpose_x_0 = const()[name = tensor("op_2222_transpose_x_0"), val = tensor(false)]; + tensor var_2222_transpose_y_0 = const()[name = tensor("op_2222_transpose_y_0"), val = tensor(false)]; + tensor transpose_179 = transpose(perm = var_2216, x = var_2215_cast); + tensor var_2222_cast = matmul(transpose_x = var_2222_transpose_x_0, transpose_y = var_2222_transpose_y_0, x = var_2220_cast, y = transpose_179); + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([0, 2, 1, 3])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([1, 1500, 1280])]; + tensor transpose_176 = transpose(perm = var_2223, x = var_2222_cast); + tensor x_239_cast = reshape(shape = concat_19, x = transpose_176); + tensor var_2228_to_fp16 = const()[name = tensor("op_2228_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(771896192)))]; + tensor var_2229_to_fp16 = const()[name = tensor("op_2229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775173056)))]; + tensor var_2230_cast = linear(bias = var_2229_to_fp16, weight = var_2228_to_fp16, x = x_239_cast); + tensor x_241_cast = add(x = x_235_cast, y = var_2230_cast); + tensor var_2236_axes_0 = const()[name = tensor("op_2236_axes_0"), val = tensor([-1])]; + tensor blocks_19_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775175680)))]; + tensor blocks_19_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775178304)))]; + tensor var_2236_cast = layer_norm(axes = var_2236_axes_0, beta = blocks_19_mlp_ln_bias_to_fp16, epsilon = var_2161_to_fp16, gamma = blocks_19_mlp_ln_weight_to_fp16, x = x_241_cast); + tensor var_2245_to_fp16 = const()[name = tensor("op_2245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775180928)))]; + tensor var_2246_to_fp16 = const()[name = tensor("op_2246_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788288192)))]; + tensor input_161_cast = linear(bias = var_2246_to_fp16, weight = var_2245_to_fp16, x = var_2236_cast); + tensor x_245_mode_0 = const()[name = tensor("x_245_mode_0"), val = tensor("EXACT")]; + tensor x_245_cast = gelu(mode = x_245_mode_0, x = input_161_cast); + tensor var_2251_to_fp16 = const()[name = tensor("op_2251_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788298496)))]; + tensor var_2252_to_fp16 = const()[name = tensor("op_2252_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801405760)))]; + tensor var_2253_cast = linear(bias = var_2252_to_fp16, weight = var_2251_to_fp16, x = x_245_cast); + tensor x_247_cast = add(x = x_241_cast, y = var_2253_cast); + tensor var_2262 = const()[name = tensor("op_2262"), val = tensor(-1)]; + tensor var_2279_axes_0 = const()[name = tensor("op_2279_axes_0"), val = tensor([-1])]; + tensor blocks_20_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_20_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801408384)))]; + tensor blocks_20_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_20_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801411008)))]; + tensor var_2268_to_fp16 = const()[name = tensor("op_2268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2279_cast = layer_norm(axes = var_2279_axes_0, beta = blocks_20_attn_ln_bias_to_fp16, epsilon = var_2268_to_fp16, gamma = blocks_20_attn_ln_weight_to_fp16, x = x_247_cast); + tensor var_2290_to_fp16 = const()[name = tensor("op_2290_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801413632)))]; + tensor var_2291_to_fp16 = const()[name = tensor("op_2291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804690496)))]; + tensor q_81_cast = linear(bias = var_2291_to_fp16, weight = var_2290_to_fp16, x = var_2279_cast); + tensor var_2294_to_fp16 = const()[name = tensor("op_2294_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804693120)))]; + tensor k_81_bias_0_to_fp16 = const()[name = tensor("k_81_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807969984)))]; + tensor k_81_cast = linear(bias = k_81_bias_0_to_fp16, weight = var_2294_to_fp16, x = var_2279_cast); + tensor var_2298_to_fp16 = const()[name = tensor("op_2298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807972608)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811249472)))]; + tensor v_81_cast = linear(bias = var_2299_to_fp16, weight = var_2298_to_fp16, x = var_2279_cast); + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 1500, 20, -1])]; + tensor var_2308_cast = reshape(shape = var_2307, x = q_81_cast); + tensor const_264_to_fp16 = const()[name = tensor("const_264_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_83_cast = mul(x = var_2308_cast, y = const_264_to_fp16); + tensor var_2314 = const()[name = tensor("op_2314"), val = tensor([1, 1500, 20, -1])]; + tensor var_2315_cast = reshape(shape = var_2314, x = k_81_cast); + tensor const_265_to_fp16 = const()[name = tensor("const_265_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_83_cast = mul(x = var_2315_cast, y = const_265_to_fp16); + tensor var_2321 = const()[name = tensor("op_2321"), val = tensor([1, 1500, 20, -1])]; + tensor var_2322_cast = reshape(shape = var_2321, x = v_81_cast); + tensor var_2323 = const()[name = tensor("op_2323"), val = tensor([0, 2, 1, 3])]; + tensor qk_41_transpose_x_0 = const()[name = tensor("qk_41_transpose_x_0"), val = tensor(false)]; + tensor qk_41_transpose_y_0 = const()[name = tensor("qk_41_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_173 = transpose(perm = transpose_105_perm_0, x = k_83_cast); + tensor transpose_174 = transpose(perm = transpose_104_perm_0, x = q_83_cast); + tensor qk_41_cast = matmul(transpose_x = qk_41_transpose_x_0, transpose_y = qk_41_transpose_y_0, x = transpose_174, y = transpose_173); + tensor var_2327_cast = softmax(axis = var_2262, x = qk_41_cast); + tensor var_2329_transpose_x_0 = const()[name = tensor("op_2329_transpose_x_0"), val = tensor(false)]; + tensor var_2329_transpose_y_0 = const()[name = tensor("op_2329_transpose_y_0"), val = tensor(false)]; + tensor transpose_175 = transpose(perm = var_2323, x = var_2322_cast); + tensor var_2329_cast = matmul(transpose_x = var_2329_transpose_x_0, transpose_y = var_2329_transpose_y_0, x = var_2327_cast, y = transpose_175); + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 1500, 1280])]; + tensor transpose_172 = transpose(perm = var_2330, x = var_2329_cast); + tensor x_251_cast = reshape(shape = concat_20, x = transpose_172); + tensor var_2335_to_fp16 = const()[name = tensor("op_2335_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811252096)))]; + tensor var_2336_to_fp16 = const()[name = tensor("op_2336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814528960)))]; + tensor var_2337_cast = linear(bias = var_2336_to_fp16, weight = var_2335_to_fp16, x = x_251_cast); + tensor x_253_cast = add(x = x_247_cast, y = var_2337_cast); + tensor var_2343_axes_0 = const()[name = tensor("op_2343_axes_0"), val = tensor([-1])]; + tensor blocks_20_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814531584)))]; + tensor blocks_20_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814534208)))]; + tensor var_2343_cast = layer_norm(axes = var_2343_axes_0, beta = blocks_20_mlp_ln_bias_to_fp16, epsilon = var_2268_to_fp16, gamma = blocks_20_mlp_ln_weight_to_fp16, x = x_253_cast); + tensor var_2352_to_fp16 = const()[name = tensor("op_2352_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814536832)))]; + tensor var_2353_to_fp16 = const()[name = tensor("op_2353_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827644096)))]; + tensor input_169_cast = linear(bias = var_2353_to_fp16, weight = var_2352_to_fp16, x = var_2343_cast); + tensor x_257_mode_0 = const()[name = tensor("x_257_mode_0"), val = tensor("EXACT")]; + tensor x_257_cast = gelu(mode = x_257_mode_0, x = input_169_cast); + tensor var_2358_to_fp16 = const()[name = tensor("op_2358_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827654400)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840761664)))]; + tensor var_2360_cast = linear(bias = var_2359_to_fp16, weight = var_2358_to_fp16, x = x_257_cast); + tensor x_259_cast = add(x = x_253_cast, y = var_2360_cast); + tensor var_2369 = const()[name = tensor("op_2369"), val = tensor(-1)]; + tensor var_2386_axes_0 = const()[name = tensor("op_2386_axes_0"), val = tensor([-1])]; + tensor blocks_21_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_21_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840764288)))]; + tensor blocks_21_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_21_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840766912)))]; + tensor var_2375_to_fp16 = const()[name = tensor("op_2375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2386_cast = layer_norm(axes = var_2386_axes_0, beta = blocks_21_attn_ln_bias_to_fp16, epsilon = var_2375_to_fp16, gamma = blocks_21_attn_ln_weight_to_fp16, x = x_259_cast); + tensor var_2397_to_fp16 = const()[name = tensor("op_2397_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840769536)))]; + tensor var_2398_to_fp16 = const()[name = tensor("op_2398_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844046400)))]; + tensor q_85_cast = linear(bias = var_2398_to_fp16, weight = var_2397_to_fp16, x = var_2386_cast); + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844049024)))]; + tensor k_85_bias_0_to_fp16 = const()[name = tensor("k_85_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847325888)))]; + tensor k_85_cast = linear(bias = k_85_bias_0_to_fp16, weight = var_2401_to_fp16, x = var_2386_cast); + tensor var_2405_to_fp16 = const()[name = tensor("op_2405_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847328512)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850605376)))]; + tensor v_85_cast = linear(bias = var_2406_to_fp16, weight = var_2405_to_fp16, x = var_2386_cast); + tensor var_2414 = const()[name = tensor("op_2414"), val = tensor([1, 1500, 20, -1])]; + tensor var_2415_cast = reshape(shape = var_2414, x = q_85_cast); + tensor const_266_to_fp16 = const()[name = tensor("const_266_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_87_cast = mul(x = var_2415_cast, y = const_266_to_fp16); + tensor var_2421 = const()[name = tensor("op_2421"), val = tensor([1, 1500, 20, -1])]; + tensor var_2422_cast = reshape(shape = var_2421, x = k_85_cast); + tensor const_267_to_fp16 = const()[name = tensor("const_267_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_87_cast = mul(x = var_2422_cast, y = const_267_to_fp16); + tensor var_2428 = const()[name = tensor("op_2428"), val = tensor([1, 1500, 20, -1])]; + tensor var_2429_cast = reshape(shape = var_2428, x = v_85_cast); + tensor var_2430 = const()[name = tensor("op_2430"), val = tensor([0, 2, 1, 3])]; + tensor qk_43_transpose_x_0 = const()[name = tensor("qk_43_transpose_x_0"), val = tensor(false)]; + tensor qk_43_transpose_y_0 = const()[name = tensor("qk_43_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_169 = transpose(perm = transpose_107_perm_0, x = k_87_cast); + tensor transpose_170 = transpose(perm = transpose_106_perm_0, x = q_87_cast); + tensor qk_43_cast = matmul(transpose_x = qk_43_transpose_x_0, transpose_y = qk_43_transpose_y_0, x = transpose_170, y = transpose_169); + tensor var_2434_cast = softmax(axis = var_2369, x = qk_43_cast); + tensor var_2436_transpose_x_0 = const()[name = tensor("op_2436_transpose_x_0"), val = tensor(false)]; + tensor var_2436_transpose_y_0 = const()[name = tensor("op_2436_transpose_y_0"), val = tensor(false)]; + tensor transpose_171 = transpose(perm = var_2430, x = var_2429_cast); + tensor var_2436_cast = matmul(transpose_x = var_2436_transpose_x_0, transpose_y = var_2436_transpose_y_0, x = var_2434_cast, y = transpose_171); + tensor var_2437 = const()[name = tensor("op_2437"), val = tensor([0, 2, 1, 3])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([1, 1500, 1280])]; + tensor transpose_168 = transpose(perm = var_2437, x = var_2436_cast); + tensor x_263_cast = reshape(shape = concat_21, x = transpose_168); + tensor var_2442_to_fp16 = const()[name = tensor("op_2442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850608000)))]; + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853884864)))]; + tensor var_2444_cast = linear(bias = var_2443_to_fp16, weight = var_2442_to_fp16, x = x_263_cast); + tensor x_265_cast = add(x = x_259_cast, y = var_2444_cast); + tensor var_2450_axes_0 = const()[name = tensor("op_2450_axes_0"), val = tensor([-1])]; + tensor blocks_21_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853887488)))]; + tensor blocks_21_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853890112)))]; + tensor var_2450_cast = layer_norm(axes = var_2450_axes_0, beta = blocks_21_mlp_ln_bias_to_fp16, epsilon = var_2375_to_fp16, gamma = blocks_21_mlp_ln_weight_to_fp16, x = x_265_cast); + tensor var_2459_to_fp16 = const()[name = tensor("op_2459_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(853892736)))]; + tensor var_2460_to_fp16 = const()[name = tensor("op_2460_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867000000)))]; + tensor input_177_cast = linear(bias = var_2460_to_fp16, weight = var_2459_to_fp16, x = var_2450_cast); + tensor x_269_mode_0 = const()[name = tensor("x_269_mode_0"), val = tensor("EXACT")]; + tensor x_269_cast = gelu(mode = x_269_mode_0, x = input_177_cast); + tensor var_2465_to_fp16 = const()[name = tensor("op_2465_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867010304)))]; + tensor var_2466_to_fp16 = const()[name = tensor("op_2466_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880117568)))]; + tensor var_2467_cast = linear(bias = var_2466_to_fp16, weight = var_2465_to_fp16, x = x_269_cast); + tensor x_271_cast = add(x = x_265_cast, y = var_2467_cast); + tensor var_2476 = const()[name = tensor("op_2476"), val = tensor(-1)]; + tensor var_2493_axes_0 = const()[name = tensor("op_2493_axes_0"), val = tensor([-1])]; + tensor blocks_22_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_22_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880120192)))]; + tensor blocks_22_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_22_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880122816)))]; + tensor var_2482_to_fp16 = const()[name = tensor("op_2482_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2493_cast = layer_norm(axes = var_2493_axes_0, beta = blocks_22_attn_ln_bias_to_fp16, epsilon = var_2482_to_fp16, gamma = blocks_22_attn_ln_weight_to_fp16, x = x_271_cast); + tensor var_2504_to_fp16 = const()[name = tensor("op_2504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880125440)))]; + tensor var_2505_to_fp16 = const()[name = tensor("op_2505_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883402304)))]; + tensor q_89_cast = linear(bias = var_2505_to_fp16, weight = var_2504_to_fp16, x = var_2493_cast); + tensor var_2508_to_fp16 = const()[name = tensor("op_2508_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883404928)))]; + tensor k_89_bias_0_to_fp16 = const()[name = tensor("k_89_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886681792)))]; + tensor k_89_cast = linear(bias = k_89_bias_0_to_fp16, weight = var_2508_to_fp16, x = var_2493_cast); + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886684416)))]; + tensor var_2513_to_fp16 = const()[name = tensor("op_2513_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889961280)))]; + tensor v_89_cast = linear(bias = var_2513_to_fp16, weight = var_2512_to_fp16, x = var_2493_cast); + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([1, 1500, 20, -1])]; + tensor var_2522_cast = reshape(shape = var_2521, x = q_89_cast); + tensor const_268_to_fp16 = const()[name = tensor("const_268_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_91_cast = mul(x = var_2522_cast, y = const_268_to_fp16); + tensor var_2528 = const()[name = tensor("op_2528"), val = tensor([1, 1500, 20, -1])]; + tensor var_2529_cast = reshape(shape = var_2528, x = k_89_cast); + tensor const_269_to_fp16 = const()[name = tensor("const_269_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_91_cast = mul(x = var_2529_cast, y = const_269_to_fp16); + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor([1, 1500, 20, -1])]; + tensor var_2536_cast = reshape(shape = var_2535, x = v_89_cast); + tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([0, 2, 1, 3])]; + tensor qk_45_transpose_x_0 = const()[name = tensor("qk_45_transpose_x_0"), val = tensor(false)]; + tensor qk_45_transpose_y_0 = const()[name = tensor("qk_45_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_165 = transpose(perm = transpose_109_perm_0, x = k_91_cast); + tensor transpose_166 = transpose(perm = transpose_108_perm_0, x = q_91_cast); + tensor qk_45_cast = matmul(transpose_x = qk_45_transpose_x_0, transpose_y = qk_45_transpose_y_0, x = transpose_166, y = transpose_165); + tensor var_2541_cast = softmax(axis = var_2476, x = qk_45_cast); + tensor var_2543_transpose_x_0 = const()[name = tensor("op_2543_transpose_x_0"), val = tensor(false)]; + tensor var_2543_transpose_y_0 = const()[name = tensor("op_2543_transpose_y_0"), val = tensor(false)]; + tensor transpose_167 = transpose(perm = var_2537, x = var_2536_cast); + tensor var_2543_cast = matmul(transpose_x = var_2543_transpose_x_0, transpose_y = var_2543_transpose_y_0, x = var_2541_cast, y = transpose_167); + tensor var_2544 = const()[name = tensor("op_2544"), val = tensor([0, 2, 1, 3])]; + tensor concat_22 = const()[name = tensor("concat_22"), val = tensor([1, 1500, 1280])]; + tensor transpose_164 = transpose(perm = var_2544, x = var_2543_cast); + tensor x_275_cast = reshape(shape = concat_22, x = transpose_164); + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(889963904)))]; + tensor var_2550_to_fp16 = const()[name = tensor("op_2550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893240768)))]; + tensor var_2551_cast = linear(bias = var_2550_to_fp16, weight = var_2549_to_fp16, x = x_275_cast); + tensor x_277_cast = add(x = x_271_cast, y = var_2551_cast); + tensor var_2557_axes_0 = const()[name = tensor("op_2557_axes_0"), val = tensor([-1])]; + tensor blocks_22_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893243392)))]; + tensor blocks_22_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893246016)))]; + tensor var_2557_cast = layer_norm(axes = var_2557_axes_0, beta = blocks_22_mlp_ln_bias_to_fp16, epsilon = var_2482_to_fp16, gamma = blocks_22_mlp_ln_weight_to_fp16, x = x_277_cast); + tensor var_2566_to_fp16 = const()[name = tensor("op_2566_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893248640)))]; + tensor var_2567_to_fp16 = const()[name = tensor("op_2567_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906355904)))]; + tensor input_185_cast = linear(bias = var_2567_to_fp16, weight = var_2566_to_fp16, x = var_2557_cast); + tensor x_281_mode_0 = const()[name = tensor("x_281_mode_0"), val = tensor("EXACT")]; + tensor x_281_cast = gelu(mode = x_281_mode_0, x = input_185_cast); + tensor var_2572_to_fp16 = const()[name = tensor("op_2572_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906366208)))]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919473472)))]; + tensor var_2574_cast = linear(bias = var_2573_to_fp16, weight = var_2572_to_fp16, x = x_281_cast); + tensor x_283_cast = add(x = x_277_cast, y = var_2574_cast); + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor var_2600_axes_0 = const()[name = tensor("op_2600_axes_0"), val = tensor([-1])]; + tensor blocks_23_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_23_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919476096)))]; + tensor blocks_23_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_23_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919478720)))]; + tensor var_2589_to_fp16 = const()[name = tensor("op_2589_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2600_cast = layer_norm(axes = var_2600_axes_0, beta = blocks_23_attn_ln_bias_to_fp16, epsilon = var_2589_to_fp16, gamma = blocks_23_attn_ln_weight_to_fp16, x = x_283_cast); + tensor var_2611_to_fp16 = const()[name = tensor("op_2611_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919481344)))]; + tensor var_2612_to_fp16 = const()[name = tensor("op_2612_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922758208)))]; + tensor q_93_cast = linear(bias = var_2612_to_fp16, weight = var_2611_to_fp16, x = var_2600_cast); + tensor var_2615_to_fp16 = const()[name = tensor("op_2615_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(922760832)))]; + tensor k_93_bias_0_to_fp16 = const()[name = tensor("k_93_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926037696)))]; + tensor k_93_cast = linear(bias = k_93_bias_0_to_fp16, weight = var_2615_to_fp16, x = var_2600_cast); + tensor var_2619_to_fp16 = const()[name = tensor("op_2619_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926040320)))]; + tensor var_2620_to_fp16 = const()[name = tensor("op_2620_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929317184)))]; + tensor v_93_cast = linear(bias = var_2620_to_fp16, weight = var_2619_to_fp16, x = var_2600_cast); + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 1500, 20, -1])]; + tensor var_2629_cast = reshape(shape = var_2628, x = q_93_cast); + tensor const_270_to_fp16 = const()[name = tensor("const_270_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_95_cast = mul(x = var_2629_cast, y = const_270_to_fp16); + tensor var_2635 = const()[name = tensor("op_2635"), val = tensor([1, 1500, 20, -1])]; + tensor var_2636_cast = reshape(shape = var_2635, x = k_93_cast); + tensor const_271_to_fp16 = const()[name = tensor("const_271_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_95_cast = mul(x = var_2636_cast, y = const_271_to_fp16); + tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 1500, 20, -1])]; + tensor var_2643_cast = reshape(shape = var_2642, x = v_93_cast); + tensor var_2644 = const()[name = tensor("op_2644"), val = tensor([0, 2, 1, 3])]; + tensor qk_47_transpose_x_0 = const()[name = tensor("qk_47_transpose_x_0"), val = tensor(false)]; + tensor qk_47_transpose_y_0 = const()[name = tensor("qk_47_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_161 = transpose(perm = transpose_111_perm_0, x = k_95_cast); + tensor transpose_162 = transpose(perm = transpose_110_perm_0, x = q_95_cast); + tensor qk_47_cast = matmul(transpose_x = qk_47_transpose_x_0, transpose_y = qk_47_transpose_y_0, x = transpose_162, y = transpose_161); + tensor var_2648_cast = softmax(axis = var_2583, x = qk_47_cast); + tensor var_2650_transpose_x_0 = const()[name = tensor("op_2650_transpose_x_0"), val = tensor(false)]; + tensor var_2650_transpose_y_0 = const()[name = tensor("op_2650_transpose_y_0"), val = tensor(false)]; + tensor transpose_163 = transpose(perm = var_2644, x = var_2643_cast); + tensor var_2650_cast = matmul(transpose_x = var_2650_transpose_x_0, transpose_y = var_2650_transpose_y_0, x = var_2648_cast, y = transpose_163); + tensor var_2651 = const()[name = tensor("op_2651"), val = tensor([0, 2, 1, 3])]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([1, 1500, 1280])]; + tensor transpose_160 = transpose(perm = var_2651, x = var_2650_cast); + tensor x_287_cast = reshape(shape = concat_23, x = transpose_160); + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929319808)))]; + tensor var_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932596672)))]; + tensor var_2658_cast = linear(bias = var_2657_to_fp16, weight = var_2656_to_fp16, x = x_287_cast); + tensor x_289_cast = add(x = x_283_cast, y = var_2658_cast); + tensor var_2664_axes_0 = const()[name = tensor("op_2664_axes_0"), val = tensor([-1])]; + tensor blocks_23_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932599296)))]; + tensor blocks_23_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932601920)))]; + tensor var_2664_cast = layer_norm(axes = var_2664_axes_0, beta = blocks_23_mlp_ln_bias_to_fp16, epsilon = var_2589_to_fp16, gamma = blocks_23_mlp_ln_weight_to_fp16, x = x_289_cast); + tensor var_2673_to_fp16 = const()[name = tensor("op_2673_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932604544)))]; + tensor var_2674_to_fp16 = const()[name = tensor("op_2674_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945711808)))]; + tensor input_193_cast = linear(bias = var_2674_to_fp16, weight = var_2673_to_fp16, x = var_2664_cast); + tensor x_293_mode_0 = const()[name = tensor("x_293_mode_0"), val = tensor("EXACT")]; + tensor x_293_cast = gelu(mode = x_293_mode_0, x = input_193_cast); + tensor var_2679_to_fp16 = const()[name = tensor("op_2679_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(945722112)))]; + tensor var_2680_to_fp16 = const()[name = tensor("op_2680_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958829376)))]; + tensor var_2681_cast = linear(bias = var_2680_to_fp16, weight = var_2679_to_fp16, x = x_293_cast); + tensor x_295_cast = add(x = x_289_cast, y = var_2681_cast); + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor(-1)]; + tensor var_2707_axes_0 = const()[name = tensor("op_2707_axes_0"), val = tensor([-1])]; + tensor blocks_24_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_24_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958832000)))]; + tensor blocks_24_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_24_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958834624)))]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2707_cast = layer_norm(axes = var_2707_axes_0, beta = blocks_24_attn_ln_bias_to_fp16, epsilon = var_2696_to_fp16, gamma = blocks_24_attn_ln_weight_to_fp16, x = x_295_cast); + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958837248)))]; + tensor var_2719_to_fp16 = const()[name = tensor("op_2719_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962114112)))]; + tensor q_97_cast = linear(bias = var_2719_to_fp16, weight = var_2718_to_fp16, x = var_2707_cast); + tensor var_2722_to_fp16 = const()[name = tensor("op_2722_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962116736)))]; + tensor k_97_bias_0_to_fp16 = const()[name = tensor("k_97_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965393600)))]; + tensor k_97_cast = linear(bias = k_97_bias_0_to_fp16, weight = var_2722_to_fp16, x = var_2707_cast); + tensor var_2726_to_fp16 = const()[name = tensor("op_2726_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965396224)))]; + tensor var_2727_to_fp16 = const()[name = tensor("op_2727_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968673088)))]; + tensor v_97_cast = linear(bias = var_2727_to_fp16, weight = var_2726_to_fp16, x = var_2707_cast); + tensor var_2735 = const()[name = tensor("op_2735"), val = tensor([1, 1500, 20, -1])]; + tensor var_2736_cast = reshape(shape = var_2735, x = q_97_cast); + tensor const_272_to_fp16 = const()[name = tensor("const_272_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_99_cast = mul(x = var_2736_cast, y = const_272_to_fp16); + tensor var_2742 = const()[name = tensor("op_2742"), val = tensor([1, 1500, 20, -1])]; + tensor var_2743_cast = reshape(shape = var_2742, x = k_97_cast); + tensor const_273_to_fp16 = const()[name = tensor("const_273_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_99_cast = mul(x = var_2743_cast, y = const_273_to_fp16); + tensor var_2749 = const()[name = tensor("op_2749"), val = tensor([1, 1500, 20, -1])]; + tensor var_2750_cast = reshape(shape = var_2749, x = v_97_cast); + tensor var_2751 = const()[name = tensor("op_2751"), val = tensor([0, 2, 1, 3])]; + tensor qk_49_transpose_x_0 = const()[name = tensor("qk_49_transpose_x_0"), val = tensor(false)]; + tensor qk_49_transpose_y_0 = const()[name = tensor("qk_49_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_157 = transpose(perm = transpose_113_perm_0, x = k_99_cast); + tensor transpose_158 = transpose(perm = transpose_112_perm_0, x = q_99_cast); + tensor qk_49_cast = matmul(transpose_x = qk_49_transpose_x_0, transpose_y = qk_49_transpose_y_0, x = transpose_158, y = transpose_157); + tensor var_2755_cast = softmax(axis = var_2690, x = qk_49_cast); + tensor var_2757_transpose_x_0 = const()[name = tensor("op_2757_transpose_x_0"), val = tensor(false)]; + tensor var_2757_transpose_y_0 = const()[name = tensor("op_2757_transpose_y_0"), val = tensor(false)]; + tensor transpose_159 = transpose(perm = var_2751, x = var_2750_cast); + tensor var_2757_cast = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = var_2755_cast, y = transpose_159); + tensor var_2758 = const()[name = tensor("op_2758"), val = tensor([0, 2, 1, 3])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([1, 1500, 1280])]; + tensor transpose_156 = transpose(perm = var_2758, x = var_2757_cast); + tensor x_299_cast = reshape(shape = concat_24, x = transpose_156); + tensor var_2763_to_fp16 = const()[name = tensor("op_2763_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968675712)))]; + tensor var_2764_to_fp16 = const()[name = tensor("op_2764_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971952576)))]; + tensor var_2765_cast = linear(bias = var_2764_to_fp16, weight = var_2763_to_fp16, x = x_299_cast); + tensor x_301_cast = add(x = x_295_cast, y = var_2765_cast); + tensor var_2771_axes_0 = const()[name = tensor("op_2771_axes_0"), val = tensor([-1])]; + tensor blocks_24_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971955200)))]; + tensor blocks_24_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971957824)))]; + tensor var_2771_cast = layer_norm(axes = var_2771_axes_0, beta = blocks_24_mlp_ln_bias_to_fp16, epsilon = var_2696_to_fp16, gamma = blocks_24_mlp_ln_weight_to_fp16, x = x_301_cast); + tensor var_2780_to_fp16 = const()[name = tensor("op_2780_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(971960448)))]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985067712)))]; + tensor input_201_cast = linear(bias = var_2781_to_fp16, weight = var_2780_to_fp16, x = var_2771_cast); + tensor x_305_mode_0 = const()[name = tensor("x_305_mode_0"), val = tensor("EXACT")]; + tensor x_305_cast = gelu(mode = x_305_mode_0, x = input_201_cast); + tensor var_2786_to_fp16 = const()[name = tensor("op_2786_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985078016)))]; + tensor var_2787_to_fp16 = const()[name = tensor("op_2787_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998185280)))]; + tensor var_2788_cast = linear(bias = var_2787_to_fp16, weight = var_2786_to_fp16, x = x_305_cast); + tensor x_307_cast = add(x = x_301_cast, y = var_2788_cast); + tensor var_2797 = const()[name = tensor("op_2797"), val = tensor(-1)]; + tensor var_2814_axes_0 = const()[name = tensor("op_2814_axes_0"), val = tensor([-1])]; + tensor blocks_25_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_25_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998187904)))]; + tensor blocks_25_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_25_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998190528)))]; + tensor var_2803_to_fp16 = const()[name = tensor("op_2803_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2814_cast = layer_norm(axes = var_2814_axes_0, beta = blocks_25_attn_ln_bias_to_fp16, epsilon = var_2803_to_fp16, gamma = blocks_25_attn_ln_weight_to_fp16, x = x_307_cast); + tensor var_2825_to_fp16 = const()[name = tensor("op_2825_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998193152)))]; + tensor var_2826_to_fp16 = const()[name = tensor("op_2826_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001470016)))]; + tensor q_101_cast = linear(bias = var_2826_to_fp16, weight = var_2825_to_fp16, x = var_2814_cast); + tensor var_2829_to_fp16 = const()[name = tensor("op_2829_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001472640)))]; + tensor k_101_bias_0_to_fp16 = const()[name = tensor("k_101_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1004749504)))]; + tensor k_101_cast = linear(bias = k_101_bias_0_to_fp16, weight = var_2829_to_fp16, x = var_2814_cast); + tensor var_2833_to_fp16 = const()[name = tensor("op_2833_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1004752128)))]; + tensor var_2834_to_fp16 = const()[name = tensor("op_2834_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008028992)))]; + tensor v_101_cast = linear(bias = var_2834_to_fp16, weight = var_2833_to_fp16, x = var_2814_cast); + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 1500, 20, -1])]; + tensor var_2843_cast = reshape(shape = var_2842, x = q_101_cast); + tensor const_274_to_fp16 = const()[name = tensor("const_274_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_103_cast = mul(x = var_2843_cast, y = const_274_to_fp16); + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1500, 20, -1])]; + tensor var_2850_cast = reshape(shape = var_2849, x = k_101_cast); + tensor const_275_to_fp16 = const()[name = tensor("const_275_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_103_cast = mul(x = var_2850_cast, y = const_275_to_fp16); + tensor var_2856 = const()[name = tensor("op_2856"), val = tensor([1, 1500, 20, -1])]; + tensor var_2857_cast = reshape(shape = var_2856, x = v_101_cast); + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([0, 2, 1, 3])]; + tensor qk_51_transpose_x_0 = const()[name = tensor("qk_51_transpose_x_0"), val = tensor(false)]; + tensor qk_51_transpose_y_0 = const()[name = tensor("qk_51_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_153 = transpose(perm = transpose_115_perm_0, x = k_103_cast); + tensor transpose_154 = transpose(perm = transpose_114_perm_0, x = q_103_cast); + tensor qk_51_cast = matmul(transpose_x = qk_51_transpose_x_0, transpose_y = qk_51_transpose_y_0, x = transpose_154, y = transpose_153); + tensor var_2862_cast = softmax(axis = var_2797, x = qk_51_cast); + tensor var_2864_transpose_x_0 = const()[name = tensor("op_2864_transpose_x_0"), val = tensor(false)]; + tensor var_2864_transpose_y_0 = const()[name = tensor("op_2864_transpose_y_0"), val = tensor(false)]; + tensor transpose_155 = transpose(perm = var_2858, x = var_2857_cast); + tensor var_2864_cast = matmul(transpose_x = var_2864_transpose_x_0, transpose_y = var_2864_transpose_y_0, x = var_2862_cast, y = transpose_155); + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor([0, 2, 1, 3])]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 1500, 1280])]; + tensor transpose_152 = transpose(perm = var_2865, x = var_2864_cast); + tensor x_311_cast = reshape(shape = concat_25, x = transpose_152); + tensor var_2870_to_fp16 = const()[name = tensor("op_2870_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008031616)))]; + tensor var_2871_to_fp16 = const()[name = tensor("op_2871_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011308480)))]; + tensor var_2872_cast = linear(bias = var_2871_to_fp16, weight = var_2870_to_fp16, x = x_311_cast); + tensor x_313_cast = add(x = x_307_cast, y = var_2872_cast); + tensor var_2878_axes_0 = const()[name = tensor("op_2878_axes_0"), val = tensor([-1])]; + tensor blocks_25_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011311104)))]; + tensor blocks_25_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011313728)))]; + tensor var_2878_cast = layer_norm(axes = var_2878_axes_0, beta = blocks_25_mlp_ln_bias_to_fp16, epsilon = var_2803_to_fp16, gamma = blocks_25_mlp_ln_weight_to_fp16, x = x_313_cast); + tensor var_2887_to_fp16 = const()[name = tensor("op_2887_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011316352)))]; + tensor var_2888_to_fp16 = const()[name = tensor("op_2888_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024423616)))]; + tensor input_209_cast = linear(bias = var_2888_to_fp16, weight = var_2887_to_fp16, x = var_2878_cast); + tensor x_317_mode_0 = const()[name = tensor("x_317_mode_0"), val = tensor("EXACT")]; + tensor x_317_cast = gelu(mode = x_317_mode_0, x = input_209_cast); + tensor var_2893_to_fp16 = const()[name = tensor("op_2893_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024433920)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037541184)))]; + tensor var_2895_cast = linear(bias = var_2894_to_fp16, weight = var_2893_to_fp16, x = x_317_cast); + tensor x_319_cast = add(x = x_313_cast, y = var_2895_cast); + tensor var_2904 = const()[name = tensor("op_2904"), val = tensor(-1)]; + tensor var_2921_axes_0 = const()[name = tensor("op_2921_axes_0"), val = tensor([-1])]; + tensor blocks_26_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_26_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037543808)))]; + tensor blocks_26_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_26_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037546432)))]; + tensor var_2910_to_fp16 = const()[name = tensor("op_2910_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2921_cast = layer_norm(axes = var_2921_axes_0, beta = blocks_26_attn_ln_bias_to_fp16, epsilon = var_2910_to_fp16, gamma = blocks_26_attn_ln_weight_to_fp16, x = x_319_cast); + tensor var_2932_to_fp16 = const()[name = tensor("op_2932_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037549056)))]; + tensor var_2933_to_fp16 = const()[name = tensor("op_2933_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040825920)))]; + tensor q_105_cast = linear(bias = var_2933_to_fp16, weight = var_2932_to_fp16, x = var_2921_cast); + tensor var_2936_to_fp16 = const()[name = tensor("op_2936_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040828544)))]; + tensor k_105_bias_0_to_fp16 = const()[name = tensor("k_105_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044105408)))]; + tensor k_105_cast = linear(bias = k_105_bias_0_to_fp16, weight = var_2936_to_fp16, x = var_2921_cast); + tensor var_2940_to_fp16 = const()[name = tensor("op_2940_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044108032)))]; + tensor var_2941_to_fp16 = const()[name = tensor("op_2941_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047384896)))]; + tensor v_105_cast = linear(bias = var_2941_to_fp16, weight = var_2940_to_fp16, x = var_2921_cast); + tensor var_2949 = const()[name = tensor("op_2949"), val = tensor([1, 1500, 20, -1])]; + tensor var_2950_cast = reshape(shape = var_2949, x = q_105_cast); + tensor const_276_to_fp16 = const()[name = tensor("const_276_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_107_cast = mul(x = var_2950_cast, y = const_276_to_fp16); + tensor var_2956 = const()[name = tensor("op_2956"), val = tensor([1, 1500, 20, -1])]; + tensor var_2957_cast = reshape(shape = var_2956, x = k_105_cast); + tensor const_277_to_fp16 = const()[name = tensor("const_277_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_107_cast = mul(x = var_2957_cast, y = const_277_to_fp16); + tensor var_2963 = const()[name = tensor("op_2963"), val = tensor([1, 1500, 20, -1])]; + tensor var_2964_cast = reshape(shape = var_2963, x = v_105_cast); + tensor var_2965 = const()[name = tensor("op_2965"), val = tensor([0, 2, 1, 3])]; + tensor qk_53_transpose_x_0 = const()[name = tensor("qk_53_transpose_x_0"), val = tensor(false)]; + tensor qk_53_transpose_y_0 = const()[name = tensor("qk_53_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_149 = transpose(perm = transpose_117_perm_0, x = k_107_cast); + tensor transpose_150 = transpose(perm = transpose_116_perm_0, x = q_107_cast); + tensor qk_53_cast = matmul(transpose_x = qk_53_transpose_x_0, transpose_y = qk_53_transpose_y_0, x = transpose_150, y = transpose_149); + tensor var_2969_cast = softmax(axis = var_2904, x = qk_53_cast); + tensor var_2971_transpose_x_0 = const()[name = tensor("op_2971_transpose_x_0"), val = tensor(false)]; + tensor var_2971_transpose_y_0 = const()[name = tensor("op_2971_transpose_y_0"), val = tensor(false)]; + tensor transpose_151 = transpose(perm = var_2965, x = var_2964_cast); + tensor var_2971_cast = matmul(transpose_x = var_2971_transpose_x_0, transpose_y = var_2971_transpose_y_0, x = var_2969_cast, y = transpose_151); + tensor var_2972 = const()[name = tensor("op_2972"), val = tensor([0, 2, 1, 3])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([1, 1500, 1280])]; + tensor transpose_148 = transpose(perm = var_2972, x = var_2971_cast); + tensor x_323_cast = reshape(shape = concat_26, x = transpose_148); + tensor var_2977_to_fp16 = const()[name = tensor("op_2977_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047387520)))]; + tensor var_2978_to_fp16 = const()[name = tensor("op_2978_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050664384)))]; + tensor var_2979_cast = linear(bias = var_2978_to_fp16, weight = var_2977_to_fp16, x = x_323_cast); + tensor x_325_cast = add(x = x_319_cast, y = var_2979_cast); + tensor var_2985_axes_0 = const()[name = tensor("op_2985_axes_0"), val = tensor([-1])]; + tensor blocks_26_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050667008)))]; + tensor blocks_26_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050669632)))]; + tensor var_2985_cast = layer_norm(axes = var_2985_axes_0, beta = blocks_26_mlp_ln_bias_to_fp16, epsilon = var_2910_to_fp16, gamma = blocks_26_mlp_ln_weight_to_fp16, x = x_325_cast); + tensor var_2994_to_fp16 = const()[name = tensor("op_2994_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050672256)))]; + tensor var_2995_to_fp16 = const()[name = tensor("op_2995_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063779520)))]; + tensor input_217_cast = linear(bias = var_2995_to_fp16, weight = var_2994_to_fp16, x = var_2985_cast); + tensor x_329_mode_0 = const()[name = tensor("x_329_mode_0"), val = tensor("EXACT")]; + tensor x_329_cast = gelu(mode = x_329_mode_0, x = input_217_cast); + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1063789824)))]; + tensor var_3001_to_fp16 = const()[name = tensor("op_3001_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076897088)))]; + tensor var_3002_cast = linear(bias = var_3001_to_fp16, weight = var_3000_to_fp16, x = x_329_cast); + tensor x_331_cast = add(x = x_325_cast, y = var_3002_cast); + tensor var_3011 = const()[name = tensor("op_3011"), val = tensor(-1)]; + tensor var_3028_axes_0 = const()[name = tensor("op_3028_axes_0"), val = tensor([-1])]; + tensor blocks_27_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_27_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076899712)))]; + tensor blocks_27_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_27_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076902336)))]; + tensor var_3017_to_fp16 = const()[name = tensor("op_3017_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3028_cast = layer_norm(axes = var_3028_axes_0, beta = blocks_27_attn_ln_bias_to_fp16, epsilon = var_3017_to_fp16, gamma = blocks_27_attn_ln_weight_to_fp16, x = x_331_cast); + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1076904960)))]; + tensor var_3040_to_fp16 = const()[name = tensor("op_3040_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080181824)))]; + tensor q_109_cast = linear(bias = var_3040_to_fp16, weight = var_3039_to_fp16, x = var_3028_cast); + tensor var_3043_to_fp16 = const()[name = tensor("op_3043_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080184448)))]; + tensor k_109_bias_0_to_fp16 = const()[name = tensor("k_109_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083461312)))]; + tensor k_109_cast = linear(bias = k_109_bias_0_to_fp16, weight = var_3043_to_fp16, x = var_3028_cast); + tensor var_3047_to_fp16 = const()[name = tensor("op_3047_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083463936)))]; + tensor var_3048_to_fp16 = const()[name = tensor("op_3048_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086740800)))]; + tensor v_109_cast = linear(bias = var_3048_to_fp16, weight = var_3047_to_fp16, x = var_3028_cast); + tensor var_3056 = const()[name = tensor("op_3056"), val = tensor([1, 1500, 20, -1])]; + tensor var_3057_cast = reshape(shape = var_3056, x = q_109_cast); + tensor const_278_to_fp16 = const()[name = tensor("const_278_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_111_cast = mul(x = var_3057_cast, y = const_278_to_fp16); + tensor var_3063 = const()[name = tensor("op_3063"), val = tensor([1, 1500, 20, -1])]; + tensor var_3064_cast = reshape(shape = var_3063, x = k_109_cast); + tensor const_279_to_fp16 = const()[name = tensor("const_279_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_111_cast = mul(x = var_3064_cast, y = const_279_to_fp16); + tensor var_3070 = const()[name = tensor("op_3070"), val = tensor([1, 1500, 20, -1])]; + tensor var_3071_cast = reshape(shape = var_3070, x = v_109_cast); + tensor var_3072 = const()[name = tensor("op_3072"), val = tensor([0, 2, 1, 3])]; + tensor qk_55_transpose_x_0 = const()[name = tensor("qk_55_transpose_x_0"), val = tensor(false)]; + tensor qk_55_transpose_y_0 = const()[name = tensor("qk_55_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_145 = transpose(perm = transpose_119_perm_0, x = k_111_cast); + tensor transpose_146 = transpose(perm = transpose_118_perm_0, x = q_111_cast); + tensor qk_55_cast = matmul(transpose_x = qk_55_transpose_x_0, transpose_y = qk_55_transpose_y_0, x = transpose_146, y = transpose_145); + tensor var_3076_cast = softmax(axis = var_3011, x = qk_55_cast); + tensor var_3078_transpose_x_0 = const()[name = tensor("op_3078_transpose_x_0"), val = tensor(false)]; + tensor var_3078_transpose_y_0 = const()[name = tensor("op_3078_transpose_y_0"), val = tensor(false)]; + tensor transpose_147 = transpose(perm = var_3072, x = var_3071_cast); + tensor var_3078_cast = matmul(transpose_x = var_3078_transpose_x_0, transpose_y = var_3078_transpose_y_0, x = var_3076_cast, y = transpose_147); + tensor var_3079 = const()[name = tensor("op_3079"), val = tensor([0, 2, 1, 3])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([1, 1500, 1280])]; + tensor transpose_144 = transpose(perm = var_3079, x = var_3078_cast); + tensor x_335_cast = reshape(shape = concat_27, x = transpose_144); + tensor var_3084_to_fp16 = const()[name = tensor("op_3084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086743424)))]; + tensor var_3085_to_fp16 = const()[name = tensor("op_3085_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090020288)))]; + tensor var_3086_cast = linear(bias = var_3085_to_fp16, weight = var_3084_to_fp16, x = x_335_cast); + tensor x_337_cast = add(x = x_331_cast, y = var_3086_cast); + tensor var_3092_axes_0 = const()[name = tensor("op_3092_axes_0"), val = tensor([-1])]; + tensor blocks_27_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090022912)))]; + tensor blocks_27_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090025536)))]; + tensor var_3092_cast = layer_norm(axes = var_3092_axes_0, beta = blocks_27_mlp_ln_bias_to_fp16, epsilon = var_3017_to_fp16, gamma = blocks_27_mlp_ln_weight_to_fp16, x = x_337_cast); + tensor var_3101_to_fp16 = const()[name = tensor("op_3101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090028160)))]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103135424)))]; + tensor input_225_cast = linear(bias = var_3102_to_fp16, weight = var_3101_to_fp16, x = var_3092_cast); + tensor x_341_mode_0 = const()[name = tensor("x_341_mode_0"), val = tensor("EXACT")]; + tensor x_341_cast = gelu(mode = x_341_mode_0, x = input_225_cast); + tensor var_3107_to_fp16 = const()[name = tensor("op_3107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103145728)))]; + tensor var_3108_to_fp16 = const()[name = tensor("op_3108_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116252992)))]; + tensor var_3109_cast = linear(bias = var_3108_to_fp16, weight = var_3107_to_fp16, x = x_341_cast); + tensor x_343_cast = add(x = x_337_cast, y = var_3109_cast); + tensor var_3118 = const()[name = tensor("op_3118"), val = tensor(-1)]; + tensor var_3135_axes_0 = const()[name = tensor("op_3135_axes_0"), val = tensor([-1])]; + tensor blocks_28_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_28_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116255616)))]; + tensor blocks_28_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_28_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116258240)))]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3135_cast = layer_norm(axes = var_3135_axes_0, beta = blocks_28_attn_ln_bias_to_fp16, epsilon = var_3124_to_fp16, gamma = blocks_28_attn_ln_weight_to_fp16, x = x_343_cast); + tensor var_3146_to_fp16 = const()[name = tensor("op_3146_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116260864)))]; + tensor var_3147_to_fp16 = const()[name = tensor("op_3147_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119537728)))]; + tensor q_113_cast = linear(bias = var_3147_to_fp16, weight = var_3146_to_fp16, x = var_3135_cast); + tensor var_3150_to_fp16 = const()[name = tensor("op_3150_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119540352)))]; + tensor k_113_bias_0_to_fp16 = const()[name = tensor("k_113_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1122817216)))]; + tensor k_113_cast = linear(bias = k_113_bias_0_to_fp16, weight = var_3150_to_fp16, x = var_3135_cast); + tensor var_3154_to_fp16 = const()[name = tensor("op_3154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1122819840)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126096704)))]; + tensor v_113_cast = linear(bias = var_3155_to_fp16, weight = var_3154_to_fp16, x = var_3135_cast); + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 1500, 20, -1])]; + tensor var_3164_cast = reshape(shape = var_3163, x = q_113_cast); + tensor const_280_to_fp16 = const()[name = tensor("const_280_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_115_cast = mul(x = var_3164_cast, y = const_280_to_fp16); + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 1500, 20, -1])]; + tensor var_3171_cast = reshape(shape = var_3170, x = k_113_cast); + tensor const_281_to_fp16 = const()[name = tensor("const_281_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_115_cast = mul(x = var_3171_cast, y = const_281_to_fp16); + tensor var_3177 = const()[name = tensor("op_3177"), val = tensor([1, 1500, 20, -1])]; + tensor var_3178_cast = reshape(shape = var_3177, x = v_113_cast); + tensor var_3179 = const()[name = tensor("op_3179"), val = tensor([0, 2, 1, 3])]; + tensor qk_57_transpose_x_0 = const()[name = tensor("qk_57_transpose_x_0"), val = tensor(false)]; + tensor qk_57_transpose_y_0 = const()[name = tensor("qk_57_transpose_y_0"), val = tensor(false)]; + tensor transpose_120_perm_0 = const()[name = tensor("transpose_120_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_121_perm_0 = const()[name = tensor("transpose_121_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_141 = transpose(perm = transpose_121_perm_0, x = k_115_cast); + tensor transpose_142 = transpose(perm = transpose_120_perm_0, x = q_115_cast); + tensor qk_57_cast = matmul(transpose_x = qk_57_transpose_x_0, transpose_y = qk_57_transpose_y_0, x = transpose_142, y = transpose_141); + tensor var_3183_cast = softmax(axis = var_3118, x = qk_57_cast); + tensor var_3185_transpose_x_0 = const()[name = tensor("op_3185_transpose_x_0"), val = tensor(false)]; + tensor var_3185_transpose_y_0 = const()[name = tensor("op_3185_transpose_y_0"), val = tensor(false)]; + tensor transpose_143 = transpose(perm = var_3179, x = var_3178_cast); + tensor var_3185_cast = matmul(transpose_x = var_3185_transpose_x_0, transpose_y = var_3185_transpose_y_0, x = var_3183_cast, y = transpose_143); + tensor var_3186 = const()[name = tensor("op_3186"), val = tensor([0, 2, 1, 3])]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 1500, 1280])]; + tensor transpose_140 = transpose(perm = var_3186, x = var_3185_cast); + tensor x_347_cast = reshape(shape = concat_28, x = transpose_140); + tensor var_3191_to_fp16 = const()[name = tensor("op_3191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126099328)))]; + tensor var_3192_to_fp16 = const()[name = tensor("op_3192_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129376192)))]; + tensor var_3193_cast = linear(bias = var_3192_to_fp16, weight = var_3191_to_fp16, x = x_347_cast); + tensor x_349_cast = add(x = x_343_cast, y = var_3193_cast); + tensor var_3199_axes_0 = const()[name = tensor("op_3199_axes_0"), val = tensor([-1])]; + tensor blocks_28_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129378816)))]; + tensor blocks_28_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129381440)))]; + tensor var_3199_cast = layer_norm(axes = var_3199_axes_0, beta = blocks_28_mlp_ln_bias_to_fp16, epsilon = var_3124_to_fp16, gamma = blocks_28_mlp_ln_weight_to_fp16, x = x_349_cast); + tensor var_3208_to_fp16 = const()[name = tensor("op_3208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129384064)))]; + tensor var_3209_to_fp16 = const()[name = tensor("op_3209_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142491328)))]; + tensor input_233_cast = linear(bias = var_3209_to_fp16, weight = var_3208_to_fp16, x = var_3199_cast); + tensor x_353_mode_0 = const()[name = tensor("x_353_mode_0"), val = tensor("EXACT")]; + tensor x_353_cast = gelu(mode = x_353_mode_0, x = input_233_cast); + tensor var_3214_to_fp16 = const()[name = tensor("op_3214_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142501632)))]; + tensor var_3215_to_fp16 = const()[name = tensor("op_3215_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155608896)))]; + tensor var_3216_cast = linear(bias = var_3215_to_fp16, weight = var_3214_to_fp16, x = x_353_cast); + tensor x_355_cast = add(x = x_349_cast, y = var_3216_cast); + tensor var_3225 = const()[name = tensor("op_3225"), val = tensor(-1)]; + tensor var_3242_axes_0 = const()[name = tensor("op_3242_axes_0"), val = tensor([-1])]; + tensor blocks_29_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_29_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155611520)))]; + tensor blocks_29_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_29_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155614144)))]; + tensor var_3231_to_fp16 = const()[name = tensor("op_3231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3242_cast = layer_norm(axes = var_3242_axes_0, beta = blocks_29_attn_ln_bias_to_fp16, epsilon = var_3231_to_fp16, gamma = blocks_29_attn_ln_weight_to_fp16, x = x_355_cast); + tensor var_3253_to_fp16 = const()[name = tensor("op_3253_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155616768)))]; + tensor var_3254_to_fp16 = const()[name = tensor("op_3254_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158893632)))]; + tensor q_117_cast = linear(bias = var_3254_to_fp16, weight = var_3253_to_fp16, x = var_3242_cast); + tensor var_3257_to_fp16 = const()[name = tensor("op_3257_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1158896256)))]; + tensor k_117_bias_0_to_fp16 = const()[name = tensor("k_117_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162173120)))]; + tensor k_117_cast = linear(bias = k_117_bias_0_to_fp16, weight = var_3257_to_fp16, x = var_3242_cast); + tensor var_3261_to_fp16 = const()[name = tensor("op_3261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162175744)))]; + tensor var_3262_to_fp16 = const()[name = tensor("op_3262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165452608)))]; + tensor v_117_cast = linear(bias = var_3262_to_fp16, weight = var_3261_to_fp16, x = var_3242_cast); + tensor var_3270 = const()[name = tensor("op_3270"), val = tensor([1, 1500, 20, -1])]; + tensor var_3271_cast = reshape(shape = var_3270, x = q_117_cast); + tensor const_282_to_fp16 = const()[name = tensor("const_282_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_119_cast = mul(x = var_3271_cast, y = const_282_to_fp16); + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1500, 20, -1])]; + tensor var_3278_cast = reshape(shape = var_3277, x = k_117_cast); + tensor const_283_to_fp16 = const()[name = tensor("const_283_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_119_cast = mul(x = var_3278_cast, y = const_283_to_fp16); + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([1, 1500, 20, -1])]; + tensor var_3285_cast = reshape(shape = var_3284, x = v_117_cast); + tensor var_3286 = const()[name = tensor("op_3286"), val = tensor([0, 2, 1, 3])]; + tensor qk_59_transpose_x_0 = const()[name = tensor("qk_59_transpose_x_0"), val = tensor(false)]; + tensor qk_59_transpose_y_0 = const()[name = tensor("qk_59_transpose_y_0"), val = tensor(false)]; + tensor transpose_122_perm_0 = const()[name = tensor("transpose_122_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_123_perm_0 = const()[name = tensor("transpose_123_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_137 = transpose(perm = transpose_123_perm_0, x = k_119_cast); + tensor transpose_138 = transpose(perm = transpose_122_perm_0, x = q_119_cast); + tensor qk_59_cast = matmul(transpose_x = qk_59_transpose_x_0, transpose_y = qk_59_transpose_y_0, x = transpose_138, y = transpose_137); + tensor var_3290_cast = softmax(axis = var_3225, x = qk_59_cast); + tensor var_3292_transpose_x_0 = const()[name = tensor("op_3292_transpose_x_0"), val = tensor(false)]; + tensor var_3292_transpose_y_0 = const()[name = tensor("op_3292_transpose_y_0"), val = tensor(false)]; + tensor transpose_139 = transpose(perm = var_3286, x = var_3285_cast); + tensor var_3292_cast = matmul(transpose_x = var_3292_transpose_x_0, transpose_y = var_3292_transpose_y_0, x = var_3290_cast, y = transpose_139); + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([0, 2, 1, 3])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([1, 1500, 1280])]; + tensor transpose_136 = transpose(perm = var_3293, x = var_3292_cast); + tensor x_359_cast = reshape(shape = concat_29, x = transpose_136); + tensor var_3298_to_fp16 = const()[name = tensor("op_3298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165455232)))]; + tensor var_3299_to_fp16 = const()[name = tensor("op_3299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168732096)))]; + tensor var_3300_cast = linear(bias = var_3299_to_fp16, weight = var_3298_to_fp16, x = x_359_cast); + tensor x_361_cast = add(x = x_355_cast, y = var_3300_cast); + tensor var_3306_axes_0 = const()[name = tensor("op_3306_axes_0"), val = tensor([-1])]; + tensor blocks_29_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168734720)))]; + tensor blocks_29_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168737344)))]; + tensor var_3306_cast = layer_norm(axes = var_3306_axes_0, beta = blocks_29_mlp_ln_bias_to_fp16, epsilon = var_3231_to_fp16, gamma = blocks_29_mlp_ln_weight_to_fp16, x = x_361_cast); + tensor var_3315_to_fp16 = const()[name = tensor("op_3315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168739968)))]; + tensor var_3316_to_fp16 = const()[name = tensor("op_3316_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181847232)))]; + tensor input_241_cast = linear(bias = var_3316_to_fp16, weight = var_3315_to_fp16, x = var_3306_cast); + tensor x_365_mode_0 = const()[name = tensor("x_365_mode_0"), val = tensor("EXACT")]; + tensor x_365_cast = gelu(mode = x_365_mode_0, x = input_241_cast); + tensor var_3321_to_fp16 = const()[name = tensor("op_3321_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1181857536)))]; + tensor var_3322_to_fp16 = const()[name = tensor("op_3322_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194964800)))]; + tensor var_3323_cast = linear(bias = var_3322_to_fp16, weight = var_3321_to_fp16, x = x_365_cast); + tensor x_367_cast = add(x = x_361_cast, y = var_3323_cast); + tensor var_3332 = const()[name = tensor("op_3332"), val = tensor(-1)]; + tensor var_3349_axes_0 = const()[name = tensor("op_3349_axes_0"), val = tensor([-1])]; + tensor blocks_30_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_30_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194967424)))]; + tensor blocks_30_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_30_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194970048)))]; + tensor var_3338_to_fp16 = const()[name = tensor("op_3338_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3349_cast = layer_norm(axes = var_3349_axes_0, beta = blocks_30_attn_ln_bias_to_fp16, epsilon = var_3338_to_fp16, gamma = blocks_30_attn_ln_weight_to_fp16, x = x_367_cast); + tensor var_3360_to_fp16 = const()[name = tensor("op_3360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1194972672)))]; + tensor var_3361_to_fp16 = const()[name = tensor("op_3361_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198249536)))]; + tensor q_121_cast = linear(bias = var_3361_to_fp16, weight = var_3360_to_fp16, x = var_3349_cast); + tensor var_3364_to_fp16 = const()[name = tensor("op_3364_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198252160)))]; + tensor k_121_bias_0_to_fp16 = const()[name = tensor("k_121_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201529024)))]; + tensor k_121_cast = linear(bias = k_121_bias_0_to_fp16, weight = var_3364_to_fp16, x = var_3349_cast); + tensor var_3368_to_fp16 = const()[name = tensor("op_3368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201531648)))]; + tensor var_3369_to_fp16 = const()[name = tensor("op_3369_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204808512)))]; + tensor v_121_cast = linear(bias = var_3369_to_fp16, weight = var_3368_to_fp16, x = var_3349_cast); + tensor var_3377 = const()[name = tensor("op_3377"), val = tensor([1, 1500, 20, -1])]; + tensor var_3378_cast = reshape(shape = var_3377, x = q_121_cast); + tensor const_284_to_fp16 = const()[name = tensor("const_284_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_123_cast = mul(x = var_3378_cast, y = const_284_to_fp16); + tensor var_3384 = const()[name = tensor("op_3384"), val = tensor([1, 1500, 20, -1])]; + tensor var_3385_cast = reshape(shape = var_3384, x = k_121_cast); + tensor const_285_to_fp16 = const()[name = tensor("const_285_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_123_cast = mul(x = var_3385_cast, y = const_285_to_fp16); + tensor var_3391 = const()[name = tensor("op_3391"), val = tensor([1, 1500, 20, -1])]; + tensor var_3392_cast = reshape(shape = var_3391, x = v_121_cast); + tensor var_3393 = const()[name = tensor("op_3393"), val = tensor([0, 2, 1, 3])]; + tensor qk_61_transpose_x_0 = const()[name = tensor("qk_61_transpose_x_0"), val = tensor(false)]; + tensor qk_61_transpose_y_0 = const()[name = tensor("qk_61_transpose_y_0"), val = tensor(false)]; + tensor transpose_124_perm_0 = const()[name = tensor("transpose_124_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_125_perm_0 = const()[name = tensor("transpose_125_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_133 = transpose(perm = transpose_125_perm_0, x = k_123_cast); + tensor transpose_134 = transpose(perm = transpose_124_perm_0, x = q_123_cast); + tensor qk_61_cast = matmul(transpose_x = qk_61_transpose_x_0, transpose_y = qk_61_transpose_y_0, x = transpose_134, y = transpose_133); + tensor var_3397_cast = softmax(axis = var_3332, x = qk_61_cast); + tensor var_3399_transpose_x_0 = const()[name = tensor("op_3399_transpose_x_0"), val = tensor(false)]; + tensor var_3399_transpose_y_0 = const()[name = tensor("op_3399_transpose_y_0"), val = tensor(false)]; + tensor transpose_135 = transpose(perm = var_3393, x = var_3392_cast); + tensor var_3399_cast = matmul(transpose_x = var_3399_transpose_x_0, transpose_y = var_3399_transpose_y_0, x = var_3397_cast, y = transpose_135); + tensor var_3400 = const()[name = tensor("op_3400"), val = tensor([0, 2, 1, 3])]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([1, 1500, 1280])]; + tensor transpose_132 = transpose(perm = var_3400, x = var_3399_cast); + tensor x_371_cast = reshape(shape = concat_30, x = transpose_132); + tensor var_3405_to_fp16 = const()[name = tensor("op_3405_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1204811136)))]; + tensor var_3406_to_fp16 = const()[name = tensor("op_3406_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208088000)))]; + tensor var_3407_cast = linear(bias = var_3406_to_fp16, weight = var_3405_to_fp16, x = x_371_cast); + tensor x_373_cast = add(x = x_367_cast, y = var_3407_cast); + tensor var_3413_axes_0 = const()[name = tensor("op_3413_axes_0"), val = tensor([-1])]; + tensor blocks_30_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208090624)))]; + tensor blocks_30_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208093248)))]; + tensor var_3413_cast = layer_norm(axes = var_3413_axes_0, beta = blocks_30_mlp_ln_bias_to_fp16, epsilon = var_3338_to_fp16, gamma = blocks_30_mlp_ln_weight_to_fp16, x = x_373_cast); + tensor var_3422_to_fp16 = const()[name = tensor("op_3422_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208095872)))]; + tensor var_3423_to_fp16 = const()[name = tensor("op_3423_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221203136)))]; + tensor input_249_cast = linear(bias = var_3423_to_fp16, weight = var_3422_to_fp16, x = var_3413_cast); + tensor x_377_mode_0 = const()[name = tensor("x_377_mode_0"), val = tensor("EXACT")]; + tensor x_377_cast = gelu(mode = x_377_mode_0, x = input_249_cast); + tensor var_3428_to_fp16 = const()[name = tensor("op_3428_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221213440)))]; + tensor var_3429_to_fp16 = const()[name = tensor("op_3429_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234320704)))]; + tensor var_3430_cast = linear(bias = var_3429_to_fp16, weight = var_3428_to_fp16, x = x_377_cast); + tensor x_379_cast = add(x = x_373_cast, y = var_3430_cast); + tensor var_3439 = const()[name = tensor("op_3439"), val = tensor(-1)]; + tensor var_3456_axes_0 = const()[name = tensor("op_3456_axes_0"), val = tensor([-1])]; + tensor blocks_31_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_31_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234323328)))]; + tensor blocks_31_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_31_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234325952)))]; + tensor var_3445_to_fp16 = const()[name = tensor("op_3445_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3456_cast = layer_norm(axes = var_3456_axes_0, beta = blocks_31_attn_ln_bias_to_fp16, epsilon = var_3445_to_fp16, gamma = blocks_31_attn_ln_weight_to_fp16, x = x_379_cast); + tensor var_3467_to_fp16 = const()[name = tensor("op_3467_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234328576)))]; + tensor var_3468_to_fp16 = const()[name = tensor("op_3468_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237605440)))]; + tensor q_125_cast = linear(bias = var_3468_to_fp16, weight = var_3467_to_fp16, x = var_3456_cast); + tensor var_3471_to_fp16 = const()[name = tensor("op_3471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237608064)))]; + tensor k_125_bias_0_to_fp16 = const()[name = tensor("k_125_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1240884928)))]; + tensor k_125_cast = linear(bias = k_125_bias_0_to_fp16, weight = var_3471_to_fp16, x = var_3456_cast); + tensor var_3475_to_fp16 = const()[name = tensor("op_3475_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1240887552)))]; + tensor var_3476_to_fp16 = const()[name = tensor("op_3476_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244164416)))]; + tensor v_125_cast = linear(bias = var_3476_to_fp16, weight = var_3475_to_fp16, x = var_3456_cast); + tensor var_3484 = const()[name = tensor("op_3484"), val = tensor([1, 1500, 20, -1])]; + tensor var_3485_cast = reshape(shape = var_3484, x = q_125_cast); + tensor const_286_to_fp16 = const()[name = tensor("const_286_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_3485_cast, y = const_286_to_fp16); + tensor var_3491 = const()[name = tensor("op_3491"), val = tensor([1, 1500, 20, -1])]; + tensor var_3492_cast = reshape(shape = var_3491, x = k_125_cast); + tensor const_287_to_fp16 = const()[name = tensor("const_287_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_3492_cast, y = const_287_to_fp16); + tensor var_3498 = const()[name = tensor("op_3498"), val = tensor([1, 1500, 20, -1])]; + tensor var_3499_cast = reshape(shape = var_3498, x = v_125_cast); + tensor var_3500 = const()[name = tensor("op_3500"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_126_perm_0 = const()[name = tensor("transpose_126_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_127_perm_0 = const()[name = tensor("transpose_127_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_129 = transpose(perm = transpose_127_perm_0, x = k_cast); + tensor transpose_130 = transpose(perm = transpose_126_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_130, y = transpose_129); + tensor var_3504_cast = softmax(axis = var_3439, x = qk_cast); + tensor var_3506_transpose_x_0 = const()[name = tensor("op_3506_transpose_x_0"), val = tensor(false)]; + tensor var_3506_transpose_y_0 = const()[name = tensor("op_3506_transpose_y_0"), val = tensor(false)]; + tensor transpose_131 = transpose(perm = var_3500, x = var_3499_cast); + tensor var_3506_cast = matmul(transpose_x = var_3506_transpose_x_0, transpose_y = var_3506_transpose_y_0, x = var_3504_cast, y = transpose_131); + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor([0, 2, 1, 3])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([1, 1500, 1280])]; + tensor transpose_128 = transpose(perm = var_3507, x = var_3506_cast); + tensor x_383_cast = reshape(shape = concat_31, x = transpose_128); + tensor var_3512_to_fp16 = const()[name = tensor("op_3512_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244167040)))]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247443904)))]; + tensor var_3514_cast = linear(bias = var_3513_to_fp16, weight = var_3512_to_fp16, x = x_383_cast); + tensor x_385_cast = add(x = x_379_cast, y = var_3514_cast); + tensor var_3520_axes_0 = const()[name = tensor("op_3520_axes_0"), val = tensor([-1])]; + tensor blocks_31_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247446528)))]; + tensor blocks_31_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247449152)))]; + tensor var_3520_cast = layer_norm(axes = var_3520_axes_0, beta = blocks_31_mlp_ln_bias_to_fp16, epsilon = var_3445_to_fp16, gamma = blocks_31_mlp_ln_weight_to_fp16, x = x_385_cast); + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247451776)))]; + tensor var_3530_to_fp16 = const()[name = tensor("op_3530_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260559040)))]; + tensor input_257_cast = linear(bias = var_3530_to_fp16, weight = var_3529_to_fp16, x = var_3520_cast); + tensor x_389_mode_0 = const()[name = tensor("x_389_mode_0"), val = tensor("EXACT")]; + tensor x_389_cast = gelu(mode = x_389_mode_0, x = input_257_cast); + tensor var_3535_to_fp16 = const()[name = tensor("op_3535_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260569344)))]; + tensor var_3536_to_fp16 = const()[name = tensor("op_3536_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273676608)))]; + tensor var_3537_cast = linear(bias = var_3536_to_fp16, weight = var_3535_to_fp16, x = x_389_cast); + tensor x_cast = add(x = x_385_cast, y = var_3537_cast); + tensor var_3550_axes_0 = const()[name = tensor("op_3550_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273679232)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273681856)))]; + tensor var_3541_to_fp16 = const()[name = tensor("op_3541_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3550_cast = layer_norm(axes = var_3550_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_3541_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_3550_cast_to_fp32_dtype_0 = const()[name = tensor("op_3550_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_3550_cast_to_fp32_dtype_0, x = var_3550_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..72260fa03b3ee9db252c676463047995330d178a --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v2-encoder.mlmodelc/ggml-large-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b68e1aa7ba7f46a3c567b432287498a827d18afe9277dc30b9b7d88ef3c4eac +size 1273684480 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..1690ea9132c7f9ee7e5c32e4d346a0539980b258 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94477dfe35f9c64b53018e113a04a733fd23e97b40edf59cdd6f097f36561094 +size 1115965184 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..7598faa7b5376925098d89d08ec071a8904ae521 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bebed3e509c43090741f8efb31fed6a628385362ed06fec850412b48fee365b5 +size 243 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fc7a2db478f834d9c58bce347e4f49bb45cdbac2 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:687f91697cd68cf9f0eed4d054bc496b6103242f0c1f6f3e0a77d4c017bf5837 +size 320 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..0fa1f456a26df941200737845af9ec5bfdc5c241 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/metadata.json @@ -0,0 +1,67 @@ +[ + { + "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" : { + "Linear" : 192, + "Matmul" : 64, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 32, + "Add" : 65, + "LayerNorm" : 65, + "Mul" : 64, + "Transpose" : 129, + "Gelu" : 34, + "Reshape" : 128 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "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==1.11.0", + "com.github.apple.coremltools.version" : "7.1" + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..ffbf467b89a68a1e199954605e80efeaf557b49d --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/model.mil @@ -0,0 +1,1896 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "1.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] +{ + func main(tensor logmel_data) { + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor([1])]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983168)))]; + tensor cast_193 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor("cast_193")]; + tensor var_84_cast_fp16 = conv(bias = bias_3_to_fp16, dilations = var_82, groups = var_72, pad = var_84_pad_0, pad_type = var_84_pad_type_0, strides = var_80, weight = weight_3_to_fp16, x = cast_193)[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_88 = const()[name = tensor("op_88"), val = tensor(1)]; + tensor var_97 = const()[name = tensor("op_97"), val = tensor([2])]; + tensor var_99 = const()[name = tensor("op_99"), val = tensor([1])]; + tensor var_101_pad_type_0 = const()[name = tensor("op_101_pad_type_0"), val = tensor("custom")]; + tensor var_101_pad_0 = const()[name = tensor("op_101_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985792)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10816256)))]; + tensor var_101_cast_fp16 = conv(bias = bias_7_to_fp16, dilations = var_99, groups = var_88, pad = var_101_pad_0, pad_type = var_101_pad_type_0, strides = var_97, weight = weight_7_to_fp16, x = input_1_cast_fp16)[name = tensor("op_101_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_101_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_106 = const()[name = tensor("op_106"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10818880)))]; + tensor transpose_320 = transpose(perm = var_106, x = x_3_cast_fp16)[name = tensor("transpose_320")]; + tensor var_109_cast_fp16 = add(x = transpose_320, y = positional_embedding_to_fp16)[name = tensor("op_109_cast_fp16")]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor(-1)]; + tensor var_139_axes_0 = const()[name = tensor("op_139_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14658944)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14661568)))]; + tensor var_128_to_fp16 = const()[name = tensor("op_128_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_139_cast_fp16 = layer_norm(axes = var_139_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_128_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_109_cast_fp16)[name = tensor("op_139_cast_fp16")]; + tensor var_150_to_fp16 = const()[name = tensor("op_150_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14664192)))]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17941056)))]; + tensor linear_0_cast_fp16 = linear(bias = var_151_to_fp16, weight = var_150_to_fp16, x = var_139_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17943680)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21220544)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_154_to_fp16, x = var_139_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_158_to_fp16 = const()[name = tensor("op_158_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21223168)))]; + tensor var_159_to_fp16 = const()[name = tensor("op_159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24500032)))]; + tensor linear_2_cast_fp16 = linear(bias = var_159_to_fp16, weight = var_158_to_fp16, x = var_139_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_167 = const()[name = tensor("op_167"), val = tensor([1, 1500, 20, -1])]; + tensor var_168_cast_fp16 = reshape(shape = var_167, x = linear_0_cast_fp16)[name = tensor("op_168_cast_fp16")]; + tensor const_224_to_fp16 = const()[name = tensor("const_224_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast_fp16 = mul(x = var_168_cast_fp16, y = const_224_to_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_174 = const()[name = tensor("op_174"), val = tensor([1, 1500, 20, -1])]; + tensor var_175_cast_fp16 = reshape(shape = var_174, x = linear_1_cast_fp16)[name = tensor("op_175_cast_fp16")]; + tensor const_225_to_fp16 = const()[name = tensor("const_225_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast_fp16 = mul(x = var_175_cast_fp16, y = const_225_to_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_181 = const()[name = tensor("op_181"), val = tensor([1, 1500, 20, -1])]; + tensor var_182_cast_fp16 = reshape(shape = var_181, x = linear_2_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_128_perm_0 = const()[name = tensor("transpose_128_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_129_perm_0 = const()[name = tensor("transpose_129_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_317 = transpose(perm = transpose_129_perm_0, x = k_3_cast_fp16)[name = tensor("transpose_317")]; + tensor transpose_318 = transpose(perm = transpose_128_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_318")]; + tensor qk_1_cast_fp16 = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_318, y = transpose_317)[name = tensor("qk_1_cast_fp16")]; + tensor var_187_cast_fp16 = softmax(axis = var_122, x = qk_1_cast_fp16)[name = tensor("op_187_cast_fp16")]; + tensor var_189_transpose_x_0 = const()[name = tensor("op_189_transpose_x_0"), val = tensor(false)]; + tensor var_189_transpose_y_0 = const()[name = tensor("op_189_transpose_y_0"), val = tensor(false)]; + tensor transpose_319 = transpose(perm = var_183, x = var_182_cast_fp16)[name = tensor("transpose_319")]; + tensor var_189_cast_fp16 = matmul(transpose_x = var_189_transpose_x_0, transpose_y = var_189_transpose_y_0, x = var_187_cast_fp16, y = transpose_319)[name = tensor("op_189_cast_fp16")]; + tensor var_190 = const()[name = tensor("op_190"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 1280])]; + tensor transpose_316 = transpose(perm = var_190, x = var_189_cast_fp16)[name = tensor("transpose_316")]; + tensor x_11_cast_fp16 = reshape(shape = concat_0, x = transpose_316)[name = tensor("x_11_cast_fp16")]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24502656)))]; + tensor var_196_to_fp16 = const()[name = tensor("op_196_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27779520)))]; + tensor linear_3_cast_fp16 = linear(bias = var_196_to_fp16, weight = var_195_to_fp16, x = x_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor x_13_cast_fp16 = add(x = var_109_cast_fp16, y = linear_3_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor var_203_axes_0 = const()[name = tensor("op_203_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27782144)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27784768)))]; + tensor var_203_cast_fp16 = layer_norm(axes = var_203_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_128_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast_fp16)[name = tensor("op_203_cast_fp16")]; + tensor var_212_to_fp16 = const()[name = tensor("op_212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27787392)))]; + tensor var_213_to_fp16 = const()[name = tensor("op_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894656)))]; + tensor linear_4_cast_fp16 = linear(bias = var_213_to_fp16, weight = var_212_to_fp16, x = var_203_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast_fp16 = gelu(mode = x_17_mode_0, x = linear_4_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor var_218_to_fp16 = const()[name = tensor("op_218_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40904960)))]; + tensor var_219_to_fp16 = const()[name = tensor("op_219_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54012224)))]; + tensor linear_5_cast_fp16 = linear(bias = var_219_to_fp16, weight = var_218_to_fp16, x = x_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor x_19_cast_fp16 = add(x = x_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor(-1)]; + tensor var_246_axes_0 = const()[name = tensor("op_246_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54014848)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54017472)))]; + tensor var_235_to_fp16 = const()[name = tensor("op_235_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_246_cast_fp16 = layer_norm(axes = var_246_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_235_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_257_to_fp16 = const()[name = tensor("op_257_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54020096)))]; + tensor var_258_to_fp16 = const()[name = tensor("op_258_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57296960)))]; + tensor linear_6_cast_fp16 = linear(bias = var_258_to_fp16, weight = var_257_to_fp16, x = var_246_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57299584)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_261_to_fp16, x = var_246_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor var_265_to_fp16 = const()[name = tensor("op_265_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60576448)))]; + tensor var_266_to_fp16 = const()[name = tensor("op_266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63853312)))]; + tensor linear_8_cast_fp16 = linear(bias = var_266_to_fp16, weight = var_265_to_fp16, x = var_246_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_274 = const()[name = tensor("op_274"), val = tensor([1, 1500, 20, -1])]; + tensor var_275_cast_fp16 = reshape(shape = var_274, x = linear_6_cast_fp16)[name = tensor("op_275_cast_fp16")]; + tensor const_226_to_fp16 = const()[name = tensor("const_226_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast_fp16 = mul(x = var_275_cast_fp16, y = const_226_to_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1500, 20, -1])]; + tensor var_282_cast_fp16 = reshape(shape = var_281, x = linear_7_cast_fp16)[name = tensor("op_282_cast_fp16")]; + tensor const_227_to_fp16 = const()[name = tensor("const_227_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast_fp16 = mul(x = var_282_cast_fp16, y = const_227_to_fp16)[name = tensor("k_7_cast_fp16")]; + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1500, 20, -1])]; + tensor var_289_cast_fp16 = reshape(shape = var_288, x = linear_8_cast_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_130_perm_0 = const()[name = tensor("transpose_130_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_131_perm_0 = const()[name = tensor("transpose_131_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_313 = transpose(perm = transpose_131_perm_0, x = k_7_cast_fp16)[name = tensor("transpose_313")]; + tensor transpose_314 = transpose(perm = transpose_130_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_314")]; + tensor qk_3_cast_fp16 = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_314, y = transpose_313)[name = tensor("qk_3_cast_fp16")]; + tensor var_294_cast_fp16 = softmax(axis = var_229, x = qk_3_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_296_transpose_x_0 = const()[name = tensor("op_296_transpose_x_0"), val = tensor(false)]; + tensor var_296_transpose_y_0 = const()[name = tensor("op_296_transpose_y_0"), val = tensor(false)]; + tensor transpose_315 = transpose(perm = var_290, x = var_289_cast_fp16)[name = tensor("transpose_315")]; + tensor var_296_cast_fp16 = matmul(transpose_x = var_296_transpose_x_0, transpose_y = var_296_transpose_y_0, x = var_294_cast_fp16, y = transpose_315)[name = tensor("op_296_cast_fp16")]; + tensor var_297 = const()[name = tensor("op_297"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 1280])]; + tensor transpose_312 = transpose(perm = var_297, x = var_296_cast_fp16)[name = tensor("transpose_312")]; + tensor x_23_cast_fp16 = reshape(shape = concat_1, x = transpose_312)[name = tensor("x_23_cast_fp16")]; + tensor var_302_to_fp16 = const()[name = tensor("op_302_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63855936)))]; + tensor var_303_to_fp16 = const()[name = tensor("op_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67132800)))]; + tensor linear_9_cast_fp16 = linear(bias = var_303_to_fp16, weight = var_302_to_fp16, x = x_23_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor x_25_cast_fp16 = add(x = x_19_cast_fp16, y = linear_9_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor var_310_axes_0 = const()[name = tensor("op_310_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67135424)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67138048)))]; + tensor var_310_cast_fp16 = layer_norm(axes = var_310_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_235_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67140672)))]; + tensor var_320_to_fp16 = const()[name = tensor("op_320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80247936)))]; + tensor linear_10_cast_fp16 = linear(bias = var_320_to_fp16, weight = var_319_to_fp16, x = var_310_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast_fp16 = gelu(mode = x_29_mode_0, x = linear_10_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor var_325_to_fp16 = const()[name = tensor("op_325_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80258240)))]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93365504)))]; + tensor linear_11_cast_fp16 = linear(bias = var_326_to_fp16, weight = var_325_to_fp16, x = x_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor x_31_cast_fp16 = add(x = x_25_cast_fp16, y = linear_11_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor var_336 = const()[name = tensor("op_336"), val = tensor(-1)]; + tensor var_353_axes_0 = const()[name = tensor("op_353_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93368128)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93370752)))]; + tensor var_342_to_fp16 = const()[name = tensor("op_342_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_353_cast_fp16 = layer_norm(axes = var_353_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_342_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast_fp16)[name = tensor("op_353_cast_fp16")]; + tensor var_364_to_fp16 = const()[name = tensor("op_364_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93373376)))]; + tensor var_365_to_fp16 = const()[name = tensor("op_365_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96650240)))]; + tensor linear_12_cast_fp16 = linear(bias = var_365_to_fp16, weight = var_364_to_fp16, x = var_353_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96652864)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_368_to_fp16, x = var_353_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_372_to_fp16 = const()[name = tensor("op_372_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99929728)))]; + tensor var_373_to_fp16 = const()[name = tensor("op_373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103206592)))]; + tensor linear_14_cast_fp16 = linear(bias = var_373_to_fp16, weight = var_372_to_fp16, x = var_353_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_381 = const()[name = tensor("op_381"), val = tensor([1, 1500, 20, -1])]; + tensor var_382_cast_fp16 = reshape(shape = var_381, x = linear_12_cast_fp16)[name = tensor("op_382_cast_fp16")]; + tensor const_228_to_fp16 = const()[name = tensor("const_228_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast_fp16 = mul(x = var_382_cast_fp16, y = const_228_to_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1500, 20, -1])]; + tensor var_389_cast_fp16 = reshape(shape = var_388, x = linear_13_cast_fp16)[name = tensor("op_389_cast_fp16")]; + tensor const_229_to_fp16 = const()[name = tensor("const_229_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast_fp16 = mul(x = var_389_cast_fp16, y = const_229_to_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_395 = const()[name = tensor("op_395"), val = tensor([1, 1500, 20, -1])]; + tensor var_396_cast_fp16 = reshape(shape = var_395, x = linear_14_cast_fp16)[name = tensor("op_396_cast_fp16")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_132_perm_0 = const()[name = tensor("transpose_132_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_133_perm_0 = const()[name = tensor("transpose_133_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_309 = transpose(perm = transpose_133_perm_0, x = k_11_cast_fp16)[name = tensor("transpose_309")]; + tensor transpose_310 = transpose(perm = transpose_132_perm_0, x = q_11_cast_fp16)[name = tensor("transpose_310")]; + tensor qk_5_cast_fp16 = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_310, y = transpose_309)[name = tensor("qk_5_cast_fp16")]; + tensor var_401_cast_fp16 = softmax(axis = var_336, x = qk_5_cast_fp16)[name = tensor("op_401_cast_fp16")]; + tensor var_403_transpose_x_0 = const()[name = tensor("op_403_transpose_x_0"), val = tensor(false)]; + tensor var_403_transpose_y_0 = const()[name = tensor("op_403_transpose_y_0"), val = tensor(false)]; + tensor transpose_311 = transpose(perm = var_397, x = var_396_cast_fp16)[name = tensor("transpose_311")]; + tensor var_403_cast_fp16 = matmul(transpose_x = var_403_transpose_x_0, transpose_y = var_403_transpose_y_0, x = var_401_cast_fp16, y = transpose_311)[name = tensor("op_403_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 1280])]; + tensor transpose_308 = transpose(perm = var_404, x = var_403_cast_fp16)[name = tensor("transpose_308")]; + tensor x_35_cast_fp16 = reshape(shape = concat_2, x = transpose_308)[name = tensor("x_35_cast_fp16")]; + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103209216)))]; + tensor var_410_to_fp16 = const()[name = tensor("op_410_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106486080)))]; + tensor linear_15_cast_fp16 = linear(bias = var_410_to_fp16, weight = var_409_to_fp16, x = x_35_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor x_37_cast_fp16 = add(x = x_31_cast_fp16, y = linear_15_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor var_417_axes_0 = const()[name = tensor("op_417_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106488704)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106491328)))]; + tensor var_417_cast_fp16 = layer_norm(axes = var_417_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_342_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106493952)))]; + tensor var_427_to_fp16 = const()[name = tensor("op_427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119601216)))]; + tensor linear_16_cast_fp16 = linear(bias = var_427_to_fp16, weight = var_426_to_fp16, x = var_417_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast_fp16 = gelu(mode = x_41_mode_0, x = linear_16_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor var_432_to_fp16 = const()[name = tensor("op_432_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119611520)))]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132718784)))]; + tensor linear_17_cast_fp16 = linear(bias = var_433_to_fp16, weight = var_432_to_fp16, x = x_41_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor x_43_cast_fp16 = add(x = x_37_cast_fp16, y = linear_17_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor(-1)]; + tensor var_460_axes_0 = const()[name = tensor("op_460_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132721408)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132724032)))]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_460_cast_fp16 = layer_norm(axes = var_460_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_449_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132726656)))]; + tensor var_472_to_fp16 = const()[name = tensor("op_472_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136003520)))]; + tensor linear_18_cast_fp16 = linear(bias = var_472_to_fp16, weight = var_471_to_fp16, x = var_460_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_475_to_fp16 = const()[name = tensor("op_475_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136006144)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_475_to_fp16, x = var_460_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139283008)))]; + tensor var_480_to_fp16 = const()[name = tensor("op_480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142559872)))]; + tensor linear_20_cast_fp16 = linear(bias = var_480_to_fp16, weight = var_479_to_fp16, x = var_460_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_488 = const()[name = tensor("op_488"), val = tensor([1, 1500, 20, -1])]; + tensor var_489_cast_fp16 = reshape(shape = var_488, x = linear_18_cast_fp16)[name = tensor("op_489_cast_fp16")]; + tensor const_230_to_fp16 = const()[name = tensor("const_230_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast_fp16 = mul(x = var_489_cast_fp16, y = const_230_to_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_495 = const()[name = tensor("op_495"), val = tensor([1, 1500, 20, -1])]; + tensor var_496_cast_fp16 = reshape(shape = var_495, x = linear_19_cast_fp16)[name = tensor("op_496_cast_fp16")]; + tensor const_231_to_fp16 = const()[name = tensor("const_231_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast_fp16 = mul(x = var_496_cast_fp16, y = const_231_to_fp16)[name = tensor("k_15_cast_fp16")]; + tensor var_502 = const()[name = tensor("op_502"), val = tensor([1, 1500, 20, -1])]; + tensor var_503_cast_fp16 = reshape(shape = var_502, x = linear_20_cast_fp16)[name = tensor("op_503_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_134_perm_0 = const()[name = tensor("transpose_134_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_135_perm_0 = const()[name = tensor("transpose_135_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_305 = transpose(perm = transpose_135_perm_0, x = k_15_cast_fp16)[name = tensor("transpose_305")]; + tensor transpose_306 = transpose(perm = transpose_134_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_306")]; + tensor qk_7_cast_fp16 = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_306, y = transpose_305)[name = tensor("qk_7_cast_fp16")]; + tensor var_508_cast_fp16 = softmax(axis = var_443, x = qk_7_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_510_transpose_x_0 = const()[name = tensor("op_510_transpose_x_0"), val = tensor(false)]; + tensor var_510_transpose_y_0 = const()[name = tensor("op_510_transpose_y_0"), val = tensor(false)]; + tensor transpose_307 = transpose(perm = var_504, x = var_503_cast_fp16)[name = tensor("transpose_307")]; + tensor var_510_cast_fp16 = matmul(transpose_x = var_510_transpose_x_0, transpose_y = var_510_transpose_y_0, x = var_508_cast_fp16, y = transpose_307)[name = tensor("op_510_cast_fp16")]; + tensor var_511 = const()[name = tensor("op_511"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 1280])]; + tensor transpose_304 = transpose(perm = var_511, x = var_510_cast_fp16)[name = tensor("transpose_304")]; + tensor x_47_cast_fp16 = reshape(shape = concat_3, x = transpose_304)[name = tensor("x_47_cast_fp16")]; + tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142562496)))]; + tensor var_517_to_fp16 = const()[name = tensor("op_517_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145839360)))]; + tensor linear_21_cast_fp16 = linear(bias = var_517_to_fp16, weight = var_516_to_fp16, x = x_47_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor x_49_cast_fp16 = add(x = x_43_cast_fp16, y = linear_21_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor var_524_axes_0 = const()[name = tensor("op_524_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145841984)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145844608)))]; + tensor var_524_cast_fp16 = layer_norm(axes = var_524_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_449_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast_fp16)[name = tensor("op_524_cast_fp16")]; + tensor var_533_to_fp16 = const()[name = tensor("op_533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145847232)))]; + tensor var_534_to_fp16 = const()[name = tensor("op_534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158954496)))]; + tensor linear_22_cast_fp16 = linear(bias = var_534_to_fp16, weight = var_533_to_fp16, x = var_524_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast_fp16 = gelu(mode = x_53_mode_0, x = linear_22_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_539_to_fp16 = const()[name = tensor("op_539_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158964800)))]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172072064)))]; + tensor linear_23_cast_fp16 = linear(bias = var_540_to_fp16, weight = var_539_to_fp16, x = x_53_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = x_49_cast_fp16, y = linear_23_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor var_550 = const()[name = tensor("op_550"), val = tensor(-1)]; + tensor var_567_axes_0 = const()[name = tensor("op_567_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172074688)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172077312)))]; + tensor var_556_to_fp16 = const()[name = tensor("op_556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_567_cast_fp16 = layer_norm(axes = var_567_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_556_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor var_578_to_fp16 = const()[name = tensor("op_578_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172079936)))]; + tensor var_579_to_fp16 = const()[name = tensor("op_579_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175356800)))]; + tensor linear_24_cast_fp16 = linear(bias = var_579_to_fp16, weight = var_578_to_fp16, x = var_567_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175359424)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_582_to_fp16, x = var_567_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178636288)))]; + tensor var_587_to_fp16 = const()[name = tensor("op_587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181913152)))]; + tensor linear_26_cast_fp16 = linear(bias = var_587_to_fp16, weight = var_586_to_fp16, x = var_567_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1500, 20, -1])]; + tensor var_596_cast_fp16 = reshape(shape = var_595, x = linear_24_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor const_232_to_fp16 = const()[name = tensor("const_232_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast_fp16 = mul(x = var_596_cast_fp16, y = const_232_to_fp16)[name = tensor("q_19_cast_fp16")]; + tensor var_602 = const()[name = tensor("op_602"), val = tensor([1, 1500, 20, -1])]; + tensor var_603_cast_fp16 = reshape(shape = var_602, x = linear_25_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor const_233_to_fp16 = const()[name = tensor("const_233_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast_fp16 = mul(x = var_603_cast_fp16, y = const_233_to_fp16)[name = tensor("k_19_cast_fp16")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1500, 20, -1])]; + tensor var_610_cast_fp16 = reshape(shape = var_609, x = linear_26_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611 = const()[name = tensor("op_611"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_136_perm_0 = const()[name = tensor("transpose_136_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_137_perm_0 = const()[name = tensor("transpose_137_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_301 = transpose(perm = transpose_137_perm_0, x = k_19_cast_fp16)[name = tensor("transpose_301")]; + tensor transpose_302 = transpose(perm = transpose_136_perm_0, x = q_19_cast_fp16)[name = tensor("transpose_302")]; + tensor qk_9_cast_fp16 = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_302, y = transpose_301)[name = tensor("qk_9_cast_fp16")]; + tensor var_615_cast_fp16 = softmax(axis = var_550, x = qk_9_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor var_617_transpose_x_0 = const()[name = tensor("op_617_transpose_x_0"), val = tensor(false)]; + tensor var_617_transpose_y_0 = const()[name = tensor("op_617_transpose_y_0"), val = tensor(false)]; + tensor transpose_303 = transpose(perm = var_611, x = var_610_cast_fp16)[name = tensor("transpose_303")]; + tensor var_617_cast_fp16 = matmul(transpose_x = var_617_transpose_x_0, transpose_y = var_617_transpose_y_0, x = var_615_cast_fp16, y = transpose_303)[name = tensor("op_617_cast_fp16")]; + tensor var_618 = const()[name = tensor("op_618"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 1280])]; + tensor transpose_300 = transpose(perm = var_618, x = var_617_cast_fp16)[name = tensor("transpose_300")]; + tensor x_59_cast_fp16 = reshape(shape = concat_4, x = transpose_300)[name = tensor("x_59_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181915776)))]; + tensor var_624_to_fp16 = const()[name = tensor("op_624_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185192640)))]; + tensor linear_27_cast_fp16 = linear(bias = var_624_to_fp16, weight = var_623_to_fp16, x = x_59_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = linear_27_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor var_631_axes_0 = const()[name = tensor("op_631_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185195264)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185197888)))]; + tensor var_631_cast_fp16 = layer_norm(axes = var_631_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_556_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor var_640_to_fp16 = const()[name = tensor("op_640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185200512)))]; + tensor var_641_to_fp16 = const()[name = tensor("op_641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198307776)))]; + tensor linear_28_cast_fp16 = linear(bias = var_641_to_fp16, weight = var_640_to_fp16, x = var_631_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast_fp16 = gelu(mode = x_65_mode_0, x = linear_28_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_646_to_fp16 = const()[name = tensor("op_646_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198318080)))]; + tensor var_647_to_fp16 = const()[name = tensor("op_647_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211425344)))]; + tensor linear_29_cast_fp16 = linear(bias = var_647_to_fp16, weight = var_646_to_fp16, x = x_65_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = linear_29_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor var_657 = const()[name = tensor("op_657"), val = tensor(-1)]; + tensor var_674_axes_0 = const()[name = tensor("op_674_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211427968)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211430592)))]; + tensor var_663_to_fp16 = const()[name = tensor("op_663_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_674_cast_fp16 = layer_norm(axes = var_674_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_663_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor var_685_to_fp16 = const()[name = tensor("op_685_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211433216)))]; + tensor var_686_to_fp16 = const()[name = tensor("op_686_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214710080)))]; + tensor linear_30_cast_fp16 = linear(bias = var_686_to_fp16, weight = var_685_to_fp16, x = var_674_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_689_to_fp16 = const()[name = tensor("op_689_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214712704)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_689_to_fp16, x = var_674_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217989568)))]; + tensor var_694_to_fp16 = const()[name = tensor("op_694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221266432)))]; + tensor linear_32_cast_fp16 = linear(bias = var_694_to_fp16, weight = var_693_to_fp16, x = var_674_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_702 = const()[name = tensor("op_702"), val = tensor([1, 1500, 20, -1])]; + tensor var_703_cast_fp16 = reshape(shape = var_702, x = linear_30_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor const_234_to_fp16 = const()[name = tensor("const_234_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast_fp16 = mul(x = var_703_cast_fp16, y = const_234_to_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1500, 20, -1])]; + tensor var_710_cast_fp16 = reshape(shape = var_709, x = linear_31_cast_fp16)[name = tensor("op_710_cast_fp16")]; + tensor const_235_to_fp16 = const()[name = tensor("const_235_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast_fp16 = mul(x = var_710_cast_fp16, y = const_235_to_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_716 = const()[name = tensor("op_716"), val = tensor([1, 1500, 20, -1])]; + tensor var_717_cast_fp16 = reshape(shape = var_716, x = linear_32_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_718 = const()[name = tensor("op_718"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_138_perm_0 = const()[name = tensor("transpose_138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_139_perm_0 = const()[name = tensor("transpose_139_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_297 = transpose(perm = transpose_139_perm_0, x = k_23_cast_fp16)[name = tensor("transpose_297")]; + tensor transpose_298 = transpose(perm = transpose_138_perm_0, x = q_23_cast_fp16)[name = tensor("transpose_298")]; + tensor qk_11_cast_fp16 = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_298, y = transpose_297)[name = tensor("qk_11_cast_fp16")]; + tensor var_722_cast_fp16 = softmax(axis = var_657, x = qk_11_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_299 = transpose(perm = var_718, x = var_717_cast_fp16)[name = tensor("transpose_299")]; + tensor var_724_cast_fp16 = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = var_722_cast_fp16, y = transpose_299)[name = tensor("op_724_cast_fp16")]; + tensor var_725 = const()[name = tensor("op_725"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 1280])]; + tensor transpose_296 = transpose(perm = var_725, x = var_724_cast_fp16)[name = tensor("transpose_296")]; + tensor x_71_cast_fp16 = reshape(shape = concat_5, x = transpose_296)[name = tensor("x_71_cast_fp16")]; + tensor var_730_to_fp16 = const()[name = tensor("op_730_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221269056)))]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224545920)))]; + tensor linear_33_cast_fp16 = linear(bias = var_731_to_fp16, weight = var_730_to_fp16, x = x_71_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = linear_33_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor var_738_axes_0 = const()[name = tensor("op_738_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224548544)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224551168)))]; + tensor var_738_cast_fp16 = layer_norm(axes = var_738_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_663_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast_fp16)[name = tensor("op_738_cast_fp16")]; + tensor var_747_to_fp16 = const()[name = tensor("op_747_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224553792)))]; + tensor var_748_to_fp16 = const()[name = tensor("op_748_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237661056)))]; + tensor linear_34_cast_fp16 = linear(bias = var_748_to_fp16, weight = var_747_to_fp16, x = var_738_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast_fp16 = gelu(mode = x_77_mode_0, x = linear_34_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_753_to_fp16 = const()[name = tensor("op_753_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237671360)))]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250778624)))]; + tensor linear_35_cast_fp16 = linear(bias = var_754_to_fp16, weight = var_753_to_fp16, x = x_77_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = linear_35_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor(-1)]; + tensor var_781_axes_0 = const()[name = tensor("op_781_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250781248)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250783872)))]; + tensor var_770_to_fp16 = const()[name = tensor("op_770_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_781_cast_fp16 = layer_norm(axes = var_781_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_770_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250786496)))]; + tensor var_793_to_fp16 = const()[name = tensor("op_793_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254063360)))]; + tensor linear_36_cast_fp16 = linear(bias = var_793_to_fp16, weight = var_792_to_fp16, x = var_781_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_796_to_fp16 = const()[name = tensor("op_796_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254065984)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_796_to_fp16, x = var_781_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_800_to_fp16 = const()[name = tensor("op_800_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257342848)))]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260619712)))]; + tensor linear_38_cast_fp16 = linear(bias = var_801_to_fp16, weight = var_800_to_fp16, x = var_781_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_809 = const()[name = tensor("op_809"), val = tensor([1, 1500, 20, -1])]; + tensor var_810_cast_fp16 = reshape(shape = var_809, x = linear_36_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor const_236_to_fp16 = const()[name = tensor("const_236_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast_fp16 = mul(x = var_810_cast_fp16, y = const_236_to_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_816 = const()[name = tensor("op_816"), val = tensor([1, 1500, 20, -1])]; + tensor var_817_cast_fp16 = reshape(shape = var_816, x = linear_37_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor const_237_to_fp16 = const()[name = tensor("const_237_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast_fp16 = mul(x = var_817_cast_fp16, y = const_237_to_fp16)[name = tensor("k_27_cast_fp16")]; + tensor var_823 = const()[name = tensor("op_823"), val = tensor([1, 1500, 20, -1])]; + tensor var_824_cast_fp16 = reshape(shape = var_823, x = linear_38_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_140_perm_0 = const()[name = tensor("transpose_140_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_141_perm_0 = const()[name = tensor("transpose_141_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_293 = transpose(perm = transpose_141_perm_0, x = k_27_cast_fp16)[name = tensor("transpose_293")]; + tensor transpose_294 = transpose(perm = transpose_140_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_294")]; + tensor qk_13_cast_fp16 = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_294, y = transpose_293)[name = tensor("qk_13_cast_fp16")]; + tensor var_829_cast_fp16 = softmax(axis = var_764, x = qk_13_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_831_transpose_x_0 = const()[name = tensor("op_831_transpose_x_0"), val = tensor(false)]; + tensor var_831_transpose_y_0 = const()[name = tensor("op_831_transpose_y_0"), val = tensor(false)]; + tensor transpose_295 = transpose(perm = var_825, x = var_824_cast_fp16)[name = tensor("transpose_295")]; + tensor var_831_cast_fp16 = matmul(transpose_x = var_831_transpose_x_0, transpose_y = var_831_transpose_y_0, x = var_829_cast_fp16, y = transpose_295)[name = tensor("op_831_cast_fp16")]; + tensor var_832 = const()[name = tensor("op_832"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 1280])]; + tensor transpose_292 = transpose(perm = var_832, x = var_831_cast_fp16)[name = tensor("transpose_292")]; + tensor x_83_cast_fp16 = reshape(shape = concat_6, x = transpose_292)[name = tensor("x_83_cast_fp16")]; + tensor var_837_to_fp16 = const()[name = tensor("op_837_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260622336)))]; + tensor var_838_to_fp16 = const()[name = tensor("op_838_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263899200)))]; + tensor linear_39_cast_fp16 = linear(bias = var_838_to_fp16, weight = var_837_to_fp16, x = x_83_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = linear_39_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor var_845_axes_0 = const()[name = tensor("op_845_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263901824)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263904448)))]; + tensor var_845_cast_fp16 = layer_norm(axes = var_845_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_770_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_854_to_fp16 = const()[name = tensor("op_854_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263907072)))]; + tensor var_855_to_fp16 = const()[name = tensor("op_855_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277014336)))]; + tensor linear_40_cast_fp16 = linear(bias = var_855_to_fp16, weight = var_854_to_fp16, x = var_845_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast_fp16 = gelu(mode = x_89_mode_0, x = linear_40_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor var_860_to_fp16 = const()[name = tensor("op_860_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277024640)))]; + tensor var_861_to_fp16 = const()[name = tensor("op_861_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290131904)))]; + tensor linear_41_cast_fp16 = linear(bias = var_861_to_fp16, weight = var_860_to_fp16, x = x_89_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor x_91_cast_fp16 = add(x = x_85_cast_fp16, y = linear_41_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor var_871 = const()[name = tensor("op_871"), val = tensor(-1)]; + tensor var_888_axes_0 = const()[name = tensor("op_888_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290134528)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290137152)))]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_888_cast_fp16 = layer_norm(axes = var_888_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_877_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast_fp16)[name = tensor("op_888_cast_fp16")]; + tensor var_899_to_fp16 = const()[name = tensor("op_899_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290139776)))]; + tensor var_900_to_fp16 = const()[name = tensor("op_900_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293416640)))]; + tensor linear_42_cast_fp16 = linear(bias = var_900_to_fp16, weight = var_899_to_fp16, x = var_888_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_903_to_fp16 = const()[name = tensor("op_903_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293419264)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_903_to_fp16, x = var_888_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_907_to_fp16 = const()[name = tensor("op_907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296696128)))]; + tensor var_908_to_fp16 = const()[name = tensor("op_908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299972992)))]; + tensor linear_44_cast_fp16 = linear(bias = var_908_to_fp16, weight = var_907_to_fp16, x = var_888_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1500, 20, -1])]; + tensor var_917_cast_fp16 = reshape(shape = var_916, x = linear_42_cast_fp16)[name = tensor("op_917_cast_fp16")]; + tensor const_238_to_fp16 = const()[name = tensor("const_238_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast_fp16 = mul(x = var_917_cast_fp16, y = const_238_to_fp16)[name = tensor("q_31_cast_fp16")]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1500, 20, -1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = linear_43_cast_fp16)[name = tensor("op_924_cast_fp16")]; + tensor const_239_to_fp16 = const()[name = tensor("const_239_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast_fp16 = mul(x = var_924_cast_fp16, y = const_239_to_fp16)[name = tensor("k_31_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 1500, 20, -1])]; + tensor var_931_cast_fp16 = reshape(shape = var_930, x = linear_44_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor var_932 = const()[name = tensor("op_932"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_142_perm_0 = const()[name = tensor("transpose_142_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_143_perm_0 = const()[name = tensor("transpose_143_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_289 = transpose(perm = transpose_143_perm_0, x = k_31_cast_fp16)[name = tensor("transpose_289")]; + tensor transpose_290 = transpose(perm = transpose_142_perm_0, x = q_31_cast_fp16)[name = tensor("transpose_290")]; + tensor qk_15_cast_fp16 = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_290, y = transpose_289)[name = tensor("qk_15_cast_fp16")]; + tensor var_936_cast_fp16 = softmax(axis = var_871, x = qk_15_cast_fp16)[name = tensor("op_936_cast_fp16")]; + tensor var_938_transpose_x_0 = const()[name = tensor("op_938_transpose_x_0"), val = tensor(false)]; + tensor var_938_transpose_y_0 = const()[name = tensor("op_938_transpose_y_0"), val = tensor(false)]; + tensor transpose_291 = transpose(perm = var_932, x = var_931_cast_fp16)[name = tensor("transpose_291")]; + tensor var_938_cast_fp16 = matmul(transpose_x = var_938_transpose_x_0, transpose_y = var_938_transpose_y_0, x = var_936_cast_fp16, y = transpose_291)[name = tensor("op_938_cast_fp16")]; + tensor var_939 = const()[name = tensor("op_939"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 1280])]; + tensor transpose_288 = transpose(perm = var_939, x = var_938_cast_fp16)[name = tensor("transpose_288")]; + tensor x_95_cast_fp16 = reshape(shape = concat_7, x = transpose_288)[name = tensor("x_95_cast_fp16")]; + tensor var_944_to_fp16 = const()[name = tensor("op_944_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299975616)))]; + tensor var_945_to_fp16 = const()[name = tensor("op_945_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303252480)))]; + tensor linear_45_cast_fp16 = linear(bias = var_945_to_fp16, weight = var_944_to_fp16, x = x_95_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = x_91_cast_fp16, y = linear_45_cast_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_952_axes_0 = const()[name = tensor("op_952_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303255104)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303257728)))]; + tensor var_952_cast_fp16 = layer_norm(axes = var_952_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_877_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast_fp16)[name = tensor("op_952_cast_fp16")]; + tensor var_961_to_fp16 = const()[name = tensor("op_961_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303260352)))]; + tensor var_962_to_fp16 = const()[name = tensor("op_962_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316367616)))]; + tensor linear_46_cast_fp16 = linear(bias = var_962_to_fp16, weight = var_961_to_fp16, x = var_952_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast_fp16 = gelu(mode = x_101_mode_0, x = linear_46_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_967_to_fp16 = const()[name = tensor("op_967_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316377920)))]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329485184)))]; + tensor linear_47_cast_fp16 = linear(bias = var_968_to_fp16, weight = var_967_to_fp16, x = x_101_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = x_97_cast_fp16, y = linear_47_cast_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_978 = const()[name = tensor("op_978"), val = tensor(-1)]; + tensor var_995_axes_0 = const()[name = tensor("op_995_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329487808)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329490432)))]; + tensor var_984_to_fp16 = const()[name = tensor("op_984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_995_cast_fp16 = layer_norm(axes = var_995_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_984_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast_fp16)[name = tensor("op_995_cast_fp16")]; + tensor var_1006_to_fp16 = const()[name = tensor("op_1006_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329493056)))]; + tensor var_1007_to_fp16 = const()[name = tensor("op_1007_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332769920)))]; + tensor linear_48_cast_fp16 = linear(bias = var_1007_to_fp16, weight = var_1006_to_fp16, x = var_995_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_1010_to_fp16 = const()[name = tensor("op_1010_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332772544)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1010_to_fp16, x = var_995_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336049408)))]; + tensor var_1015_to_fp16 = const()[name = tensor("op_1015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339326272)))]; + tensor linear_50_cast_fp16 = linear(bias = var_1015_to_fp16, weight = var_1014_to_fp16, x = var_995_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 1500, 20, -1])]; + tensor var_1024_cast_fp16 = reshape(shape = var_1023, x = linear_48_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor const_240_to_fp16 = const()[name = tensor("const_240_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast_fp16 = mul(x = var_1024_cast_fp16, y = const_240_to_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([1, 1500, 20, -1])]; + tensor var_1031_cast_fp16 = reshape(shape = var_1030, x = linear_49_cast_fp16)[name = tensor("op_1031_cast_fp16")]; + tensor const_241_to_fp16 = const()[name = tensor("const_241_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast_fp16 = mul(x = var_1031_cast_fp16, y = const_241_to_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_1037 = const()[name = tensor("op_1037"), val = tensor([1, 1500, 20, -1])]; + tensor var_1038_cast_fp16 = reshape(shape = var_1037, x = linear_50_cast_fp16)[name = tensor("op_1038_cast_fp16")]; + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_144_perm_0 = const()[name = tensor("transpose_144_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_145_perm_0 = const()[name = tensor("transpose_145_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_285 = transpose(perm = transpose_145_perm_0, x = k_35_cast_fp16)[name = tensor("transpose_285")]; + tensor transpose_286 = transpose(perm = transpose_144_perm_0, x = q_35_cast_fp16)[name = tensor("transpose_286")]; + tensor qk_17_cast_fp16 = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_286, y = transpose_285)[name = tensor("qk_17_cast_fp16")]; + tensor var_1043_cast_fp16 = softmax(axis = var_978, x = qk_17_cast_fp16)[name = tensor("op_1043_cast_fp16")]; + tensor var_1045_transpose_x_0 = const()[name = tensor("op_1045_transpose_x_0"), val = tensor(false)]; + tensor var_1045_transpose_y_0 = const()[name = tensor("op_1045_transpose_y_0"), val = tensor(false)]; + tensor transpose_287 = transpose(perm = var_1039, x = var_1038_cast_fp16)[name = tensor("transpose_287")]; + tensor var_1045_cast_fp16 = matmul(transpose_x = var_1045_transpose_x_0, transpose_y = var_1045_transpose_y_0, x = var_1043_cast_fp16, y = transpose_287)[name = tensor("op_1045_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 1280])]; + tensor transpose_284 = transpose(perm = var_1046, x = var_1045_cast_fp16)[name = tensor("transpose_284")]; + tensor x_107_cast_fp16 = reshape(shape = concat_8, x = transpose_284)[name = tensor("x_107_cast_fp16")]; + tensor var_1051_to_fp16 = const()[name = tensor("op_1051_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339328896)))]; + tensor var_1052_to_fp16 = const()[name = tensor("op_1052_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342605760)))]; + tensor linear_51_cast_fp16 = linear(bias = var_1052_to_fp16, weight = var_1051_to_fp16, x = x_107_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = x_103_cast_fp16, y = linear_51_cast_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1059_axes_0 = const()[name = tensor("op_1059_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342608384)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342611008)))]; + tensor var_1059_cast_fp16 = layer_norm(axes = var_1059_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_984_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor var_1068_to_fp16 = const()[name = tensor("op_1068_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342613632)))]; + tensor var_1069_to_fp16 = const()[name = tensor("op_1069_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355720896)))]; + tensor linear_52_cast_fp16 = linear(bias = var_1069_to_fp16, weight = var_1068_to_fp16, x = var_1059_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast_fp16 = gelu(mode = x_113_mode_0, x = linear_52_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355731200)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368838464)))]; + tensor linear_53_cast_fp16 = linear(bias = var_1075_to_fp16, weight = var_1074_to_fp16, x = x_113_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = x_109_cast_fp16, y = linear_53_cast_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1085 = const()[name = tensor("op_1085"), val = tensor(-1)]; + tensor var_1102_axes_0 = const()[name = tensor("op_1102_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368841088)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368843712)))]; + tensor var_1091_to_fp16 = const()[name = tensor("op_1091_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1102_cast_fp16 = layer_norm(axes = var_1102_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1091_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor var_1113_to_fp16 = const()[name = tensor("op_1113_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368846336)))]; + tensor var_1114_to_fp16 = const()[name = tensor("op_1114_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372123200)))]; + tensor linear_54_cast_fp16 = linear(bias = var_1114_to_fp16, weight = var_1113_to_fp16, x = var_1102_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_1117_to_fp16 = const()[name = tensor("op_1117_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372125824)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1117_to_fp16, x = var_1102_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor var_1121_to_fp16 = const()[name = tensor("op_1121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375402688)))]; + tensor var_1122_to_fp16 = const()[name = tensor("op_1122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378679552)))]; + tensor linear_56_cast_fp16 = linear(bias = var_1122_to_fp16, weight = var_1121_to_fp16, x = var_1102_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1500, 20, -1])]; + tensor var_1131_cast_fp16 = reshape(shape = var_1130, x = linear_54_cast_fp16)[name = tensor("op_1131_cast_fp16")]; + tensor const_242_to_fp16 = const()[name = tensor("const_242_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast_fp16 = mul(x = var_1131_cast_fp16, y = const_242_to_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([1, 1500, 20, -1])]; + tensor var_1138_cast_fp16 = reshape(shape = var_1137, x = linear_55_cast_fp16)[name = tensor("op_1138_cast_fp16")]; + tensor const_243_to_fp16 = const()[name = tensor("const_243_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast_fp16 = mul(x = var_1138_cast_fp16, y = const_243_to_fp16)[name = tensor("k_39_cast_fp16")]; + tensor var_1144 = const()[name = tensor("op_1144"), val = tensor([1, 1500, 20, -1])]; + tensor var_1145_cast_fp16 = reshape(shape = var_1144, x = linear_56_cast_fp16)[name = tensor("op_1145_cast_fp16")]; + tensor var_1146 = const()[name = tensor("op_1146"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_146_perm_0 = const()[name = tensor("transpose_146_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_147_perm_0 = const()[name = tensor("transpose_147_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_281 = transpose(perm = transpose_147_perm_0, x = k_39_cast_fp16)[name = tensor("transpose_281")]; + tensor transpose_282 = transpose(perm = transpose_146_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_282")]; + tensor qk_19_cast_fp16 = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_282, y = transpose_281)[name = tensor("qk_19_cast_fp16")]; + tensor var_1150_cast_fp16 = softmax(axis = var_1085, x = qk_19_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1152_transpose_x_0 = const()[name = tensor("op_1152_transpose_x_0"), val = tensor(false)]; + tensor var_1152_transpose_y_0 = const()[name = tensor("op_1152_transpose_y_0"), val = tensor(false)]; + tensor transpose_283 = transpose(perm = var_1146, x = var_1145_cast_fp16)[name = tensor("transpose_283")]; + tensor var_1152_cast_fp16 = matmul(transpose_x = var_1152_transpose_x_0, transpose_y = var_1152_transpose_y_0, x = var_1150_cast_fp16, y = transpose_283)[name = tensor("op_1152_cast_fp16")]; + tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 1280])]; + tensor transpose_280 = transpose(perm = var_1153, x = var_1152_cast_fp16)[name = tensor("transpose_280")]; + tensor x_119_cast_fp16 = reshape(shape = concat_9, x = transpose_280)[name = tensor("x_119_cast_fp16")]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378682176)))]; + tensor var_1159_to_fp16 = const()[name = tensor("op_1159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381959040)))]; + tensor linear_57_cast_fp16 = linear(bias = var_1159_to_fp16, weight = var_1158_to_fp16, x = x_119_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = x_115_cast_fp16, y = linear_57_cast_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381961664)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381964288)))]; + tensor var_1166_cast_fp16 = layer_norm(axes = var_1166_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1091_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast_fp16)[name = tensor("op_1166_cast_fp16")]; + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381966912)))]; + tensor var_1176_to_fp16 = const()[name = tensor("op_1176_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395074176)))]; + tensor linear_58_cast_fp16 = linear(bias = var_1176_to_fp16, weight = var_1175_to_fp16, x = var_1166_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast_fp16 = gelu(mode = x_125_mode_0, x = linear_58_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395084480)))]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408191744)))]; + tensor linear_59_cast_fp16 = linear(bias = var_1182_to_fp16, weight = var_1181_to_fp16, x = x_125_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = x_121_cast_fp16, y = linear_59_cast_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor(-1)]; + tensor var_1209_axes_0 = const()[name = tensor("op_1209_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408194368)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408196992)))]; + tensor var_1198_to_fp16 = const()[name = tensor("op_1198_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1209_cast_fp16 = layer_norm(axes = var_1209_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast_fp16)[name = tensor("op_1209_cast_fp16")]; + tensor var_1220_to_fp16 = const()[name = tensor("op_1220_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408199616)))]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411476480)))]; + tensor linear_60_cast_fp16 = linear(bias = var_1221_to_fp16, weight = var_1220_to_fp16, x = var_1209_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_1224_to_fp16 = const()[name = tensor("op_1224_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411479104)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1224_to_fp16, x = var_1209_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_1228_to_fp16 = const()[name = tensor("op_1228_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(414755968)))]; + tensor var_1229_to_fp16 = const()[name = tensor("op_1229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418032832)))]; + tensor linear_62_cast_fp16 = linear(bias = var_1229_to_fp16, weight = var_1228_to_fp16, x = var_1209_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1500, 20, -1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = linear_60_cast_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor const_244_to_fp16 = const()[name = tensor("const_244_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast_fp16 = mul(x = var_1238_cast_fp16, y = const_244_to_fp16)[name = tensor("q_43_cast_fp16")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([1, 1500, 20, -1])]; + tensor var_1245_cast_fp16 = reshape(shape = var_1244, x = linear_61_cast_fp16)[name = tensor("op_1245_cast_fp16")]; + tensor const_245_to_fp16 = const()[name = tensor("const_245_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast_fp16 = mul(x = var_1245_cast_fp16, y = const_245_to_fp16)[name = tensor("k_43_cast_fp16")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 1500, 20, -1])]; + tensor var_1252_cast_fp16 = reshape(shape = var_1251, x = linear_62_cast_fp16)[name = tensor("op_1252_cast_fp16")]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_148_perm_0 = const()[name = tensor("transpose_148_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_149_perm_0 = const()[name = tensor("transpose_149_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_277 = transpose(perm = transpose_149_perm_0, x = k_43_cast_fp16)[name = tensor("transpose_277")]; + tensor transpose_278 = transpose(perm = transpose_148_perm_0, x = q_43_cast_fp16)[name = tensor("transpose_278")]; + tensor qk_21_cast_fp16 = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_278, y = transpose_277)[name = tensor("qk_21_cast_fp16")]; + tensor var_1257_cast_fp16 = softmax(axis = var_1192, x = qk_21_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor var_1259_transpose_x_0 = const()[name = tensor("op_1259_transpose_x_0"), val = tensor(false)]; + tensor var_1259_transpose_y_0 = const()[name = tensor("op_1259_transpose_y_0"), val = tensor(false)]; + tensor transpose_279 = transpose(perm = var_1253, x = var_1252_cast_fp16)[name = tensor("transpose_279")]; + tensor var_1259_cast_fp16 = matmul(transpose_x = var_1259_transpose_x_0, transpose_y = var_1259_transpose_y_0, x = var_1257_cast_fp16, y = transpose_279)[name = tensor("op_1259_cast_fp16")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 1280])]; + tensor transpose_276 = transpose(perm = var_1260, x = var_1259_cast_fp16)[name = tensor("transpose_276")]; + tensor x_131_cast_fp16 = reshape(shape = concat_10, x = transpose_276)[name = tensor("x_131_cast_fp16")]; + tensor var_1265_to_fp16 = const()[name = tensor("op_1265_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(418035456)))]; + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421312320)))]; + tensor linear_63_cast_fp16 = linear(bias = var_1266_to_fp16, weight = var_1265_to_fp16, x = x_131_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor x_133_cast_fp16 = add(x = x_127_cast_fp16, y = linear_63_cast_fp16)[name = tensor("x_133_cast_fp16")]; + tensor var_1273_axes_0 = const()[name = tensor("op_1273_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421314944)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421317568)))]; + tensor var_1273_cast_fp16 = layer_norm(axes = var_1273_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1198_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast_fp16)[name = tensor("op_1273_cast_fp16")]; + tensor var_1282_to_fp16 = const()[name = tensor("op_1282_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421320192)))]; + tensor var_1283_to_fp16 = const()[name = tensor("op_1283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434427456)))]; + tensor linear_64_cast_fp16 = linear(bias = var_1283_to_fp16, weight = var_1282_to_fp16, x = var_1273_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast_fp16 = gelu(mode = x_137_mode_0, x = linear_64_cast_fp16)[name = tensor("x_137_cast_fp16")]; + tensor var_1288_to_fp16 = const()[name = tensor("op_1288_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434437760)))]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447545024)))]; + tensor linear_65_cast_fp16 = linear(bias = var_1289_to_fp16, weight = var_1288_to_fp16, x = x_137_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor x_139_cast_fp16 = add(x = x_133_cast_fp16, y = linear_65_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor(-1)]; + tensor var_1316_axes_0 = const()[name = tensor("op_1316_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447547648)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447550272)))]; + tensor var_1305_to_fp16 = const()[name = tensor("op_1305_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1316_cast_fp16 = layer_norm(axes = var_1316_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1305_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1327_to_fp16 = const()[name = tensor("op_1327_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447552896)))]; + tensor var_1328_to_fp16 = const()[name = tensor("op_1328_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450829760)))]; + tensor linear_66_cast_fp16 = linear(bias = var_1328_to_fp16, weight = var_1327_to_fp16, x = var_1316_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450832384)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1331_to_fp16, x = var_1316_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1335_to_fp16 = const()[name = tensor("op_1335_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454109248)))]; + tensor var_1336_to_fp16 = const()[name = tensor("op_1336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457386112)))]; + tensor linear_68_cast_fp16 = linear(bias = var_1336_to_fp16, weight = var_1335_to_fp16, x = var_1316_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([1, 1500, 20, -1])]; + tensor var_1345_cast_fp16 = reshape(shape = var_1344, x = linear_66_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor const_246_to_fp16 = const()[name = tensor("const_246_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_47_cast_fp16 = mul(x = var_1345_cast_fp16, y = const_246_to_fp16)[name = tensor("q_47_cast_fp16")]; + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([1, 1500, 20, -1])]; + tensor var_1352_cast_fp16 = reshape(shape = var_1351, x = linear_67_cast_fp16)[name = tensor("op_1352_cast_fp16")]; + tensor const_247_to_fp16 = const()[name = tensor("const_247_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_47_cast_fp16 = mul(x = var_1352_cast_fp16, y = const_247_to_fp16)[name = tensor("k_47_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 1500, 20, -1])]; + tensor var_1359_cast_fp16 = reshape(shape = var_1358, x = linear_68_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor var_1360 = const()[name = tensor("op_1360"), val = tensor([0, 2, 1, 3])]; + tensor qk_23_transpose_x_0 = const()[name = tensor("qk_23_transpose_x_0"), val = tensor(false)]; + tensor qk_23_transpose_y_0 = const()[name = tensor("qk_23_transpose_y_0"), val = tensor(false)]; + tensor transpose_150_perm_0 = const()[name = tensor("transpose_150_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_151_perm_0 = const()[name = tensor("transpose_151_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_273 = transpose(perm = transpose_151_perm_0, x = k_47_cast_fp16)[name = tensor("transpose_273")]; + tensor transpose_274 = transpose(perm = transpose_150_perm_0, x = q_47_cast_fp16)[name = tensor("transpose_274")]; + tensor qk_23_cast_fp16 = matmul(transpose_x = qk_23_transpose_x_0, transpose_y = qk_23_transpose_y_0, x = transpose_274, y = transpose_273)[name = tensor("qk_23_cast_fp16")]; + tensor var_1364_cast_fp16 = softmax(axis = var_1299, x = qk_23_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor var_1366_transpose_x_0 = const()[name = tensor("op_1366_transpose_x_0"), val = tensor(false)]; + tensor var_1366_transpose_y_0 = const()[name = tensor("op_1366_transpose_y_0"), val = tensor(false)]; + tensor transpose_275 = transpose(perm = var_1360, x = var_1359_cast_fp16)[name = tensor("transpose_275")]; + tensor var_1366_cast_fp16 = matmul(transpose_x = var_1366_transpose_x_0, transpose_y = var_1366_transpose_y_0, x = var_1364_cast_fp16, y = transpose_275)[name = tensor("op_1366_cast_fp16")]; + tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 1280])]; + tensor transpose_272 = transpose(perm = var_1367, x = var_1366_cast_fp16)[name = tensor("transpose_272")]; + tensor x_143_cast_fp16 = reshape(shape = concat_11, x = transpose_272)[name = tensor("x_143_cast_fp16")]; + tensor var_1372_to_fp16 = const()[name = tensor("op_1372_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457388736)))]; + tensor var_1373_to_fp16 = const()[name = tensor("op_1373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460665600)))]; + tensor linear_69_cast_fp16 = linear(bias = var_1373_to_fp16, weight = var_1372_to_fp16, x = x_143_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor x_145_cast_fp16 = add(x = x_139_cast_fp16, y = linear_69_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1380_axes_0 = const()[name = tensor("op_1380_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460668224)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460670848)))]; + tensor var_1380_cast_fp16 = layer_norm(axes = var_1380_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1305_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1389_to_fp16 = const()[name = tensor("op_1389_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460673472)))]; + tensor var_1390_to_fp16 = const()[name = tensor("op_1390_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473780736)))]; + tensor linear_70_cast_fp16 = linear(bias = var_1390_to_fp16, weight = var_1389_to_fp16, x = var_1380_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast_fp16 = gelu(mode = x_149_mode_0, x = linear_70_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor var_1395_to_fp16 = const()[name = tensor("op_1395_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473791040)))]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486898304)))]; + tensor linear_71_cast_fp16 = linear(bias = var_1396_to_fp16, weight = var_1395_to_fp16, x = x_149_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor x_151_cast_fp16 = add(x = x_145_cast_fp16, y = linear_71_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor var_1423_axes_0 = const()[name = tensor("op_1423_axes_0"), val = tensor([-1])]; + tensor blocks_12_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_12_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486900928)))]; + tensor blocks_12_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_12_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486903552)))]; + tensor var_1412_to_fp16 = const()[name = tensor("op_1412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1423_cast_fp16 = layer_norm(axes = var_1423_axes_0, beta = blocks_12_attn_ln_bias_to_fp16, epsilon = var_1412_to_fp16, gamma = blocks_12_attn_ln_weight_to_fp16, x = x_151_cast_fp16)[name = tensor("op_1423_cast_fp16")]; + tensor var_1434_to_fp16 = const()[name = tensor("op_1434_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486906176)))]; + tensor var_1435_to_fp16 = const()[name = tensor("op_1435_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490183040)))]; + tensor linear_72_cast_fp16 = linear(bias = var_1435_to_fp16, weight = var_1434_to_fp16, x = var_1423_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1438_to_fp16 = const()[name = tensor("op_1438_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490185664)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1438_to_fp16, x = var_1423_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1442_to_fp16 = const()[name = tensor("op_1442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493462528)))]; + tensor var_1443_to_fp16 = const()[name = tensor("op_1443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496739392)))]; + tensor linear_74_cast_fp16 = linear(bias = var_1443_to_fp16, weight = var_1442_to_fp16, x = var_1423_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([1, 1500, 20, -1])]; + tensor var_1452_cast_fp16 = reshape(shape = var_1451, x = linear_72_cast_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor const_248_to_fp16 = const()[name = tensor("const_248_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_51_cast_fp16 = mul(x = var_1452_cast_fp16, y = const_248_to_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([1, 1500, 20, -1])]; + tensor var_1459_cast_fp16 = reshape(shape = var_1458, x = linear_73_cast_fp16)[name = tensor("op_1459_cast_fp16")]; + tensor const_249_to_fp16 = const()[name = tensor("const_249_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_51_cast_fp16 = mul(x = var_1459_cast_fp16, y = const_249_to_fp16)[name = tensor("k_51_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1500, 20, -1])]; + tensor var_1466_cast_fp16 = reshape(shape = var_1465, x = linear_74_cast_fp16)[name = tensor("op_1466_cast_fp16")]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([0, 2, 1, 3])]; + tensor qk_25_transpose_x_0 = const()[name = tensor("qk_25_transpose_x_0"), val = tensor(false)]; + tensor qk_25_transpose_y_0 = const()[name = tensor("qk_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_152_perm_0 = const()[name = tensor("transpose_152_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_153_perm_0 = const()[name = tensor("transpose_153_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_269 = transpose(perm = transpose_153_perm_0, x = k_51_cast_fp16)[name = tensor("transpose_269")]; + tensor transpose_270 = transpose(perm = transpose_152_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_270")]; + tensor qk_25_cast_fp16 = matmul(transpose_x = qk_25_transpose_x_0, transpose_y = qk_25_transpose_y_0, x = transpose_270, y = transpose_269)[name = tensor("qk_25_cast_fp16")]; + tensor var_1471_cast_fp16 = softmax(axis = var_1406, x = qk_25_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor var_1473_transpose_x_0 = const()[name = tensor("op_1473_transpose_x_0"), val = tensor(false)]; + tensor var_1473_transpose_y_0 = const()[name = tensor("op_1473_transpose_y_0"), val = tensor(false)]; + tensor transpose_271 = transpose(perm = var_1467, x = var_1466_cast_fp16)[name = tensor("transpose_271")]; + tensor var_1473_cast_fp16 = matmul(transpose_x = var_1473_transpose_x_0, transpose_y = var_1473_transpose_y_0, x = var_1471_cast_fp16, y = transpose_271)[name = tensor("op_1473_cast_fp16")]; + tensor var_1474 = const()[name = tensor("op_1474"), val = tensor([0, 2, 1, 3])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 1500, 1280])]; + tensor transpose_268 = transpose(perm = var_1474, x = var_1473_cast_fp16)[name = tensor("transpose_268")]; + tensor x_155_cast_fp16 = reshape(shape = concat_12, x = transpose_268)[name = tensor("x_155_cast_fp16")]; + tensor var_1479_to_fp16 = const()[name = tensor("op_1479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496742016)))]; + tensor var_1480_to_fp16 = const()[name = tensor("op_1480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500018880)))]; + tensor linear_75_cast_fp16 = linear(bias = var_1480_to_fp16, weight = var_1479_to_fp16, x = x_155_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor x_157_cast_fp16 = add(x = x_151_cast_fp16, y = linear_75_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1487_axes_0 = const()[name = tensor("op_1487_axes_0"), val = tensor([-1])]; + tensor blocks_12_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500021504)))]; + tensor blocks_12_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500024128)))]; + tensor var_1487_cast_fp16 = layer_norm(axes = var_1487_axes_0, beta = blocks_12_mlp_ln_bias_to_fp16, epsilon = var_1412_to_fp16, gamma = blocks_12_mlp_ln_weight_to_fp16, x = x_157_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1496_to_fp16 = const()[name = tensor("op_1496_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(500026752)))]; + tensor var_1497_to_fp16 = const()[name = tensor("op_1497_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513134016)))]; + tensor linear_76_cast_fp16 = linear(bias = var_1497_to_fp16, weight = var_1496_to_fp16, x = var_1487_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor x_161_mode_0 = const()[name = tensor("x_161_mode_0"), val = tensor("EXACT")]; + tensor x_161_cast_fp16 = gelu(mode = x_161_mode_0, x = linear_76_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor var_1502_to_fp16 = const()[name = tensor("op_1502_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513144320)))]; + tensor var_1503_to_fp16 = const()[name = tensor("op_1503_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526251584)))]; + tensor linear_77_cast_fp16 = linear(bias = var_1503_to_fp16, weight = var_1502_to_fp16, x = x_161_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor x_163_cast_fp16 = add(x = x_157_cast_fp16, y = linear_77_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1513 = const()[name = tensor("op_1513"), val = tensor(-1)]; + tensor var_1530_axes_0 = const()[name = tensor("op_1530_axes_0"), val = tensor([-1])]; + tensor blocks_13_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_13_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526254208)))]; + tensor blocks_13_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_13_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526256832)))]; + tensor var_1519_to_fp16 = const()[name = tensor("op_1519_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1530_cast_fp16 = layer_norm(axes = var_1530_axes_0, beta = blocks_13_attn_ln_bias_to_fp16, epsilon = var_1519_to_fp16, gamma = blocks_13_attn_ln_weight_to_fp16, x = x_163_cast_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor var_1541_to_fp16 = const()[name = tensor("op_1541_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526259456)))]; + tensor var_1542_to_fp16 = const()[name = tensor("op_1542_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529536320)))]; + tensor linear_78_cast_fp16 = linear(bias = var_1542_to_fp16, weight = var_1541_to_fp16, x = var_1530_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_1545_to_fp16 = const()[name = tensor("op_1545_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529538944)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1545_to_fp16, x = var_1530_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor var_1549_to_fp16 = const()[name = tensor("op_1549_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532815808)))]; + tensor var_1550_to_fp16 = const()[name = tensor("op_1550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536092672)))]; + tensor linear_80_cast_fp16 = linear(bias = var_1550_to_fp16, weight = var_1549_to_fp16, x = var_1530_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_1558 = const()[name = tensor("op_1558"), val = tensor([1, 1500, 20, -1])]; + tensor var_1559_cast_fp16 = reshape(shape = var_1558, x = linear_78_cast_fp16)[name = tensor("op_1559_cast_fp16")]; + tensor const_250_to_fp16 = const()[name = tensor("const_250_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_55_cast_fp16 = mul(x = var_1559_cast_fp16, y = const_250_to_fp16)[name = tensor("q_55_cast_fp16")]; + tensor var_1565 = const()[name = tensor("op_1565"), val = tensor([1, 1500, 20, -1])]; + tensor var_1566_cast_fp16 = reshape(shape = var_1565, x = linear_79_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor const_251_to_fp16 = const()[name = tensor("const_251_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_55_cast_fp16 = mul(x = var_1566_cast_fp16, y = const_251_to_fp16)[name = tensor("k_55_cast_fp16")]; + tensor var_1572 = const()[name = tensor("op_1572"), val = tensor([1, 1500, 20, -1])]; + tensor var_1573_cast_fp16 = reshape(shape = var_1572, x = linear_80_cast_fp16)[name = tensor("op_1573_cast_fp16")]; + tensor var_1574 = const()[name = tensor("op_1574"), val = tensor([0, 2, 1, 3])]; + tensor qk_27_transpose_x_0 = const()[name = tensor("qk_27_transpose_x_0"), val = tensor(false)]; + tensor qk_27_transpose_y_0 = const()[name = tensor("qk_27_transpose_y_0"), val = tensor(false)]; + tensor transpose_154_perm_0 = const()[name = tensor("transpose_154_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_155_perm_0 = const()[name = tensor("transpose_155_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_265 = transpose(perm = transpose_155_perm_0, x = k_55_cast_fp16)[name = tensor("transpose_265")]; + tensor transpose_266 = transpose(perm = transpose_154_perm_0, x = q_55_cast_fp16)[name = tensor("transpose_266")]; + tensor qk_27_cast_fp16 = matmul(transpose_x = qk_27_transpose_x_0, transpose_y = qk_27_transpose_y_0, x = transpose_266, y = transpose_265)[name = tensor("qk_27_cast_fp16")]; + tensor var_1578_cast_fp16 = softmax(axis = var_1513, x = qk_27_cast_fp16)[name = tensor("op_1578_cast_fp16")]; + tensor var_1580_transpose_x_0 = const()[name = tensor("op_1580_transpose_x_0"), val = tensor(false)]; + tensor var_1580_transpose_y_0 = const()[name = tensor("op_1580_transpose_y_0"), val = tensor(false)]; + tensor transpose_267 = transpose(perm = var_1574, x = var_1573_cast_fp16)[name = tensor("transpose_267")]; + tensor var_1580_cast_fp16 = matmul(transpose_x = var_1580_transpose_x_0, transpose_y = var_1580_transpose_y_0, x = var_1578_cast_fp16, y = transpose_267)[name = tensor("op_1580_cast_fp16")]; + tensor var_1581 = const()[name = tensor("op_1581"), val = tensor([0, 2, 1, 3])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([1, 1500, 1280])]; + tensor transpose_264 = transpose(perm = var_1581, x = var_1580_cast_fp16)[name = tensor("transpose_264")]; + tensor x_167_cast_fp16 = reshape(shape = concat_13, x = transpose_264)[name = tensor("x_167_cast_fp16")]; + tensor var_1586_to_fp16 = const()[name = tensor("op_1586_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095296)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539372160)))]; + tensor linear_81_cast_fp16 = linear(bias = var_1587_to_fp16, weight = var_1586_to_fp16, x = x_167_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor x_169_cast_fp16 = add(x = x_163_cast_fp16, y = linear_81_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1594_axes_0 = const()[name = tensor("op_1594_axes_0"), val = tensor([-1])]; + tensor blocks_13_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539374784)))]; + tensor blocks_13_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539377408)))]; + tensor var_1594_cast_fp16 = layer_norm(axes = var_1594_axes_0, beta = blocks_13_mlp_ln_bias_to_fp16, epsilon = var_1519_to_fp16, gamma = blocks_13_mlp_ln_weight_to_fp16, x = x_169_cast_fp16)[name = tensor("op_1594_cast_fp16")]; + tensor var_1603_to_fp16 = const()[name = tensor("op_1603_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539380032)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552487296)))]; + tensor linear_82_cast_fp16 = linear(bias = var_1604_to_fp16, weight = var_1603_to_fp16, x = var_1594_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor x_173_mode_0 = const()[name = tensor("x_173_mode_0"), val = tensor("EXACT")]; + tensor x_173_cast_fp16 = gelu(mode = x_173_mode_0, x = linear_82_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor var_1609_to_fp16 = const()[name = tensor("op_1609_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552497600)))]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565604864)))]; + tensor linear_83_cast_fp16 = linear(bias = var_1610_to_fp16, weight = var_1609_to_fp16, x = x_173_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor x_175_cast_fp16 = add(x = x_169_cast_fp16, y = linear_83_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1620 = const()[name = tensor("op_1620"), val = tensor(-1)]; + tensor var_1637_axes_0 = const()[name = tensor("op_1637_axes_0"), val = tensor([-1])]; + tensor blocks_14_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_14_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565607488)))]; + tensor blocks_14_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_14_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565610112)))]; + tensor var_1626_to_fp16 = const()[name = tensor("op_1626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1637_cast_fp16 = layer_norm(axes = var_1637_axes_0, beta = blocks_14_attn_ln_bias_to_fp16, epsilon = var_1626_to_fp16, gamma = blocks_14_attn_ln_weight_to_fp16, x = x_175_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor var_1648_to_fp16 = const()[name = tensor("op_1648_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565612736)))]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568889600)))]; + tensor linear_84_cast_fp16 = linear(bias = var_1649_to_fp16, weight = var_1648_to_fp16, x = var_1637_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1652_to_fp16 = const()[name = tensor("op_1652_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568892224)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1652_to_fp16, x = var_1637_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_1656_to_fp16 = const()[name = tensor("op_1656_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572169088)))]; + tensor var_1657_to_fp16 = const()[name = tensor("op_1657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575445952)))]; + tensor linear_86_cast_fp16 = linear(bias = var_1657_to_fp16, weight = var_1656_to_fp16, x = var_1637_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1665 = const()[name = tensor("op_1665"), val = tensor([1, 1500, 20, -1])]; + tensor var_1666_cast_fp16 = reshape(shape = var_1665, x = linear_84_cast_fp16)[name = tensor("op_1666_cast_fp16")]; + tensor const_252_to_fp16 = const()[name = tensor("const_252_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_59_cast_fp16 = mul(x = var_1666_cast_fp16, y = const_252_to_fp16)[name = tensor("q_59_cast_fp16")]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 1500, 20, -1])]; + tensor var_1673_cast_fp16 = reshape(shape = var_1672, x = linear_85_cast_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor const_253_to_fp16 = const()[name = tensor("const_253_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_59_cast_fp16 = mul(x = var_1673_cast_fp16, y = const_253_to_fp16)[name = tensor("k_59_cast_fp16")]; + tensor var_1679 = const()[name = tensor("op_1679"), val = tensor([1, 1500, 20, -1])]; + tensor var_1680_cast_fp16 = reshape(shape = var_1679, x = linear_86_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([0, 2, 1, 3])]; + tensor qk_29_transpose_x_0 = const()[name = tensor("qk_29_transpose_x_0"), val = tensor(false)]; + tensor qk_29_transpose_y_0 = const()[name = tensor("qk_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_156_perm_0 = const()[name = tensor("transpose_156_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_157_perm_0 = const()[name = tensor("transpose_157_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_261 = transpose(perm = transpose_157_perm_0, x = k_59_cast_fp16)[name = tensor("transpose_261")]; + tensor transpose_262 = transpose(perm = transpose_156_perm_0, x = q_59_cast_fp16)[name = tensor("transpose_262")]; + tensor qk_29_cast_fp16 = matmul(transpose_x = qk_29_transpose_x_0, transpose_y = qk_29_transpose_y_0, x = transpose_262, y = transpose_261)[name = tensor("qk_29_cast_fp16")]; + tensor var_1685_cast_fp16 = softmax(axis = var_1620, x = qk_29_cast_fp16)[name = tensor("op_1685_cast_fp16")]; + tensor var_1687_transpose_x_0 = const()[name = tensor("op_1687_transpose_x_0"), val = tensor(false)]; + tensor var_1687_transpose_y_0 = const()[name = tensor("op_1687_transpose_y_0"), val = tensor(false)]; + tensor transpose_263 = transpose(perm = var_1681, x = var_1680_cast_fp16)[name = tensor("transpose_263")]; + tensor var_1687_cast_fp16 = matmul(transpose_x = var_1687_transpose_x_0, transpose_y = var_1687_transpose_y_0, x = var_1685_cast_fp16, y = transpose_263)[name = tensor("op_1687_cast_fp16")]; + tensor var_1688 = const()[name = tensor("op_1688"), val = tensor([0, 2, 1, 3])]; + tensor concat_14 = const()[name = tensor("concat_14"), val = tensor([1, 1500, 1280])]; + tensor transpose_260 = transpose(perm = var_1688, x = var_1687_cast_fp16)[name = tensor("transpose_260")]; + tensor x_179_cast_fp16 = reshape(shape = concat_14, x = transpose_260)[name = tensor("x_179_cast_fp16")]; + tensor var_1693_to_fp16 = const()[name = tensor("op_1693_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(575448576)))]; + tensor var_1694_to_fp16 = const()[name = tensor("op_1694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578725440)))]; + tensor linear_87_cast_fp16 = linear(bias = var_1694_to_fp16, weight = var_1693_to_fp16, x = x_179_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor x_181_cast_fp16 = add(x = x_175_cast_fp16, y = linear_87_cast_fp16)[name = tensor("x_181_cast_fp16")]; + tensor var_1701_axes_0 = const()[name = tensor("op_1701_axes_0"), val = tensor([-1])]; + tensor blocks_14_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578728064)))]; + tensor blocks_14_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578730688)))]; + tensor var_1701_cast_fp16 = layer_norm(axes = var_1701_axes_0, beta = blocks_14_mlp_ln_bias_to_fp16, epsilon = var_1626_to_fp16, gamma = blocks_14_mlp_ln_weight_to_fp16, x = x_181_cast_fp16)[name = tensor("op_1701_cast_fp16")]; + tensor var_1710_to_fp16 = const()[name = tensor("op_1710_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578733312)))]; + tensor var_1711_to_fp16 = const()[name = tensor("op_1711_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591840576)))]; + tensor linear_88_cast_fp16 = linear(bias = var_1711_to_fp16, weight = var_1710_to_fp16, x = var_1701_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor x_185_mode_0 = const()[name = tensor("x_185_mode_0"), val = tensor("EXACT")]; + tensor x_185_cast_fp16 = gelu(mode = x_185_mode_0, x = linear_88_cast_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1716_to_fp16 = const()[name = tensor("op_1716_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591850880)))]; + tensor var_1717_to_fp16 = const()[name = tensor("op_1717_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604958144)))]; + tensor linear_89_cast_fp16 = linear(bias = var_1717_to_fp16, weight = var_1716_to_fp16, x = x_185_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = x_181_cast_fp16, y = linear_89_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor var_1727 = const()[name = tensor("op_1727"), val = tensor(-1)]; + tensor var_1744_axes_0 = const()[name = tensor("op_1744_axes_0"), val = tensor([-1])]; + tensor blocks_15_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_15_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604960768)))]; + tensor blocks_15_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_15_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604963392)))]; + tensor var_1733_to_fp16 = const()[name = tensor("op_1733_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1744_cast_fp16 = layer_norm(axes = var_1744_axes_0, beta = blocks_15_attn_ln_bias_to_fp16, epsilon = var_1733_to_fp16, gamma = blocks_15_attn_ln_weight_to_fp16, x = x_187_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1755_to_fp16 = const()[name = tensor("op_1755_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604966016)))]; + tensor var_1756_to_fp16 = const()[name = tensor("op_1756_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(608242880)))]; + tensor linear_90_cast_fp16 = linear(bias = var_1756_to_fp16, weight = var_1755_to_fp16, x = var_1744_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1759_to_fp16 = const()[name = tensor("op_1759_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(608245504)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1759_to_fp16, x = var_1744_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611522368)))]; + tensor var_1764_to_fp16 = const()[name = tensor("op_1764_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614799232)))]; + tensor linear_92_cast_fp16 = linear(bias = var_1764_to_fp16, weight = var_1763_to_fp16, x = var_1744_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1500, 20, -1])]; + tensor var_1773_cast_fp16 = reshape(shape = var_1772, x = linear_90_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor const_254_to_fp16 = const()[name = tensor("const_254_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_63_cast_fp16 = mul(x = var_1773_cast_fp16, y = const_254_to_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 1500, 20, -1])]; + tensor var_1780_cast_fp16 = reshape(shape = var_1779, x = linear_91_cast_fp16)[name = tensor("op_1780_cast_fp16")]; + tensor const_255_to_fp16 = const()[name = tensor("const_255_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_63_cast_fp16 = mul(x = var_1780_cast_fp16, y = const_255_to_fp16)[name = tensor("k_63_cast_fp16")]; + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 1500, 20, -1])]; + tensor var_1787_cast_fp16 = reshape(shape = var_1786, x = linear_92_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([0, 2, 1, 3])]; + tensor qk_31_transpose_x_0 = const()[name = tensor("qk_31_transpose_x_0"), val = tensor(false)]; + tensor qk_31_transpose_y_0 = const()[name = tensor("qk_31_transpose_y_0"), val = tensor(false)]; + tensor transpose_158_perm_0 = const()[name = tensor("transpose_158_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_159_perm_0 = const()[name = tensor("transpose_159_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_257 = transpose(perm = transpose_159_perm_0, x = k_63_cast_fp16)[name = tensor("transpose_257")]; + tensor transpose_258 = transpose(perm = transpose_158_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_258")]; + tensor qk_31_cast_fp16 = matmul(transpose_x = qk_31_transpose_x_0, transpose_y = qk_31_transpose_y_0, x = transpose_258, y = transpose_257)[name = tensor("qk_31_cast_fp16")]; + tensor var_1792_cast_fp16 = softmax(axis = var_1727, x = qk_31_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor var_1794_transpose_x_0 = const()[name = tensor("op_1794_transpose_x_0"), val = tensor(false)]; + tensor var_1794_transpose_y_0 = const()[name = tensor("op_1794_transpose_y_0"), val = tensor(false)]; + tensor transpose_259 = transpose(perm = var_1788, x = var_1787_cast_fp16)[name = tensor("transpose_259")]; + tensor var_1794_cast_fp16 = matmul(transpose_x = var_1794_transpose_x_0, transpose_y = var_1794_transpose_y_0, x = var_1792_cast_fp16, y = transpose_259)[name = tensor("op_1794_cast_fp16")]; + tensor var_1795 = const()[name = tensor("op_1795"), val = tensor([0, 2, 1, 3])]; + tensor concat_15 = const()[name = tensor("concat_15"), val = tensor([1, 1500, 1280])]; + tensor transpose_256 = transpose(perm = var_1795, x = var_1794_cast_fp16)[name = tensor("transpose_256")]; + tensor x_191_cast_fp16 = reshape(shape = concat_15, x = transpose_256)[name = tensor("x_191_cast_fp16")]; + tensor var_1800_to_fp16 = const()[name = tensor("op_1800_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614801856)))]; + tensor var_1801_to_fp16 = const()[name = tensor("op_1801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618078720)))]; + tensor linear_93_cast_fp16 = linear(bias = var_1801_to_fp16, weight = var_1800_to_fp16, x = x_191_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = linear_93_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor var_1808_axes_0 = const()[name = tensor("op_1808_axes_0"), val = tensor([-1])]; + tensor blocks_15_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618081344)))]; + tensor blocks_15_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618083968)))]; + tensor var_1808_cast_fp16 = layer_norm(axes = var_1808_axes_0, beta = blocks_15_mlp_ln_bias_to_fp16, epsilon = var_1733_to_fp16, gamma = blocks_15_mlp_ln_weight_to_fp16, x = x_193_cast_fp16)[name = tensor("op_1808_cast_fp16")]; + tensor var_1817_to_fp16 = const()[name = tensor("op_1817_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(618086592)))]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(631193856)))]; + tensor linear_94_cast_fp16 = linear(bias = var_1818_to_fp16, weight = var_1817_to_fp16, x = var_1808_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor x_197_mode_0 = const()[name = tensor("x_197_mode_0"), val = tensor("EXACT")]; + tensor x_197_cast_fp16 = gelu(mode = x_197_mode_0, x = linear_94_cast_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1823_to_fp16 = const()[name = tensor("op_1823_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(631204160)))]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644311424)))]; + tensor linear_95_cast_fp16 = linear(bias = var_1824_to_fp16, weight = var_1823_to_fp16, x = x_197_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = linear_95_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor var_1834 = const()[name = tensor("op_1834"), val = tensor(-1)]; + tensor var_1851_axes_0 = const()[name = tensor("op_1851_axes_0"), val = tensor([-1])]; + tensor blocks_16_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_16_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644314048)))]; + tensor blocks_16_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_16_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644316672)))]; + tensor var_1840_to_fp16 = const()[name = tensor("op_1840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1851_cast_fp16 = layer_norm(axes = var_1851_axes_0, beta = blocks_16_attn_ln_bias_to_fp16, epsilon = var_1840_to_fp16, gamma = blocks_16_attn_ln_weight_to_fp16, x = x_199_cast_fp16)[name = tensor("op_1851_cast_fp16")]; + tensor var_1862_to_fp16 = const()[name = tensor("op_1862_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644319296)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647596160)))]; + tensor linear_96_cast_fp16 = linear(bias = var_1863_to_fp16, weight = var_1862_to_fp16, x = var_1851_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1866_to_fp16 = const()[name = tensor("op_1866_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647598784)))]; + tensor linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1866_to_fp16, x = var_1851_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor var_1870_to_fp16 = const()[name = tensor("op_1870_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650875648)))]; + tensor var_1871_to_fp16 = const()[name = tensor("op_1871_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(654152512)))]; + tensor linear_98_cast_fp16 = linear(bias = var_1871_to_fp16, weight = var_1870_to_fp16, x = var_1851_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([1, 1500, 20, -1])]; + tensor var_1880_cast_fp16 = reshape(shape = var_1879, x = linear_96_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor const_256_to_fp16 = const()[name = tensor("const_256_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_67_cast_fp16 = mul(x = var_1880_cast_fp16, y = const_256_to_fp16)[name = tensor("q_67_cast_fp16")]; + tensor var_1886 = const()[name = tensor("op_1886"), val = tensor([1, 1500, 20, -1])]; + tensor var_1887_cast_fp16 = reshape(shape = var_1886, x = linear_97_cast_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor const_257_to_fp16 = const()[name = tensor("const_257_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_67_cast_fp16 = mul(x = var_1887_cast_fp16, y = const_257_to_fp16)[name = tensor("k_67_cast_fp16")]; + tensor var_1893 = const()[name = tensor("op_1893"), val = tensor([1, 1500, 20, -1])]; + tensor var_1894_cast_fp16 = reshape(shape = var_1893, x = linear_98_cast_fp16)[name = tensor("op_1894_cast_fp16")]; + tensor var_1895 = const()[name = tensor("op_1895"), val = tensor([0, 2, 1, 3])]; + tensor qk_33_transpose_x_0 = const()[name = tensor("qk_33_transpose_x_0"), val = tensor(false)]; + tensor qk_33_transpose_y_0 = const()[name = tensor("qk_33_transpose_y_0"), val = tensor(false)]; + tensor transpose_160_perm_0 = const()[name = tensor("transpose_160_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_161_perm_0 = const()[name = tensor("transpose_161_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_253 = transpose(perm = transpose_161_perm_0, x = k_67_cast_fp16)[name = tensor("transpose_253")]; + tensor transpose_254 = transpose(perm = transpose_160_perm_0, x = q_67_cast_fp16)[name = tensor("transpose_254")]; + tensor qk_33_cast_fp16 = matmul(transpose_x = qk_33_transpose_x_0, transpose_y = qk_33_transpose_y_0, x = transpose_254, y = transpose_253)[name = tensor("qk_33_cast_fp16")]; + tensor var_1899_cast_fp16 = softmax(axis = var_1834, x = qk_33_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor var_1901_transpose_x_0 = const()[name = tensor("op_1901_transpose_x_0"), val = tensor(false)]; + tensor var_1901_transpose_y_0 = const()[name = tensor("op_1901_transpose_y_0"), val = tensor(false)]; + tensor transpose_255 = transpose(perm = var_1895, x = var_1894_cast_fp16)[name = tensor("transpose_255")]; + tensor var_1901_cast_fp16 = matmul(transpose_x = var_1901_transpose_x_0, transpose_y = var_1901_transpose_y_0, x = var_1899_cast_fp16, y = transpose_255)[name = tensor("op_1901_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([1, 1500, 1280])]; + tensor transpose_252 = transpose(perm = var_1902, x = var_1901_cast_fp16)[name = tensor("transpose_252")]; + tensor x_203_cast_fp16 = reshape(shape = concat_16, x = transpose_252)[name = tensor("x_203_cast_fp16")]; + tensor var_1907_to_fp16 = const()[name = tensor("op_1907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(654155136)))]; + tensor var_1908_to_fp16 = const()[name = tensor("op_1908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657432000)))]; + tensor linear_99_cast_fp16 = linear(bias = var_1908_to_fp16, weight = var_1907_to_fp16, x = x_203_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = linear_99_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor var_1915_axes_0 = const()[name = tensor("op_1915_axes_0"), val = tensor([-1])]; + tensor blocks_16_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657434624)))]; + tensor blocks_16_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657437248)))]; + tensor var_1915_cast_fp16 = layer_norm(axes = var_1915_axes_0, beta = blocks_16_mlp_ln_bias_to_fp16, epsilon = var_1840_to_fp16, gamma = blocks_16_mlp_ln_weight_to_fp16, x = x_205_cast_fp16)[name = tensor("op_1915_cast_fp16")]; + tensor var_1924_to_fp16 = const()[name = tensor("op_1924_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657439872)))]; + tensor var_1925_to_fp16 = const()[name = tensor("op_1925_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670547136)))]; + tensor linear_100_cast_fp16 = linear(bias = var_1925_to_fp16, weight = var_1924_to_fp16, x = var_1915_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor x_209_mode_0 = const()[name = tensor("x_209_mode_0"), val = tensor("EXACT")]; + tensor x_209_cast_fp16 = gelu(mode = x_209_mode_0, x = linear_100_cast_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1930_to_fp16 = const()[name = tensor("op_1930_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(670557440)))]; + tensor var_1931_to_fp16 = const()[name = tensor("op_1931_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683664704)))]; + tensor linear_101_cast_fp16 = linear(bias = var_1931_to_fp16, weight = var_1930_to_fp16, x = x_209_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = linear_101_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor var_1941 = const()[name = tensor("op_1941"), val = tensor(-1)]; + tensor var_1958_axes_0 = const()[name = tensor("op_1958_axes_0"), val = tensor([-1])]; + tensor blocks_17_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_17_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683667328)))]; + tensor blocks_17_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_17_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683669952)))]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1958_cast_fp16 = layer_norm(axes = var_1958_axes_0, beta = blocks_17_attn_ln_bias_to_fp16, epsilon = var_1947_to_fp16, gamma = blocks_17_attn_ln_weight_to_fp16, x = x_211_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor var_1969_to_fp16 = const()[name = tensor("op_1969_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683672576)))]; + tensor var_1970_to_fp16 = const()[name = tensor("op_1970_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686949440)))]; + tensor linear_102_cast_fp16 = linear(bias = var_1970_to_fp16, weight = var_1969_to_fp16, x = var_1958_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686952064)))]; + tensor linear_103_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_1973_to_fp16, x = var_1958_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1977_to_fp16 = const()[name = tensor("op_1977_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(690228928)))]; + tensor var_1978_to_fp16 = const()[name = tensor("op_1978_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693505792)))]; + tensor linear_104_cast_fp16 = linear(bias = var_1978_to_fp16, weight = var_1977_to_fp16, x = var_1958_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 1500, 20, -1])]; + tensor var_1987_cast_fp16 = reshape(shape = var_1986, x = linear_102_cast_fp16)[name = tensor("op_1987_cast_fp16")]; + tensor const_258_to_fp16 = const()[name = tensor("const_258_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_71_cast_fp16 = mul(x = var_1987_cast_fp16, y = const_258_to_fp16)[name = tensor("q_71_cast_fp16")]; + tensor var_1993 = const()[name = tensor("op_1993"), val = tensor([1, 1500, 20, -1])]; + tensor var_1994_cast_fp16 = reshape(shape = var_1993, x = linear_103_cast_fp16)[name = tensor("op_1994_cast_fp16")]; + tensor const_259_to_fp16 = const()[name = tensor("const_259_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_71_cast_fp16 = mul(x = var_1994_cast_fp16, y = const_259_to_fp16)[name = tensor("k_71_cast_fp16")]; + tensor var_2000 = const()[name = tensor("op_2000"), val = tensor([1, 1500, 20, -1])]; + tensor var_2001_cast_fp16 = reshape(shape = var_2000, x = linear_104_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor var_2002 = const()[name = tensor("op_2002"), val = tensor([0, 2, 1, 3])]; + tensor qk_35_transpose_x_0 = const()[name = tensor("qk_35_transpose_x_0"), val = tensor(false)]; + tensor qk_35_transpose_y_0 = const()[name = tensor("qk_35_transpose_y_0"), val = tensor(false)]; + tensor transpose_162_perm_0 = const()[name = tensor("transpose_162_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_163_perm_0 = const()[name = tensor("transpose_163_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_249 = transpose(perm = transpose_163_perm_0, x = k_71_cast_fp16)[name = tensor("transpose_249")]; + tensor transpose_250 = transpose(perm = transpose_162_perm_0, x = q_71_cast_fp16)[name = tensor("transpose_250")]; + tensor qk_35_cast_fp16 = matmul(transpose_x = qk_35_transpose_x_0, transpose_y = qk_35_transpose_y_0, x = transpose_250, y = transpose_249)[name = tensor("qk_35_cast_fp16")]; + tensor var_2006_cast_fp16 = softmax(axis = var_1941, x = qk_35_cast_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2008_transpose_x_0 = const()[name = tensor("op_2008_transpose_x_0"), val = tensor(false)]; + tensor var_2008_transpose_y_0 = const()[name = tensor("op_2008_transpose_y_0"), val = tensor(false)]; + tensor transpose_251 = transpose(perm = var_2002, x = var_2001_cast_fp16)[name = tensor("transpose_251")]; + tensor var_2008_cast_fp16 = matmul(transpose_x = var_2008_transpose_x_0, transpose_y = var_2008_transpose_y_0, x = var_2006_cast_fp16, y = transpose_251)[name = tensor("op_2008_cast_fp16")]; + tensor var_2009 = const()[name = tensor("op_2009"), val = tensor([0, 2, 1, 3])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([1, 1500, 1280])]; + tensor transpose_248 = transpose(perm = var_2009, x = var_2008_cast_fp16)[name = tensor("transpose_248")]; + tensor x_215_cast_fp16 = reshape(shape = concat_17, x = transpose_248)[name = tensor("x_215_cast_fp16")]; + tensor var_2014_to_fp16 = const()[name = tensor("op_2014_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(693508416)))]; + tensor var_2015_to_fp16 = const()[name = tensor("op_2015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696785280)))]; + tensor linear_105_cast_fp16 = linear(bias = var_2015_to_fp16, weight = var_2014_to_fp16, x = x_215_cast_fp16)[name = tensor("linear_105_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = linear_105_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor var_2022_axes_0 = const()[name = tensor("op_2022_axes_0"), val = tensor([-1])]; + tensor blocks_17_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696787904)))]; + tensor blocks_17_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696790528)))]; + tensor var_2022_cast_fp16 = layer_norm(axes = var_2022_axes_0, beta = blocks_17_mlp_ln_bias_to_fp16, epsilon = var_1947_to_fp16, gamma = blocks_17_mlp_ln_weight_to_fp16, x = x_217_cast_fp16)[name = tensor("op_2022_cast_fp16")]; + tensor var_2031_to_fp16 = const()[name = tensor("op_2031_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(696793152)))]; + tensor var_2032_to_fp16 = const()[name = tensor("op_2032_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709900416)))]; + tensor linear_106_cast_fp16 = linear(bias = var_2032_to_fp16, weight = var_2031_to_fp16, x = var_2022_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor x_221_mode_0 = const()[name = tensor("x_221_mode_0"), val = tensor("EXACT")]; + tensor x_221_cast_fp16 = gelu(mode = x_221_mode_0, x = linear_106_cast_fp16)[name = tensor("x_221_cast_fp16")]; + tensor var_2037_to_fp16 = const()[name = tensor("op_2037_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(709910720)))]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723017984)))]; + tensor linear_107_cast_fp16 = linear(bias = var_2038_to_fp16, weight = var_2037_to_fp16, x = x_221_cast_fp16)[name = tensor("linear_107_cast_fp16")]; + tensor x_223_cast_fp16 = add(x = x_217_cast_fp16, y = linear_107_cast_fp16)[name = tensor("x_223_cast_fp16")]; + tensor var_2048 = const()[name = tensor("op_2048"), val = tensor(-1)]; + tensor var_2065_axes_0 = const()[name = tensor("op_2065_axes_0"), val = tensor([-1])]; + tensor blocks_18_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_18_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723020608)))]; + tensor blocks_18_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_18_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723023232)))]; + tensor var_2054_to_fp16 = const()[name = tensor("op_2054_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2065_cast_fp16 = layer_norm(axes = var_2065_axes_0, beta = blocks_18_attn_ln_bias_to_fp16, epsilon = var_2054_to_fp16, gamma = blocks_18_attn_ln_weight_to_fp16, x = x_223_cast_fp16)[name = tensor("op_2065_cast_fp16")]; + tensor var_2076_to_fp16 = const()[name = tensor("op_2076_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723025856)))]; + tensor var_2077_to_fp16 = const()[name = tensor("op_2077_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726302720)))]; + tensor linear_108_cast_fp16 = linear(bias = var_2077_to_fp16, weight = var_2076_to_fp16, x = var_2065_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor var_2080_to_fp16 = const()[name = tensor("op_2080_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726305344)))]; + tensor linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2080_to_fp16, x = var_2065_cast_fp16)[name = tensor("linear_109_cast_fp16")]; + tensor var_2084_to_fp16 = const()[name = tensor("op_2084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(729582208)))]; + tensor var_2085_to_fp16 = const()[name = tensor("op_2085_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732859072)))]; + tensor linear_110_cast_fp16 = linear(bias = var_2085_to_fp16, weight = var_2084_to_fp16, x = var_2065_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([1, 1500, 20, -1])]; + tensor var_2094_cast_fp16 = reshape(shape = var_2093, x = linear_108_cast_fp16)[name = tensor("op_2094_cast_fp16")]; + tensor const_260_to_fp16 = const()[name = tensor("const_260_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_75_cast_fp16 = mul(x = var_2094_cast_fp16, y = const_260_to_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_2100 = const()[name = tensor("op_2100"), val = tensor([1, 1500, 20, -1])]; + tensor var_2101_cast_fp16 = reshape(shape = var_2100, x = linear_109_cast_fp16)[name = tensor("op_2101_cast_fp16")]; + tensor const_261_to_fp16 = const()[name = tensor("const_261_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_75_cast_fp16 = mul(x = var_2101_cast_fp16, y = const_261_to_fp16)[name = tensor("k_75_cast_fp16")]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 1500, 20, -1])]; + tensor var_2108_cast_fp16 = reshape(shape = var_2107, x = linear_110_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor var_2109 = const()[name = tensor("op_2109"), val = tensor([0, 2, 1, 3])]; + tensor qk_37_transpose_x_0 = const()[name = tensor("qk_37_transpose_x_0"), val = tensor(false)]; + tensor qk_37_transpose_y_0 = const()[name = tensor("qk_37_transpose_y_0"), val = tensor(false)]; + tensor transpose_164_perm_0 = const()[name = tensor("transpose_164_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_165_perm_0 = const()[name = tensor("transpose_165_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_245 = transpose(perm = transpose_165_perm_0, x = k_75_cast_fp16)[name = tensor("transpose_245")]; + tensor transpose_246 = transpose(perm = transpose_164_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_246")]; + tensor qk_37_cast_fp16 = matmul(transpose_x = qk_37_transpose_x_0, transpose_y = qk_37_transpose_y_0, x = transpose_246, y = transpose_245)[name = tensor("qk_37_cast_fp16")]; + tensor var_2113_cast_fp16 = softmax(axis = var_2048, x = qk_37_cast_fp16)[name = tensor("op_2113_cast_fp16")]; + tensor var_2115_transpose_x_0 = const()[name = tensor("op_2115_transpose_x_0"), val = tensor(false)]; + tensor var_2115_transpose_y_0 = const()[name = tensor("op_2115_transpose_y_0"), val = tensor(false)]; + tensor transpose_247 = transpose(perm = var_2109, x = var_2108_cast_fp16)[name = tensor("transpose_247")]; + tensor var_2115_cast_fp16 = matmul(transpose_x = var_2115_transpose_x_0, transpose_y = var_2115_transpose_y_0, x = var_2113_cast_fp16, y = transpose_247)[name = tensor("op_2115_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor([0, 2, 1, 3])]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 1500, 1280])]; + tensor transpose_244 = transpose(perm = var_2116, x = var_2115_cast_fp16)[name = tensor("transpose_244")]; + tensor x_227_cast_fp16 = reshape(shape = concat_18, x = transpose_244)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_to_fp16 = const()[name = tensor("op_2121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(732861696)))]; + tensor var_2122_to_fp16 = const()[name = tensor("op_2122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736138560)))]; + tensor linear_111_cast_fp16 = linear(bias = var_2122_to_fp16, weight = var_2121_to_fp16, x = x_227_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = x_223_cast_fp16, y = linear_111_cast_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2129_axes_0 = const()[name = tensor("op_2129_axes_0"), val = tensor([-1])]; + tensor blocks_18_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736141184)))]; + tensor blocks_18_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736143808)))]; + tensor var_2129_cast_fp16 = layer_norm(axes = var_2129_axes_0, beta = blocks_18_mlp_ln_bias_to_fp16, epsilon = var_2054_to_fp16, gamma = blocks_18_mlp_ln_weight_to_fp16, x = x_229_cast_fp16)[name = tensor("op_2129_cast_fp16")]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736146432)))]; + tensor var_2139_to_fp16 = const()[name = tensor("op_2139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749253696)))]; + tensor linear_112_cast_fp16 = linear(bias = var_2139_to_fp16, weight = var_2138_to_fp16, x = var_2129_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor x_233_mode_0 = const()[name = tensor("x_233_mode_0"), val = tensor("EXACT")]; + tensor x_233_cast_fp16 = gelu(mode = x_233_mode_0, x = linear_112_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2144_to_fp16 = const()[name = tensor("op_2144_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749264000)))]; + tensor var_2145_to_fp16 = const()[name = tensor("op_2145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762371264)))]; + tensor linear_113_cast_fp16 = linear(bias = var_2145_to_fp16, weight = var_2144_to_fp16, x = x_233_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = x_229_cast_fp16, y = linear_113_cast_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2155 = const()[name = tensor("op_2155"), val = tensor(-1)]; + tensor var_2172_axes_0 = const()[name = tensor("op_2172_axes_0"), val = tensor([-1])]; + tensor blocks_19_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_19_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762373888)))]; + tensor blocks_19_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_19_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762376512)))]; + tensor var_2161_to_fp16 = const()[name = tensor("op_2161_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2172_cast_fp16 = layer_norm(axes = var_2172_axes_0, beta = blocks_19_attn_ln_bias_to_fp16, epsilon = var_2161_to_fp16, gamma = blocks_19_attn_ln_weight_to_fp16, x = x_235_cast_fp16)[name = tensor("op_2172_cast_fp16")]; + tensor var_2183_to_fp16 = const()[name = tensor("op_2183_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762379136)))]; + tensor var_2184_to_fp16 = const()[name = tensor("op_2184_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765656000)))]; + tensor linear_114_cast_fp16 = linear(bias = var_2184_to_fp16, weight = var_2183_to_fp16, x = var_2172_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor var_2187_to_fp16 = const()[name = tensor("op_2187_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765658624)))]; + tensor linear_115_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2187_to_fp16, x = var_2172_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_2191_to_fp16 = const()[name = tensor("op_2191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768935488)))]; + tensor var_2192_to_fp16 = const()[name = tensor("op_2192_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(772212352)))]; + tensor linear_116_cast_fp16 = linear(bias = var_2192_to_fp16, weight = var_2191_to_fp16, x = var_2172_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_2200 = const()[name = tensor("op_2200"), val = tensor([1, 1500, 20, -1])]; + tensor var_2201_cast_fp16 = reshape(shape = var_2200, x = linear_114_cast_fp16)[name = tensor("op_2201_cast_fp16")]; + tensor const_262_to_fp16 = const()[name = tensor("const_262_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_79_cast_fp16 = mul(x = var_2201_cast_fp16, y = const_262_to_fp16)[name = tensor("q_79_cast_fp16")]; + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([1, 1500, 20, -1])]; + tensor var_2208_cast_fp16 = reshape(shape = var_2207, x = linear_115_cast_fp16)[name = tensor("op_2208_cast_fp16")]; + tensor const_263_to_fp16 = const()[name = tensor("const_263_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_79_cast_fp16 = mul(x = var_2208_cast_fp16, y = const_263_to_fp16)[name = tensor("k_79_cast_fp16")]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1500, 20, -1])]; + tensor var_2215_cast_fp16 = reshape(shape = var_2214, x = linear_116_cast_fp16)[name = tensor("op_2215_cast_fp16")]; + tensor var_2216 = const()[name = tensor("op_2216"), val = tensor([0, 2, 1, 3])]; + tensor qk_39_transpose_x_0 = const()[name = tensor("qk_39_transpose_x_0"), val = tensor(false)]; + tensor qk_39_transpose_y_0 = const()[name = tensor("qk_39_transpose_y_0"), val = tensor(false)]; + tensor transpose_166_perm_0 = const()[name = tensor("transpose_166_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_167_perm_0 = const()[name = tensor("transpose_167_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_241 = transpose(perm = transpose_167_perm_0, x = k_79_cast_fp16)[name = tensor("transpose_241")]; + tensor transpose_242 = transpose(perm = transpose_166_perm_0, x = q_79_cast_fp16)[name = tensor("transpose_242")]; + tensor qk_39_cast_fp16 = matmul(transpose_x = qk_39_transpose_x_0, transpose_y = qk_39_transpose_y_0, x = transpose_242, y = transpose_241)[name = tensor("qk_39_cast_fp16")]; + tensor var_2220_cast_fp16 = softmax(axis = var_2155, x = qk_39_cast_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2222_transpose_x_0 = const()[name = tensor("op_2222_transpose_x_0"), val = tensor(false)]; + tensor var_2222_transpose_y_0 = const()[name = tensor("op_2222_transpose_y_0"), val = tensor(false)]; + tensor transpose_243 = transpose(perm = var_2216, x = var_2215_cast_fp16)[name = tensor("transpose_243")]; + tensor var_2222_cast_fp16 = matmul(transpose_x = var_2222_transpose_x_0, transpose_y = var_2222_transpose_y_0, x = var_2220_cast_fp16, y = transpose_243)[name = tensor("op_2222_cast_fp16")]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([0, 2, 1, 3])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([1, 1500, 1280])]; + tensor transpose_240 = transpose(perm = var_2223, x = var_2222_cast_fp16)[name = tensor("transpose_240")]; + tensor x_239_cast_fp16 = reshape(shape = concat_19, x = transpose_240)[name = tensor("x_239_cast_fp16")]; + tensor var_2228_to_fp16 = const()[name = tensor("op_2228_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(772214976)))]; + tensor var_2229_to_fp16 = const()[name = tensor("op_2229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775491840)))]; + tensor linear_117_cast_fp16 = linear(bias = var_2229_to_fp16, weight = var_2228_to_fp16, x = x_239_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = x_235_cast_fp16, y = linear_117_cast_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2236_axes_0 = const()[name = tensor("op_2236_axes_0"), val = tensor([-1])]; + tensor blocks_19_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775494464)))]; + tensor blocks_19_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775497088)))]; + tensor var_2236_cast_fp16 = layer_norm(axes = var_2236_axes_0, beta = blocks_19_mlp_ln_bias_to_fp16, epsilon = var_2161_to_fp16, gamma = blocks_19_mlp_ln_weight_to_fp16, x = x_241_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor var_2245_to_fp16 = const()[name = tensor("op_2245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(775499712)))]; + tensor var_2246_to_fp16 = const()[name = tensor("op_2246_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788606976)))]; + tensor linear_118_cast_fp16 = linear(bias = var_2246_to_fp16, weight = var_2245_to_fp16, x = var_2236_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor x_245_mode_0 = const()[name = tensor("x_245_mode_0"), val = tensor("EXACT")]; + tensor x_245_cast_fp16 = gelu(mode = x_245_mode_0, x = linear_118_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2251_to_fp16 = const()[name = tensor("op_2251_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788617280)))]; + tensor var_2252_to_fp16 = const()[name = tensor("op_2252_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801724544)))]; + tensor linear_119_cast_fp16 = linear(bias = var_2252_to_fp16, weight = var_2251_to_fp16, x = x_245_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = x_241_cast_fp16, y = linear_119_cast_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2262 = const()[name = tensor("op_2262"), val = tensor(-1)]; + tensor var_2279_axes_0 = const()[name = tensor("op_2279_axes_0"), val = tensor([-1])]; + tensor blocks_20_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_20_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801727168)))]; + tensor blocks_20_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_20_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801729792)))]; + tensor var_2268_to_fp16 = const()[name = tensor("op_2268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2279_cast_fp16 = layer_norm(axes = var_2279_axes_0, beta = blocks_20_attn_ln_bias_to_fp16, epsilon = var_2268_to_fp16, gamma = blocks_20_attn_ln_weight_to_fp16, x = x_247_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor var_2290_to_fp16 = const()[name = tensor("op_2290_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(801732416)))]; + tensor var_2291_to_fp16 = const()[name = tensor("op_2291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805009280)))]; + tensor linear_120_cast_fp16 = linear(bias = var_2291_to_fp16, weight = var_2290_to_fp16, x = var_2279_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor var_2294_to_fp16 = const()[name = tensor("op_2294_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(805011904)))]; + tensor linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2294_to_fp16, x = var_2279_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_2298_to_fp16 = const()[name = tensor("op_2298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(808288768)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811565632)))]; + tensor linear_122_cast_fp16 = linear(bias = var_2299_to_fp16, weight = var_2298_to_fp16, x = var_2279_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 1500, 20, -1])]; + tensor var_2308_cast_fp16 = reshape(shape = var_2307, x = linear_120_cast_fp16)[name = tensor("op_2308_cast_fp16")]; + tensor const_264_to_fp16 = const()[name = tensor("const_264_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_83_cast_fp16 = mul(x = var_2308_cast_fp16, y = const_264_to_fp16)[name = tensor("q_83_cast_fp16")]; + tensor var_2314 = const()[name = tensor("op_2314"), val = tensor([1, 1500, 20, -1])]; + tensor var_2315_cast_fp16 = reshape(shape = var_2314, x = linear_121_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor const_265_to_fp16 = const()[name = tensor("const_265_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_83_cast_fp16 = mul(x = var_2315_cast_fp16, y = const_265_to_fp16)[name = tensor("k_83_cast_fp16")]; + tensor var_2321 = const()[name = tensor("op_2321"), val = tensor([1, 1500, 20, -1])]; + tensor var_2322_cast_fp16 = reshape(shape = var_2321, x = linear_122_cast_fp16)[name = tensor("op_2322_cast_fp16")]; + tensor var_2323 = const()[name = tensor("op_2323"), val = tensor([0, 2, 1, 3])]; + tensor qk_41_transpose_x_0 = const()[name = tensor("qk_41_transpose_x_0"), val = tensor(false)]; + tensor qk_41_transpose_y_0 = const()[name = tensor("qk_41_transpose_y_0"), val = tensor(false)]; + tensor transpose_168_perm_0 = const()[name = tensor("transpose_168_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_169_perm_0 = const()[name = tensor("transpose_169_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_237 = transpose(perm = transpose_169_perm_0, x = k_83_cast_fp16)[name = tensor("transpose_237")]; + tensor transpose_238 = transpose(perm = transpose_168_perm_0, x = q_83_cast_fp16)[name = tensor("transpose_238")]; + tensor qk_41_cast_fp16 = matmul(transpose_x = qk_41_transpose_x_0, transpose_y = qk_41_transpose_y_0, x = transpose_238, y = transpose_237)[name = tensor("qk_41_cast_fp16")]; + tensor var_2327_cast_fp16 = softmax(axis = var_2262, x = qk_41_cast_fp16)[name = tensor("op_2327_cast_fp16")]; + tensor var_2329_transpose_x_0 = const()[name = tensor("op_2329_transpose_x_0"), val = tensor(false)]; + tensor var_2329_transpose_y_0 = const()[name = tensor("op_2329_transpose_y_0"), val = tensor(false)]; + tensor transpose_239 = transpose(perm = var_2323, x = var_2322_cast_fp16)[name = tensor("transpose_239")]; + tensor var_2329_cast_fp16 = matmul(transpose_x = var_2329_transpose_x_0, transpose_y = var_2329_transpose_y_0, x = var_2327_cast_fp16, y = transpose_239)[name = tensor("op_2329_cast_fp16")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 1500, 1280])]; + tensor transpose_236 = transpose(perm = var_2330, x = var_2329_cast_fp16)[name = tensor("transpose_236")]; + tensor x_251_cast_fp16 = reshape(shape = concat_20, x = transpose_236)[name = tensor("x_251_cast_fp16")]; + tensor var_2335_to_fp16 = const()[name = tensor("op_2335_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(811568256)))]; + tensor var_2336_to_fp16 = const()[name = tensor("op_2336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814845120)))]; + tensor linear_123_cast_fp16 = linear(bias = var_2336_to_fp16, weight = var_2335_to_fp16, x = x_251_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = x_247_cast_fp16, y = linear_123_cast_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2343_axes_0 = const()[name = tensor("op_2343_axes_0"), val = tensor([-1])]; + tensor blocks_20_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814847744)))]; + tensor blocks_20_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814850368)))]; + tensor var_2343_cast_fp16 = layer_norm(axes = var_2343_axes_0, beta = blocks_20_mlp_ln_bias_to_fp16, epsilon = var_2268_to_fp16, gamma = blocks_20_mlp_ln_weight_to_fp16, x = x_253_cast_fp16)[name = tensor("op_2343_cast_fp16")]; + tensor var_2352_to_fp16 = const()[name = tensor("op_2352_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814852992)))]; + tensor var_2353_to_fp16 = const()[name = tensor("op_2353_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827960256)))]; + tensor linear_124_cast_fp16 = linear(bias = var_2353_to_fp16, weight = var_2352_to_fp16, x = var_2343_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor x_257_mode_0 = const()[name = tensor("x_257_mode_0"), val = tensor("EXACT")]; + tensor x_257_cast_fp16 = gelu(mode = x_257_mode_0, x = linear_124_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2358_to_fp16 = const()[name = tensor("op_2358_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827970560)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841077824)))]; + tensor linear_125_cast_fp16 = linear(bias = var_2359_to_fp16, weight = var_2358_to_fp16, x = x_257_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = x_253_cast_fp16, y = linear_125_cast_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2369 = const()[name = tensor("op_2369"), val = tensor(-1)]; + tensor var_2386_axes_0 = const()[name = tensor("op_2386_axes_0"), val = tensor([-1])]; + tensor blocks_21_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_21_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841080448)))]; + tensor blocks_21_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_21_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841083072)))]; + tensor var_2375_to_fp16 = const()[name = tensor("op_2375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2386_cast_fp16 = layer_norm(axes = var_2386_axes_0, beta = blocks_21_attn_ln_bias_to_fp16, epsilon = var_2375_to_fp16, gamma = blocks_21_attn_ln_weight_to_fp16, x = x_259_cast_fp16)[name = tensor("op_2386_cast_fp16")]; + tensor var_2397_to_fp16 = const()[name = tensor("op_2397_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(841085696)))]; + tensor var_2398_to_fp16 = const()[name = tensor("op_2398_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844362560)))]; + tensor linear_126_cast_fp16 = linear(bias = var_2398_to_fp16, weight = var_2397_to_fp16, x = var_2386_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(844365184)))]; + tensor linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2401_to_fp16, x = var_2386_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_2405_to_fp16 = const()[name = tensor("op_2405_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(847642048)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850918912)))]; + tensor linear_128_cast_fp16 = linear(bias = var_2406_to_fp16, weight = var_2405_to_fp16, x = var_2386_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_2414 = const()[name = tensor("op_2414"), val = tensor([1, 1500, 20, -1])]; + tensor var_2415_cast_fp16 = reshape(shape = var_2414, x = linear_126_cast_fp16)[name = tensor("op_2415_cast_fp16")]; + tensor const_266_to_fp16 = const()[name = tensor("const_266_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_87_cast_fp16 = mul(x = var_2415_cast_fp16, y = const_266_to_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_2421 = const()[name = tensor("op_2421"), val = tensor([1, 1500, 20, -1])]; + tensor var_2422_cast_fp16 = reshape(shape = var_2421, x = linear_127_cast_fp16)[name = tensor("op_2422_cast_fp16")]; + tensor const_267_to_fp16 = const()[name = tensor("const_267_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_87_cast_fp16 = mul(x = var_2422_cast_fp16, y = const_267_to_fp16)[name = tensor("k_87_cast_fp16")]; + tensor var_2428 = const()[name = tensor("op_2428"), val = tensor([1, 1500, 20, -1])]; + tensor var_2429_cast_fp16 = reshape(shape = var_2428, x = linear_128_cast_fp16)[name = tensor("op_2429_cast_fp16")]; + tensor var_2430 = const()[name = tensor("op_2430"), val = tensor([0, 2, 1, 3])]; + tensor qk_43_transpose_x_0 = const()[name = tensor("qk_43_transpose_x_0"), val = tensor(false)]; + tensor qk_43_transpose_y_0 = const()[name = tensor("qk_43_transpose_y_0"), val = tensor(false)]; + tensor transpose_170_perm_0 = const()[name = tensor("transpose_170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_171_perm_0 = const()[name = tensor("transpose_171_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_233 = transpose(perm = transpose_171_perm_0, x = k_87_cast_fp16)[name = tensor("transpose_233")]; + tensor transpose_234 = transpose(perm = transpose_170_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_234")]; + tensor qk_43_cast_fp16 = matmul(transpose_x = qk_43_transpose_x_0, transpose_y = qk_43_transpose_y_0, x = transpose_234, y = transpose_233)[name = tensor("qk_43_cast_fp16")]; + tensor var_2434_cast_fp16 = softmax(axis = var_2369, x = qk_43_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2436_transpose_x_0 = const()[name = tensor("op_2436_transpose_x_0"), val = tensor(false)]; + tensor var_2436_transpose_y_0 = const()[name = tensor("op_2436_transpose_y_0"), val = tensor(false)]; + tensor transpose_235 = transpose(perm = var_2430, x = var_2429_cast_fp16)[name = tensor("transpose_235")]; + tensor var_2436_cast_fp16 = matmul(transpose_x = var_2436_transpose_x_0, transpose_y = var_2436_transpose_y_0, x = var_2434_cast_fp16, y = transpose_235)[name = tensor("op_2436_cast_fp16")]; + tensor var_2437 = const()[name = tensor("op_2437"), val = tensor([0, 2, 1, 3])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([1, 1500, 1280])]; + tensor transpose_232 = transpose(perm = var_2437, x = var_2436_cast_fp16)[name = tensor("transpose_232")]; + tensor x_263_cast_fp16 = reshape(shape = concat_21, x = transpose_232)[name = tensor("x_263_cast_fp16")]; + tensor var_2442_to_fp16 = const()[name = tensor("op_2442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850921536)))]; + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854198400)))]; + tensor linear_129_cast_fp16 = linear(bias = var_2443_to_fp16, weight = var_2442_to_fp16, x = x_263_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor x_265_cast_fp16 = add(x = x_259_cast_fp16, y = linear_129_cast_fp16)[name = tensor("x_265_cast_fp16")]; + tensor var_2450_axes_0 = const()[name = tensor("op_2450_axes_0"), val = tensor([-1])]; + tensor blocks_21_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854201024)))]; + tensor blocks_21_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854203648)))]; + tensor var_2450_cast_fp16 = layer_norm(axes = var_2450_axes_0, beta = blocks_21_mlp_ln_bias_to_fp16, epsilon = var_2375_to_fp16, gamma = blocks_21_mlp_ln_weight_to_fp16, x = x_265_cast_fp16)[name = tensor("op_2450_cast_fp16")]; + tensor var_2459_to_fp16 = const()[name = tensor("op_2459_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854206272)))]; + tensor var_2460_to_fp16 = const()[name = tensor("op_2460_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867313536)))]; + tensor linear_130_cast_fp16 = linear(bias = var_2460_to_fp16, weight = var_2459_to_fp16, x = var_2450_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor x_269_mode_0 = const()[name = tensor("x_269_mode_0"), val = tensor("EXACT")]; + tensor x_269_cast_fp16 = gelu(mode = x_269_mode_0, x = linear_130_cast_fp16)[name = tensor("x_269_cast_fp16")]; + tensor var_2465_to_fp16 = const()[name = tensor("op_2465_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(867323840)))]; + tensor var_2466_to_fp16 = const()[name = tensor("op_2466_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880431104)))]; + tensor linear_131_cast_fp16 = linear(bias = var_2466_to_fp16, weight = var_2465_to_fp16, x = x_269_cast_fp16)[name = tensor("linear_131_cast_fp16")]; + tensor x_271_cast_fp16 = add(x = x_265_cast_fp16, y = linear_131_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2476 = const()[name = tensor("op_2476"), val = tensor(-1)]; + tensor var_2493_axes_0 = const()[name = tensor("op_2493_axes_0"), val = tensor([-1])]; + tensor blocks_22_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_22_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880433728)))]; + tensor blocks_22_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_22_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880436352)))]; + tensor var_2482_to_fp16 = const()[name = tensor("op_2482_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2493_cast_fp16 = layer_norm(axes = var_2493_axes_0, beta = blocks_22_attn_ln_bias_to_fp16, epsilon = var_2482_to_fp16, gamma = blocks_22_attn_ln_weight_to_fp16, x = x_271_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor var_2504_to_fp16 = const()[name = tensor("op_2504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(880438976)))]; + tensor var_2505_to_fp16 = const()[name = tensor("op_2505_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883715840)))]; + tensor linear_132_cast_fp16 = linear(bias = var_2505_to_fp16, weight = var_2504_to_fp16, x = var_2493_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor var_2508_to_fp16 = const()[name = tensor("op_2508_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883718464)))]; + tensor linear_133_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2508_to_fp16, x = var_2493_cast_fp16)[name = tensor("linear_133_cast_fp16")]; + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886995328)))]; + tensor var_2513_to_fp16 = const()[name = tensor("op_2513_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890272192)))]; + tensor linear_134_cast_fp16 = linear(bias = var_2513_to_fp16, weight = var_2512_to_fp16, x = var_2493_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([1, 1500, 20, -1])]; + tensor var_2522_cast_fp16 = reshape(shape = var_2521, x = linear_132_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor const_268_to_fp16 = const()[name = tensor("const_268_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_91_cast_fp16 = mul(x = var_2522_cast_fp16, y = const_268_to_fp16)[name = tensor("q_91_cast_fp16")]; + tensor var_2528 = const()[name = tensor("op_2528"), val = tensor([1, 1500, 20, -1])]; + tensor var_2529_cast_fp16 = reshape(shape = var_2528, x = linear_133_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor const_269_to_fp16 = const()[name = tensor("const_269_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_91_cast_fp16 = mul(x = var_2529_cast_fp16, y = const_269_to_fp16)[name = tensor("k_91_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor([1, 1500, 20, -1])]; + tensor var_2536_cast_fp16 = reshape(shape = var_2535, x = linear_134_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([0, 2, 1, 3])]; + tensor qk_45_transpose_x_0 = const()[name = tensor("qk_45_transpose_x_0"), val = tensor(false)]; + tensor qk_45_transpose_y_0 = const()[name = tensor("qk_45_transpose_y_0"), val = tensor(false)]; + tensor transpose_172_perm_0 = const()[name = tensor("transpose_172_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_173_perm_0 = const()[name = tensor("transpose_173_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_229 = transpose(perm = transpose_173_perm_0, x = k_91_cast_fp16)[name = tensor("transpose_229")]; + tensor transpose_230 = transpose(perm = transpose_172_perm_0, x = q_91_cast_fp16)[name = tensor("transpose_230")]; + tensor qk_45_cast_fp16 = matmul(transpose_x = qk_45_transpose_x_0, transpose_y = qk_45_transpose_y_0, x = transpose_230, y = transpose_229)[name = tensor("qk_45_cast_fp16")]; + tensor var_2541_cast_fp16 = softmax(axis = var_2476, x = qk_45_cast_fp16)[name = tensor("op_2541_cast_fp16")]; + tensor var_2543_transpose_x_0 = const()[name = tensor("op_2543_transpose_x_0"), val = tensor(false)]; + tensor var_2543_transpose_y_0 = const()[name = tensor("op_2543_transpose_y_0"), val = tensor(false)]; + tensor transpose_231 = transpose(perm = var_2537, x = var_2536_cast_fp16)[name = tensor("transpose_231")]; + tensor var_2543_cast_fp16 = matmul(transpose_x = var_2543_transpose_x_0, transpose_y = var_2543_transpose_y_0, x = var_2541_cast_fp16, y = transpose_231)[name = tensor("op_2543_cast_fp16")]; + tensor var_2544 = const()[name = tensor("op_2544"), val = tensor([0, 2, 1, 3])]; + tensor concat_22 = const()[name = tensor("concat_22"), val = tensor([1, 1500, 1280])]; + tensor transpose_228 = transpose(perm = var_2544, x = var_2543_cast_fp16)[name = tensor("transpose_228")]; + tensor x_275_cast_fp16 = reshape(shape = concat_22, x = transpose_228)[name = tensor("x_275_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890274816)))]; + tensor var_2550_to_fp16 = const()[name = tensor("op_2550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893551680)))]; + tensor linear_135_cast_fp16 = linear(bias = var_2550_to_fp16, weight = var_2549_to_fp16, x = x_275_cast_fp16)[name = tensor("linear_135_cast_fp16")]; + tensor x_277_cast_fp16 = add(x = x_271_cast_fp16, y = linear_135_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2557_axes_0 = const()[name = tensor("op_2557_axes_0"), val = tensor([-1])]; + tensor blocks_22_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893554304)))]; + tensor blocks_22_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893556928)))]; + tensor var_2557_cast_fp16 = layer_norm(axes = var_2557_axes_0, beta = blocks_22_mlp_ln_bias_to_fp16, epsilon = var_2482_to_fp16, gamma = blocks_22_mlp_ln_weight_to_fp16, x = x_277_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor var_2566_to_fp16 = const()[name = tensor("op_2566_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893559552)))]; + tensor var_2567_to_fp16 = const()[name = tensor("op_2567_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906666816)))]; + tensor linear_136_cast_fp16 = linear(bias = var_2567_to_fp16, weight = var_2566_to_fp16, x = var_2557_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor x_281_mode_0 = const()[name = tensor("x_281_mode_0"), val = tensor("EXACT")]; + tensor x_281_cast_fp16 = gelu(mode = x_281_mode_0, x = linear_136_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor var_2572_to_fp16 = const()[name = tensor("op_2572_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(906677120)))]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919784384)))]; + tensor linear_137_cast_fp16 = linear(bias = var_2573_to_fp16, weight = var_2572_to_fp16, x = x_281_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor x_283_cast_fp16 = add(x = x_277_cast_fp16, y = linear_137_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor var_2600_axes_0 = const()[name = tensor("op_2600_axes_0"), val = tensor([-1])]; + tensor blocks_23_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_23_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919787008)))]; + tensor blocks_23_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_23_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919789632)))]; + tensor var_2589_to_fp16 = const()[name = tensor("op_2589_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2600_cast_fp16 = layer_norm(axes = var_2600_axes_0, beta = blocks_23_attn_ln_bias_to_fp16, epsilon = var_2589_to_fp16, gamma = blocks_23_attn_ln_weight_to_fp16, x = x_283_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2611_to_fp16 = const()[name = tensor("op_2611_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(919792256)))]; + tensor var_2612_to_fp16 = const()[name = tensor("op_2612_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(923069120)))]; + tensor linear_138_cast_fp16 = linear(bias = var_2612_to_fp16, weight = var_2611_to_fp16, x = var_2600_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2615_to_fp16 = const()[name = tensor("op_2615_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(923071744)))]; + tensor linear_139_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2615_to_fp16, x = var_2600_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2619_to_fp16 = const()[name = tensor("op_2619_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(926348608)))]; + tensor var_2620_to_fp16 = const()[name = tensor("op_2620_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929625472)))]; + tensor linear_140_cast_fp16 = linear(bias = var_2620_to_fp16, weight = var_2619_to_fp16, x = var_2600_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 1500, 20, -1])]; + tensor var_2629_cast_fp16 = reshape(shape = var_2628, x = linear_138_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor const_270_to_fp16 = const()[name = tensor("const_270_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_95_cast_fp16 = mul(x = var_2629_cast_fp16, y = const_270_to_fp16)[name = tensor("q_95_cast_fp16")]; + tensor var_2635 = const()[name = tensor("op_2635"), val = tensor([1, 1500, 20, -1])]; + tensor var_2636_cast_fp16 = reshape(shape = var_2635, x = linear_139_cast_fp16)[name = tensor("op_2636_cast_fp16")]; + tensor const_271_to_fp16 = const()[name = tensor("const_271_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_95_cast_fp16 = mul(x = var_2636_cast_fp16, y = const_271_to_fp16)[name = tensor("k_95_cast_fp16")]; + tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 1500, 20, -1])]; + tensor var_2643_cast_fp16 = reshape(shape = var_2642, x = linear_140_cast_fp16)[name = tensor("op_2643_cast_fp16")]; + tensor var_2644 = const()[name = tensor("op_2644"), val = tensor([0, 2, 1, 3])]; + tensor qk_47_transpose_x_0 = const()[name = tensor("qk_47_transpose_x_0"), val = tensor(false)]; + tensor qk_47_transpose_y_0 = const()[name = tensor("qk_47_transpose_y_0"), val = tensor(false)]; + tensor transpose_174_perm_0 = const()[name = tensor("transpose_174_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_175_perm_0 = const()[name = tensor("transpose_175_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_225 = transpose(perm = transpose_175_perm_0, x = k_95_cast_fp16)[name = tensor("transpose_225")]; + tensor transpose_226 = transpose(perm = transpose_174_perm_0, x = q_95_cast_fp16)[name = tensor("transpose_226")]; + tensor qk_47_cast_fp16 = matmul(transpose_x = qk_47_transpose_x_0, transpose_y = qk_47_transpose_y_0, x = transpose_226, y = transpose_225)[name = tensor("qk_47_cast_fp16")]; + tensor var_2648_cast_fp16 = softmax(axis = var_2583, x = qk_47_cast_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor var_2650_transpose_x_0 = const()[name = tensor("op_2650_transpose_x_0"), val = tensor(false)]; + tensor var_2650_transpose_y_0 = const()[name = tensor("op_2650_transpose_y_0"), val = tensor(false)]; + tensor transpose_227 = transpose(perm = var_2644, x = var_2643_cast_fp16)[name = tensor("transpose_227")]; + tensor var_2650_cast_fp16 = matmul(transpose_x = var_2650_transpose_x_0, transpose_y = var_2650_transpose_y_0, x = var_2648_cast_fp16, y = transpose_227)[name = tensor("op_2650_cast_fp16")]; + tensor var_2651 = const()[name = tensor("op_2651"), val = tensor([0, 2, 1, 3])]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([1, 1500, 1280])]; + tensor transpose_224 = transpose(perm = var_2651, x = var_2650_cast_fp16)[name = tensor("transpose_224")]; + tensor x_287_cast_fp16 = reshape(shape = concat_23, x = transpose_224)[name = tensor("x_287_cast_fp16")]; + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(929628096)))]; + tensor var_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932904960)))]; + tensor linear_141_cast_fp16 = linear(bias = var_2657_to_fp16, weight = var_2656_to_fp16, x = x_287_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor x_289_cast_fp16 = add(x = x_283_cast_fp16, y = linear_141_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2664_axes_0 = const()[name = tensor("op_2664_axes_0"), val = tensor([-1])]; + tensor blocks_23_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932907584)))]; + tensor blocks_23_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932910208)))]; + tensor var_2664_cast_fp16 = layer_norm(axes = var_2664_axes_0, beta = blocks_23_mlp_ln_bias_to_fp16, epsilon = var_2589_to_fp16, gamma = blocks_23_mlp_ln_weight_to_fp16, x = x_289_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor var_2673_to_fp16 = const()[name = tensor("op_2673_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(932912832)))]; + tensor var_2674_to_fp16 = const()[name = tensor("op_2674_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(946020096)))]; + tensor linear_142_cast_fp16 = linear(bias = var_2674_to_fp16, weight = var_2673_to_fp16, x = var_2664_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor x_293_mode_0 = const()[name = tensor("x_293_mode_0"), val = tensor("EXACT")]; + tensor x_293_cast_fp16 = gelu(mode = x_293_mode_0, x = linear_142_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor var_2679_to_fp16 = const()[name = tensor("op_2679_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(946030400)))]; + tensor var_2680_to_fp16 = const()[name = tensor("op_2680_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959137664)))]; + tensor linear_143_cast_fp16 = linear(bias = var_2680_to_fp16, weight = var_2679_to_fp16, x = x_293_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor x_295_cast_fp16 = add(x = x_289_cast_fp16, y = linear_143_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor(-1)]; + tensor var_2707_axes_0 = const()[name = tensor("op_2707_axes_0"), val = tensor([-1])]; + tensor blocks_24_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_24_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959140288)))]; + tensor blocks_24_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_24_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959142912)))]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2707_cast_fp16 = layer_norm(axes = var_2707_axes_0, beta = blocks_24_attn_ln_bias_to_fp16, epsilon = var_2696_to_fp16, gamma = blocks_24_attn_ln_weight_to_fp16, x = x_295_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(959145536)))]; + tensor var_2719_to_fp16 = const()[name = tensor("op_2719_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962422400)))]; + tensor linear_144_cast_fp16 = linear(bias = var_2719_to_fp16, weight = var_2718_to_fp16, x = var_2707_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2722_to_fp16 = const()[name = tensor("op_2722_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(962425024)))]; + tensor linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2722_to_fp16, x = var_2707_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2726_to_fp16 = const()[name = tensor("op_2726_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(965701888)))]; + tensor var_2727_to_fp16 = const()[name = tensor("op_2727_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968978752)))]; + tensor linear_146_cast_fp16 = linear(bias = var_2727_to_fp16, weight = var_2726_to_fp16, x = var_2707_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor var_2735 = const()[name = tensor("op_2735"), val = tensor([1, 1500, 20, -1])]; + tensor var_2736_cast_fp16 = reshape(shape = var_2735, x = linear_144_cast_fp16)[name = tensor("op_2736_cast_fp16")]; + tensor const_272_to_fp16 = const()[name = tensor("const_272_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_99_cast_fp16 = mul(x = var_2736_cast_fp16, y = const_272_to_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_2742 = const()[name = tensor("op_2742"), val = tensor([1, 1500, 20, -1])]; + tensor var_2743_cast_fp16 = reshape(shape = var_2742, x = linear_145_cast_fp16)[name = tensor("op_2743_cast_fp16")]; + tensor const_273_to_fp16 = const()[name = tensor("const_273_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_99_cast_fp16 = mul(x = var_2743_cast_fp16, y = const_273_to_fp16)[name = tensor("k_99_cast_fp16")]; + tensor var_2749 = const()[name = tensor("op_2749"), val = tensor([1, 1500, 20, -1])]; + tensor var_2750_cast_fp16 = reshape(shape = var_2749, x = linear_146_cast_fp16)[name = tensor("op_2750_cast_fp16")]; + tensor var_2751 = const()[name = tensor("op_2751"), val = tensor([0, 2, 1, 3])]; + tensor qk_49_transpose_x_0 = const()[name = tensor("qk_49_transpose_x_0"), val = tensor(false)]; + tensor qk_49_transpose_y_0 = const()[name = tensor("qk_49_transpose_y_0"), val = tensor(false)]; + tensor transpose_176_perm_0 = const()[name = tensor("transpose_176_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_177_perm_0 = const()[name = tensor("transpose_177_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_221 = transpose(perm = transpose_177_perm_0, x = k_99_cast_fp16)[name = tensor("transpose_221")]; + tensor transpose_222 = transpose(perm = transpose_176_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_222")]; + tensor qk_49_cast_fp16 = matmul(transpose_x = qk_49_transpose_x_0, transpose_y = qk_49_transpose_y_0, x = transpose_222, y = transpose_221)[name = tensor("qk_49_cast_fp16")]; + tensor var_2755_cast_fp16 = softmax(axis = var_2690, x = qk_49_cast_fp16)[name = tensor("op_2755_cast_fp16")]; + tensor var_2757_transpose_x_0 = const()[name = tensor("op_2757_transpose_x_0"), val = tensor(false)]; + tensor var_2757_transpose_y_0 = const()[name = tensor("op_2757_transpose_y_0"), val = tensor(false)]; + tensor transpose_223 = transpose(perm = var_2751, x = var_2750_cast_fp16)[name = tensor("transpose_223")]; + tensor var_2757_cast_fp16 = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = var_2755_cast_fp16, y = transpose_223)[name = tensor("op_2757_cast_fp16")]; + tensor var_2758 = const()[name = tensor("op_2758"), val = tensor([0, 2, 1, 3])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([1, 1500, 1280])]; + tensor transpose_220 = transpose(perm = var_2758, x = var_2757_cast_fp16)[name = tensor("transpose_220")]; + tensor x_299_cast_fp16 = reshape(shape = concat_24, x = transpose_220)[name = tensor("x_299_cast_fp16")]; + tensor var_2763_to_fp16 = const()[name = tensor("op_2763_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(968981376)))]; + tensor var_2764_to_fp16 = const()[name = tensor("op_2764_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972258240)))]; + tensor linear_147_cast_fp16 = linear(bias = var_2764_to_fp16, weight = var_2763_to_fp16, x = x_299_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor x_301_cast_fp16 = add(x = x_295_cast_fp16, y = linear_147_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2771_axes_0 = const()[name = tensor("op_2771_axes_0"), val = tensor([-1])]; + tensor blocks_24_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_24_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972260864)))]; + tensor blocks_24_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_24_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972263488)))]; + tensor var_2771_cast_fp16 = layer_norm(axes = var_2771_axes_0, beta = blocks_24_mlp_ln_bias_to_fp16, epsilon = var_2696_to_fp16, gamma = blocks_24_mlp_ln_weight_to_fp16, x = x_301_cast_fp16)[name = tensor("op_2771_cast_fp16")]; + tensor var_2780_to_fp16 = const()[name = tensor("op_2780_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(972266112)))]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985373376)))]; + tensor linear_148_cast_fp16 = linear(bias = var_2781_to_fp16, weight = var_2780_to_fp16, x = var_2771_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor x_305_mode_0 = const()[name = tensor("x_305_mode_0"), val = tensor("EXACT")]; + tensor x_305_cast_fp16 = gelu(mode = x_305_mode_0, x = linear_148_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor var_2786_to_fp16 = const()[name = tensor("op_2786_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985383680)))]; + tensor var_2787_to_fp16 = const()[name = tensor("op_2787_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998490944)))]; + tensor linear_149_cast_fp16 = linear(bias = var_2787_to_fp16, weight = var_2786_to_fp16, x = x_305_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor x_307_cast_fp16 = add(x = x_301_cast_fp16, y = linear_149_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2797 = const()[name = tensor("op_2797"), val = tensor(-1)]; + tensor var_2814_axes_0 = const()[name = tensor("op_2814_axes_0"), val = tensor([-1])]; + tensor blocks_25_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_25_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998493568)))]; + tensor blocks_25_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_25_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998496192)))]; + tensor var_2803_to_fp16 = const()[name = tensor("op_2803_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2814_cast_fp16 = layer_norm(axes = var_2814_axes_0, beta = blocks_25_attn_ln_bias_to_fp16, epsilon = var_2803_to_fp16, gamma = blocks_25_attn_ln_weight_to_fp16, x = x_307_cast_fp16)[name = tensor("op_2814_cast_fp16")]; + tensor var_2825_to_fp16 = const()[name = tensor("op_2825_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998498816)))]; + tensor var_2826_to_fp16 = const()[name = tensor("op_2826_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001775680)))]; + tensor linear_150_cast_fp16 = linear(bias = var_2826_to_fp16, weight = var_2825_to_fp16, x = var_2814_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2829_to_fp16 = const()[name = tensor("op_2829_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1001778304)))]; + tensor linear_151_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2829_to_fp16, x = var_2814_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2833_to_fp16 = const()[name = tensor("op_2833_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1005055168)))]; + tensor var_2834_to_fp16 = const()[name = tensor("op_2834_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008332032)))]; + tensor linear_152_cast_fp16 = linear(bias = var_2834_to_fp16, weight = var_2833_to_fp16, x = var_2814_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 1500, 20, -1])]; + tensor var_2843_cast_fp16 = reshape(shape = var_2842, x = linear_150_cast_fp16)[name = tensor("op_2843_cast_fp16")]; + tensor const_274_to_fp16 = const()[name = tensor("const_274_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_103_cast_fp16 = mul(x = var_2843_cast_fp16, y = const_274_to_fp16)[name = tensor("q_103_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1500, 20, -1])]; + tensor var_2850_cast_fp16 = reshape(shape = var_2849, x = linear_151_cast_fp16)[name = tensor("op_2850_cast_fp16")]; + tensor const_275_to_fp16 = const()[name = tensor("const_275_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_103_cast_fp16 = mul(x = var_2850_cast_fp16, y = const_275_to_fp16)[name = tensor("k_103_cast_fp16")]; + tensor var_2856 = const()[name = tensor("op_2856"), val = tensor([1, 1500, 20, -1])]; + tensor var_2857_cast_fp16 = reshape(shape = var_2856, x = linear_152_cast_fp16)[name = tensor("op_2857_cast_fp16")]; + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([0, 2, 1, 3])]; + tensor qk_51_transpose_x_0 = const()[name = tensor("qk_51_transpose_x_0"), val = tensor(false)]; + tensor qk_51_transpose_y_0 = const()[name = tensor("qk_51_transpose_y_0"), val = tensor(false)]; + tensor transpose_178_perm_0 = const()[name = tensor("transpose_178_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_179_perm_0 = const()[name = tensor("transpose_179_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_217 = transpose(perm = transpose_179_perm_0, x = k_103_cast_fp16)[name = tensor("transpose_217")]; + tensor transpose_218 = transpose(perm = transpose_178_perm_0, x = q_103_cast_fp16)[name = tensor("transpose_218")]; + tensor qk_51_cast_fp16 = matmul(transpose_x = qk_51_transpose_x_0, transpose_y = qk_51_transpose_y_0, x = transpose_218, y = transpose_217)[name = tensor("qk_51_cast_fp16")]; + tensor var_2862_cast_fp16 = softmax(axis = var_2797, x = qk_51_cast_fp16)[name = tensor("op_2862_cast_fp16")]; + tensor var_2864_transpose_x_0 = const()[name = tensor("op_2864_transpose_x_0"), val = tensor(false)]; + tensor var_2864_transpose_y_0 = const()[name = tensor("op_2864_transpose_y_0"), val = tensor(false)]; + tensor transpose_219 = transpose(perm = var_2858, x = var_2857_cast_fp16)[name = tensor("transpose_219")]; + tensor var_2864_cast_fp16 = matmul(transpose_x = var_2864_transpose_x_0, transpose_y = var_2864_transpose_y_0, x = var_2862_cast_fp16, y = transpose_219)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor([0, 2, 1, 3])]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 1500, 1280])]; + tensor transpose_216 = transpose(perm = var_2865, x = var_2864_cast_fp16)[name = tensor("transpose_216")]; + tensor x_311_cast_fp16 = reshape(shape = concat_25, x = transpose_216)[name = tensor("x_311_cast_fp16")]; + tensor var_2870_to_fp16 = const()[name = tensor("op_2870_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008334656)))]; + tensor var_2871_to_fp16 = const()[name = tensor("op_2871_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011611520)))]; + tensor linear_153_cast_fp16 = linear(bias = var_2871_to_fp16, weight = var_2870_to_fp16, x = x_311_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor x_313_cast_fp16 = add(x = x_307_cast_fp16, y = linear_153_cast_fp16)[name = tensor("x_313_cast_fp16")]; + tensor var_2878_axes_0 = const()[name = tensor("op_2878_axes_0"), val = tensor([-1])]; + tensor blocks_25_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_25_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011614144)))]; + tensor blocks_25_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_25_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011616768)))]; + tensor var_2878_cast_fp16 = layer_norm(axes = var_2878_axes_0, beta = blocks_25_mlp_ln_bias_to_fp16, epsilon = var_2803_to_fp16, gamma = blocks_25_mlp_ln_weight_to_fp16, x = x_313_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2887_to_fp16 = const()[name = tensor("op_2887_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1011619392)))]; + tensor var_2888_to_fp16 = const()[name = tensor("op_2888_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024726656)))]; + tensor linear_154_cast_fp16 = linear(bias = var_2888_to_fp16, weight = var_2887_to_fp16, x = var_2878_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor x_317_mode_0 = const()[name = tensor("x_317_mode_0"), val = tensor("EXACT")]; + tensor x_317_cast_fp16 = gelu(mode = x_317_mode_0, x = linear_154_cast_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2893_to_fp16 = const()[name = tensor("op_2893_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1024736960)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037844224)))]; + tensor linear_155_cast_fp16 = linear(bias = var_2894_to_fp16, weight = var_2893_to_fp16, x = x_317_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = x_313_cast_fp16, y = linear_155_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor var_2904 = const()[name = tensor("op_2904"), val = tensor(-1)]; + tensor var_2921_axes_0 = const()[name = tensor("op_2921_axes_0"), val = tensor([-1])]; + tensor blocks_26_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_26_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037846848)))]; + tensor blocks_26_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_26_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037849472)))]; + tensor var_2910_to_fp16 = const()[name = tensor("op_2910_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2921_cast_fp16 = layer_norm(axes = var_2921_axes_0, beta = blocks_26_attn_ln_bias_to_fp16, epsilon = var_2910_to_fp16, gamma = blocks_26_attn_ln_weight_to_fp16, x = x_319_cast_fp16)[name = tensor("op_2921_cast_fp16")]; + tensor var_2932_to_fp16 = const()[name = tensor("op_2932_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037852096)))]; + tensor var_2933_to_fp16 = const()[name = tensor("op_2933_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1041128960)))]; + tensor linear_156_cast_fp16 = linear(bias = var_2933_to_fp16, weight = var_2932_to_fp16, x = var_2921_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor var_2936_to_fp16 = const()[name = tensor("op_2936_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1041131584)))]; + tensor linear_157_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_2936_to_fp16, x = var_2921_cast_fp16)[name = tensor("linear_157_cast_fp16")]; + tensor var_2940_to_fp16 = const()[name = tensor("op_2940_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044408448)))]; + tensor var_2941_to_fp16 = const()[name = tensor("op_2941_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047685312)))]; + tensor linear_158_cast_fp16 = linear(bias = var_2941_to_fp16, weight = var_2940_to_fp16, x = var_2921_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor var_2949 = const()[name = tensor("op_2949"), val = tensor([1, 1500, 20, -1])]; + tensor var_2950_cast_fp16 = reshape(shape = var_2949, x = linear_156_cast_fp16)[name = tensor("op_2950_cast_fp16")]; + tensor const_276_to_fp16 = const()[name = tensor("const_276_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_107_cast_fp16 = mul(x = var_2950_cast_fp16, y = const_276_to_fp16)[name = tensor("q_107_cast_fp16")]; + tensor var_2956 = const()[name = tensor("op_2956"), val = tensor([1, 1500, 20, -1])]; + tensor var_2957_cast_fp16 = reshape(shape = var_2956, x = linear_157_cast_fp16)[name = tensor("op_2957_cast_fp16")]; + tensor const_277_to_fp16 = const()[name = tensor("const_277_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_107_cast_fp16 = mul(x = var_2957_cast_fp16, y = const_277_to_fp16)[name = tensor("k_107_cast_fp16")]; + tensor var_2963 = const()[name = tensor("op_2963"), val = tensor([1, 1500, 20, -1])]; + tensor var_2964_cast_fp16 = reshape(shape = var_2963, x = linear_158_cast_fp16)[name = tensor("op_2964_cast_fp16")]; + tensor var_2965 = const()[name = tensor("op_2965"), val = tensor([0, 2, 1, 3])]; + tensor qk_53_transpose_x_0 = const()[name = tensor("qk_53_transpose_x_0"), val = tensor(false)]; + tensor qk_53_transpose_y_0 = const()[name = tensor("qk_53_transpose_y_0"), val = tensor(false)]; + tensor transpose_180_perm_0 = const()[name = tensor("transpose_180_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_181_perm_0 = const()[name = tensor("transpose_181_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_213 = transpose(perm = transpose_181_perm_0, x = k_107_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_214 = transpose(perm = transpose_180_perm_0, x = q_107_cast_fp16)[name = tensor("transpose_214")]; + tensor qk_53_cast_fp16 = matmul(transpose_x = qk_53_transpose_x_0, transpose_y = qk_53_transpose_y_0, x = transpose_214, y = transpose_213)[name = tensor("qk_53_cast_fp16")]; + tensor var_2969_cast_fp16 = softmax(axis = var_2904, x = qk_53_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor var_2971_transpose_x_0 = const()[name = tensor("op_2971_transpose_x_0"), val = tensor(false)]; + tensor var_2971_transpose_y_0 = const()[name = tensor("op_2971_transpose_y_0"), val = tensor(false)]; + tensor transpose_215 = transpose(perm = var_2965, x = var_2964_cast_fp16)[name = tensor("transpose_215")]; + tensor var_2971_cast_fp16 = matmul(transpose_x = var_2971_transpose_x_0, transpose_y = var_2971_transpose_y_0, x = var_2969_cast_fp16, y = transpose_215)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972 = const()[name = tensor("op_2972"), val = tensor([0, 2, 1, 3])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([1, 1500, 1280])]; + tensor transpose_212 = transpose(perm = var_2972, x = var_2971_cast_fp16)[name = tensor("transpose_212")]; + tensor x_323_cast_fp16 = reshape(shape = concat_26, x = transpose_212)[name = tensor("x_323_cast_fp16")]; + tensor var_2977_to_fp16 = const()[name = tensor("op_2977_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1047687936)))]; + tensor var_2978_to_fp16 = const()[name = tensor("op_2978_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050964800)))]; + tensor linear_159_cast_fp16 = linear(bias = var_2978_to_fp16, weight = var_2977_to_fp16, x = x_323_cast_fp16)[name = tensor("linear_159_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = linear_159_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor var_2985_axes_0 = const()[name = tensor("op_2985_axes_0"), val = tensor([-1])]; + tensor blocks_26_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_26_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050967424)))]; + tensor blocks_26_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_26_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050970048)))]; + tensor var_2985_cast_fp16 = layer_norm(axes = var_2985_axes_0, beta = blocks_26_mlp_ln_bias_to_fp16, epsilon = var_2910_to_fp16, gamma = blocks_26_mlp_ln_weight_to_fp16, x = x_325_cast_fp16)[name = tensor("op_2985_cast_fp16")]; + tensor var_2994_to_fp16 = const()[name = tensor("op_2994_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050972672)))]; + tensor var_2995_to_fp16 = const()[name = tensor("op_2995_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1064079936)))]; + tensor linear_160_cast_fp16 = linear(bias = var_2995_to_fp16, weight = var_2994_to_fp16, x = var_2985_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor x_329_mode_0 = const()[name = tensor("x_329_mode_0"), val = tensor("EXACT")]; + tensor x_329_cast_fp16 = gelu(mode = x_329_mode_0, x = linear_160_cast_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1064090240)))]; + tensor var_3001_to_fp16 = const()[name = tensor("op_3001_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077197504)))]; + tensor linear_161_cast_fp16 = linear(bias = var_3001_to_fp16, weight = var_3000_to_fp16, x = x_329_cast_fp16)[name = tensor("linear_161_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = linear_161_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor var_3011 = const()[name = tensor("op_3011"), val = tensor(-1)]; + tensor var_3028_axes_0 = const()[name = tensor("op_3028_axes_0"), val = tensor([-1])]; + tensor blocks_27_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_27_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077200128)))]; + tensor blocks_27_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_27_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077202752)))]; + tensor var_3017_to_fp16 = const()[name = tensor("op_3017_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3028_cast_fp16 = layer_norm(axes = var_3028_axes_0, beta = blocks_27_attn_ln_bias_to_fp16, epsilon = var_3017_to_fp16, gamma = blocks_27_attn_ln_weight_to_fp16, x = x_331_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1077205376)))]; + tensor var_3040_to_fp16 = const()[name = tensor("op_3040_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080482240)))]; + tensor linear_162_cast_fp16 = linear(bias = var_3040_to_fp16, weight = var_3039_to_fp16, x = var_3028_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_3043_to_fp16 = const()[name = tensor("op_3043_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1080484864)))]; + tensor linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3043_to_fp16, x = var_3028_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor var_3047_to_fp16 = const()[name = tensor("op_3047_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1083761728)))]; + tensor var_3048_to_fp16 = const()[name = tensor("op_3048_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087038592)))]; + tensor linear_164_cast_fp16 = linear(bias = var_3048_to_fp16, weight = var_3047_to_fp16, x = var_3028_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_3056 = const()[name = tensor("op_3056"), val = tensor([1, 1500, 20, -1])]; + tensor var_3057_cast_fp16 = reshape(shape = var_3056, x = linear_162_cast_fp16)[name = tensor("op_3057_cast_fp16")]; + tensor const_278_to_fp16 = const()[name = tensor("const_278_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_111_cast_fp16 = mul(x = var_3057_cast_fp16, y = const_278_to_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_3063 = const()[name = tensor("op_3063"), val = tensor([1, 1500, 20, -1])]; + tensor var_3064_cast_fp16 = reshape(shape = var_3063, x = linear_163_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor const_279_to_fp16 = const()[name = tensor("const_279_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_111_cast_fp16 = mul(x = var_3064_cast_fp16, y = const_279_to_fp16)[name = tensor("k_111_cast_fp16")]; + tensor var_3070 = const()[name = tensor("op_3070"), val = tensor([1, 1500, 20, -1])]; + tensor var_3071_cast_fp16 = reshape(shape = var_3070, x = linear_164_cast_fp16)[name = tensor("op_3071_cast_fp16")]; + tensor var_3072 = const()[name = tensor("op_3072"), val = tensor([0, 2, 1, 3])]; + tensor qk_55_transpose_x_0 = const()[name = tensor("qk_55_transpose_x_0"), val = tensor(false)]; + tensor qk_55_transpose_y_0 = const()[name = tensor("qk_55_transpose_y_0"), val = tensor(false)]; + tensor transpose_182_perm_0 = const()[name = tensor("transpose_182_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_183_perm_0 = const()[name = tensor("transpose_183_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_209 = transpose(perm = transpose_183_perm_0, x = k_111_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_210 = transpose(perm = transpose_182_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_210")]; + tensor qk_55_cast_fp16 = matmul(transpose_x = qk_55_transpose_x_0, transpose_y = qk_55_transpose_y_0, x = transpose_210, y = transpose_209)[name = tensor("qk_55_cast_fp16")]; + tensor var_3076_cast_fp16 = softmax(axis = var_3011, x = qk_55_cast_fp16)[name = tensor("op_3076_cast_fp16")]; + tensor var_3078_transpose_x_0 = const()[name = tensor("op_3078_transpose_x_0"), val = tensor(false)]; + tensor var_3078_transpose_y_0 = const()[name = tensor("op_3078_transpose_y_0"), val = tensor(false)]; + tensor transpose_211 = transpose(perm = var_3072, x = var_3071_cast_fp16)[name = tensor("transpose_211")]; + tensor var_3078_cast_fp16 = matmul(transpose_x = var_3078_transpose_x_0, transpose_y = var_3078_transpose_y_0, x = var_3076_cast_fp16, y = transpose_211)[name = tensor("op_3078_cast_fp16")]; + tensor var_3079 = const()[name = tensor("op_3079"), val = tensor([0, 2, 1, 3])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([1, 1500, 1280])]; + tensor transpose_208 = transpose(perm = var_3079, x = var_3078_cast_fp16)[name = tensor("transpose_208")]; + tensor x_335_cast_fp16 = reshape(shape = concat_27, x = transpose_208)[name = tensor("x_335_cast_fp16")]; + tensor var_3084_to_fp16 = const()[name = tensor("op_3084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087041216)))]; + tensor var_3085_to_fp16 = const()[name = tensor("op_3085_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090318080)))]; + tensor linear_165_cast_fp16 = linear(bias = var_3085_to_fp16, weight = var_3084_to_fp16, x = x_335_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = linear_165_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor var_3092_axes_0 = const()[name = tensor("op_3092_axes_0"), val = tensor([-1])]; + tensor blocks_27_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_27_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090320704)))]; + tensor blocks_27_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_27_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090323328)))]; + tensor var_3092_cast_fp16 = layer_norm(axes = var_3092_axes_0, beta = blocks_27_mlp_ln_bias_to_fp16, epsilon = var_3017_to_fp16, gamma = blocks_27_mlp_ln_weight_to_fp16, x = x_337_cast_fp16)[name = tensor("op_3092_cast_fp16")]; + tensor var_3101_to_fp16 = const()[name = tensor("op_3101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090325952)))]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103433216)))]; + tensor linear_166_cast_fp16 = linear(bias = var_3102_to_fp16, weight = var_3101_to_fp16, x = var_3092_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor x_341_mode_0 = const()[name = tensor("x_341_mode_0"), val = tensor("EXACT")]; + tensor x_341_cast_fp16 = gelu(mode = x_341_mode_0, x = linear_166_cast_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3107_to_fp16 = const()[name = tensor("op_3107_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1103443520)))]; + tensor var_3108_to_fp16 = const()[name = tensor("op_3108_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116550784)))]; + tensor linear_167_cast_fp16 = linear(bias = var_3108_to_fp16, weight = var_3107_to_fp16, x = x_341_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = linear_167_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor var_3118 = const()[name = tensor("op_3118"), val = tensor(-1)]; + tensor var_3135_axes_0 = const()[name = tensor("op_3135_axes_0"), val = tensor([-1])]; + tensor blocks_28_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_28_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116553408)))]; + tensor blocks_28_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_28_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116556032)))]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3135_cast_fp16 = layer_norm(axes = var_3135_axes_0, beta = blocks_28_attn_ln_bias_to_fp16, epsilon = var_3124_to_fp16, gamma = blocks_28_attn_ln_weight_to_fp16, x = x_343_cast_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3146_to_fp16 = const()[name = tensor("op_3146_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1116558656)))]; + tensor var_3147_to_fp16 = const()[name = tensor("op_3147_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119835520)))]; + tensor linear_168_cast_fp16 = linear(bias = var_3147_to_fp16, weight = var_3146_to_fp16, x = var_3135_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_3150_to_fp16 = const()[name = tensor("op_3150_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1119838144)))]; + tensor linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3150_to_fp16, x = var_3135_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor var_3154_to_fp16 = const()[name = tensor("op_3154_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1123115008)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126391872)))]; + tensor linear_170_cast_fp16 = linear(bias = var_3155_to_fp16, weight = var_3154_to_fp16, x = var_3135_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 1500, 20, -1])]; + tensor var_3164_cast_fp16 = reshape(shape = var_3163, x = linear_168_cast_fp16)[name = tensor("op_3164_cast_fp16")]; + tensor const_280_to_fp16 = const()[name = tensor("const_280_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_115_cast_fp16 = mul(x = var_3164_cast_fp16, y = const_280_to_fp16)[name = tensor("q_115_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 1500, 20, -1])]; + tensor var_3171_cast_fp16 = reshape(shape = var_3170, x = linear_169_cast_fp16)[name = tensor("op_3171_cast_fp16")]; + tensor const_281_to_fp16 = const()[name = tensor("const_281_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_115_cast_fp16 = mul(x = var_3171_cast_fp16, y = const_281_to_fp16)[name = tensor("k_115_cast_fp16")]; + tensor var_3177 = const()[name = tensor("op_3177"), val = tensor([1, 1500, 20, -1])]; + tensor var_3178_cast_fp16 = reshape(shape = var_3177, x = linear_170_cast_fp16)[name = tensor("op_3178_cast_fp16")]; + tensor var_3179 = const()[name = tensor("op_3179"), val = tensor([0, 2, 1, 3])]; + tensor qk_57_transpose_x_0 = const()[name = tensor("qk_57_transpose_x_0"), val = tensor(false)]; + tensor qk_57_transpose_y_0 = const()[name = tensor("qk_57_transpose_y_0"), val = tensor(false)]; + tensor transpose_184_perm_0 = const()[name = tensor("transpose_184_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_185_perm_0 = const()[name = tensor("transpose_185_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_205 = transpose(perm = transpose_185_perm_0, x = k_115_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_206 = transpose(perm = transpose_184_perm_0, x = q_115_cast_fp16)[name = tensor("transpose_206")]; + tensor qk_57_cast_fp16 = matmul(transpose_x = qk_57_transpose_x_0, transpose_y = qk_57_transpose_y_0, x = transpose_206, y = transpose_205)[name = tensor("qk_57_cast_fp16")]; + tensor var_3183_cast_fp16 = softmax(axis = var_3118, x = qk_57_cast_fp16)[name = tensor("op_3183_cast_fp16")]; + tensor var_3185_transpose_x_0 = const()[name = tensor("op_3185_transpose_x_0"), val = tensor(false)]; + tensor var_3185_transpose_y_0 = const()[name = tensor("op_3185_transpose_y_0"), val = tensor(false)]; + tensor transpose_207 = transpose(perm = var_3179, x = var_3178_cast_fp16)[name = tensor("transpose_207")]; + tensor var_3185_cast_fp16 = matmul(transpose_x = var_3185_transpose_x_0, transpose_y = var_3185_transpose_y_0, x = var_3183_cast_fp16, y = transpose_207)[name = tensor("op_3185_cast_fp16")]; + tensor var_3186 = const()[name = tensor("op_3186"), val = tensor([0, 2, 1, 3])]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 1500, 1280])]; + tensor transpose_204 = transpose(perm = var_3186, x = var_3185_cast_fp16)[name = tensor("transpose_204")]; + tensor x_347_cast_fp16 = reshape(shape = concat_28, x = transpose_204)[name = tensor("x_347_cast_fp16")]; + tensor var_3191_to_fp16 = const()[name = tensor("op_3191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126394496)))]; + tensor var_3192_to_fp16 = const()[name = tensor("op_3192_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129671360)))]; + tensor linear_171_cast_fp16 = linear(bias = var_3192_to_fp16, weight = var_3191_to_fp16, x = x_347_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = linear_171_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor var_3199_axes_0 = const()[name = tensor("op_3199_axes_0"), val = tensor([-1])]; + tensor blocks_28_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_28_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129673984)))]; + tensor blocks_28_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_28_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129676608)))]; + tensor var_3199_cast_fp16 = layer_norm(axes = var_3199_axes_0, beta = blocks_28_mlp_ln_bias_to_fp16, epsilon = var_3124_to_fp16, gamma = blocks_28_mlp_ln_weight_to_fp16, x = x_349_cast_fp16)[name = tensor("op_3199_cast_fp16")]; + tensor var_3208_to_fp16 = const()[name = tensor("op_3208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1129679232)))]; + tensor var_3209_to_fp16 = const()[name = tensor("op_3209_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142786496)))]; + tensor linear_172_cast_fp16 = linear(bias = var_3209_to_fp16, weight = var_3208_to_fp16, x = var_3199_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor x_353_mode_0 = const()[name = tensor("x_353_mode_0"), val = tensor("EXACT")]; + tensor x_353_cast_fp16 = gelu(mode = x_353_mode_0, x = linear_172_cast_fp16)[name = tensor("x_353_cast_fp16")]; + tensor var_3214_to_fp16 = const()[name = tensor("op_3214_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142796800)))]; + tensor var_3215_to_fp16 = const()[name = tensor("op_3215_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155904064)))]; + tensor linear_173_cast_fp16 = linear(bias = var_3215_to_fp16, weight = var_3214_to_fp16, x = x_353_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor x_355_cast_fp16 = add(x = x_349_cast_fp16, y = linear_173_cast_fp16)[name = tensor("x_355_cast_fp16")]; + tensor var_3225 = const()[name = tensor("op_3225"), val = tensor(-1)]; + tensor var_3242_axes_0 = const()[name = tensor("op_3242_axes_0"), val = tensor([-1])]; + tensor blocks_29_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_29_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155906688)))]; + tensor blocks_29_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_29_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155909312)))]; + tensor var_3231_to_fp16 = const()[name = tensor("op_3231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3242_cast_fp16 = layer_norm(axes = var_3242_axes_0, beta = blocks_29_attn_ln_bias_to_fp16, epsilon = var_3231_to_fp16, gamma = blocks_29_attn_ln_weight_to_fp16, x = x_355_cast_fp16)[name = tensor("op_3242_cast_fp16")]; + tensor var_3253_to_fp16 = const()[name = tensor("op_3253_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1155911936)))]; + tensor var_3254_to_fp16 = const()[name = tensor("op_3254_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159188800)))]; + tensor linear_174_cast_fp16 = linear(bias = var_3254_to_fp16, weight = var_3253_to_fp16, x = var_3242_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_3257_to_fp16 = const()[name = tensor("op_3257_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159191424)))]; + tensor linear_175_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3257_to_fp16, x = var_3242_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor var_3261_to_fp16 = const()[name = tensor("op_3261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1162468288)))]; + tensor var_3262_to_fp16 = const()[name = tensor("op_3262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165745152)))]; + tensor linear_176_cast_fp16 = linear(bias = var_3262_to_fp16, weight = var_3261_to_fp16, x = var_3242_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_3270 = const()[name = tensor("op_3270"), val = tensor([1, 1500, 20, -1])]; + tensor var_3271_cast_fp16 = reshape(shape = var_3270, x = linear_174_cast_fp16)[name = tensor("op_3271_cast_fp16")]; + tensor const_282_to_fp16 = const()[name = tensor("const_282_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_119_cast_fp16 = mul(x = var_3271_cast_fp16, y = const_282_to_fp16)[name = tensor("q_119_cast_fp16")]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1500, 20, -1])]; + tensor var_3278_cast_fp16 = reshape(shape = var_3277, x = linear_175_cast_fp16)[name = tensor("op_3278_cast_fp16")]; + tensor const_283_to_fp16 = const()[name = tensor("const_283_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_119_cast_fp16 = mul(x = var_3278_cast_fp16, y = const_283_to_fp16)[name = tensor("k_119_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([1, 1500, 20, -1])]; + tensor var_3285_cast_fp16 = reshape(shape = var_3284, x = linear_176_cast_fp16)[name = tensor("op_3285_cast_fp16")]; + tensor var_3286 = const()[name = tensor("op_3286"), val = tensor([0, 2, 1, 3])]; + tensor qk_59_transpose_x_0 = const()[name = tensor("qk_59_transpose_x_0"), val = tensor(false)]; + tensor qk_59_transpose_y_0 = const()[name = tensor("qk_59_transpose_y_0"), val = tensor(false)]; + tensor transpose_186_perm_0 = const()[name = tensor("transpose_186_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_187_perm_0 = const()[name = tensor("transpose_187_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_201 = transpose(perm = transpose_187_perm_0, x = k_119_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_202 = transpose(perm = transpose_186_perm_0, x = q_119_cast_fp16)[name = tensor("transpose_202")]; + tensor qk_59_cast_fp16 = matmul(transpose_x = qk_59_transpose_x_0, transpose_y = qk_59_transpose_y_0, x = transpose_202, y = transpose_201)[name = tensor("qk_59_cast_fp16")]; + tensor var_3290_cast_fp16 = softmax(axis = var_3225, x = qk_59_cast_fp16)[name = tensor("op_3290_cast_fp16")]; + tensor var_3292_transpose_x_0 = const()[name = tensor("op_3292_transpose_x_0"), val = tensor(false)]; + tensor var_3292_transpose_y_0 = const()[name = tensor("op_3292_transpose_y_0"), val = tensor(false)]; + tensor transpose_203 = transpose(perm = var_3286, x = var_3285_cast_fp16)[name = tensor("transpose_203")]; + tensor var_3292_cast_fp16 = matmul(transpose_x = var_3292_transpose_x_0, transpose_y = var_3292_transpose_y_0, x = var_3290_cast_fp16, y = transpose_203)[name = tensor("op_3292_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([0, 2, 1, 3])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([1, 1500, 1280])]; + tensor transpose_200 = transpose(perm = var_3293, x = var_3292_cast_fp16)[name = tensor("transpose_200")]; + tensor x_359_cast_fp16 = reshape(shape = concat_29, x = transpose_200)[name = tensor("x_359_cast_fp16")]; + tensor var_3298_to_fp16 = const()[name = tensor("op_3298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1165747776)))]; + tensor var_3299_to_fp16 = const()[name = tensor("op_3299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169024640)))]; + tensor linear_177_cast_fp16 = linear(bias = var_3299_to_fp16, weight = var_3298_to_fp16, x = x_359_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor x_361_cast_fp16 = add(x = x_355_cast_fp16, y = linear_177_cast_fp16)[name = tensor("x_361_cast_fp16")]; + tensor var_3306_axes_0 = const()[name = tensor("op_3306_axes_0"), val = tensor([-1])]; + tensor blocks_29_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_29_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169027264)))]; + tensor blocks_29_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_29_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169029888)))]; + tensor var_3306_cast_fp16 = layer_norm(axes = var_3306_axes_0, beta = blocks_29_mlp_ln_bias_to_fp16, epsilon = var_3231_to_fp16, gamma = blocks_29_mlp_ln_weight_to_fp16, x = x_361_cast_fp16)[name = tensor("op_3306_cast_fp16")]; + tensor var_3315_to_fp16 = const()[name = tensor("op_3315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1169032512)))]; + tensor var_3316_to_fp16 = const()[name = tensor("op_3316_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1182139776)))]; + tensor linear_178_cast_fp16 = linear(bias = var_3316_to_fp16, weight = var_3315_to_fp16, x = var_3306_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor x_365_mode_0 = const()[name = tensor("x_365_mode_0"), val = tensor("EXACT")]; + tensor x_365_cast_fp16 = gelu(mode = x_365_mode_0, x = linear_178_cast_fp16)[name = tensor("x_365_cast_fp16")]; + tensor var_3321_to_fp16 = const()[name = tensor("op_3321_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1182150080)))]; + tensor var_3322_to_fp16 = const()[name = tensor("op_3322_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195257344)))]; + tensor linear_179_cast_fp16 = linear(bias = var_3322_to_fp16, weight = var_3321_to_fp16, x = x_365_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor x_367_cast_fp16 = add(x = x_361_cast_fp16, y = linear_179_cast_fp16)[name = tensor("x_367_cast_fp16")]; + tensor var_3332 = const()[name = tensor("op_3332"), val = tensor(-1)]; + tensor var_3349_axes_0 = const()[name = tensor("op_3349_axes_0"), val = tensor([-1])]; + tensor blocks_30_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_30_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195259968)))]; + tensor blocks_30_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_30_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195262592)))]; + tensor var_3338_to_fp16 = const()[name = tensor("op_3338_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3349_cast_fp16 = layer_norm(axes = var_3349_axes_0, beta = blocks_30_attn_ln_bias_to_fp16, epsilon = var_3338_to_fp16, gamma = blocks_30_attn_ln_weight_to_fp16, x = x_367_cast_fp16)[name = tensor("op_3349_cast_fp16")]; + tensor var_3360_to_fp16 = const()[name = tensor("op_3360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195265216)))]; + tensor var_3361_to_fp16 = const()[name = tensor("op_3361_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198542080)))]; + tensor linear_180_cast_fp16 = linear(bias = var_3361_to_fp16, weight = var_3360_to_fp16, x = var_3349_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_3364_to_fp16 = const()[name = tensor("op_3364_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1198544704)))]; + tensor linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3364_to_fp16, x = var_3349_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_3368_to_fp16 = const()[name = tensor("op_3368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201821568)))]; + tensor var_3369_to_fp16 = const()[name = tensor("op_3369_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1205098432)))]; + tensor linear_182_cast_fp16 = linear(bias = var_3369_to_fp16, weight = var_3368_to_fp16, x = var_3349_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor var_3377 = const()[name = tensor("op_3377"), val = tensor([1, 1500, 20, -1])]; + tensor var_3378_cast_fp16 = reshape(shape = var_3377, x = linear_180_cast_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor const_284_to_fp16 = const()[name = tensor("const_284_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_123_cast_fp16 = mul(x = var_3378_cast_fp16, y = const_284_to_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_3384 = const()[name = tensor("op_3384"), val = tensor([1, 1500, 20, -1])]; + tensor var_3385_cast_fp16 = reshape(shape = var_3384, x = linear_181_cast_fp16)[name = tensor("op_3385_cast_fp16")]; + tensor const_285_to_fp16 = const()[name = tensor("const_285_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_123_cast_fp16 = mul(x = var_3385_cast_fp16, y = const_285_to_fp16)[name = tensor("k_123_cast_fp16")]; + tensor var_3391 = const()[name = tensor("op_3391"), val = tensor([1, 1500, 20, -1])]; + tensor var_3392_cast_fp16 = reshape(shape = var_3391, x = linear_182_cast_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor var_3393 = const()[name = tensor("op_3393"), val = tensor([0, 2, 1, 3])]; + tensor qk_61_transpose_x_0 = const()[name = tensor("qk_61_transpose_x_0"), val = tensor(false)]; + tensor qk_61_transpose_y_0 = const()[name = tensor("qk_61_transpose_y_0"), val = tensor(false)]; + tensor transpose_188_perm_0 = const()[name = tensor("transpose_188_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_189_perm_0 = const()[name = tensor("transpose_189_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_197 = transpose(perm = transpose_189_perm_0, x = k_123_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_198 = transpose(perm = transpose_188_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_198")]; + tensor qk_61_cast_fp16 = matmul(transpose_x = qk_61_transpose_x_0, transpose_y = qk_61_transpose_y_0, x = transpose_198, y = transpose_197)[name = tensor("qk_61_cast_fp16")]; + tensor var_3397_cast_fp16 = softmax(axis = var_3332, x = qk_61_cast_fp16)[name = tensor("op_3397_cast_fp16")]; + tensor var_3399_transpose_x_0 = const()[name = tensor("op_3399_transpose_x_0"), val = tensor(false)]; + tensor var_3399_transpose_y_0 = const()[name = tensor("op_3399_transpose_y_0"), val = tensor(false)]; + tensor transpose_199 = transpose(perm = var_3393, x = var_3392_cast_fp16)[name = tensor("transpose_199")]; + tensor var_3399_cast_fp16 = matmul(transpose_x = var_3399_transpose_x_0, transpose_y = var_3399_transpose_y_0, x = var_3397_cast_fp16, y = transpose_199)[name = tensor("op_3399_cast_fp16")]; + tensor var_3400 = const()[name = tensor("op_3400"), val = tensor([0, 2, 1, 3])]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([1, 1500, 1280])]; + tensor transpose_196 = transpose(perm = var_3400, x = var_3399_cast_fp16)[name = tensor("transpose_196")]; + tensor x_371_cast_fp16 = reshape(shape = concat_30, x = transpose_196)[name = tensor("x_371_cast_fp16")]; + tensor var_3405_to_fp16 = const()[name = tensor("op_3405_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1205101056)))]; + tensor var_3406_to_fp16 = const()[name = tensor("op_3406_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208377920)))]; + tensor linear_183_cast_fp16 = linear(bias = var_3406_to_fp16, weight = var_3405_to_fp16, x = x_371_cast_fp16)[name = tensor("linear_183_cast_fp16")]; + tensor x_373_cast_fp16 = add(x = x_367_cast_fp16, y = linear_183_cast_fp16)[name = tensor("x_373_cast_fp16")]; + tensor var_3413_axes_0 = const()[name = tensor("op_3413_axes_0"), val = tensor([-1])]; + tensor blocks_30_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_30_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208380544)))]; + tensor blocks_30_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_30_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208383168)))]; + tensor var_3413_cast_fp16 = layer_norm(axes = var_3413_axes_0, beta = blocks_30_mlp_ln_bias_to_fp16, epsilon = var_3338_to_fp16, gamma = blocks_30_mlp_ln_weight_to_fp16, x = x_373_cast_fp16)[name = tensor("op_3413_cast_fp16")]; + tensor var_3422_to_fp16 = const()[name = tensor("op_3422_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1208385792)))]; + tensor var_3423_to_fp16 = const()[name = tensor("op_3423_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221493056)))]; + tensor linear_184_cast_fp16 = linear(bias = var_3423_to_fp16, weight = var_3422_to_fp16, x = var_3413_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor x_377_mode_0 = const()[name = tensor("x_377_mode_0"), val = tensor("EXACT")]; + tensor x_377_cast_fp16 = gelu(mode = x_377_mode_0, x = linear_184_cast_fp16)[name = tensor("x_377_cast_fp16")]; + tensor var_3428_to_fp16 = const()[name = tensor("op_3428_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1221503360)))]; + tensor var_3429_to_fp16 = const()[name = tensor("op_3429_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234610624)))]; + tensor linear_185_cast_fp16 = linear(bias = var_3429_to_fp16, weight = var_3428_to_fp16, x = x_377_cast_fp16)[name = tensor("linear_185_cast_fp16")]; + tensor x_379_cast_fp16 = add(x = x_373_cast_fp16, y = linear_185_cast_fp16)[name = tensor("x_379_cast_fp16")]; + tensor var_3439 = const()[name = tensor("op_3439"), val = tensor(-1)]; + tensor var_3456_axes_0 = const()[name = tensor("op_3456_axes_0"), val = tensor([-1])]; + tensor blocks_31_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_31_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234613248)))]; + tensor blocks_31_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_31_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234615872)))]; + tensor var_3445_to_fp16 = const()[name = tensor("op_3445_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3456_cast_fp16 = layer_norm(axes = var_3456_axes_0, beta = blocks_31_attn_ln_bias_to_fp16, epsilon = var_3445_to_fp16, gamma = blocks_31_attn_ln_weight_to_fp16, x = x_379_cast_fp16)[name = tensor("op_3456_cast_fp16")]; + tensor var_3467_to_fp16 = const()[name = tensor("op_3467_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1234618496)))]; + tensor var_3468_to_fp16 = const()[name = tensor("op_3468_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237895360)))]; + tensor linear_186_cast_fp16 = linear(bias = var_3468_to_fp16, weight = var_3467_to_fp16, x = var_3456_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor var_3471_to_fp16 = const()[name = tensor("op_3471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1237897984)))]; + tensor linear_187_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = var_3471_to_fp16, x = var_3456_cast_fp16)[name = tensor("linear_187_cast_fp16")]; + tensor var_3475_to_fp16 = const()[name = tensor("op_3475_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1241174848)))]; + tensor var_3476_to_fp16 = const()[name = tensor("op_3476_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244451712)))]; + tensor linear_188_cast_fp16 = linear(bias = var_3476_to_fp16, weight = var_3475_to_fp16, x = var_3456_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_3484 = const()[name = tensor("op_3484"), val = tensor([1, 1500, 20, -1])]; + tensor var_3485_cast_fp16 = reshape(shape = var_3484, x = linear_186_cast_fp16)[name = tensor("op_3485_cast_fp16")]; + tensor const_286_to_fp16 = const()[name = tensor("const_286_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast_fp16 = mul(x = var_3485_cast_fp16, y = const_286_to_fp16)[name = tensor("q_cast_fp16")]; + tensor var_3491 = const()[name = tensor("op_3491"), val = tensor([1, 1500, 20, -1])]; + tensor var_3492_cast_fp16 = reshape(shape = var_3491, x = linear_187_cast_fp16)[name = tensor("op_3492_cast_fp16")]; + tensor const_287_to_fp16 = const()[name = tensor("const_287_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast_fp16 = mul(x = var_3492_cast_fp16, y = const_287_to_fp16)[name = tensor("k_cast_fp16")]; + tensor var_3498 = const()[name = tensor("op_3498"), val = tensor([1, 1500, 20, -1])]; + tensor var_3499_cast_fp16 = reshape(shape = var_3498, x = linear_188_cast_fp16)[name = tensor("op_3499_cast_fp16")]; + tensor var_3500 = const()[name = tensor("op_3500"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_190_perm_0 = const()[name = tensor("transpose_190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_191_perm_0 = const()[name = tensor("transpose_191_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_193 = transpose(perm = transpose_191_perm_0, x = k_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_194 = transpose(perm = transpose_190_perm_0, x = q_cast_fp16)[name = tensor("transpose_194")]; + tensor qk_cast_fp16 = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_194, y = transpose_193)[name = tensor("qk_cast_fp16")]; + tensor var_3504_cast_fp16 = softmax(axis = var_3439, x = qk_cast_fp16)[name = tensor("op_3504_cast_fp16")]; + tensor var_3506_transpose_x_0 = const()[name = tensor("op_3506_transpose_x_0"), val = tensor(false)]; + tensor var_3506_transpose_y_0 = const()[name = tensor("op_3506_transpose_y_0"), val = tensor(false)]; + tensor transpose_195 = transpose(perm = var_3500, x = var_3499_cast_fp16)[name = tensor("transpose_195")]; + tensor var_3506_cast_fp16 = matmul(transpose_x = var_3506_transpose_x_0, transpose_y = var_3506_transpose_y_0, x = var_3504_cast_fp16, y = transpose_195)[name = tensor("op_3506_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor([0, 2, 1, 3])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([1, 1500, 1280])]; + tensor transpose_192 = transpose(perm = var_3507, x = var_3506_cast_fp16)[name = tensor("transpose_192")]; + tensor x_383_cast_fp16 = reshape(shape = concat_31, x = transpose_192)[name = tensor("x_383_cast_fp16")]; + tensor var_3512_to_fp16 = const()[name = tensor("op_3512_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1244454336)))]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247731200)))]; + tensor linear_189_cast_fp16 = linear(bias = var_3513_to_fp16, weight = var_3512_to_fp16, x = x_383_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor x_385_cast_fp16 = add(x = x_379_cast_fp16, y = linear_189_cast_fp16)[name = tensor("x_385_cast_fp16")]; + tensor var_3520_axes_0 = const()[name = tensor("op_3520_axes_0"), val = tensor([-1])]; + tensor blocks_31_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_31_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247733824)))]; + tensor blocks_31_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_31_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247736448)))]; + tensor var_3520_cast_fp16 = layer_norm(axes = var_3520_axes_0, beta = blocks_31_mlp_ln_bias_to_fp16, epsilon = var_3445_to_fp16, gamma = blocks_31_mlp_ln_weight_to_fp16, x = x_385_cast_fp16)[name = tensor("op_3520_cast_fp16")]; + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1247739072)))]; + tensor var_3530_to_fp16 = const()[name = tensor("op_3530_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260846336)))]; + tensor linear_190_cast_fp16 = linear(bias = var_3530_to_fp16, weight = var_3529_to_fp16, x = var_3520_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor x_389_mode_0 = const()[name = tensor("x_389_mode_0"), val = tensor("EXACT")]; + tensor x_389_cast_fp16 = gelu(mode = x_389_mode_0, x = linear_190_cast_fp16)[name = tensor("x_389_cast_fp16")]; + tensor var_3535_to_fp16 = const()[name = tensor("op_3535_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1260856640)))]; + tensor var_3536_to_fp16 = const()[name = tensor("op_3536_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273963904)))]; + tensor linear_191_cast_fp16 = linear(bias = var_3536_to_fp16, weight = var_3535_to_fp16, x = x_389_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor x_cast_fp16 = add(x = x_385_cast_fp16, y = linear_191_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3550_axes_0 = const()[name = tensor("op_3550_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273966528)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1273969152)))]; + tensor var_3541_to_fp16 = const()[name = tensor("op_3541_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3550_cast_fp16 = layer_norm(axes = var_3550_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_3541_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast_fp16)[name = tensor("op_3550_cast_fp16")]; + tensor var_3550_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3550_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_3550_cast_fp16_to_fp32_dtype_0, x = var_3550_cast_fp16)[name = tensor("cast_192")]; + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f89c1f6aec683cb19e696fb1daf448c6b1a4f0cf --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b56493c9c35558415d78a3f8b71f05f3999cb20be3f1a6f18e5afa31e9ff25b +size 1273971776 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..2b88756038fef26f808846c99d72668f83cf8b6d --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e54422207be483c358c35b76e95b2acaeaed8fcd01b606c40b6f1f2a6c4e21a +size 1106374306 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c39617d775a352f453a0b05a97803df77cb7a81b --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:311e822db8601dd4f6051f276975a410f77290e20058815f0bbc2d3fe6339f86 +size 243 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2de4dd3f09da42374c939e1517ba00806601681f --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53adfc091caf04e1f1cf9f42215860bd1f9481d2e0116a0b71e78b9e87003045 +size 319 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..70d176282527ae944c34d2f3e56f140fae82bbfe --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/metadata.json @@ -0,0 +1,68 @@ +[ + { + "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", + "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==2.1.0", + "com.github.apple.coremltools.version" : "8.0" + }, + "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_turbo", + "method" : "predict" + } +] \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8756e62aec14e45e1dd595ed47aae6a520cc5300 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/model.mil @@ -0,0 +1,5643 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.1.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})] +{ + 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 weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_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_131")]; + tensor var_84_cast_fp16 = conv(bias = bias_3_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 = weight_3_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 weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985792)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10816256)))]; + tensor var_102_cast_fp16 = conv(bias = bias_7_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 = weight_7_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_130")]; + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1e7696784b2c34070bafe9f2fb581415a0ade6b6 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-large-v3-turbo-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcc450fb244d55335f6df82a41558de1b07d44acaf67c7b7b3040da44f94bdd3 +size 1273969152 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..d3016446da726b38fd5f805cd6684cf4ba5e446f --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fded8206831d8707a4ee2b24d1fe777ca1acf1cb6c0532cfdb7a5d0109a7da1c +size 551522554 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..cfebf363e7217ce13431621718bc293f474504b9 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adbe456375e7eb3407732a426ecb65bbda86860e4aa801f3a696b70b8a533cdd +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..5e8a17765c9a9adbc7ce77736778897c9543219c --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 144, + "Matmul" : 48, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 24, + "Add" : 49, + "LayerNorm" : 49, + "Mul" : 48, + "Transpose" : 97, + "Gelu" : 26, + "Reshape" : 96 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..4771e10615202abab5bfc7b155ba0872539682da --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/model.mil @@ -0,0 +1,1455 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_56 = const()[name = tensor("op_56"), val = tensor(1)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor([1])]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491648)))]; + tensor cast_727 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_68_cast = conv(bias = bias_3_to_fp16, dilations = var_66, groups = var_56, pad = var_68_pad_0, pad_type = var_68_pad_type_0, strides = var_64, weight = weight_3_to_fp16, x = cast_727); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_68_cast); + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor var_81 = const()[name = tensor("op_81"), val = tensor([2])]; + tensor var_83 = const()[name = tensor("op_83"), val = tensor([1])]; + tensor var_85_pad_type_0 = const()[name = tensor("op_85_pad_type_0"), val = tensor("custom")]; + tensor var_85_pad_0 = const()[name = tensor("op_85_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493760)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6785280)))]; + tensor var_85_cast = conv(bias = bias_7_to_fp16, dilations = var_83, groups = var_72, pad = var_85_pad_0, pad_type = var_85_pad_type_0, strides = var_81, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_85_cast); + tensor var_90 = const()[name = tensor("op_90"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6787392)))]; + tensor transpose_192 = transpose(perm = var_90, x = x_3_cast); + tensor var_93_cast = add(x = transpose_192, y = positional_embedding_to_fp16); + tensor var_106 = const()[name = tensor("op_106"), val = tensor(-1)]; + tensor var_123_axes_0 = const()[name = tensor("op_123_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9859456)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9861568)))]; + tensor var_112_to_fp16 = const()[name = tensor("op_112_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_123_cast = layer_norm(axes = var_123_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_112_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_93_cast); + tensor var_134_to_fp16 = const()[name = tensor("op_134_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9863680)))]; + tensor var_135_to_fp16 = const()[name = tensor("op_135_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11960896)))]; + tensor q_1_cast = linear(bias = var_135_to_fp16, weight = var_134_to_fp16, x = var_123_cast); + tensor var_138_to_fp16 = const()[name = tensor("op_138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11963008)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14060224)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_138_to_fp16, x = var_123_cast); + tensor var_142_to_fp16 = const()[name = tensor("op_142_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14062336)))]; + tensor var_143_to_fp16 = const()[name = tensor("op_143_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16159552)))]; + tensor v_1_cast = linear(bias = var_143_to_fp16, weight = var_142_to_fp16, x = var_123_cast); + tensor var_151 = const()[name = tensor("op_151"), val = tensor([1, 1500, 16, -1])]; + tensor var_152_cast = reshape(shape = var_151, x = q_1_cast); + tensor const_168_to_fp16 = const()[name = tensor("const_168_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_152_cast, y = const_168_to_fp16); + tensor var_158 = const()[name = tensor("op_158"), val = tensor([1, 1500, 16, -1])]; + tensor var_159_cast = reshape(shape = var_158, x = k_1_cast); + tensor const_169_to_fp16 = const()[name = tensor("const_169_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_159_cast, y = const_169_to_fp16); + tensor var_165 = const()[name = tensor("op_165"), val = tensor([1, 1500, 16, -1])]; + tensor var_166_cast = reshape(shape = var_165, x = v_1_cast); + tensor var_167 = const()[name = tensor("op_167"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_48_perm_0 = const()[name = tensor("transpose_48_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_49_perm_0 = const()[name = tensor("transpose_49_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_189 = transpose(perm = transpose_49_perm_0, x = k_3_cast); + tensor transpose_190 = transpose(perm = transpose_48_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_190, y = transpose_189); + tensor var_171_cast = softmax(axis = var_106, x = qk_1_cast); + tensor var_173_transpose_x_0 = const()[name = tensor("op_173_transpose_x_0"), val = tensor(false)]; + tensor var_173_transpose_y_0 = const()[name = tensor("op_173_transpose_y_0"), val = tensor(false)]; + tensor transpose_191 = transpose(perm = var_167, x = var_166_cast); + tensor var_173_cast = matmul(transpose_x = var_173_transpose_x_0, transpose_y = var_173_transpose_y_0, x = var_171_cast, y = transpose_191); + tensor var_174 = const()[name = tensor("op_174"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 1024])]; + tensor transpose_188 = transpose(perm = var_174, x = var_173_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_188); + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16161664)))]; + tensor var_180_to_fp16 = const()[name = tensor("op_180_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18258880)))]; + tensor var_181_cast = linear(bias = var_180_to_fp16, weight = var_179_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_93_cast, y = var_181_cast); + tensor var_187_axes_0 = const()[name = tensor("op_187_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18260992)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18263104)))]; + tensor var_187_cast = layer_norm(axes = var_187_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_112_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_196_to_fp16 = const()[name = tensor("op_196_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18265216)))]; + tensor var_197_to_fp16 = const()[name = tensor("op_197_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26653888)))]; + tensor input_9_cast = linear(bias = var_197_to_fp16, weight = var_196_to_fp16, x = var_187_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26662144)))]; + tensor var_203_to_fp16 = const()[name = tensor("op_203_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35050816)))]; + tensor var_204_cast = linear(bias = var_203_to_fp16, weight = var_202_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_204_cast); + tensor var_213 = const()[name = tensor("op_213"), val = tensor(-1)]; + tensor var_230_axes_0 = const()[name = tensor("op_230_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35052928)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35055040)))]; + tensor var_219_to_fp16 = const()[name = tensor("op_219_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_230_cast = layer_norm(axes = var_230_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_219_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_241_to_fp16 = const()[name = tensor("op_241_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35057152)))]; + tensor var_242_to_fp16 = const()[name = tensor("op_242_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37154368)))]; + tensor q_5_cast = linear(bias = var_242_to_fp16, weight = var_241_to_fp16, x = var_230_cast); + tensor var_245_to_fp16 = const()[name = tensor("op_245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37156480)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39253696)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_245_to_fp16, x = var_230_cast); + tensor var_249_to_fp16 = const()[name = tensor("op_249_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39255808)))]; + tensor var_250_to_fp16 = const()[name = tensor("op_250_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41353024)))]; + tensor v_5_cast = linear(bias = var_250_to_fp16, weight = var_249_to_fp16, x = var_230_cast); + tensor var_258 = const()[name = tensor("op_258"), val = tensor([1, 1500, 16, -1])]; + tensor var_259_cast = reshape(shape = var_258, x = q_5_cast); + tensor const_170_to_fp16 = const()[name = tensor("const_170_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_259_cast, y = const_170_to_fp16); + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1500, 16, -1])]; + tensor var_266_cast = reshape(shape = var_265, x = k_5_cast); + tensor const_171_to_fp16 = const()[name = tensor("const_171_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_266_cast, y = const_171_to_fp16); + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 1500, 16, -1])]; + tensor var_273_cast = reshape(shape = var_272, x = v_5_cast); + tensor var_274 = const()[name = tensor("op_274"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_50_perm_0 = const()[name = tensor("transpose_50_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_51_perm_0 = const()[name = tensor("transpose_51_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_185 = transpose(perm = transpose_51_perm_0, x = k_7_cast); + tensor transpose_186 = transpose(perm = transpose_50_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_186, y = transpose_185); + tensor var_278_cast = softmax(axis = var_213, x = qk_3_cast); + tensor var_280_transpose_x_0 = const()[name = tensor("op_280_transpose_x_0"), val = tensor(false)]; + tensor var_280_transpose_y_0 = const()[name = tensor("op_280_transpose_y_0"), val = tensor(false)]; + tensor transpose_187 = transpose(perm = var_274, x = var_273_cast); + tensor var_280_cast = matmul(transpose_x = var_280_transpose_x_0, transpose_y = var_280_transpose_y_0, x = var_278_cast, y = transpose_187); + tensor var_281 = const()[name = tensor("op_281"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 1024])]; + tensor transpose_184 = transpose(perm = var_281, x = var_280_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_184); + tensor var_286_to_fp16 = const()[name = tensor("op_286_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41355136)))]; + tensor var_287_to_fp16 = const()[name = tensor("op_287_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43452352)))]; + tensor var_288_cast = linear(bias = var_287_to_fp16, weight = var_286_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_288_cast); + tensor var_294_axes_0 = const()[name = tensor("op_294_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43454464)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43456576)))]; + tensor var_294_cast = layer_norm(axes = var_294_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_219_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_303_to_fp16 = const()[name = tensor("op_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43458688)))]; + tensor var_304_to_fp16 = const()[name = tensor("op_304_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51847360)))]; + tensor input_17_cast = linear(bias = var_304_to_fp16, weight = var_303_to_fp16, x = var_294_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51855616)))]; + tensor var_310_to_fp16 = const()[name = tensor("op_310_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60244288)))]; + tensor var_311_cast = linear(bias = var_310_to_fp16, weight = var_309_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_311_cast); + tensor var_320 = const()[name = tensor("op_320"), val = tensor(-1)]; + tensor var_337_axes_0 = const()[name = tensor("op_337_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60246400)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60248512)))]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_337_cast = layer_norm(axes = var_337_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_326_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_348_to_fp16 = const()[name = tensor("op_348_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60250624)))]; + tensor var_349_to_fp16 = const()[name = tensor("op_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62347840)))]; + tensor q_9_cast = linear(bias = var_349_to_fp16, weight = var_348_to_fp16, x = var_337_cast); + tensor var_352_to_fp16 = const()[name = tensor("op_352_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62349952)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64447168)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_352_to_fp16, x = var_337_cast); + tensor var_356_to_fp16 = const()[name = tensor("op_356_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64449280)))]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66546496)))]; + tensor v_9_cast = linear(bias = var_357_to_fp16, weight = var_356_to_fp16, x = var_337_cast); + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1500, 16, -1])]; + tensor var_366_cast = reshape(shape = var_365, x = q_9_cast); + tensor const_172_to_fp16 = const()[name = tensor("const_172_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_366_cast, y = const_172_to_fp16); + tensor var_372 = const()[name = tensor("op_372"), val = tensor([1, 1500, 16, -1])]; + tensor var_373_cast = reshape(shape = var_372, x = k_9_cast); + tensor const_173_to_fp16 = const()[name = tensor("const_173_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_373_cast, y = const_173_to_fp16); + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1500, 16, -1])]; + tensor var_380_cast = reshape(shape = var_379, x = v_9_cast); + tensor var_381 = const()[name = tensor("op_381"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_52_perm_0 = const()[name = tensor("transpose_52_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_53_perm_0 = const()[name = tensor("transpose_53_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_181 = transpose(perm = transpose_53_perm_0, x = k_11_cast); + tensor transpose_182 = transpose(perm = transpose_52_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_182, y = transpose_181); + tensor var_385_cast = softmax(axis = var_320, x = qk_5_cast); + tensor var_387_transpose_x_0 = const()[name = tensor("op_387_transpose_x_0"), val = tensor(false)]; + tensor var_387_transpose_y_0 = const()[name = tensor("op_387_transpose_y_0"), val = tensor(false)]; + tensor transpose_183 = transpose(perm = var_381, x = var_380_cast); + tensor var_387_cast = matmul(transpose_x = var_387_transpose_x_0, transpose_y = var_387_transpose_y_0, x = var_385_cast, y = transpose_183); + tensor var_388 = const()[name = tensor("op_388"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 1024])]; + tensor transpose_180 = transpose(perm = var_388, x = var_387_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_180); + tensor var_393_to_fp16 = const()[name = tensor("op_393_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66548608)))]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68645824)))]; + tensor var_395_cast = linear(bias = var_394_to_fp16, weight = var_393_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_395_cast); + tensor var_401_axes_0 = const()[name = tensor("op_401_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68647936)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68650048)))]; + tensor var_401_cast = layer_norm(axes = var_401_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_326_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_410_to_fp16 = const()[name = tensor("op_410_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68652160)))]; + tensor var_411_to_fp16 = const()[name = tensor("op_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77040832)))]; + tensor input_25_cast = linear(bias = var_411_to_fp16, weight = var_410_to_fp16, x = var_401_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_416_to_fp16 = const()[name = tensor("op_416_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77049088)))]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85437760)))]; + tensor var_418_cast = linear(bias = var_417_to_fp16, weight = var_416_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_418_cast); + tensor var_427 = const()[name = tensor("op_427"), val = tensor(-1)]; + tensor var_444_axes_0 = const()[name = tensor("op_444_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85439872)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85441984)))]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_444_cast = layer_norm(axes = var_444_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_433_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_455_to_fp16 = const()[name = tensor("op_455_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85444096)))]; + tensor var_456_to_fp16 = const()[name = tensor("op_456_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87541312)))]; + tensor q_13_cast = linear(bias = var_456_to_fp16, weight = var_455_to_fp16, x = var_444_cast); + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87543424)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89640640)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_459_to_fp16, x = var_444_cast); + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89642752)))]; + tensor var_464_to_fp16 = const()[name = tensor("op_464_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91739968)))]; + tensor v_13_cast = linear(bias = var_464_to_fp16, weight = var_463_to_fp16, x = var_444_cast); + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 1500, 16, -1])]; + tensor var_473_cast = reshape(shape = var_472, x = q_13_cast); + tensor const_174_to_fp16 = const()[name = tensor("const_174_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_473_cast, y = const_174_to_fp16); + tensor var_479 = const()[name = tensor("op_479"), val = tensor([1, 1500, 16, -1])]; + tensor var_480_cast = reshape(shape = var_479, x = k_13_cast); + tensor const_175_to_fp16 = const()[name = tensor("const_175_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_480_cast, y = const_175_to_fp16); + tensor var_486 = const()[name = tensor("op_486"), val = tensor([1, 1500, 16, -1])]; + tensor var_487_cast = reshape(shape = var_486, x = v_13_cast); + tensor var_488 = const()[name = tensor("op_488"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_54_perm_0 = const()[name = tensor("transpose_54_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_55_perm_0 = const()[name = tensor("transpose_55_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_177 = transpose(perm = transpose_55_perm_0, x = k_15_cast); + tensor transpose_178 = transpose(perm = transpose_54_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_178, y = transpose_177); + tensor var_492_cast = softmax(axis = var_427, x = qk_7_cast); + tensor var_494_transpose_x_0 = const()[name = tensor("op_494_transpose_x_0"), val = tensor(false)]; + tensor var_494_transpose_y_0 = const()[name = tensor("op_494_transpose_y_0"), val = tensor(false)]; + tensor transpose_179 = transpose(perm = var_488, x = var_487_cast); + tensor var_494_cast = matmul(transpose_x = var_494_transpose_x_0, transpose_y = var_494_transpose_y_0, x = var_492_cast, y = transpose_179); + tensor var_495 = const()[name = tensor("op_495"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 1024])]; + tensor transpose_176 = transpose(perm = var_495, x = var_494_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_176); + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91742080)))]; + tensor var_501_to_fp16 = const()[name = tensor("op_501_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93839296)))]; + tensor var_502_cast = linear(bias = var_501_to_fp16, weight = var_500_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_502_cast); + tensor var_508_axes_0 = const()[name = tensor("op_508_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93841408)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93843520)))]; + tensor var_508_cast = layer_norm(axes = var_508_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_433_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_517_to_fp16 = const()[name = tensor("op_517_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93845632)))]; + tensor var_518_to_fp16 = const()[name = tensor("op_518_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102234304)))]; + tensor input_33_cast = linear(bias = var_518_to_fp16, weight = var_517_to_fp16, x = var_508_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_523_to_fp16 = const()[name = tensor("op_523_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102242560)))]; + tensor var_524_to_fp16 = const()[name = tensor("op_524_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110631232)))]; + tensor var_525_cast = linear(bias = var_524_to_fp16, weight = var_523_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_525_cast); + tensor var_534 = const()[name = tensor("op_534"), val = tensor(-1)]; + tensor var_551_axes_0 = const()[name = tensor("op_551_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110633344)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110635456)))]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_551_cast = layer_norm(axes = var_551_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_540_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_562_to_fp16 = const()[name = tensor("op_562_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110637568)))]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112734784)))]; + tensor q_17_cast = linear(bias = var_563_to_fp16, weight = var_562_to_fp16, x = var_551_cast); + tensor var_566_to_fp16 = const()[name = tensor("op_566_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112736896)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114834112)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_566_to_fp16, x = var_551_cast); + tensor var_570_to_fp16 = const()[name = tensor("op_570_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114836224)))]; + tensor var_571_to_fp16 = const()[name = tensor("op_571_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116933440)))]; + tensor v_17_cast = linear(bias = var_571_to_fp16, weight = var_570_to_fp16, x = var_551_cast); + tensor var_579 = const()[name = tensor("op_579"), val = tensor([1, 1500, 16, -1])]; + tensor var_580_cast = reshape(shape = var_579, x = q_17_cast); + tensor const_176_to_fp16 = const()[name = tensor("const_176_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_580_cast, y = const_176_to_fp16); + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 1500, 16, -1])]; + tensor var_587_cast = reshape(shape = var_586, x = k_17_cast); + tensor const_177_to_fp16 = const()[name = tensor("const_177_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_587_cast, y = const_177_to_fp16); + tensor var_593 = const()[name = tensor("op_593"), val = tensor([1, 1500, 16, -1])]; + tensor var_594_cast = reshape(shape = var_593, x = v_17_cast); + tensor var_595 = const()[name = tensor("op_595"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_56_perm_0 = const()[name = tensor("transpose_56_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_57_perm_0 = const()[name = tensor("transpose_57_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_173 = transpose(perm = transpose_57_perm_0, x = k_19_cast); + tensor transpose_174 = transpose(perm = transpose_56_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_174, y = transpose_173); + tensor var_599_cast = softmax(axis = var_534, x = qk_9_cast); + tensor var_601_transpose_x_0 = const()[name = tensor("op_601_transpose_x_0"), val = tensor(false)]; + tensor var_601_transpose_y_0 = const()[name = tensor("op_601_transpose_y_0"), val = tensor(false)]; + tensor transpose_175 = transpose(perm = var_595, x = var_594_cast); + tensor var_601_cast = matmul(transpose_x = var_601_transpose_x_0, transpose_y = var_601_transpose_y_0, x = var_599_cast, y = transpose_175); + tensor var_602 = const()[name = tensor("op_602"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 1024])]; + tensor transpose_172 = transpose(perm = var_602, x = var_601_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_172); + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116935552)))]; + tensor var_608_to_fp16 = const()[name = tensor("op_608_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119032768)))]; + tensor var_609_cast = linear(bias = var_608_to_fp16, weight = var_607_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_609_cast); + tensor var_615_axes_0 = const()[name = tensor("op_615_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119034880)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119036992)))]; + tensor var_615_cast = layer_norm(axes = var_615_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_540_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_624_to_fp16 = const()[name = tensor("op_624_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119039104)))]; + tensor var_625_to_fp16 = const()[name = tensor("op_625_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127427776)))]; + tensor input_41_cast = linear(bias = var_625_to_fp16, weight = var_624_to_fp16, x = var_615_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_630_to_fp16 = const()[name = tensor("op_630_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127436032)))]; + tensor var_631_to_fp16 = const()[name = tensor("op_631_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135824704)))]; + tensor var_632_cast = linear(bias = var_631_to_fp16, weight = var_630_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_632_cast); + tensor var_641 = const()[name = tensor("op_641"), val = tensor(-1)]; + tensor var_658_axes_0 = const()[name = tensor("op_658_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135826816)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135828928)))]; + tensor var_647_to_fp16 = const()[name = tensor("op_647_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_658_cast = layer_norm(axes = var_658_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_647_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_669_to_fp16 = const()[name = tensor("op_669_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135831040)))]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137928256)))]; + tensor q_21_cast = linear(bias = var_670_to_fp16, weight = var_669_to_fp16, x = var_658_cast); + tensor var_673_to_fp16 = const()[name = tensor("op_673_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137930368)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140027584)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_673_to_fp16, x = var_658_cast); + tensor var_677_to_fp16 = const()[name = tensor("op_677_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140029696)))]; + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142126912)))]; + tensor v_21_cast = linear(bias = var_678_to_fp16, weight = var_677_to_fp16, x = var_658_cast); + tensor var_686 = const()[name = tensor("op_686"), val = tensor([1, 1500, 16, -1])]; + tensor var_687_cast = reshape(shape = var_686, x = q_21_cast); + tensor const_178_to_fp16 = const()[name = tensor("const_178_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast = mul(x = var_687_cast, y = const_178_to_fp16); + tensor var_693 = const()[name = tensor("op_693"), val = tensor([1, 1500, 16, -1])]; + tensor var_694_cast = reshape(shape = var_693, x = k_21_cast); + tensor const_179_to_fp16 = const()[name = tensor("const_179_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast = mul(x = var_694_cast, y = const_179_to_fp16); + tensor var_700 = const()[name = tensor("op_700"), val = tensor([1, 1500, 16, -1])]; + tensor var_701_cast = reshape(shape = var_700, x = v_21_cast); + tensor var_702 = const()[name = tensor("op_702"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_58_perm_0 = const()[name = tensor("transpose_58_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_59_perm_0 = const()[name = tensor("transpose_59_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_169 = transpose(perm = transpose_59_perm_0, x = k_23_cast); + tensor transpose_170 = transpose(perm = transpose_58_perm_0, x = q_23_cast); + tensor qk_11_cast = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_170, y = transpose_169); + tensor var_706_cast = softmax(axis = var_641, x = qk_11_cast); + tensor var_708_transpose_x_0 = const()[name = tensor("op_708_transpose_x_0"), val = tensor(false)]; + tensor var_708_transpose_y_0 = const()[name = tensor("op_708_transpose_y_0"), val = tensor(false)]; + tensor transpose_171 = transpose(perm = var_702, x = var_701_cast); + tensor var_708_cast = matmul(transpose_x = var_708_transpose_x_0, transpose_y = var_708_transpose_y_0, x = var_706_cast, y = transpose_171); + tensor var_709 = const()[name = tensor("op_709"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 1024])]; + tensor transpose_168 = transpose(perm = var_709, x = var_708_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_168); + tensor var_714_to_fp16 = const()[name = tensor("op_714_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142129024)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144226240)))]; + tensor var_716_cast = linear(bias = var_715_to_fp16, weight = var_714_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_716_cast); + tensor var_722_axes_0 = const()[name = tensor("op_722_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144228352)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144230464)))]; + tensor var_722_cast = layer_norm(axes = var_722_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_647_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144232576)))]; + tensor var_732_to_fp16 = const()[name = tensor("op_732_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152621248)))]; + tensor input_49_cast = linear(bias = var_732_to_fp16, weight = var_731_to_fp16, x = var_722_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_737_to_fp16 = const()[name = tensor("op_737_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152629504)))]; + tensor var_738_to_fp16 = const()[name = tensor("op_738_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161018176)))]; + tensor var_739_cast = linear(bias = var_738_to_fp16, weight = var_737_to_fp16, x = x_77_cast); + tensor x_79_cast = add(x = x_73_cast, y = var_739_cast); + tensor var_748 = const()[name = tensor("op_748"), val = tensor(-1)]; + tensor var_765_axes_0 = const()[name = tensor("op_765_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161020288)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161022400)))]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_765_cast = layer_norm(axes = var_765_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_754_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast); + tensor var_776_to_fp16 = const()[name = tensor("op_776_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161024512)))]; + tensor var_777_to_fp16 = const()[name = tensor("op_777_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163121728)))]; + tensor q_25_cast = linear(bias = var_777_to_fp16, weight = var_776_to_fp16, x = var_765_cast); + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163123840)))]; + tensor k_25_bias_0_to_fp16 = const()[name = tensor("k_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165221056)))]; + tensor k_25_cast = linear(bias = k_25_bias_0_to_fp16, weight = var_780_to_fp16, x = var_765_cast); + tensor var_784_to_fp16 = const()[name = tensor("op_784_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165223168)))]; + tensor var_785_to_fp16 = const()[name = tensor("op_785_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167320384)))]; + tensor v_25_cast = linear(bias = var_785_to_fp16, weight = var_784_to_fp16, x = var_765_cast); + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 1500, 16, -1])]; + tensor var_794_cast = reshape(shape = var_793, x = q_25_cast); + tensor const_180_to_fp16 = const()[name = tensor("const_180_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast = mul(x = var_794_cast, y = const_180_to_fp16); + tensor var_800 = const()[name = tensor("op_800"), val = tensor([1, 1500, 16, -1])]; + tensor var_801_cast = reshape(shape = var_800, x = k_25_cast); + tensor const_181_to_fp16 = const()[name = tensor("const_181_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast = mul(x = var_801_cast, y = const_181_to_fp16); + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1500, 16, -1])]; + tensor var_808_cast = reshape(shape = var_807, x = v_25_cast); + tensor var_809 = const()[name = tensor("op_809"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_60_perm_0 = const()[name = tensor("transpose_60_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_61_perm_0 = const()[name = tensor("transpose_61_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_165 = transpose(perm = transpose_61_perm_0, x = k_27_cast); + tensor transpose_166 = transpose(perm = transpose_60_perm_0, x = q_27_cast); + tensor qk_13_cast = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_166, y = transpose_165); + tensor var_813_cast = softmax(axis = var_748, x = qk_13_cast); + tensor var_815_transpose_x_0 = const()[name = tensor("op_815_transpose_x_0"), val = tensor(false)]; + tensor var_815_transpose_y_0 = const()[name = tensor("op_815_transpose_y_0"), val = tensor(false)]; + tensor transpose_167 = transpose(perm = var_809, x = var_808_cast); + tensor var_815_cast = matmul(transpose_x = var_815_transpose_x_0, transpose_y = var_815_transpose_y_0, x = var_813_cast, y = transpose_167); + tensor var_816 = const()[name = tensor("op_816"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 1024])]; + tensor transpose_164 = transpose(perm = var_816, x = var_815_cast); + tensor x_83_cast = reshape(shape = concat_6, x = transpose_164); + tensor var_821_to_fp16 = const()[name = tensor("op_821_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167322496)))]; + tensor var_822_to_fp16 = const()[name = tensor("op_822_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169419712)))]; + tensor var_823_cast = linear(bias = var_822_to_fp16, weight = var_821_to_fp16, x = x_83_cast); + tensor x_85_cast = add(x = x_79_cast, y = var_823_cast); + tensor var_829_axes_0 = const()[name = tensor("op_829_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169421824)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169423936)))]; + tensor var_829_cast = layer_norm(axes = var_829_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_754_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast); + tensor var_838_to_fp16 = const()[name = tensor("op_838_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169426048)))]; + tensor var_839_to_fp16 = const()[name = tensor("op_839_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177814720)))]; + tensor input_57_cast = linear(bias = var_839_to_fp16, weight = var_838_to_fp16, x = var_829_cast); + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast = gelu(mode = x_89_mode_0, x = input_57_cast); + tensor var_844_to_fp16 = const()[name = tensor("op_844_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177822976)))]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186211648)))]; + tensor var_846_cast = linear(bias = var_845_to_fp16, weight = var_844_to_fp16, x = x_89_cast); + tensor x_91_cast = add(x = x_85_cast, y = var_846_cast); + tensor var_855 = const()[name = tensor("op_855"), val = tensor(-1)]; + tensor var_872_axes_0 = const()[name = tensor("op_872_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186213760)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186215872)))]; + tensor var_861_to_fp16 = const()[name = tensor("op_861_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_872_cast = layer_norm(axes = var_872_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_861_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast); + tensor var_883_to_fp16 = const()[name = tensor("op_883_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186217984)))]; + tensor var_884_to_fp16 = const()[name = tensor("op_884_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188315200)))]; + tensor q_29_cast = linear(bias = var_884_to_fp16, weight = var_883_to_fp16, x = var_872_cast); + tensor var_887_to_fp16 = const()[name = tensor("op_887_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188317312)))]; + tensor k_29_bias_0_to_fp16 = const()[name = tensor("k_29_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190414528)))]; + tensor k_29_cast = linear(bias = k_29_bias_0_to_fp16, weight = var_887_to_fp16, x = var_872_cast); + tensor var_891_to_fp16 = const()[name = tensor("op_891_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190416640)))]; + tensor var_892_to_fp16 = const()[name = tensor("op_892_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192513856)))]; + tensor v_29_cast = linear(bias = var_892_to_fp16, weight = var_891_to_fp16, x = var_872_cast); + tensor var_900 = const()[name = tensor("op_900"), val = tensor([1, 1500, 16, -1])]; + tensor var_901_cast = reshape(shape = var_900, x = q_29_cast); + tensor const_182_to_fp16 = const()[name = tensor("const_182_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast = mul(x = var_901_cast, y = const_182_to_fp16); + tensor var_907 = const()[name = tensor("op_907"), val = tensor([1, 1500, 16, -1])]; + tensor var_908_cast = reshape(shape = var_907, x = k_29_cast); + tensor const_183_to_fp16 = const()[name = tensor("const_183_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast = mul(x = var_908_cast, y = const_183_to_fp16); + tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 1500, 16, -1])]; + tensor var_915_cast = reshape(shape = var_914, x = v_29_cast); + tensor var_916 = const()[name = tensor("op_916"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_62_perm_0 = const()[name = tensor("transpose_62_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_63_perm_0 = const()[name = tensor("transpose_63_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_161 = transpose(perm = transpose_63_perm_0, x = k_31_cast); + tensor transpose_162 = transpose(perm = transpose_62_perm_0, x = q_31_cast); + tensor qk_15_cast = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_162, y = transpose_161); + tensor var_920_cast = softmax(axis = var_855, x = qk_15_cast); + tensor var_922_transpose_x_0 = const()[name = tensor("op_922_transpose_x_0"), val = tensor(false)]; + tensor var_922_transpose_y_0 = const()[name = tensor("op_922_transpose_y_0"), val = tensor(false)]; + tensor transpose_163 = transpose(perm = var_916, x = var_915_cast); + tensor var_922_cast = matmul(transpose_x = var_922_transpose_x_0, transpose_y = var_922_transpose_y_0, x = var_920_cast, y = transpose_163); + tensor var_923 = const()[name = tensor("op_923"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 1024])]; + tensor transpose_160 = transpose(perm = var_923, x = var_922_cast); + tensor x_95_cast = reshape(shape = concat_7, x = transpose_160); + tensor var_928_to_fp16 = const()[name = tensor("op_928_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192515968)))]; + tensor var_929_to_fp16 = const()[name = tensor("op_929_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194613184)))]; + tensor var_930_cast = linear(bias = var_929_to_fp16, weight = var_928_to_fp16, x = x_95_cast); + tensor x_97_cast = add(x = x_91_cast, y = var_930_cast); + tensor var_936_axes_0 = const()[name = tensor("op_936_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194615296)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194617408)))]; + tensor var_936_cast = layer_norm(axes = var_936_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_861_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast); + tensor var_945_to_fp16 = const()[name = tensor("op_945_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194619520)))]; + tensor var_946_to_fp16 = const()[name = tensor("op_946_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203008192)))]; + tensor input_65_cast = linear(bias = var_946_to_fp16, weight = var_945_to_fp16, x = var_936_cast); + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast = gelu(mode = x_101_mode_0, x = input_65_cast); + tensor var_951_to_fp16 = const()[name = tensor("op_951_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203016448)))]; + tensor var_952_to_fp16 = const()[name = tensor("op_952_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211405120)))]; + tensor var_953_cast = linear(bias = var_952_to_fp16, weight = var_951_to_fp16, x = x_101_cast); + tensor x_103_cast = add(x = x_97_cast, y = var_953_cast); + tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; + tensor var_979_axes_0 = const()[name = tensor("op_979_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211407232)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211409344)))]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_979_cast = layer_norm(axes = var_979_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_968_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast); + tensor var_990_to_fp16 = const()[name = tensor("op_990_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211411456)))]; + tensor var_991_to_fp16 = const()[name = tensor("op_991_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213508672)))]; + tensor q_33_cast = linear(bias = var_991_to_fp16, weight = var_990_to_fp16, x = var_979_cast); + tensor var_994_to_fp16 = const()[name = tensor("op_994_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213510784)))]; + tensor k_33_bias_0_to_fp16 = const()[name = tensor("k_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215608000)))]; + tensor k_33_cast = linear(bias = k_33_bias_0_to_fp16, weight = var_994_to_fp16, x = var_979_cast); + tensor var_998_to_fp16 = const()[name = tensor("op_998_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215610112)))]; + tensor var_999_to_fp16 = const()[name = tensor("op_999_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217707328)))]; + tensor v_33_cast = linear(bias = var_999_to_fp16, weight = var_998_to_fp16, x = var_979_cast); + tensor var_1007 = const()[name = tensor("op_1007"), val = tensor([1, 1500, 16, -1])]; + tensor var_1008_cast = reshape(shape = var_1007, x = q_33_cast); + tensor const_184_to_fp16 = const()[name = tensor("const_184_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast = mul(x = var_1008_cast, y = const_184_to_fp16); + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1500, 16, -1])]; + tensor var_1015_cast = reshape(shape = var_1014, x = k_33_cast); + tensor const_185_to_fp16 = const()[name = tensor("const_185_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast = mul(x = var_1015_cast, y = const_185_to_fp16); + tensor var_1021 = const()[name = tensor("op_1021"), val = tensor([1, 1500, 16, -1])]; + tensor var_1022_cast = reshape(shape = var_1021, x = v_33_cast); + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_64_perm_0 = const()[name = tensor("transpose_64_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_65_perm_0 = const()[name = tensor("transpose_65_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_157 = transpose(perm = transpose_65_perm_0, x = k_35_cast); + tensor transpose_158 = transpose(perm = transpose_64_perm_0, x = q_35_cast); + tensor qk_17_cast = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_158, y = transpose_157); + tensor var_1027_cast = softmax(axis = var_962, x = qk_17_cast); + tensor var_1029_transpose_x_0 = const()[name = tensor("op_1029_transpose_x_0"), val = tensor(false)]; + tensor var_1029_transpose_y_0 = const()[name = tensor("op_1029_transpose_y_0"), val = tensor(false)]; + tensor transpose_159 = transpose(perm = var_1023, x = var_1022_cast); + tensor var_1029_cast = matmul(transpose_x = var_1029_transpose_x_0, transpose_y = var_1029_transpose_y_0, x = var_1027_cast, y = transpose_159); + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 1024])]; + tensor transpose_156 = transpose(perm = var_1030, x = var_1029_cast); + tensor x_107_cast = reshape(shape = concat_8, x = transpose_156); + tensor var_1035_to_fp16 = const()[name = tensor("op_1035_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217709440)))]; + tensor var_1036_to_fp16 = const()[name = tensor("op_1036_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219806656)))]; + tensor var_1037_cast = linear(bias = var_1036_to_fp16, weight = var_1035_to_fp16, x = x_107_cast); + tensor x_109_cast = add(x = x_103_cast, y = var_1037_cast); + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219808768)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219810880)))]; + tensor var_1043_cast = layer_norm(axes = var_1043_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_968_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast); + tensor var_1052_to_fp16 = const()[name = tensor("op_1052_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219812992)))]; + tensor var_1053_to_fp16 = const()[name = tensor("op_1053_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228201664)))]; + tensor input_73_cast = linear(bias = var_1053_to_fp16, weight = var_1052_to_fp16, x = var_1043_cast); + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast = gelu(mode = x_113_mode_0, x = input_73_cast); + tensor var_1058_to_fp16 = const()[name = tensor("op_1058_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228209920)))]; + tensor var_1059_to_fp16 = const()[name = tensor("op_1059_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236598592)))]; + tensor var_1060_cast = linear(bias = var_1059_to_fp16, weight = var_1058_to_fp16, x = x_113_cast); + tensor x_115_cast = add(x = x_109_cast, y = var_1060_cast); + tensor var_1069 = const()[name = tensor("op_1069"), val = tensor(-1)]; + tensor var_1086_axes_0 = const()[name = tensor("op_1086_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236600704)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236602816)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1086_cast = layer_norm(axes = var_1086_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1075_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast); + tensor var_1097_to_fp16 = const()[name = tensor("op_1097_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236604928)))]; + tensor var_1098_to_fp16 = const()[name = tensor("op_1098_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238702144)))]; + tensor q_37_cast = linear(bias = var_1098_to_fp16, weight = var_1097_to_fp16, x = var_1086_cast); + tensor var_1101_to_fp16 = const()[name = tensor("op_1101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238704256)))]; + tensor k_37_bias_0_to_fp16 = const()[name = tensor("k_37_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240801472)))]; + tensor k_37_cast = linear(bias = k_37_bias_0_to_fp16, weight = var_1101_to_fp16, x = var_1086_cast); + tensor var_1105_to_fp16 = const()[name = tensor("op_1105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240803584)))]; + tensor var_1106_to_fp16 = const()[name = tensor("op_1106_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242900800)))]; + tensor v_37_cast = linear(bias = var_1106_to_fp16, weight = var_1105_to_fp16, x = var_1086_cast); + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([1, 1500, 16, -1])]; + tensor var_1115_cast = reshape(shape = var_1114, x = q_37_cast); + tensor const_186_to_fp16 = const()[name = tensor("const_186_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast = mul(x = var_1115_cast, y = const_186_to_fp16); + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor([1, 1500, 16, -1])]; + tensor var_1122_cast = reshape(shape = var_1121, x = k_37_cast); + tensor const_187_to_fp16 = const()[name = tensor("const_187_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast = mul(x = var_1122_cast, y = const_187_to_fp16); + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1500, 16, -1])]; + tensor var_1129_cast = reshape(shape = var_1128, x = v_37_cast); + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_66_perm_0 = const()[name = tensor("transpose_66_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_67_perm_0 = const()[name = tensor("transpose_67_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_153 = transpose(perm = transpose_67_perm_0, x = k_39_cast); + tensor transpose_154 = transpose(perm = transpose_66_perm_0, x = q_39_cast); + tensor qk_19_cast = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_154, y = transpose_153); + tensor var_1134_cast = softmax(axis = var_1069, x = qk_19_cast); + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_155 = transpose(perm = var_1130, x = var_1129_cast); + tensor var_1136_cast = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = var_1134_cast, y = transpose_155); + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 1024])]; + tensor transpose_152 = transpose(perm = var_1137, x = var_1136_cast); + tensor x_119_cast = reshape(shape = concat_9, x = transpose_152); + tensor var_1142_to_fp16 = const()[name = tensor("op_1142_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242902912)))]; + tensor var_1143_to_fp16 = const()[name = tensor("op_1143_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245000128)))]; + tensor var_1144_cast = linear(bias = var_1143_to_fp16, weight = var_1142_to_fp16, x = x_119_cast); + tensor x_121_cast = add(x = x_115_cast, y = var_1144_cast); + tensor var_1150_axes_0 = const()[name = tensor("op_1150_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245002240)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245004352)))]; + tensor var_1150_cast = layer_norm(axes = var_1150_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1075_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast); + tensor var_1159_to_fp16 = const()[name = tensor("op_1159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245006464)))]; + tensor var_1160_to_fp16 = const()[name = tensor("op_1160_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253395136)))]; + tensor input_81_cast = linear(bias = var_1160_to_fp16, weight = var_1159_to_fp16, x = var_1150_cast); + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast = gelu(mode = x_125_mode_0, x = input_81_cast); + tensor var_1165_to_fp16 = const()[name = tensor("op_1165_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253403392)))]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261792064)))]; + tensor var_1167_cast = linear(bias = var_1166_to_fp16, weight = var_1165_to_fp16, x = x_125_cast); + tensor x_127_cast = add(x = x_121_cast, y = var_1167_cast); + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor(-1)]; + tensor var_1193_axes_0 = const()[name = tensor("op_1193_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261794176)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261796288)))]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1193_cast = layer_norm(axes = var_1193_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1182_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast); + tensor var_1204_to_fp16 = const()[name = tensor("op_1204_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261798400)))]; + tensor var_1205_to_fp16 = const()[name = tensor("op_1205_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263895616)))]; + tensor q_41_cast = linear(bias = var_1205_to_fp16, weight = var_1204_to_fp16, x = var_1193_cast); + tensor var_1208_to_fp16 = const()[name = tensor("op_1208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263897728)))]; + tensor k_41_bias_0_to_fp16 = const()[name = tensor("k_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265994944)))]; + tensor k_41_cast = linear(bias = k_41_bias_0_to_fp16, weight = var_1208_to_fp16, x = var_1193_cast); + tensor var_1212_to_fp16 = const()[name = tensor("op_1212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265997056)))]; + tensor var_1213_to_fp16 = const()[name = tensor("op_1213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268094272)))]; + tensor v_41_cast = linear(bias = var_1213_to_fp16, weight = var_1212_to_fp16, x = var_1193_cast); + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 1500, 16, -1])]; + tensor var_1222_cast = reshape(shape = var_1221, x = q_41_cast); + tensor const_188_to_fp16 = const()[name = tensor("const_188_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast = mul(x = var_1222_cast, y = const_188_to_fp16); + tensor var_1228 = const()[name = tensor("op_1228"), val = tensor([1, 1500, 16, -1])]; + tensor var_1229_cast = reshape(shape = var_1228, x = k_41_cast); + tensor const_189_to_fp16 = const()[name = tensor("const_189_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast = mul(x = var_1229_cast, y = const_189_to_fp16); + tensor var_1235 = const()[name = tensor("op_1235"), val = tensor([1, 1500, 16, -1])]; + tensor var_1236_cast = reshape(shape = var_1235, x = v_41_cast); + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_68_perm_0 = const()[name = tensor("transpose_68_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_149 = transpose(perm = transpose_69_perm_0, x = k_43_cast); + tensor transpose_150 = transpose(perm = transpose_68_perm_0, x = q_43_cast); + tensor qk_21_cast = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_150, y = transpose_149); + tensor var_1241_cast = softmax(axis = var_1176, x = qk_21_cast); + tensor var_1243_transpose_x_0 = const()[name = tensor("op_1243_transpose_x_0"), val = tensor(false)]; + tensor var_1243_transpose_y_0 = const()[name = tensor("op_1243_transpose_y_0"), val = tensor(false)]; + tensor transpose_151 = transpose(perm = var_1237, x = var_1236_cast); + tensor var_1243_cast = matmul(transpose_x = var_1243_transpose_x_0, transpose_y = var_1243_transpose_y_0, x = var_1241_cast, y = transpose_151); + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 1024])]; + tensor transpose_148 = transpose(perm = var_1244, x = var_1243_cast); + tensor x_131_cast = reshape(shape = concat_10, x = transpose_148); + tensor var_1249_to_fp16 = const()[name = tensor("op_1249_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268096384)))]; + tensor var_1250_to_fp16 = const()[name = tensor("op_1250_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270193600)))]; + tensor var_1251_cast = linear(bias = var_1250_to_fp16, weight = var_1249_to_fp16, x = x_131_cast); + tensor x_133_cast = add(x = x_127_cast, y = var_1251_cast); + tensor var_1257_axes_0 = const()[name = tensor("op_1257_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270195712)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270197824)))]; + tensor var_1257_cast = layer_norm(axes = var_1257_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1182_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast); + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270199936)))]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278588608)))]; + tensor input_89_cast = linear(bias = var_1267_to_fp16, weight = var_1266_to_fp16, x = var_1257_cast); + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast = gelu(mode = x_137_mode_0, x = input_89_cast); + tensor var_1272_to_fp16 = const()[name = tensor("op_1272_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278596864)))]; + tensor var_1273_to_fp16 = const()[name = tensor("op_1273_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286985536)))]; + tensor var_1274_cast = linear(bias = var_1273_to_fp16, weight = var_1272_to_fp16, x = x_137_cast); + tensor x_139_cast = add(x = x_133_cast, y = var_1274_cast); + tensor var_1283 = const()[name = tensor("op_1283"), val = tensor(-1)]; + tensor var_1300_axes_0 = const()[name = tensor("op_1300_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286987648)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286989760)))]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1300_cast = layer_norm(axes = var_1300_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1289_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast); + tensor var_1311_to_fp16 = const()[name = tensor("op_1311_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286991872)))]; + tensor var_1312_to_fp16 = const()[name = tensor("op_1312_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289089088)))]; + tensor q_45_cast = linear(bias = var_1312_to_fp16, weight = var_1311_to_fp16, x = var_1300_cast); + tensor var_1315_to_fp16 = const()[name = tensor("op_1315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289091200)))]; + tensor k_45_bias_0_to_fp16 = const()[name = tensor("k_45_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291188416)))]; + tensor k_45_cast = linear(bias = k_45_bias_0_to_fp16, weight = var_1315_to_fp16, x = var_1300_cast); + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291190528)))]; + tensor var_1320_to_fp16 = const()[name = tensor("op_1320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293287744)))]; + tensor v_45_cast = linear(bias = var_1320_to_fp16, weight = var_1319_to_fp16, x = var_1300_cast); + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([1, 1500, 16, -1])]; + tensor var_1329_cast = reshape(shape = var_1328, x = q_45_cast); + tensor const_190_to_fp16 = const()[name = tensor("const_190_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_47_cast = mul(x = var_1329_cast, y = const_190_to_fp16); + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 1500, 16, -1])]; + tensor var_1336_cast = reshape(shape = var_1335, x = k_45_cast); + tensor const_191_to_fp16 = const()[name = tensor("const_191_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_47_cast = mul(x = var_1336_cast, y = const_191_to_fp16); + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([1, 1500, 16, -1])]; + tensor var_1343_cast = reshape(shape = var_1342, x = v_45_cast); + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([0, 2, 1, 3])]; + tensor qk_23_transpose_x_0 = const()[name = tensor("qk_23_transpose_x_0"), val = tensor(false)]; + tensor qk_23_transpose_y_0 = const()[name = tensor("qk_23_transpose_y_0"), val = tensor(false)]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_145 = transpose(perm = transpose_71_perm_0, x = k_47_cast); + tensor transpose_146 = transpose(perm = transpose_70_perm_0, x = q_47_cast); + tensor qk_23_cast = matmul(transpose_x = qk_23_transpose_x_0, transpose_y = qk_23_transpose_y_0, x = transpose_146, y = transpose_145); + tensor var_1348_cast = softmax(axis = var_1283, x = qk_23_cast); + tensor var_1350_transpose_x_0 = const()[name = tensor("op_1350_transpose_x_0"), val = tensor(false)]; + tensor var_1350_transpose_y_0 = const()[name = tensor("op_1350_transpose_y_0"), val = tensor(false)]; + tensor transpose_147 = transpose(perm = var_1344, x = var_1343_cast); + tensor var_1350_cast = matmul(transpose_x = var_1350_transpose_x_0, transpose_y = var_1350_transpose_y_0, x = var_1348_cast, y = transpose_147); + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 1024])]; + tensor transpose_144 = transpose(perm = var_1351, x = var_1350_cast); + tensor x_143_cast = reshape(shape = concat_11, x = transpose_144); + tensor var_1356_to_fp16 = const()[name = tensor("op_1356_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293289856)))]; + tensor var_1357_to_fp16 = const()[name = tensor("op_1357_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295387072)))]; + tensor var_1358_cast = linear(bias = var_1357_to_fp16, weight = var_1356_to_fp16, x = x_143_cast); + tensor x_145_cast = add(x = x_139_cast, y = var_1358_cast); + tensor var_1364_axes_0 = const()[name = tensor("op_1364_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295389184)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295391296)))]; + tensor var_1364_cast = layer_norm(axes = var_1364_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1289_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast); + tensor var_1373_to_fp16 = const()[name = tensor("op_1373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295393408)))]; + tensor var_1374_to_fp16 = const()[name = tensor("op_1374_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782080)))]; + tensor input_97_cast = linear(bias = var_1374_to_fp16, weight = var_1373_to_fp16, x = var_1364_cast); + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast = gelu(mode = x_149_mode_0, x = input_97_cast); + tensor var_1379_to_fp16 = const()[name = tensor("op_1379_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790336)))]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312179008)))]; + tensor var_1381_cast = linear(bias = var_1380_to_fp16, weight = var_1379_to_fp16, x = x_149_cast); + tensor x_151_cast = add(x = x_145_cast, y = var_1381_cast); + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor(-1)]; + tensor var_1407_axes_0 = const()[name = tensor("op_1407_axes_0"), val = tensor([-1])]; + tensor blocks_12_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_12_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312181120)))]; + tensor blocks_12_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_12_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312183232)))]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1407_cast = layer_norm(axes = var_1407_axes_0, beta = blocks_12_attn_ln_bias_to_fp16, epsilon = var_1396_to_fp16, gamma = blocks_12_attn_ln_weight_to_fp16, x = x_151_cast); + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312185344)))]; + tensor var_1419_to_fp16 = const()[name = tensor("op_1419_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314282560)))]; + tensor q_49_cast = linear(bias = var_1419_to_fp16, weight = var_1418_to_fp16, x = var_1407_cast); + tensor var_1422_to_fp16 = const()[name = tensor("op_1422_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314284672)))]; + tensor k_49_bias_0_to_fp16 = const()[name = tensor("k_49_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316381888)))]; + tensor k_49_cast = linear(bias = k_49_bias_0_to_fp16, weight = var_1422_to_fp16, x = var_1407_cast); + tensor var_1426_to_fp16 = const()[name = tensor("op_1426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316384000)))]; + tensor var_1427_to_fp16 = const()[name = tensor("op_1427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318481216)))]; + tensor v_49_cast = linear(bias = var_1427_to_fp16, weight = var_1426_to_fp16, x = var_1407_cast); + tensor var_1435 = const()[name = tensor("op_1435"), val = tensor([1, 1500, 16, -1])]; + tensor var_1436_cast = reshape(shape = var_1435, x = q_49_cast); + tensor const_192_to_fp16 = const()[name = tensor("const_192_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_51_cast = mul(x = var_1436_cast, y = const_192_to_fp16); + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1500, 16, -1])]; + tensor var_1443_cast = reshape(shape = var_1442, x = k_49_cast); + tensor const_193_to_fp16 = const()[name = tensor("const_193_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_51_cast = mul(x = var_1443_cast, y = const_193_to_fp16); + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1500, 16, -1])]; + tensor var_1450_cast = reshape(shape = var_1449, x = v_49_cast); + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([0, 2, 1, 3])]; + tensor qk_25_transpose_x_0 = const()[name = tensor("qk_25_transpose_x_0"), val = tensor(false)]; + tensor qk_25_transpose_y_0 = const()[name = tensor("qk_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_141 = transpose(perm = transpose_73_perm_0, x = k_51_cast); + tensor transpose_142 = transpose(perm = transpose_72_perm_0, x = q_51_cast); + tensor qk_25_cast = matmul(transpose_x = qk_25_transpose_x_0, transpose_y = qk_25_transpose_y_0, x = transpose_142, y = transpose_141); + tensor var_1455_cast = softmax(axis = var_1390, x = qk_25_cast); + tensor var_1457_transpose_x_0 = const()[name = tensor("op_1457_transpose_x_0"), val = tensor(false)]; + tensor var_1457_transpose_y_0 = const()[name = tensor("op_1457_transpose_y_0"), val = tensor(false)]; + tensor transpose_143 = transpose(perm = var_1451, x = var_1450_cast); + tensor var_1457_cast = matmul(transpose_x = var_1457_transpose_x_0, transpose_y = var_1457_transpose_y_0, x = var_1455_cast, y = transpose_143); + tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([0, 2, 1, 3])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 1500, 1024])]; + tensor transpose_140 = transpose(perm = var_1458, x = var_1457_cast); + tensor x_155_cast = reshape(shape = concat_12, x = transpose_140); + tensor var_1463_to_fp16 = const()[name = tensor("op_1463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318483328)))]; + tensor var_1464_to_fp16 = const()[name = tensor("op_1464_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320580544)))]; + tensor var_1465_cast = linear(bias = var_1464_to_fp16, weight = var_1463_to_fp16, x = x_155_cast); + tensor x_157_cast = add(x = x_151_cast, y = var_1465_cast); + tensor var_1471_axes_0 = const()[name = tensor("op_1471_axes_0"), val = tensor([-1])]; + tensor blocks_12_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320582656)))]; + tensor blocks_12_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320584768)))]; + tensor var_1471_cast = layer_norm(axes = var_1471_axes_0, beta = blocks_12_mlp_ln_bias_to_fp16, epsilon = var_1396_to_fp16, gamma = blocks_12_mlp_ln_weight_to_fp16, x = x_157_cast); + tensor var_1480_to_fp16 = const()[name = tensor("op_1480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320586880)))]; + tensor var_1481_to_fp16 = const()[name = tensor("op_1481_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328975552)))]; + tensor input_105_cast = linear(bias = var_1481_to_fp16, weight = var_1480_to_fp16, x = var_1471_cast); + tensor x_161_mode_0 = const()[name = tensor("x_161_mode_0"), val = tensor("EXACT")]; + tensor x_161_cast = gelu(mode = x_161_mode_0, x = input_105_cast); + tensor var_1486_to_fp16 = const()[name = tensor("op_1486_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328983808)))]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337372480)))]; + tensor var_1488_cast = linear(bias = var_1487_to_fp16, weight = var_1486_to_fp16, x = x_161_cast); + tensor x_163_cast = add(x = x_157_cast, y = var_1488_cast); + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor(-1)]; + tensor var_1514_axes_0 = const()[name = tensor("op_1514_axes_0"), val = tensor([-1])]; + tensor blocks_13_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_13_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337374592)))]; + tensor blocks_13_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_13_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337376704)))]; + tensor var_1503_to_fp16 = const()[name = tensor("op_1503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1514_cast = layer_norm(axes = var_1514_axes_0, beta = blocks_13_attn_ln_bias_to_fp16, epsilon = var_1503_to_fp16, gamma = blocks_13_attn_ln_weight_to_fp16, x = x_163_cast); + tensor var_1525_to_fp16 = const()[name = tensor("op_1525_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337378816)))]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339476032)))]; + tensor q_53_cast = linear(bias = var_1526_to_fp16, weight = var_1525_to_fp16, x = var_1514_cast); + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339478144)))]; + tensor k_53_bias_0_to_fp16 = const()[name = tensor("k_53_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341575360)))]; + tensor k_53_cast = linear(bias = k_53_bias_0_to_fp16, weight = var_1529_to_fp16, x = var_1514_cast); + tensor var_1533_to_fp16 = const()[name = tensor("op_1533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341577472)))]; + tensor var_1534_to_fp16 = const()[name = tensor("op_1534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343674688)))]; + tensor v_53_cast = linear(bias = var_1534_to_fp16, weight = var_1533_to_fp16, x = var_1514_cast); + tensor var_1542 = const()[name = tensor("op_1542"), val = tensor([1, 1500, 16, -1])]; + tensor var_1543_cast = reshape(shape = var_1542, x = q_53_cast); + tensor const_194_to_fp16 = const()[name = tensor("const_194_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_55_cast = mul(x = var_1543_cast, y = const_194_to_fp16); + tensor var_1549 = const()[name = tensor("op_1549"), val = tensor([1, 1500, 16, -1])]; + tensor var_1550_cast = reshape(shape = var_1549, x = k_53_cast); + tensor const_195_to_fp16 = const()[name = tensor("const_195_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_55_cast = mul(x = var_1550_cast, y = const_195_to_fp16); + tensor var_1556 = const()[name = tensor("op_1556"), val = tensor([1, 1500, 16, -1])]; + tensor var_1557_cast = reshape(shape = var_1556, x = v_53_cast); + tensor var_1558 = const()[name = tensor("op_1558"), val = tensor([0, 2, 1, 3])]; + tensor qk_27_transpose_x_0 = const()[name = tensor("qk_27_transpose_x_0"), val = tensor(false)]; + tensor qk_27_transpose_y_0 = const()[name = tensor("qk_27_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_137 = transpose(perm = transpose_75_perm_0, x = k_55_cast); + tensor transpose_138 = transpose(perm = transpose_74_perm_0, x = q_55_cast); + tensor qk_27_cast = matmul(transpose_x = qk_27_transpose_x_0, transpose_y = qk_27_transpose_y_0, x = transpose_138, y = transpose_137); + tensor var_1562_cast = softmax(axis = var_1497, x = qk_27_cast); + tensor var_1564_transpose_x_0 = const()[name = tensor("op_1564_transpose_x_0"), val = tensor(false)]; + tensor var_1564_transpose_y_0 = const()[name = tensor("op_1564_transpose_y_0"), val = tensor(false)]; + tensor transpose_139 = transpose(perm = var_1558, x = var_1557_cast); + tensor var_1564_cast = matmul(transpose_x = var_1564_transpose_x_0, transpose_y = var_1564_transpose_y_0, x = var_1562_cast, y = transpose_139); + tensor var_1565 = const()[name = tensor("op_1565"), val = tensor([0, 2, 1, 3])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([1, 1500, 1024])]; + tensor transpose_136 = transpose(perm = var_1565, x = var_1564_cast); + tensor x_167_cast = reshape(shape = concat_13, x = transpose_136); + tensor var_1570_to_fp16 = const()[name = tensor("op_1570_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343676800)))]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345774016)))]; + tensor var_1572_cast = linear(bias = var_1571_to_fp16, weight = var_1570_to_fp16, x = x_167_cast); + tensor x_169_cast = add(x = x_163_cast, y = var_1572_cast); + tensor var_1578_axes_0 = const()[name = tensor("op_1578_axes_0"), val = tensor([-1])]; + tensor blocks_13_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345776128)))]; + tensor blocks_13_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345778240)))]; + tensor var_1578_cast = layer_norm(axes = var_1578_axes_0, beta = blocks_13_mlp_ln_bias_to_fp16, epsilon = var_1503_to_fp16, gamma = blocks_13_mlp_ln_weight_to_fp16, x = x_169_cast); + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345780352)))]; + tensor var_1588_to_fp16 = const()[name = tensor("op_1588_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354169024)))]; + tensor input_113_cast = linear(bias = var_1588_to_fp16, weight = var_1587_to_fp16, x = var_1578_cast); + tensor x_173_mode_0 = const()[name = tensor("x_173_mode_0"), val = tensor("EXACT")]; + tensor x_173_cast = gelu(mode = x_173_mode_0, x = input_113_cast); + tensor var_1593_to_fp16 = const()[name = tensor("op_1593_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354177280)))]; + tensor var_1594_to_fp16 = const()[name = tensor("op_1594_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362565952)))]; + tensor var_1595_cast = linear(bias = var_1594_to_fp16, weight = var_1593_to_fp16, x = x_173_cast); + tensor x_175_cast = add(x = x_169_cast, y = var_1595_cast); + tensor var_1604 = const()[name = tensor("op_1604"), val = tensor(-1)]; + tensor var_1621_axes_0 = const()[name = tensor("op_1621_axes_0"), val = tensor([-1])]; + tensor blocks_14_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_14_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362568064)))]; + tensor blocks_14_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_14_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362570176)))]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1621_cast = layer_norm(axes = var_1621_axes_0, beta = blocks_14_attn_ln_bias_to_fp16, epsilon = var_1610_to_fp16, gamma = blocks_14_attn_ln_weight_to_fp16, x = x_175_cast); + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362572288)))]; + tensor var_1633_to_fp16 = const()[name = tensor("op_1633_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364669504)))]; + tensor q_57_cast = linear(bias = var_1633_to_fp16, weight = var_1632_to_fp16, x = var_1621_cast); + tensor var_1636_to_fp16 = const()[name = tensor("op_1636_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364671616)))]; + tensor k_57_bias_0_to_fp16 = const()[name = tensor("k_57_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366768832)))]; + tensor k_57_cast = linear(bias = k_57_bias_0_to_fp16, weight = var_1636_to_fp16, x = var_1621_cast); + tensor var_1640_to_fp16 = const()[name = tensor("op_1640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366770944)))]; + tensor var_1641_to_fp16 = const()[name = tensor("op_1641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368868160)))]; + tensor v_57_cast = linear(bias = var_1641_to_fp16, weight = var_1640_to_fp16, x = var_1621_cast); + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor([1, 1500, 16, -1])]; + tensor var_1650_cast = reshape(shape = var_1649, x = q_57_cast); + tensor const_196_to_fp16 = const()[name = tensor("const_196_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_59_cast = mul(x = var_1650_cast, y = const_196_to_fp16); + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1500, 16, -1])]; + tensor var_1657_cast = reshape(shape = var_1656, x = k_57_cast); + tensor const_197_to_fp16 = const()[name = tensor("const_197_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_59_cast = mul(x = var_1657_cast, y = const_197_to_fp16); + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1500, 16, -1])]; + tensor var_1664_cast = reshape(shape = var_1663, x = v_57_cast); + tensor var_1665 = const()[name = tensor("op_1665"), val = tensor([0, 2, 1, 3])]; + tensor qk_29_transpose_x_0 = const()[name = tensor("qk_29_transpose_x_0"), val = tensor(false)]; + tensor qk_29_transpose_y_0 = const()[name = tensor("qk_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_133 = transpose(perm = transpose_77_perm_0, x = k_59_cast); + tensor transpose_134 = transpose(perm = transpose_76_perm_0, x = q_59_cast); + tensor qk_29_cast = matmul(transpose_x = qk_29_transpose_x_0, transpose_y = qk_29_transpose_y_0, x = transpose_134, y = transpose_133); + tensor var_1669_cast = softmax(axis = var_1604, x = qk_29_cast); + tensor var_1671_transpose_x_0 = const()[name = tensor("op_1671_transpose_x_0"), val = tensor(false)]; + tensor var_1671_transpose_y_0 = const()[name = tensor("op_1671_transpose_y_0"), val = tensor(false)]; + tensor transpose_135 = transpose(perm = var_1665, x = var_1664_cast); + tensor var_1671_cast = matmul(transpose_x = var_1671_transpose_x_0, transpose_y = var_1671_transpose_y_0, x = var_1669_cast, y = transpose_135); + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([0, 2, 1, 3])]; + tensor concat_14 = const()[name = tensor("concat_14"), val = tensor([1, 1500, 1024])]; + tensor transpose_132 = transpose(perm = var_1672, x = var_1671_cast); + tensor x_179_cast = reshape(shape = concat_14, x = transpose_132); + tensor var_1677_to_fp16 = const()[name = tensor("op_1677_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368870272)))]; + tensor var_1678_to_fp16 = const()[name = tensor("op_1678_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370967488)))]; + tensor var_1679_cast = linear(bias = var_1678_to_fp16, weight = var_1677_to_fp16, x = x_179_cast); + tensor x_181_cast = add(x = x_175_cast, y = var_1679_cast); + tensor var_1685_axes_0 = const()[name = tensor("op_1685_axes_0"), val = tensor([-1])]; + tensor blocks_14_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370969600)))]; + tensor blocks_14_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370971712)))]; + tensor var_1685_cast = layer_norm(axes = var_1685_axes_0, beta = blocks_14_mlp_ln_bias_to_fp16, epsilon = var_1610_to_fp16, gamma = blocks_14_mlp_ln_weight_to_fp16, x = x_181_cast); + tensor var_1694_to_fp16 = const()[name = tensor("op_1694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370973824)))]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379362496)))]; + tensor input_121_cast = linear(bias = var_1695_to_fp16, weight = var_1694_to_fp16, x = var_1685_cast); + tensor x_185_mode_0 = const()[name = tensor("x_185_mode_0"), val = tensor("EXACT")]; + tensor x_185_cast = gelu(mode = x_185_mode_0, x = input_121_cast); + tensor var_1700_to_fp16 = const()[name = tensor("op_1700_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379370752)))]; + tensor var_1701_to_fp16 = const()[name = tensor("op_1701_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387759424)))]; + tensor var_1702_cast = linear(bias = var_1701_to_fp16, weight = var_1700_to_fp16, x = x_185_cast); + tensor x_187_cast = add(x = x_181_cast, y = var_1702_cast); + tensor var_1711 = const()[name = tensor("op_1711"), val = tensor(-1)]; + tensor var_1728_axes_0 = const()[name = tensor("op_1728_axes_0"), val = tensor([-1])]; + tensor blocks_15_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_15_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387761536)))]; + tensor blocks_15_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_15_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387763648)))]; + tensor var_1717_to_fp16 = const()[name = tensor("op_1717_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1728_cast = layer_norm(axes = var_1728_axes_0, beta = blocks_15_attn_ln_bias_to_fp16, epsilon = var_1717_to_fp16, gamma = blocks_15_attn_ln_weight_to_fp16, x = x_187_cast); + tensor var_1739_to_fp16 = const()[name = tensor("op_1739_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387765760)))]; + tensor var_1740_to_fp16 = const()[name = tensor("op_1740_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389862976)))]; + tensor q_61_cast = linear(bias = var_1740_to_fp16, weight = var_1739_to_fp16, x = var_1728_cast); + tensor var_1743_to_fp16 = const()[name = tensor("op_1743_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389865088)))]; + tensor k_61_bias_0_to_fp16 = const()[name = tensor("k_61_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391962304)))]; + tensor k_61_cast = linear(bias = k_61_bias_0_to_fp16, weight = var_1743_to_fp16, x = var_1728_cast); + tensor var_1747_to_fp16 = const()[name = tensor("op_1747_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391964416)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394061632)))]; + tensor v_61_cast = linear(bias = var_1748_to_fp16, weight = var_1747_to_fp16, x = var_1728_cast); + tensor var_1756 = const()[name = tensor("op_1756"), val = tensor([1, 1500, 16, -1])]; + tensor var_1757_cast = reshape(shape = var_1756, x = q_61_cast); + tensor const_198_to_fp16 = const()[name = tensor("const_198_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_63_cast = mul(x = var_1757_cast, y = const_198_to_fp16); + tensor var_1763 = const()[name = tensor("op_1763"), val = tensor([1, 1500, 16, -1])]; + tensor var_1764_cast = reshape(shape = var_1763, x = k_61_cast); + tensor const_199_to_fp16 = const()[name = tensor("const_199_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_63_cast = mul(x = var_1764_cast, y = const_199_to_fp16); + tensor var_1770 = const()[name = tensor("op_1770"), val = tensor([1, 1500, 16, -1])]; + tensor var_1771_cast = reshape(shape = var_1770, x = v_61_cast); + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([0, 2, 1, 3])]; + tensor qk_31_transpose_x_0 = const()[name = tensor("qk_31_transpose_x_0"), val = tensor(false)]; + tensor qk_31_transpose_y_0 = const()[name = tensor("qk_31_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_129 = transpose(perm = transpose_79_perm_0, x = k_63_cast); + tensor transpose_130 = transpose(perm = transpose_78_perm_0, x = q_63_cast); + tensor qk_31_cast = matmul(transpose_x = qk_31_transpose_x_0, transpose_y = qk_31_transpose_y_0, x = transpose_130, y = transpose_129); + tensor var_1776_cast = softmax(axis = var_1711, x = qk_31_cast); + tensor var_1778_transpose_x_0 = const()[name = tensor("op_1778_transpose_x_0"), val = tensor(false)]; + tensor var_1778_transpose_y_0 = const()[name = tensor("op_1778_transpose_y_0"), val = tensor(false)]; + tensor transpose_131 = transpose(perm = var_1772, x = var_1771_cast); + tensor var_1778_cast = matmul(transpose_x = var_1778_transpose_x_0, transpose_y = var_1778_transpose_y_0, x = var_1776_cast, y = transpose_131); + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([0, 2, 1, 3])]; + tensor concat_15 = const()[name = tensor("concat_15"), val = tensor([1, 1500, 1024])]; + tensor transpose_128 = transpose(perm = var_1779, x = var_1778_cast); + tensor x_191_cast = reshape(shape = concat_15, x = transpose_128); + tensor var_1784_to_fp16 = const()[name = tensor("op_1784_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394063744)))]; + tensor var_1785_to_fp16 = const()[name = tensor("op_1785_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396160960)))]; + tensor var_1786_cast = linear(bias = var_1785_to_fp16, weight = var_1784_to_fp16, x = x_191_cast); + tensor x_193_cast = add(x = x_187_cast, y = var_1786_cast); + tensor var_1792_axes_0 = const()[name = tensor("op_1792_axes_0"), val = tensor([-1])]; + tensor blocks_15_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396163072)))]; + tensor blocks_15_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396165184)))]; + tensor var_1792_cast = layer_norm(axes = var_1792_axes_0, beta = blocks_15_mlp_ln_bias_to_fp16, epsilon = var_1717_to_fp16, gamma = blocks_15_mlp_ln_weight_to_fp16, x = x_193_cast); + tensor var_1801_to_fp16 = const()[name = tensor("op_1801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396167296)))]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404555968)))]; + tensor input_129_cast = linear(bias = var_1802_to_fp16, weight = var_1801_to_fp16, x = var_1792_cast); + tensor x_197_mode_0 = const()[name = tensor("x_197_mode_0"), val = tensor("EXACT")]; + tensor x_197_cast = gelu(mode = x_197_mode_0, x = input_129_cast); + tensor var_1807_to_fp16 = const()[name = tensor("op_1807_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404564224)))]; + tensor var_1808_to_fp16 = const()[name = tensor("op_1808_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412952896)))]; + tensor var_1809_cast = linear(bias = var_1808_to_fp16, weight = var_1807_to_fp16, x = x_197_cast); + tensor x_199_cast = add(x = x_193_cast, y = var_1809_cast); + tensor var_1818 = const()[name = tensor("op_1818"), val = tensor(-1)]; + tensor var_1835_axes_0 = const()[name = tensor("op_1835_axes_0"), val = tensor([-1])]; + tensor blocks_16_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_16_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412955008)))]; + tensor blocks_16_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_16_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412957120)))]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1835_cast = layer_norm(axes = var_1835_axes_0, beta = blocks_16_attn_ln_bias_to_fp16, epsilon = var_1824_to_fp16, gamma = blocks_16_attn_ln_weight_to_fp16, x = x_199_cast); + tensor var_1846_to_fp16 = const()[name = tensor("op_1846_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412959232)))]; + tensor var_1847_to_fp16 = const()[name = tensor("op_1847_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415056448)))]; + tensor q_65_cast = linear(bias = var_1847_to_fp16, weight = var_1846_to_fp16, x = var_1835_cast); + tensor var_1850_to_fp16 = const()[name = tensor("op_1850_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415058560)))]; + tensor k_65_bias_0_to_fp16 = const()[name = tensor("k_65_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417155776)))]; + tensor k_65_cast = linear(bias = k_65_bias_0_to_fp16, weight = var_1850_to_fp16, x = var_1835_cast); + tensor var_1854_to_fp16 = const()[name = tensor("op_1854_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417157888)))]; + tensor var_1855_to_fp16 = const()[name = tensor("op_1855_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419255104)))]; + tensor v_65_cast = linear(bias = var_1855_to_fp16, weight = var_1854_to_fp16, x = var_1835_cast); + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 1500, 16, -1])]; + tensor var_1864_cast = reshape(shape = var_1863, x = q_65_cast); + tensor const_200_to_fp16 = const()[name = tensor("const_200_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_67_cast = mul(x = var_1864_cast, y = const_200_to_fp16); + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 1500, 16, -1])]; + tensor var_1871_cast = reshape(shape = var_1870, x = k_65_cast); + tensor const_201_to_fp16 = const()[name = tensor("const_201_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_67_cast = mul(x = var_1871_cast, y = const_201_to_fp16); + tensor var_1877 = const()[name = tensor("op_1877"), val = tensor([1, 1500, 16, -1])]; + tensor var_1878_cast = reshape(shape = var_1877, x = v_65_cast); + tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([0, 2, 1, 3])]; + tensor qk_33_transpose_x_0 = const()[name = tensor("qk_33_transpose_x_0"), val = tensor(false)]; + tensor qk_33_transpose_y_0 = const()[name = tensor("qk_33_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_125 = transpose(perm = transpose_81_perm_0, x = k_67_cast); + tensor transpose_126 = transpose(perm = transpose_80_perm_0, x = q_67_cast); + tensor qk_33_cast = matmul(transpose_x = qk_33_transpose_x_0, transpose_y = qk_33_transpose_y_0, x = transpose_126, y = transpose_125); + tensor var_1883_cast = softmax(axis = var_1818, x = qk_33_cast); + tensor var_1885_transpose_x_0 = const()[name = tensor("op_1885_transpose_x_0"), val = tensor(false)]; + tensor var_1885_transpose_y_0 = const()[name = tensor("op_1885_transpose_y_0"), val = tensor(false)]; + tensor transpose_127 = transpose(perm = var_1879, x = var_1878_cast); + tensor var_1885_cast = matmul(transpose_x = var_1885_transpose_x_0, transpose_y = var_1885_transpose_y_0, x = var_1883_cast, y = transpose_127); + tensor var_1886 = const()[name = tensor("op_1886"), val = tensor([0, 2, 1, 3])]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([1, 1500, 1024])]; + tensor transpose_124 = transpose(perm = var_1886, x = var_1885_cast); + tensor x_203_cast = reshape(shape = concat_16, x = transpose_124); + tensor var_1891_to_fp16 = const()[name = tensor("op_1891_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419257216)))]; + tensor var_1892_to_fp16 = const()[name = tensor("op_1892_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421354432)))]; + tensor var_1893_cast = linear(bias = var_1892_to_fp16, weight = var_1891_to_fp16, x = x_203_cast); + tensor x_205_cast = add(x = x_199_cast, y = var_1893_cast); + tensor var_1899_axes_0 = const()[name = tensor("op_1899_axes_0"), val = tensor([-1])]; + tensor blocks_16_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421356544)))]; + tensor blocks_16_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421358656)))]; + tensor var_1899_cast = layer_norm(axes = var_1899_axes_0, beta = blocks_16_mlp_ln_bias_to_fp16, epsilon = var_1824_to_fp16, gamma = blocks_16_mlp_ln_weight_to_fp16, x = x_205_cast); + tensor var_1908_to_fp16 = const()[name = tensor("op_1908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421360768)))]; + tensor var_1909_to_fp16 = const()[name = tensor("op_1909_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429749440)))]; + tensor input_137_cast = linear(bias = var_1909_to_fp16, weight = var_1908_to_fp16, x = var_1899_cast); + tensor x_209_mode_0 = const()[name = tensor("x_209_mode_0"), val = tensor("EXACT")]; + tensor x_209_cast = gelu(mode = x_209_mode_0, x = input_137_cast); + tensor var_1914_to_fp16 = const()[name = tensor("op_1914_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429757696)))]; + tensor var_1915_to_fp16 = const()[name = tensor("op_1915_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438146368)))]; + tensor var_1916_cast = linear(bias = var_1915_to_fp16, weight = var_1914_to_fp16, x = x_209_cast); + tensor x_211_cast = add(x = x_205_cast, y = var_1916_cast); + tensor var_1925 = const()[name = tensor("op_1925"), val = tensor(-1)]; + tensor var_1942_axes_0 = const()[name = tensor("op_1942_axes_0"), val = tensor([-1])]; + tensor blocks_17_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_17_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438148480)))]; + tensor blocks_17_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_17_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438150592)))]; + tensor var_1931_to_fp16 = const()[name = tensor("op_1931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1942_cast = layer_norm(axes = var_1942_axes_0, beta = blocks_17_attn_ln_bias_to_fp16, epsilon = var_1931_to_fp16, gamma = blocks_17_attn_ln_weight_to_fp16, x = x_211_cast); + tensor var_1953_to_fp16 = const()[name = tensor("op_1953_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438152704)))]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440249920)))]; + tensor q_69_cast = linear(bias = var_1954_to_fp16, weight = var_1953_to_fp16, x = var_1942_cast); + tensor var_1957_to_fp16 = const()[name = tensor("op_1957_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440252032)))]; + tensor k_69_bias_0_to_fp16 = const()[name = tensor("k_69_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442349248)))]; + tensor k_69_cast = linear(bias = k_69_bias_0_to_fp16, weight = var_1957_to_fp16, x = var_1942_cast); + tensor var_1961_to_fp16 = const()[name = tensor("op_1961_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442351360)))]; + tensor var_1962_to_fp16 = const()[name = tensor("op_1962_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444448576)))]; + tensor v_69_cast = linear(bias = var_1962_to_fp16, weight = var_1961_to_fp16, x = var_1942_cast); + tensor var_1970 = const()[name = tensor("op_1970"), val = tensor([1, 1500, 16, -1])]; + tensor var_1971_cast = reshape(shape = var_1970, x = q_69_cast); + tensor const_202_to_fp16 = const()[name = tensor("const_202_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_71_cast = mul(x = var_1971_cast, y = const_202_to_fp16); + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1500, 16, -1])]; + tensor var_1978_cast = reshape(shape = var_1977, x = k_69_cast); + tensor const_203_to_fp16 = const()[name = tensor("const_203_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_71_cast = mul(x = var_1978_cast, y = const_203_to_fp16); + tensor var_1984 = const()[name = tensor("op_1984"), val = tensor([1, 1500, 16, -1])]; + tensor var_1985_cast = reshape(shape = var_1984, x = v_69_cast); + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([0, 2, 1, 3])]; + tensor qk_35_transpose_x_0 = const()[name = tensor("qk_35_transpose_x_0"), val = tensor(false)]; + tensor qk_35_transpose_y_0 = const()[name = tensor("qk_35_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_121 = transpose(perm = transpose_83_perm_0, x = k_71_cast); + tensor transpose_122 = transpose(perm = transpose_82_perm_0, x = q_71_cast); + tensor qk_35_cast = matmul(transpose_x = qk_35_transpose_x_0, transpose_y = qk_35_transpose_y_0, x = transpose_122, y = transpose_121); + tensor var_1990_cast = softmax(axis = var_1925, x = qk_35_cast); + tensor var_1992_transpose_x_0 = const()[name = tensor("op_1992_transpose_x_0"), val = tensor(false)]; + tensor var_1992_transpose_y_0 = const()[name = tensor("op_1992_transpose_y_0"), val = tensor(false)]; + tensor transpose_123 = transpose(perm = var_1986, x = var_1985_cast); + tensor var_1992_cast = matmul(transpose_x = var_1992_transpose_x_0, transpose_y = var_1992_transpose_y_0, x = var_1990_cast, y = transpose_123); + tensor var_1993 = const()[name = tensor("op_1993"), val = tensor([0, 2, 1, 3])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([1, 1500, 1024])]; + tensor transpose_120 = transpose(perm = var_1993, x = var_1992_cast); + tensor x_215_cast = reshape(shape = concat_17, x = transpose_120); + tensor var_1998_to_fp16 = const()[name = tensor("op_1998_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444450688)))]; + tensor var_1999_to_fp16 = const()[name = tensor("op_1999_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446547904)))]; + tensor var_2000_cast = linear(bias = var_1999_to_fp16, weight = var_1998_to_fp16, x = x_215_cast); + tensor x_217_cast = add(x = x_211_cast, y = var_2000_cast); + tensor var_2006_axes_0 = const()[name = tensor("op_2006_axes_0"), val = tensor([-1])]; + tensor blocks_17_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446550016)))]; + tensor blocks_17_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446552128)))]; + tensor var_2006_cast = layer_norm(axes = var_2006_axes_0, beta = blocks_17_mlp_ln_bias_to_fp16, epsilon = var_1931_to_fp16, gamma = blocks_17_mlp_ln_weight_to_fp16, x = x_217_cast); + tensor var_2015_to_fp16 = const()[name = tensor("op_2015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446554240)))]; + tensor var_2016_to_fp16 = const()[name = tensor("op_2016_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454942912)))]; + tensor input_145_cast = linear(bias = var_2016_to_fp16, weight = var_2015_to_fp16, x = var_2006_cast); + tensor x_221_mode_0 = const()[name = tensor("x_221_mode_0"), val = tensor("EXACT")]; + tensor x_221_cast = gelu(mode = x_221_mode_0, x = input_145_cast); + tensor var_2021_to_fp16 = const()[name = tensor("op_2021_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454951168)))]; + tensor var_2022_to_fp16 = const()[name = tensor("op_2022_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463339840)))]; + tensor var_2023_cast = linear(bias = var_2022_to_fp16, weight = var_2021_to_fp16, x = x_221_cast); + tensor x_223_cast = add(x = x_217_cast, y = var_2023_cast); + tensor var_2032 = const()[name = tensor("op_2032"), val = tensor(-1)]; + tensor var_2049_axes_0 = const()[name = tensor("op_2049_axes_0"), val = tensor([-1])]; + tensor blocks_18_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_18_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463341952)))]; + tensor blocks_18_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_18_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463344064)))]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2049_cast = layer_norm(axes = var_2049_axes_0, beta = blocks_18_attn_ln_bias_to_fp16, epsilon = var_2038_to_fp16, gamma = blocks_18_attn_ln_weight_to_fp16, x = x_223_cast); + tensor var_2060_to_fp16 = const()[name = tensor("op_2060_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463346176)))]; + tensor var_2061_to_fp16 = const()[name = tensor("op_2061_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465443392)))]; + tensor q_73_cast = linear(bias = var_2061_to_fp16, weight = var_2060_to_fp16, x = var_2049_cast); + tensor var_2064_to_fp16 = const()[name = tensor("op_2064_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465445504)))]; + tensor k_73_bias_0_to_fp16 = const()[name = tensor("k_73_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467542720)))]; + tensor k_73_cast = linear(bias = k_73_bias_0_to_fp16, weight = var_2064_to_fp16, x = var_2049_cast); + tensor var_2068_to_fp16 = const()[name = tensor("op_2068_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467544832)))]; + tensor var_2069_to_fp16 = const()[name = tensor("op_2069_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469642048)))]; + tensor v_73_cast = linear(bias = var_2069_to_fp16, weight = var_2068_to_fp16, x = var_2049_cast); + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1500, 16, -1])]; + tensor var_2078_cast = reshape(shape = var_2077, x = q_73_cast); + tensor const_204_to_fp16 = const()[name = tensor("const_204_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_75_cast = mul(x = var_2078_cast, y = const_204_to_fp16); + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1500, 16, -1])]; + tensor var_2085_cast = reshape(shape = var_2084, x = k_73_cast); + tensor const_205_to_fp16 = const()[name = tensor("const_205_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_75_cast = mul(x = var_2085_cast, y = const_205_to_fp16); + tensor var_2091 = const()[name = tensor("op_2091"), val = tensor([1, 1500, 16, -1])]; + tensor var_2092_cast = reshape(shape = var_2091, x = v_73_cast); + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([0, 2, 1, 3])]; + tensor qk_37_transpose_x_0 = const()[name = tensor("qk_37_transpose_x_0"), val = tensor(false)]; + tensor qk_37_transpose_y_0 = const()[name = tensor("qk_37_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_117 = transpose(perm = transpose_85_perm_0, x = k_75_cast); + tensor transpose_118 = transpose(perm = transpose_84_perm_0, x = q_75_cast); + tensor qk_37_cast = matmul(transpose_x = qk_37_transpose_x_0, transpose_y = qk_37_transpose_y_0, x = transpose_118, y = transpose_117); + tensor var_2097_cast = softmax(axis = var_2032, x = qk_37_cast); + tensor var_2099_transpose_x_0 = const()[name = tensor("op_2099_transpose_x_0"), val = tensor(false)]; + tensor var_2099_transpose_y_0 = const()[name = tensor("op_2099_transpose_y_0"), val = tensor(false)]; + tensor transpose_119 = transpose(perm = var_2093, x = var_2092_cast); + tensor var_2099_cast = matmul(transpose_x = var_2099_transpose_x_0, transpose_y = var_2099_transpose_y_0, x = var_2097_cast, y = transpose_119); + tensor var_2100 = const()[name = tensor("op_2100"), val = tensor([0, 2, 1, 3])]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 1500, 1024])]; + tensor transpose_116 = transpose(perm = var_2100, x = var_2099_cast); + tensor x_227_cast = reshape(shape = concat_18, x = transpose_116); + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469644160)))]; + tensor var_2106_to_fp16 = const()[name = tensor("op_2106_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471741376)))]; + tensor var_2107_cast = linear(bias = var_2106_to_fp16, weight = var_2105_to_fp16, x = x_227_cast); + tensor x_229_cast = add(x = x_223_cast, y = var_2107_cast); + tensor var_2113_axes_0 = const()[name = tensor("op_2113_axes_0"), val = tensor([-1])]; + tensor blocks_18_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471743488)))]; + tensor blocks_18_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471745600)))]; + tensor var_2113_cast = layer_norm(axes = var_2113_axes_0, beta = blocks_18_mlp_ln_bias_to_fp16, epsilon = var_2038_to_fp16, gamma = blocks_18_mlp_ln_weight_to_fp16, x = x_229_cast); + tensor var_2122_to_fp16 = const()[name = tensor("op_2122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471747712)))]; + tensor var_2123_to_fp16 = const()[name = tensor("op_2123_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480136384)))]; + tensor input_153_cast = linear(bias = var_2123_to_fp16, weight = var_2122_to_fp16, x = var_2113_cast); + tensor x_233_mode_0 = const()[name = tensor("x_233_mode_0"), val = tensor("EXACT")]; + tensor x_233_cast = gelu(mode = x_233_mode_0, x = input_153_cast); + tensor var_2128_to_fp16 = const()[name = tensor("op_2128_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480144640)))]; + tensor var_2129_to_fp16 = const()[name = tensor("op_2129_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488533312)))]; + tensor var_2130_cast = linear(bias = var_2129_to_fp16, weight = var_2128_to_fp16, x = x_233_cast); + tensor x_235_cast = add(x = x_229_cast, y = var_2130_cast); + tensor var_2139 = const()[name = tensor("op_2139"), val = tensor(-1)]; + tensor var_2156_axes_0 = const()[name = tensor("op_2156_axes_0"), val = tensor([-1])]; + tensor blocks_19_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_19_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488535424)))]; + tensor blocks_19_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_19_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488537536)))]; + tensor var_2145_to_fp16 = const()[name = tensor("op_2145_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2156_cast = layer_norm(axes = var_2156_axes_0, beta = blocks_19_attn_ln_bias_to_fp16, epsilon = var_2145_to_fp16, gamma = blocks_19_attn_ln_weight_to_fp16, x = x_235_cast); + tensor var_2167_to_fp16 = const()[name = tensor("op_2167_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488539648)))]; + tensor var_2168_to_fp16 = const()[name = tensor("op_2168_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490636864)))]; + tensor q_77_cast = linear(bias = var_2168_to_fp16, weight = var_2167_to_fp16, x = var_2156_cast); + tensor var_2171_to_fp16 = const()[name = tensor("op_2171_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490638976)))]; + tensor k_77_bias_0_to_fp16 = const()[name = tensor("k_77_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492736192)))]; + tensor k_77_cast = linear(bias = k_77_bias_0_to_fp16, weight = var_2171_to_fp16, x = var_2156_cast); + tensor var_2175_to_fp16 = const()[name = tensor("op_2175_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492738304)))]; + tensor var_2176_to_fp16 = const()[name = tensor("op_2176_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494835520)))]; + tensor v_77_cast = linear(bias = var_2176_to_fp16, weight = var_2175_to_fp16, x = var_2156_cast); + tensor var_2184 = const()[name = tensor("op_2184"), val = tensor([1, 1500, 16, -1])]; + tensor var_2185_cast = reshape(shape = var_2184, x = q_77_cast); + tensor const_206_to_fp16 = const()[name = tensor("const_206_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_79_cast = mul(x = var_2185_cast, y = const_206_to_fp16); + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1500, 16, -1])]; + tensor var_2192_cast = reshape(shape = var_2191, x = k_77_cast); + tensor const_207_to_fp16 = const()[name = tensor("const_207_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_79_cast = mul(x = var_2192_cast, y = const_207_to_fp16); + tensor var_2198 = const()[name = tensor("op_2198"), val = tensor([1, 1500, 16, -1])]; + tensor var_2199_cast = reshape(shape = var_2198, x = v_77_cast); + tensor var_2200 = const()[name = tensor("op_2200"), val = tensor([0, 2, 1, 3])]; + tensor qk_39_transpose_x_0 = const()[name = tensor("qk_39_transpose_x_0"), val = tensor(false)]; + tensor qk_39_transpose_y_0 = const()[name = tensor("qk_39_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_113 = transpose(perm = transpose_87_perm_0, x = k_79_cast); + tensor transpose_114 = transpose(perm = transpose_86_perm_0, x = q_79_cast); + tensor qk_39_cast = matmul(transpose_x = qk_39_transpose_x_0, transpose_y = qk_39_transpose_y_0, x = transpose_114, y = transpose_113); + tensor var_2204_cast = softmax(axis = var_2139, x = qk_39_cast); + tensor var_2206_transpose_x_0 = const()[name = tensor("op_2206_transpose_x_0"), val = tensor(false)]; + tensor var_2206_transpose_y_0 = const()[name = tensor("op_2206_transpose_y_0"), val = tensor(false)]; + tensor transpose_115 = transpose(perm = var_2200, x = var_2199_cast); + tensor var_2206_cast = matmul(transpose_x = var_2206_transpose_x_0, transpose_y = var_2206_transpose_y_0, x = var_2204_cast, y = transpose_115); + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([0, 2, 1, 3])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([1, 1500, 1024])]; + tensor transpose_112 = transpose(perm = var_2207, x = var_2206_cast); + tensor x_239_cast = reshape(shape = concat_19, x = transpose_112); + tensor var_2212_to_fp16 = const()[name = tensor("op_2212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494837632)))]; + tensor var_2213_to_fp16 = const()[name = tensor("op_2213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496934848)))]; + tensor var_2214_cast = linear(bias = var_2213_to_fp16, weight = var_2212_to_fp16, x = x_239_cast); + tensor x_241_cast = add(x = x_235_cast, y = var_2214_cast); + tensor var_2220_axes_0 = const()[name = tensor("op_2220_axes_0"), val = tensor([-1])]; + tensor blocks_19_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496936960)))]; + tensor blocks_19_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496939072)))]; + tensor var_2220_cast = layer_norm(axes = var_2220_axes_0, beta = blocks_19_mlp_ln_bias_to_fp16, epsilon = var_2145_to_fp16, gamma = blocks_19_mlp_ln_weight_to_fp16, x = x_241_cast); + tensor var_2229_to_fp16 = const()[name = tensor("op_2229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496941184)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505329856)))]; + tensor input_161_cast = linear(bias = var_2230_to_fp16, weight = var_2229_to_fp16, x = var_2220_cast); + tensor x_245_mode_0 = const()[name = tensor("x_245_mode_0"), val = tensor("EXACT")]; + tensor x_245_cast = gelu(mode = x_245_mode_0, x = input_161_cast); + tensor var_2235_to_fp16 = const()[name = tensor("op_2235_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505338112)))]; + tensor var_2236_to_fp16 = const()[name = tensor("op_2236_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513726784)))]; + tensor var_2237_cast = linear(bias = var_2236_to_fp16, weight = var_2235_to_fp16, x = x_245_cast); + tensor x_247_cast = add(x = x_241_cast, y = var_2237_cast); + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor(-1)]; + tensor var_2263_axes_0 = const()[name = tensor("op_2263_axes_0"), val = tensor([-1])]; + tensor blocks_20_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_20_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513728896)))]; + tensor blocks_20_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_20_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513731008)))]; + tensor var_2252_to_fp16 = const()[name = tensor("op_2252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2263_cast = layer_norm(axes = var_2263_axes_0, beta = blocks_20_attn_ln_bias_to_fp16, epsilon = var_2252_to_fp16, gamma = blocks_20_attn_ln_weight_to_fp16, x = x_247_cast); + tensor var_2274_to_fp16 = const()[name = tensor("op_2274_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513733120)))]; + tensor var_2275_to_fp16 = const()[name = tensor("op_2275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515830336)))]; + tensor q_81_cast = linear(bias = var_2275_to_fp16, weight = var_2274_to_fp16, x = var_2263_cast); + tensor var_2278_to_fp16 = const()[name = tensor("op_2278_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515832448)))]; + tensor k_81_bias_0_to_fp16 = const()[name = tensor("k_81_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517929664)))]; + tensor k_81_cast = linear(bias = k_81_bias_0_to_fp16, weight = var_2278_to_fp16, x = var_2263_cast); + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517931776)))]; + tensor var_2283_to_fp16 = const()[name = tensor("op_2283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520028992)))]; + tensor v_81_cast = linear(bias = var_2283_to_fp16, weight = var_2282_to_fp16, x = var_2263_cast); + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1500, 16, -1])]; + tensor var_2292_cast = reshape(shape = var_2291, x = q_81_cast); + tensor const_208_to_fp16 = const()[name = tensor("const_208_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_83_cast = mul(x = var_2292_cast, y = const_208_to_fp16); + tensor var_2298 = const()[name = tensor("op_2298"), val = tensor([1, 1500, 16, -1])]; + tensor var_2299_cast = reshape(shape = var_2298, x = k_81_cast); + tensor const_209_to_fp16 = const()[name = tensor("const_209_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_83_cast = mul(x = var_2299_cast, y = const_209_to_fp16); + tensor var_2305 = const()[name = tensor("op_2305"), val = tensor([1, 1500, 16, -1])]; + tensor var_2306_cast = reshape(shape = var_2305, x = v_81_cast); + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([0, 2, 1, 3])]; + tensor qk_41_transpose_x_0 = const()[name = tensor("qk_41_transpose_x_0"), val = tensor(false)]; + tensor qk_41_transpose_y_0 = const()[name = tensor("qk_41_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_109 = transpose(perm = transpose_89_perm_0, x = k_83_cast); + tensor transpose_110 = transpose(perm = transpose_88_perm_0, x = q_83_cast); + tensor qk_41_cast = matmul(transpose_x = qk_41_transpose_x_0, transpose_y = qk_41_transpose_y_0, x = transpose_110, y = transpose_109); + tensor var_2311_cast = softmax(axis = var_2246, x = qk_41_cast); + tensor var_2313_transpose_x_0 = const()[name = tensor("op_2313_transpose_x_0"), val = tensor(false)]; + tensor var_2313_transpose_y_0 = const()[name = tensor("op_2313_transpose_y_0"), val = tensor(false)]; + tensor transpose_111 = transpose(perm = var_2307, x = var_2306_cast); + tensor var_2313_cast = matmul(transpose_x = var_2313_transpose_x_0, transpose_y = var_2313_transpose_y_0, x = var_2311_cast, y = transpose_111); + tensor var_2314 = const()[name = tensor("op_2314"), val = tensor([0, 2, 1, 3])]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 1500, 1024])]; + tensor transpose_108 = transpose(perm = var_2314, x = var_2313_cast); + tensor x_251_cast = reshape(shape = concat_20, x = transpose_108); + tensor var_2319_to_fp16 = const()[name = tensor("op_2319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520031104)))]; + tensor var_2320_to_fp16 = const()[name = tensor("op_2320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522128320)))]; + tensor var_2321_cast = linear(bias = var_2320_to_fp16, weight = var_2319_to_fp16, x = x_251_cast); + tensor x_253_cast = add(x = x_247_cast, y = var_2321_cast); + tensor var_2327_axes_0 = const()[name = tensor("op_2327_axes_0"), val = tensor([-1])]; + tensor blocks_20_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522130432)))]; + tensor blocks_20_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522132544)))]; + tensor var_2327_cast = layer_norm(axes = var_2327_axes_0, beta = blocks_20_mlp_ln_bias_to_fp16, epsilon = var_2252_to_fp16, gamma = blocks_20_mlp_ln_weight_to_fp16, x = x_253_cast); + tensor var_2336_to_fp16 = const()[name = tensor("op_2336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522134656)))]; + tensor var_2337_to_fp16 = const()[name = tensor("op_2337_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530523328)))]; + tensor input_169_cast = linear(bias = var_2337_to_fp16, weight = var_2336_to_fp16, x = var_2327_cast); + tensor x_257_mode_0 = const()[name = tensor("x_257_mode_0"), val = tensor("EXACT")]; + tensor x_257_cast = gelu(mode = x_257_mode_0, x = input_169_cast); + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530531584)))]; + tensor var_2343_to_fp16 = const()[name = tensor("op_2343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538920256)))]; + tensor var_2344_cast = linear(bias = var_2343_to_fp16, weight = var_2342_to_fp16, x = x_257_cast); + tensor x_259_cast = add(x = x_253_cast, y = var_2344_cast); + tensor var_2353 = const()[name = tensor("op_2353"), val = tensor(-1)]; + tensor var_2370_axes_0 = const()[name = tensor("op_2370_axes_0"), val = tensor([-1])]; + tensor blocks_21_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_21_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538922368)))]; + tensor blocks_21_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_21_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538924480)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2370_cast = layer_norm(axes = var_2370_axes_0, beta = blocks_21_attn_ln_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = blocks_21_attn_ln_weight_to_fp16, x = x_259_cast); + tensor var_2381_to_fp16 = const()[name = tensor("op_2381_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538926592)))]; + tensor var_2382_to_fp16 = const()[name = tensor("op_2382_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541023808)))]; + tensor q_85_cast = linear(bias = var_2382_to_fp16, weight = var_2381_to_fp16, x = var_2370_cast); + tensor var_2385_to_fp16 = const()[name = tensor("op_2385_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541025920)))]; + tensor k_85_bias_0_to_fp16 = const()[name = tensor("k_85_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543123136)))]; + tensor k_85_cast = linear(bias = k_85_bias_0_to_fp16, weight = var_2385_to_fp16, x = var_2370_cast); + tensor var_2389_to_fp16 = const()[name = tensor("op_2389_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543125248)))]; + tensor var_2390_to_fp16 = const()[name = tensor("op_2390_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545222464)))]; + tensor v_85_cast = linear(bias = var_2390_to_fp16, weight = var_2389_to_fp16, x = var_2370_cast); + tensor var_2398 = const()[name = tensor("op_2398"), val = tensor([1, 1500, 16, -1])]; + tensor var_2399_cast = reshape(shape = var_2398, x = q_85_cast); + tensor const_210_to_fp16 = const()[name = tensor("const_210_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_87_cast = mul(x = var_2399_cast, y = const_210_to_fp16); + tensor var_2405 = const()[name = tensor("op_2405"), val = tensor([1, 1500, 16, -1])]; + tensor var_2406_cast = reshape(shape = var_2405, x = k_85_cast); + tensor const_211_to_fp16 = const()[name = tensor("const_211_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_87_cast = mul(x = var_2406_cast, y = const_211_to_fp16); + tensor var_2412 = const()[name = tensor("op_2412"), val = tensor([1, 1500, 16, -1])]; + tensor var_2413_cast = reshape(shape = var_2412, x = v_85_cast); + tensor var_2414 = const()[name = tensor("op_2414"), val = tensor([0, 2, 1, 3])]; + tensor qk_43_transpose_x_0 = const()[name = tensor("qk_43_transpose_x_0"), val = tensor(false)]; + tensor qk_43_transpose_y_0 = const()[name = tensor("qk_43_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_105 = transpose(perm = transpose_91_perm_0, x = k_87_cast); + tensor transpose_106 = transpose(perm = transpose_90_perm_0, x = q_87_cast); + tensor qk_43_cast = matmul(transpose_x = qk_43_transpose_x_0, transpose_y = qk_43_transpose_y_0, x = transpose_106, y = transpose_105); + tensor var_2418_cast = softmax(axis = var_2353, x = qk_43_cast); + tensor var_2420_transpose_x_0 = const()[name = tensor("op_2420_transpose_x_0"), val = tensor(false)]; + tensor var_2420_transpose_y_0 = const()[name = tensor("op_2420_transpose_y_0"), val = tensor(false)]; + tensor transpose_107 = transpose(perm = var_2414, x = var_2413_cast); + tensor var_2420_cast = matmul(transpose_x = var_2420_transpose_x_0, transpose_y = var_2420_transpose_y_0, x = var_2418_cast, y = transpose_107); + tensor var_2421 = const()[name = tensor("op_2421"), val = tensor([0, 2, 1, 3])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([1, 1500, 1024])]; + tensor transpose_104 = transpose(perm = var_2421, x = var_2420_cast); + tensor x_263_cast = reshape(shape = concat_21, x = transpose_104); + tensor var_2426_to_fp16 = const()[name = tensor("op_2426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545224576)))]; + tensor var_2427_to_fp16 = const()[name = tensor("op_2427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547321792)))]; + tensor var_2428_cast = linear(bias = var_2427_to_fp16, weight = var_2426_to_fp16, x = x_263_cast); + tensor x_265_cast = add(x = x_259_cast, y = var_2428_cast); + tensor var_2434_axes_0 = const()[name = tensor("op_2434_axes_0"), val = tensor([-1])]; + tensor blocks_21_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547323904)))]; + tensor blocks_21_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547326016)))]; + tensor var_2434_cast = layer_norm(axes = var_2434_axes_0, beta = blocks_21_mlp_ln_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = blocks_21_mlp_ln_weight_to_fp16, x = x_265_cast); + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547328128)))]; + tensor var_2444_to_fp16 = const()[name = tensor("op_2444_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555716800)))]; + tensor input_177_cast = linear(bias = var_2444_to_fp16, weight = var_2443_to_fp16, x = var_2434_cast); + tensor x_269_mode_0 = const()[name = tensor("x_269_mode_0"), val = tensor("EXACT")]; + tensor x_269_cast = gelu(mode = x_269_mode_0, x = input_177_cast); + tensor var_2449_to_fp16 = const()[name = tensor("op_2449_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555725056)))]; + tensor var_2450_to_fp16 = const()[name = tensor("op_2450_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564113728)))]; + tensor var_2451_cast = linear(bias = var_2450_to_fp16, weight = var_2449_to_fp16, x = x_269_cast); + tensor x_271_cast = add(x = x_265_cast, y = var_2451_cast); + tensor var_2460 = const()[name = tensor("op_2460"), val = tensor(-1)]; + tensor var_2477_axes_0 = const()[name = tensor("op_2477_axes_0"), val = tensor([-1])]; + tensor blocks_22_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_22_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564115840)))]; + tensor blocks_22_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_22_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564117952)))]; + tensor var_2466_to_fp16 = const()[name = tensor("op_2466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2477_cast = layer_norm(axes = var_2477_axes_0, beta = blocks_22_attn_ln_bias_to_fp16, epsilon = var_2466_to_fp16, gamma = blocks_22_attn_ln_weight_to_fp16, x = x_271_cast); + tensor var_2488_to_fp16 = const()[name = tensor("op_2488_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564120064)))]; + tensor var_2489_to_fp16 = const()[name = tensor("op_2489_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566217280)))]; + tensor q_89_cast = linear(bias = var_2489_to_fp16, weight = var_2488_to_fp16, x = var_2477_cast); + tensor var_2492_to_fp16 = const()[name = tensor("op_2492_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566219392)))]; + tensor k_89_bias_0_to_fp16 = const()[name = tensor("k_89_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568316608)))]; + tensor k_89_cast = linear(bias = k_89_bias_0_to_fp16, weight = var_2492_to_fp16, x = var_2477_cast); + tensor var_2496_to_fp16 = const()[name = tensor("op_2496_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568318720)))]; + tensor var_2497_to_fp16 = const()[name = tensor("op_2497_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570415936)))]; + tensor v_89_cast = linear(bias = var_2497_to_fp16, weight = var_2496_to_fp16, x = var_2477_cast); + tensor var_2505 = const()[name = tensor("op_2505"), val = tensor([1, 1500, 16, -1])]; + tensor var_2506_cast = reshape(shape = var_2505, x = q_89_cast); + tensor const_212_to_fp16 = const()[name = tensor("const_212_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_91_cast = mul(x = var_2506_cast, y = const_212_to_fp16); + tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 1500, 16, -1])]; + tensor var_2513_cast = reshape(shape = var_2512, x = k_89_cast); + tensor const_213_to_fp16 = const()[name = tensor("const_213_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_91_cast = mul(x = var_2513_cast, y = const_213_to_fp16); + tensor var_2519 = const()[name = tensor("op_2519"), val = tensor([1, 1500, 16, -1])]; + tensor var_2520_cast = reshape(shape = var_2519, x = v_89_cast); + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([0, 2, 1, 3])]; + tensor qk_45_transpose_x_0 = const()[name = tensor("qk_45_transpose_x_0"), val = tensor(false)]; + tensor qk_45_transpose_y_0 = const()[name = tensor("qk_45_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_101 = transpose(perm = transpose_93_perm_0, x = k_91_cast); + tensor transpose_102 = transpose(perm = transpose_92_perm_0, x = q_91_cast); + tensor qk_45_cast = matmul(transpose_x = qk_45_transpose_x_0, transpose_y = qk_45_transpose_y_0, x = transpose_102, y = transpose_101); + tensor var_2525_cast = softmax(axis = var_2460, x = qk_45_cast); + tensor var_2527_transpose_x_0 = const()[name = tensor("op_2527_transpose_x_0"), val = tensor(false)]; + tensor var_2527_transpose_y_0 = const()[name = tensor("op_2527_transpose_y_0"), val = tensor(false)]; + tensor transpose_103 = transpose(perm = var_2521, x = var_2520_cast); + tensor var_2527_cast = matmul(transpose_x = var_2527_transpose_x_0, transpose_y = var_2527_transpose_y_0, x = var_2525_cast, y = transpose_103); + tensor var_2528 = const()[name = tensor("op_2528"), val = tensor([0, 2, 1, 3])]; + tensor concat_22 = const()[name = tensor("concat_22"), val = tensor([1, 1500, 1024])]; + tensor transpose_100 = transpose(perm = var_2528, x = var_2527_cast); + tensor x_275_cast = reshape(shape = concat_22, x = transpose_100); + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570418048)))]; + tensor var_2534_to_fp16 = const()[name = tensor("op_2534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572515264)))]; + tensor var_2535_cast = linear(bias = var_2534_to_fp16, weight = var_2533_to_fp16, x = x_275_cast); + tensor x_277_cast = add(x = x_271_cast, y = var_2535_cast); + tensor var_2541_axes_0 = const()[name = tensor("op_2541_axes_0"), val = tensor([-1])]; + tensor blocks_22_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572517376)))]; + tensor blocks_22_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572519488)))]; + tensor var_2541_cast = layer_norm(axes = var_2541_axes_0, beta = blocks_22_mlp_ln_bias_to_fp16, epsilon = var_2466_to_fp16, gamma = blocks_22_mlp_ln_weight_to_fp16, x = x_277_cast); + tensor var_2550_to_fp16 = const()[name = tensor("op_2550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572521600)))]; + tensor var_2551_to_fp16 = const()[name = tensor("op_2551_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580910272)))]; + tensor input_185_cast = linear(bias = var_2551_to_fp16, weight = var_2550_to_fp16, x = var_2541_cast); + tensor x_281_mode_0 = const()[name = tensor("x_281_mode_0"), val = tensor("EXACT")]; + tensor x_281_cast = gelu(mode = x_281_mode_0, x = input_185_cast); + tensor var_2556_to_fp16 = const()[name = tensor("op_2556_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580918528)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589307200)))]; + tensor var_2558_cast = linear(bias = var_2557_to_fp16, weight = var_2556_to_fp16, x = x_281_cast); + tensor x_283_cast = add(x = x_277_cast, y = var_2558_cast); + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor(-1)]; + tensor var_2584_axes_0 = const()[name = tensor("op_2584_axes_0"), val = tensor([-1])]; + tensor blocks_23_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_23_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589309312)))]; + tensor blocks_23_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_23_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589311424)))]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2584_cast = layer_norm(axes = var_2584_axes_0, beta = blocks_23_attn_ln_bias_to_fp16, epsilon = var_2573_to_fp16, gamma = blocks_23_attn_ln_weight_to_fp16, x = x_283_cast); + tensor var_2595_to_fp16 = const()[name = tensor("op_2595_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589313536)))]; + tensor var_2596_to_fp16 = const()[name = tensor("op_2596_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591410752)))]; + tensor q_93_cast = linear(bias = var_2596_to_fp16, weight = var_2595_to_fp16, x = var_2584_cast); + tensor var_2599_to_fp16 = const()[name = tensor("op_2599_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591412864)))]; + tensor k_93_bias_0_to_fp16 = const()[name = tensor("k_93_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593510080)))]; + tensor k_93_cast = linear(bias = k_93_bias_0_to_fp16, weight = var_2599_to_fp16, x = var_2584_cast); + tensor var_2603_to_fp16 = const()[name = tensor("op_2603_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593512192)))]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595609408)))]; + tensor v_93_cast = linear(bias = var_2604_to_fp16, weight = var_2603_to_fp16, x = var_2584_cast); + tensor var_2612 = const()[name = tensor("op_2612"), val = tensor([1, 1500, 16, -1])]; + tensor var_2613_cast = reshape(shape = var_2612, x = q_93_cast); + tensor const_214_to_fp16 = const()[name = tensor("const_214_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_2613_cast, y = const_214_to_fp16); + tensor var_2619 = const()[name = tensor("op_2619"), val = tensor([1, 1500, 16, -1])]; + tensor var_2620_cast = reshape(shape = var_2619, x = k_93_cast); + tensor const_215_to_fp16 = const()[name = tensor("const_215_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_2620_cast, y = const_215_to_fp16); + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor([1, 1500, 16, -1])]; + tensor var_2627_cast = reshape(shape = var_2626, x = v_93_cast); + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_97 = transpose(perm = transpose_95_perm_0, x = k_cast); + tensor transpose_98 = transpose(perm = transpose_94_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_98, y = transpose_97); + tensor var_2632_cast = softmax(axis = var_2567, x = qk_cast); + tensor var_2634_transpose_x_0 = const()[name = tensor("op_2634_transpose_x_0"), val = tensor(false)]; + tensor var_2634_transpose_y_0 = const()[name = tensor("op_2634_transpose_y_0"), val = tensor(false)]; + tensor transpose_99 = transpose(perm = var_2628, x = var_2627_cast); + tensor var_2634_cast = matmul(transpose_x = var_2634_transpose_x_0, transpose_y = var_2634_transpose_y_0, x = var_2632_cast, y = transpose_99); + tensor var_2635 = const()[name = tensor("op_2635"), val = tensor([0, 2, 1, 3])]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([1, 1500, 1024])]; + tensor transpose_96 = transpose(perm = var_2635, x = var_2634_cast); + tensor x_287_cast = reshape(shape = concat_23, x = transpose_96); + tensor var_2640_to_fp16 = const()[name = tensor("op_2640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595611520)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597708736)))]; + tensor var_2642_cast = linear(bias = var_2641_to_fp16, weight = var_2640_to_fp16, x = x_287_cast); + tensor x_289_cast = add(x = x_283_cast, y = var_2642_cast); + tensor var_2648_axes_0 = const()[name = tensor("op_2648_axes_0"), val = tensor([-1])]; + tensor blocks_23_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597710848)))]; + tensor blocks_23_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597712960)))]; + tensor var_2648_cast = layer_norm(axes = var_2648_axes_0, beta = blocks_23_mlp_ln_bias_to_fp16, epsilon = var_2573_to_fp16, gamma = blocks_23_mlp_ln_weight_to_fp16, x = x_289_cast); + tensor var_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597715072)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606103744)))]; + tensor input_193_cast = linear(bias = var_2658_to_fp16, weight = var_2657_to_fp16, x = var_2648_cast); + tensor x_293_mode_0 = const()[name = tensor("x_293_mode_0"), val = tensor("EXACT")]; + tensor x_293_cast = gelu(mode = x_293_mode_0, x = input_193_cast); + tensor var_2663_to_fp16 = const()[name = tensor("op_2663_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606112000)))]; + tensor var_2664_to_fp16 = const()[name = tensor("op_2664_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614500672)))]; + tensor var_2665_cast = linear(bias = var_2664_to_fp16, weight = var_2663_to_fp16, x = x_293_cast); + tensor x_cast = add(x = x_289_cast, y = var_2665_cast); + tensor var_2678_axes_0 = const()[name = tensor("op_2678_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614502784)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614504896)))]; + tensor var_2669_to_fp16 = const()[name = tensor("op_2669_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2678_cast = layer_norm(axes = var_2678_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_2669_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_2678_cast_to_fp32_dtype_0 = const()[name = tensor("op_2678_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_2678_cast_to_fp32_dtype_0, x = var_2678_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..fb9924fc07bf1eae79440b58362feaa6a60639c7 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a188b0e4e3109f28f38f1f47ea2497ffe623923419df8e1ae12cb5f809a1815 +size 614507008 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..d1aecfd7776e8b30c6d1c19176649e67bae2983f --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea5175add1ad27873a1fc28ade228cc24ab283659f4a3267ba4ced7be23343eb +size 546740951 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..cfebf363e7217ce13431621718bc293f474504b9 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adbe456375e7eb3407732a426ecb65bbda86860e4aa801f3a696b70b8a533cdd +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..a83f08a4fb2b982fbeeece957015e7401dfa6f10 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 144, + "Matmul" : 48, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 24, + "Add" : 49, + "LayerNorm" : 49, + "Mul" : 48, + "Transpose" : 97, + "Gelu" : 26, + "Reshape" : 96 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..4771e10615202abab5bfc7b155ba0872539682da --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/model.mil @@ -0,0 +1,1455 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_56 = const()[name = tensor("op_56"), val = tensor(1)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor([1])]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491648)))]; + tensor cast_727 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_68_cast = conv(bias = bias_3_to_fp16, dilations = var_66, groups = var_56, pad = var_68_pad_0, pad_type = var_68_pad_type_0, strides = var_64, weight = weight_3_to_fp16, x = cast_727); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_68_cast); + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor var_81 = const()[name = tensor("op_81"), val = tensor([2])]; + tensor var_83 = const()[name = tensor("op_83"), val = tensor([1])]; + tensor var_85_pad_type_0 = const()[name = tensor("op_85_pad_type_0"), val = tensor("custom")]; + tensor var_85_pad_0 = const()[name = tensor("op_85_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493760)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6785280)))]; + tensor var_85_cast = conv(bias = bias_7_to_fp16, dilations = var_83, groups = var_72, pad = var_85_pad_0, pad_type = var_85_pad_type_0, strides = var_81, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_85_cast); + tensor var_90 = const()[name = tensor("op_90"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6787392)))]; + tensor transpose_192 = transpose(perm = var_90, x = x_3_cast); + tensor var_93_cast = add(x = transpose_192, y = positional_embedding_to_fp16); + tensor var_106 = const()[name = tensor("op_106"), val = tensor(-1)]; + tensor var_123_axes_0 = const()[name = tensor("op_123_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9859456)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9861568)))]; + tensor var_112_to_fp16 = const()[name = tensor("op_112_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_123_cast = layer_norm(axes = var_123_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_112_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_93_cast); + tensor var_134_to_fp16 = const()[name = tensor("op_134_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9863680)))]; + tensor var_135_to_fp16 = const()[name = tensor("op_135_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11960896)))]; + tensor q_1_cast = linear(bias = var_135_to_fp16, weight = var_134_to_fp16, x = var_123_cast); + tensor var_138_to_fp16 = const()[name = tensor("op_138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11963008)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14060224)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_138_to_fp16, x = var_123_cast); + tensor var_142_to_fp16 = const()[name = tensor("op_142_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14062336)))]; + tensor var_143_to_fp16 = const()[name = tensor("op_143_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16159552)))]; + tensor v_1_cast = linear(bias = var_143_to_fp16, weight = var_142_to_fp16, x = var_123_cast); + tensor var_151 = const()[name = tensor("op_151"), val = tensor([1, 1500, 16, -1])]; + tensor var_152_cast = reshape(shape = var_151, x = q_1_cast); + tensor const_168_to_fp16 = const()[name = tensor("const_168_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_152_cast, y = const_168_to_fp16); + tensor var_158 = const()[name = tensor("op_158"), val = tensor([1, 1500, 16, -1])]; + tensor var_159_cast = reshape(shape = var_158, x = k_1_cast); + tensor const_169_to_fp16 = const()[name = tensor("const_169_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_159_cast, y = const_169_to_fp16); + tensor var_165 = const()[name = tensor("op_165"), val = tensor([1, 1500, 16, -1])]; + tensor var_166_cast = reshape(shape = var_165, x = v_1_cast); + tensor var_167 = const()[name = tensor("op_167"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_48_perm_0 = const()[name = tensor("transpose_48_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_49_perm_0 = const()[name = tensor("transpose_49_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_189 = transpose(perm = transpose_49_perm_0, x = k_3_cast); + tensor transpose_190 = transpose(perm = transpose_48_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_190, y = transpose_189); + tensor var_171_cast = softmax(axis = var_106, x = qk_1_cast); + tensor var_173_transpose_x_0 = const()[name = tensor("op_173_transpose_x_0"), val = tensor(false)]; + tensor var_173_transpose_y_0 = const()[name = tensor("op_173_transpose_y_0"), val = tensor(false)]; + tensor transpose_191 = transpose(perm = var_167, x = var_166_cast); + tensor var_173_cast = matmul(transpose_x = var_173_transpose_x_0, transpose_y = var_173_transpose_y_0, x = var_171_cast, y = transpose_191); + tensor var_174 = const()[name = tensor("op_174"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 1024])]; + tensor transpose_188 = transpose(perm = var_174, x = var_173_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_188); + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16161664)))]; + tensor var_180_to_fp16 = const()[name = tensor("op_180_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18258880)))]; + tensor var_181_cast = linear(bias = var_180_to_fp16, weight = var_179_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_93_cast, y = var_181_cast); + tensor var_187_axes_0 = const()[name = tensor("op_187_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18260992)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18263104)))]; + tensor var_187_cast = layer_norm(axes = var_187_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_112_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_196_to_fp16 = const()[name = tensor("op_196_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18265216)))]; + tensor var_197_to_fp16 = const()[name = tensor("op_197_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26653888)))]; + tensor input_9_cast = linear(bias = var_197_to_fp16, weight = var_196_to_fp16, x = var_187_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26662144)))]; + tensor var_203_to_fp16 = const()[name = tensor("op_203_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35050816)))]; + tensor var_204_cast = linear(bias = var_203_to_fp16, weight = var_202_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_204_cast); + tensor var_213 = const()[name = tensor("op_213"), val = tensor(-1)]; + tensor var_230_axes_0 = const()[name = tensor("op_230_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35052928)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35055040)))]; + tensor var_219_to_fp16 = const()[name = tensor("op_219_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_230_cast = layer_norm(axes = var_230_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_219_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_241_to_fp16 = const()[name = tensor("op_241_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35057152)))]; + tensor var_242_to_fp16 = const()[name = tensor("op_242_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37154368)))]; + tensor q_5_cast = linear(bias = var_242_to_fp16, weight = var_241_to_fp16, x = var_230_cast); + tensor var_245_to_fp16 = const()[name = tensor("op_245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37156480)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39253696)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_245_to_fp16, x = var_230_cast); + tensor var_249_to_fp16 = const()[name = tensor("op_249_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39255808)))]; + tensor var_250_to_fp16 = const()[name = tensor("op_250_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41353024)))]; + tensor v_5_cast = linear(bias = var_250_to_fp16, weight = var_249_to_fp16, x = var_230_cast); + tensor var_258 = const()[name = tensor("op_258"), val = tensor([1, 1500, 16, -1])]; + tensor var_259_cast = reshape(shape = var_258, x = q_5_cast); + tensor const_170_to_fp16 = const()[name = tensor("const_170_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_259_cast, y = const_170_to_fp16); + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1500, 16, -1])]; + tensor var_266_cast = reshape(shape = var_265, x = k_5_cast); + tensor const_171_to_fp16 = const()[name = tensor("const_171_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_266_cast, y = const_171_to_fp16); + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 1500, 16, -1])]; + tensor var_273_cast = reshape(shape = var_272, x = v_5_cast); + tensor var_274 = const()[name = tensor("op_274"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_50_perm_0 = const()[name = tensor("transpose_50_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_51_perm_0 = const()[name = tensor("transpose_51_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_185 = transpose(perm = transpose_51_perm_0, x = k_7_cast); + tensor transpose_186 = transpose(perm = transpose_50_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_186, y = transpose_185); + tensor var_278_cast = softmax(axis = var_213, x = qk_3_cast); + tensor var_280_transpose_x_0 = const()[name = tensor("op_280_transpose_x_0"), val = tensor(false)]; + tensor var_280_transpose_y_0 = const()[name = tensor("op_280_transpose_y_0"), val = tensor(false)]; + tensor transpose_187 = transpose(perm = var_274, x = var_273_cast); + tensor var_280_cast = matmul(transpose_x = var_280_transpose_x_0, transpose_y = var_280_transpose_y_0, x = var_278_cast, y = transpose_187); + tensor var_281 = const()[name = tensor("op_281"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 1024])]; + tensor transpose_184 = transpose(perm = var_281, x = var_280_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_184); + tensor var_286_to_fp16 = const()[name = tensor("op_286_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41355136)))]; + tensor var_287_to_fp16 = const()[name = tensor("op_287_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43452352)))]; + tensor var_288_cast = linear(bias = var_287_to_fp16, weight = var_286_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_288_cast); + tensor var_294_axes_0 = const()[name = tensor("op_294_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43454464)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43456576)))]; + tensor var_294_cast = layer_norm(axes = var_294_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_219_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_303_to_fp16 = const()[name = tensor("op_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43458688)))]; + tensor var_304_to_fp16 = const()[name = tensor("op_304_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51847360)))]; + tensor input_17_cast = linear(bias = var_304_to_fp16, weight = var_303_to_fp16, x = var_294_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51855616)))]; + tensor var_310_to_fp16 = const()[name = tensor("op_310_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60244288)))]; + tensor var_311_cast = linear(bias = var_310_to_fp16, weight = var_309_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_311_cast); + tensor var_320 = const()[name = tensor("op_320"), val = tensor(-1)]; + tensor var_337_axes_0 = const()[name = tensor("op_337_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60246400)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60248512)))]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_337_cast = layer_norm(axes = var_337_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_326_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_348_to_fp16 = const()[name = tensor("op_348_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60250624)))]; + tensor var_349_to_fp16 = const()[name = tensor("op_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62347840)))]; + tensor q_9_cast = linear(bias = var_349_to_fp16, weight = var_348_to_fp16, x = var_337_cast); + tensor var_352_to_fp16 = const()[name = tensor("op_352_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62349952)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64447168)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_352_to_fp16, x = var_337_cast); + tensor var_356_to_fp16 = const()[name = tensor("op_356_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64449280)))]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66546496)))]; + tensor v_9_cast = linear(bias = var_357_to_fp16, weight = var_356_to_fp16, x = var_337_cast); + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1500, 16, -1])]; + tensor var_366_cast = reshape(shape = var_365, x = q_9_cast); + tensor const_172_to_fp16 = const()[name = tensor("const_172_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_366_cast, y = const_172_to_fp16); + tensor var_372 = const()[name = tensor("op_372"), val = tensor([1, 1500, 16, -1])]; + tensor var_373_cast = reshape(shape = var_372, x = k_9_cast); + tensor const_173_to_fp16 = const()[name = tensor("const_173_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_373_cast, y = const_173_to_fp16); + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1500, 16, -1])]; + tensor var_380_cast = reshape(shape = var_379, x = v_9_cast); + tensor var_381 = const()[name = tensor("op_381"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_52_perm_0 = const()[name = tensor("transpose_52_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_53_perm_0 = const()[name = tensor("transpose_53_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_181 = transpose(perm = transpose_53_perm_0, x = k_11_cast); + tensor transpose_182 = transpose(perm = transpose_52_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_182, y = transpose_181); + tensor var_385_cast = softmax(axis = var_320, x = qk_5_cast); + tensor var_387_transpose_x_0 = const()[name = tensor("op_387_transpose_x_0"), val = tensor(false)]; + tensor var_387_transpose_y_0 = const()[name = tensor("op_387_transpose_y_0"), val = tensor(false)]; + tensor transpose_183 = transpose(perm = var_381, x = var_380_cast); + tensor var_387_cast = matmul(transpose_x = var_387_transpose_x_0, transpose_y = var_387_transpose_y_0, x = var_385_cast, y = transpose_183); + tensor var_388 = const()[name = tensor("op_388"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 1024])]; + tensor transpose_180 = transpose(perm = var_388, x = var_387_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_180); + tensor var_393_to_fp16 = const()[name = tensor("op_393_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66548608)))]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68645824)))]; + tensor var_395_cast = linear(bias = var_394_to_fp16, weight = var_393_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_395_cast); + tensor var_401_axes_0 = const()[name = tensor("op_401_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68647936)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68650048)))]; + tensor var_401_cast = layer_norm(axes = var_401_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_326_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_410_to_fp16 = const()[name = tensor("op_410_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68652160)))]; + tensor var_411_to_fp16 = const()[name = tensor("op_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77040832)))]; + tensor input_25_cast = linear(bias = var_411_to_fp16, weight = var_410_to_fp16, x = var_401_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_416_to_fp16 = const()[name = tensor("op_416_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77049088)))]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85437760)))]; + tensor var_418_cast = linear(bias = var_417_to_fp16, weight = var_416_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_418_cast); + tensor var_427 = const()[name = tensor("op_427"), val = tensor(-1)]; + tensor var_444_axes_0 = const()[name = tensor("op_444_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85439872)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85441984)))]; + tensor var_433_to_fp16 = const()[name = tensor("op_433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_444_cast = layer_norm(axes = var_444_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_433_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_455_to_fp16 = const()[name = tensor("op_455_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85444096)))]; + tensor var_456_to_fp16 = const()[name = tensor("op_456_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87541312)))]; + tensor q_13_cast = linear(bias = var_456_to_fp16, weight = var_455_to_fp16, x = var_444_cast); + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87543424)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89640640)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_459_to_fp16, x = var_444_cast); + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89642752)))]; + tensor var_464_to_fp16 = const()[name = tensor("op_464_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91739968)))]; + tensor v_13_cast = linear(bias = var_464_to_fp16, weight = var_463_to_fp16, x = var_444_cast); + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 1500, 16, -1])]; + tensor var_473_cast = reshape(shape = var_472, x = q_13_cast); + tensor const_174_to_fp16 = const()[name = tensor("const_174_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_473_cast, y = const_174_to_fp16); + tensor var_479 = const()[name = tensor("op_479"), val = tensor([1, 1500, 16, -1])]; + tensor var_480_cast = reshape(shape = var_479, x = k_13_cast); + tensor const_175_to_fp16 = const()[name = tensor("const_175_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_480_cast, y = const_175_to_fp16); + tensor var_486 = const()[name = tensor("op_486"), val = tensor([1, 1500, 16, -1])]; + tensor var_487_cast = reshape(shape = var_486, x = v_13_cast); + tensor var_488 = const()[name = tensor("op_488"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_54_perm_0 = const()[name = tensor("transpose_54_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_55_perm_0 = const()[name = tensor("transpose_55_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_177 = transpose(perm = transpose_55_perm_0, x = k_15_cast); + tensor transpose_178 = transpose(perm = transpose_54_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_178, y = transpose_177); + tensor var_492_cast = softmax(axis = var_427, x = qk_7_cast); + tensor var_494_transpose_x_0 = const()[name = tensor("op_494_transpose_x_0"), val = tensor(false)]; + tensor var_494_transpose_y_0 = const()[name = tensor("op_494_transpose_y_0"), val = tensor(false)]; + tensor transpose_179 = transpose(perm = var_488, x = var_487_cast); + tensor var_494_cast = matmul(transpose_x = var_494_transpose_x_0, transpose_y = var_494_transpose_y_0, x = var_492_cast, y = transpose_179); + tensor var_495 = const()[name = tensor("op_495"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 1024])]; + tensor transpose_176 = transpose(perm = var_495, x = var_494_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_176); + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91742080)))]; + tensor var_501_to_fp16 = const()[name = tensor("op_501_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93839296)))]; + tensor var_502_cast = linear(bias = var_501_to_fp16, weight = var_500_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_502_cast); + tensor var_508_axes_0 = const()[name = tensor("op_508_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93841408)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93843520)))]; + tensor var_508_cast = layer_norm(axes = var_508_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_433_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_517_to_fp16 = const()[name = tensor("op_517_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93845632)))]; + tensor var_518_to_fp16 = const()[name = tensor("op_518_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102234304)))]; + tensor input_33_cast = linear(bias = var_518_to_fp16, weight = var_517_to_fp16, x = var_508_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_523_to_fp16 = const()[name = tensor("op_523_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102242560)))]; + tensor var_524_to_fp16 = const()[name = tensor("op_524_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110631232)))]; + tensor var_525_cast = linear(bias = var_524_to_fp16, weight = var_523_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_525_cast); + tensor var_534 = const()[name = tensor("op_534"), val = tensor(-1)]; + tensor var_551_axes_0 = const()[name = tensor("op_551_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110633344)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110635456)))]; + tensor var_540_to_fp16 = const()[name = tensor("op_540_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_551_cast = layer_norm(axes = var_551_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_540_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_562_to_fp16 = const()[name = tensor("op_562_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110637568)))]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112734784)))]; + tensor q_17_cast = linear(bias = var_563_to_fp16, weight = var_562_to_fp16, x = var_551_cast); + tensor var_566_to_fp16 = const()[name = tensor("op_566_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112736896)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114834112)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_566_to_fp16, x = var_551_cast); + tensor var_570_to_fp16 = const()[name = tensor("op_570_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114836224)))]; + tensor var_571_to_fp16 = const()[name = tensor("op_571_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116933440)))]; + tensor v_17_cast = linear(bias = var_571_to_fp16, weight = var_570_to_fp16, x = var_551_cast); + tensor var_579 = const()[name = tensor("op_579"), val = tensor([1, 1500, 16, -1])]; + tensor var_580_cast = reshape(shape = var_579, x = q_17_cast); + tensor const_176_to_fp16 = const()[name = tensor("const_176_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_580_cast, y = const_176_to_fp16); + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 1500, 16, -1])]; + tensor var_587_cast = reshape(shape = var_586, x = k_17_cast); + tensor const_177_to_fp16 = const()[name = tensor("const_177_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_587_cast, y = const_177_to_fp16); + tensor var_593 = const()[name = tensor("op_593"), val = tensor([1, 1500, 16, -1])]; + tensor var_594_cast = reshape(shape = var_593, x = v_17_cast); + tensor var_595 = const()[name = tensor("op_595"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_56_perm_0 = const()[name = tensor("transpose_56_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_57_perm_0 = const()[name = tensor("transpose_57_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_173 = transpose(perm = transpose_57_perm_0, x = k_19_cast); + tensor transpose_174 = transpose(perm = transpose_56_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_174, y = transpose_173); + tensor var_599_cast = softmax(axis = var_534, x = qk_9_cast); + tensor var_601_transpose_x_0 = const()[name = tensor("op_601_transpose_x_0"), val = tensor(false)]; + tensor var_601_transpose_y_0 = const()[name = tensor("op_601_transpose_y_0"), val = tensor(false)]; + tensor transpose_175 = transpose(perm = var_595, x = var_594_cast); + tensor var_601_cast = matmul(transpose_x = var_601_transpose_x_0, transpose_y = var_601_transpose_y_0, x = var_599_cast, y = transpose_175); + tensor var_602 = const()[name = tensor("op_602"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 1024])]; + tensor transpose_172 = transpose(perm = var_602, x = var_601_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_172); + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116935552)))]; + tensor var_608_to_fp16 = const()[name = tensor("op_608_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119032768)))]; + tensor var_609_cast = linear(bias = var_608_to_fp16, weight = var_607_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_609_cast); + tensor var_615_axes_0 = const()[name = tensor("op_615_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119034880)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119036992)))]; + tensor var_615_cast = layer_norm(axes = var_615_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_540_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_624_to_fp16 = const()[name = tensor("op_624_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119039104)))]; + tensor var_625_to_fp16 = const()[name = tensor("op_625_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127427776)))]; + tensor input_41_cast = linear(bias = var_625_to_fp16, weight = var_624_to_fp16, x = var_615_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_630_to_fp16 = const()[name = tensor("op_630_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127436032)))]; + tensor var_631_to_fp16 = const()[name = tensor("op_631_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135824704)))]; + tensor var_632_cast = linear(bias = var_631_to_fp16, weight = var_630_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_632_cast); + tensor var_641 = const()[name = tensor("op_641"), val = tensor(-1)]; + tensor var_658_axes_0 = const()[name = tensor("op_658_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135826816)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135828928)))]; + tensor var_647_to_fp16 = const()[name = tensor("op_647_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_658_cast = layer_norm(axes = var_658_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_647_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_669_to_fp16 = const()[name = tensor("op_669_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135831040)))]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137928256)))]; + tensor q_21_cast = linear(bias = var_670_to_fp16, weight = var_669_to_fp16, x = var_658_cast); + tensor var_673_to_fp16 = const()[name = tensor("op_673_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137930368)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140027584)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_673_to_fp16, x = var_658_cast); + tensor var_677_to_fp16 = const()[name = tensor("op_677_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140029696)))]; + tensor var_678_to_fp16 = const()[name = tensor("op_678_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142126912)))]; + tensor v_21_cast = linear(bias = var_678_to_fp16, weight = var_677_to_fp16, x = var_658_cast); + tensor var_686 = const()[name = tensor("op_686"), val = tensor([1, 1500, 16, -1])]; + tensor var_687_cast = reshape(shape = var_686, x = q_21_cast); + tensor const_178_to_fp16 = const()[name = tensor("const_178_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast = mul(x = var_687_cast, y = const_178_to_fp16); + tensor var_693 = const()[name = tensor("op_693"), val = tensor([1, 1500, 16, -1])]; + tensor var_694_cast = reshape(shape = var_693, x = k_21_cast); + tensor const_179_to_fp16 = const()[name = tensor("const_179_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast = mul(x = var_694_cast, y = const_179_to_fp16); + tensor var_700 = const()[name = tensor("op_700"), val = tensor([1, 1500, 16, -1])]; + tensor var_701_cast = reshape(shape = var_700, x = v_21_cast); + tensor var_702 = const()[name = tensor("op_702"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_58_perm_0 = const()[name = tensor("transpose_58_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_59_perm_0 = const()[name = tensor("transpose_59_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_169 = transpose(perm = transpose_59_perm_0, x = k_23_cast); + tensor transpose_170 = transpose(perm = transpose_58_perm_0, x = q_23_cast); + tensor qk_11_cast = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_170, y = transpose_169); + tensor var_706_cast = softmax(axis = var_641, x = qk_11_cast); + tensor var_708_transpose_x_0 = const()[name = tensor("op_708_transpose_x_0"), val = tensor(false)]; + tensor var_708_transpose_y_0 = const()[name = tensor("op_708_transpose_y_0"), val = tensor(false)]; + tensor transpose_171 = transpose(perm = var_702, x = var_701_cast); + tensor var_708_cast = matmul(transpose_x = var_708_transpose_x_0, transpose_y = var_708_transpose_y_0, x = var_706_cast, y = transpose_171); + tensor var_709 = const()[name = tensor("op_709"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 1024])]; + tensor transpose_168 = transpose(perm = var_709, x = var_708_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_168); + tensor var_714_to_fp16 = const()[name = tensor("op_714_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142129024)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144226240)))]; + tensor var_716_cast = linear(bias = var_715_to_fp16, weight = var_714_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_716_cast); + tensor var_722_axes_0 = const()[name = tensor("op_722_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144228352)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144230464)))]; + tensor var_722_cast = layer_norm(axes = var_722_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_647_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144232576)))]; + tensor var_732_to_fp16 = const()[name = tensor("op_732_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152621248)))]; + tensor input_49_cast = linear(bias = var_732_to_fp16, weight = var_731_to_fp16, x = var_722_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_737_to_fp16 = const()[name = tensor("op_737_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152629504)))]; + tensor var_738_to_fp16 = const()[name = tensor("op_738_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161018176)))]; + tensor var_739_cast = linear(bias = var_738_to_fp16, weight = var_737_to_fp16, x = x_77_cast); + tensor x_79_cast = add(x = x_73_cast, y = var_739_cast); + tensor var_748 = const()[name = tensor("op_748"), val = tensor(-1)]; + tensor var_765_axes_0 = const()[name = tensor("op_765_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161020288)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161022400)))]; + tensor var_754_to_fp16 = const()[name = tensor("op_754_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_765_cast = layer_norm(axes = var_765_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_754_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast); + tensor var_776_to_fp16 = const()[name = tensor("op_776_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161024512)))]; + tensor var_777_to_fp16 = const()[name = tensor("op_777_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163121728)))]; + tensor q_25_cast = linear(bias = var_777_to_fp16, weight = var_776_to_fp16, x = var_765_cast); + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163123840)))]; + tensor k_25_bias_0_to_fp16 = const()[name = tensor("k_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165221056)))]; + tensor k_25_cast = linear(bias = k_25_bias_0_to_fp16, weight = var_780_to_fp16, x = var_765_cast); + tensor var_784_to_fp16 = const()[name = tensor("op_784_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165223168)))]; + tensor var_785_to_fp16 = const()[name = tensor("op_785_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167320384)))]; + tensor v_25_cast = linear(bias = var_785_to_fp16, weight = var_784_to_fp16, x = var_765_cast); + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 1500, 16, -1])]; + tensor var_794_cast = reshape(shape = var_793, x = q_25_cast); + tensor const_180_to_fp16 = const()[name = tensor("const_180_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast = mul(x = var_794_cast, y = const_180_to_fp16); + tensor var_800 = const()[name = tensor("op_800"), val = tensor([1, 1500, 16, -1])]; + tensor var_801_cast = reshape(shape = var_800, x = k_25_cast); + tensor const_181_to_fp16 = const()[name = tensor("const_181_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast = mul(x = var_801_cast, y = const_181_to_fp16); + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1500, 16, -1])]; + tensor var_808_cast = reshape(shape = var_807, x = v_25_cast); + tensor var_809 = const()[name = tensor("op_809"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_60_perm_0 = const()[name = tensor("transpose_60_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_61_perm_0 = const()[name = tensor("transpose_61_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_165 = transpose(perm = transpose_61_perm_0, x = k_27_cast); + tensor transpose_166 = transpose(perm = transpose_60_perm_0, x = q_27_cast); + tensor qk_13_cast = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_166, y = transpose_165); + tensor var_813_cast = softmax(axis = var_748, x = qk_13_cast); + tensor var_815_transpose_x_0 = const()[name = tensor("op_815_transpose_x_0"), val = tensor(false)]; + tensor var_815_transpose_y_0 = const()[name = tensor("op_815_transpose_y_0"), val = tensor(false)]; + tensor transpose_167 = transpose(perm = var_809, x = var_808_cast); + tensor var_815_cast = matmul(transpose_x = var_815_transpose_x_0, transpose_y = var_815_transpose_y_0, x = var_813_cast, y = transpose_167); + tensor var_816 = const()[name = tensor("op_816"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 1024])]; + tensor transpose_164 = transpose(perm = var_816, x = var_815_cast); + tensor x_83_cast = reshape(shape = concat_6, x = transpose_164); + tensor var_821_to_fp16 = const()[name = tensor("op_821_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167322496)))]; + tensor var_822_to_fp16 = const()[name = tensor("op_822_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169419712)))]; + tensor var_823_cast = linear(bias = var_822_to_fp16, weight = var_821_to_fp16, x = x_83_cast); + tensor x_85_cast = add(x = x_79_cast, y = var_823_cast); + tensor var_829_axes_0 = const()[name = tensor("op_829_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169421824)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169423936)))]; + tensor var_829_cast = layer_norm(axes = var_829_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_754_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast); + tensor var_838_to_fp16 = const()[name = tensor("op_838_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169426048)))]; + tensor var_839_to_fp16 = const()[name = tensor("op_839_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177814720)))]; + tensor input_57_cast = linear(bias = var_839_to_fp16, weight = var_838_to_fp16, x = var_829_cast); + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast = gelu(mode = x_89_mode_0, x = input_57_cast); + tensor var_844_to_fp16 = const()[name = tensor("op_844_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177822976)))]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186211648)))]; + tensor var_846_cast = linear(bias = var_845_to_fp16, weight = var_844_to_fp16, x = x_89_cast); + tensor x_91_cast = add(x = x_85_cast, y = var_846_cast); + tensor var_855 = const()[name = tensor("op_855"), val = tensor(-1)]; + tensor var_872_axes_0 = const()[name = tensor("op_872_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186213760)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186215872)))]; + tensor var_861_to_fp16 = const()[name = tensor("op_861_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_872_cast = layer_norm(axes = var_872_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_861_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast); + tensor var_883_to_fp16 = const()[name = tensor("op_883_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186217984)))]; + tensor var_884_to_fp16 = const()[name = tensor("op_884_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188315200)))]; + tensor q_29_cast = linear(bias = var_884_to_fp16, weight = var_883_to_fp16, x = var_872_cast); + tensor var_887_to_fp16 = const()[name = tensor("op_887_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(188317312)))]; + tensor k_29_bias_0_to_fp16 = const()[name = tensor("k_29_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190414528)))]; + tensor k_29_cast = linear(bias = k_29_bias_0_to_fp16, weight = var_887_to_fp16, x = var_872_cast); + tensor var_891_to_fp16 = const()[name = tensor("op_891_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190416640)))]; + tensor var_892_to_fp16 = const()[name = tensor("op_892_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192513856)))]; + tensor v_29_cast = linear(bias = var_892_to_fp16, weight = var_891_to_fp16, x = var_872_cast); + tensor var_900 = const()[name = tensor("op_900"), val = tensor([1, 1500, 16, -1])]; + tensor var_901_cast = reshape(shape = var_900, x = q_29_cast); + tensor const_182_to_fp16 = const()[name = tensor("const_182_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast = mul(x = var_901_cast, y = const_182_to_fp16); + tensor var_907 = const()[name = tensor("op_907"), val = tensor([1, 1500, 16, -1])]; + tensor var_908_cast = reshape(shape = var_907, x = k_29_cast); + tensor const_183_to_fp16 = const()[name = tensor("const_183_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast = mul(x = var_908_cast, y = const_183_to_fp16); + tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 1500, 16, -1])]; + tensor var_915_cast = reshape(shape = var_914, x = v_29_cast); + tensor var_916 = const()[name = tensor("op_916"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_62_perm_0 = const()[name = tensor("transpose_62_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_63_perm_0 = const()[name = tensor("transpose_63_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_161 = transpose(perm = transpose_63_perm_0, x = k_31_cast); + tensor transpose_162 = transpose(perm = transpose_62_perm_0, x = q_31_cast); + tensor qk_15_cast = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_162, y = transpose_161); + tensor var_920_cast = softmax(axis = var_855, x = qk_15_cast); + tensor var_922_transpose_x_0 = const()[name = tensor("op_922_transpose_x_0"), val = tensor(false)]; + tensor var_922_transpose_y_0 = const()[name = tensor("op_922_transpose_y_0"), val = tensor(false)]; + tensor transpose_163 = transpose(perm = var_916, x = var_915_cast); + tensor var_922_cast = matmul(transpose_x = var_922_transpose_x_0, transpose_y = var_922_transpose_y_0, x = var_920_cast, y = transpose_163); + tensor var_923 = const()[name = tensor("op_923"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 1024])]; + tensor transpose_160 = transpose(perm = var_923, x = var_922_cast); + tensor x_95_cast = reshape(shape = concat_7, x = transpose_160); + tensor var_928_to_fp16 = const()[name = tensor("op_928_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192515968)))]; + tensor var_929_to_fp16 = const()[name = tensor("op_929_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194613184)))]; + tensor var_930_cast = linear(bias = var_929_to_fp16, weight = var_928_to_fp16, x = x_95_cast); + tensor x_97_cast = add(x = x_91_cast, y = var_930_cast); + tensor var_936_axes_0 = const()[name = tensor("op_936_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194615296)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194617408)))]; + tensor var_936_cast = layer_norm(axes = var_936_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_861_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast); + tensor var_945_to_fp16 = const()[name = tensor("op_945_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194619520)))]; + tensor var_946_to_fp16 = const()[name = tensor("op_946_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203008192)))]; + tensor input_65_cast = linear(bias = var_946_to_fp16, weight = var_945_to_fp16, x = var_936_cast); + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast = gelu(mode = x_101_mode_0, x = input_65_cast); + tensor var_951_to_fp16 = const()[name = tensor("op_951_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203016448)))]; + tensor var_952_to_fp16 = const()[name = tensor("op_952_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211405120)))]; + tensor var_953_cast = linear(bias = var_952_to_fp16, weight = var_951_to_fp16, x = x_101_cast); + tensor x_103_cast = add(x = x_97_cast, y = var_953_cast); + tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; + tensor var_979_axes_0 = const()[name = tensor("op_979_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211407232)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211409344)))]; + tensor var_968_to_fp16 = const()[name = tensor("op_968_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_979_cast = layer_norm(axes = var_979_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_968_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast); + tensor var_990_to_fp16 = const()[name = tensor("op_990_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211411456)))]; + tensor var_991_to_fp16 = const()[name = tensor("op_991_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213508672)))]; + tensor q_33_cast = linear(bias = var_991_to_fp16, weight = var_990_to_fp16, x = var_979_cast); + tensor var_994_to_fp16 = const()[name = tensor("op_994_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(213510784)))]; + tensor k_33_bias_0_to_fp16 = const()[name = tensor("k_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215608000)))]; + tensor k_33_cast = linear(bias = k_33_bias_0_to_fp16, weight = var_994_to_fp16, x = var_979_cast); + tensor var_998_to_fp16 = const()[name = tensor("op_998_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215610112)))]; + tensor var_999_to_fp16 = const()[name = tensor("op_999_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217707328)))]; + tensor v_33_cast = linear(bias = var_999_to_fp16, weight = var_998_to_fp16, x = var_979_cast); + tensor var_1007 = const()[name = tensor("op_1007"), val = tensor([1, 1500, 16, -1])]; + tensor var_1008_cast = reshape(shape = var_1007, x = q_33_cast); + tensor const_184_to_fp16 = const()[name = tensor("const_184_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast = mul(x = var_1008_cast, y = const_184_to_fp16); + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1500, 16, -1])]; + tensor var_1015_cast = reshape(shape = var_1014, x = k_33_cast); + tensor const_185_to_fp16 = const()[name = tensor("const_185_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast = mul(x = var_1015_cast, y = const_185_to_fp16); + tensor var_1021 = const()[name = tensor("op_1021"), val = tensor([1, 1500, 16, -1])]; + tensor var_1022_cast = reshape(shape = var_1021, x = v_33_cast); + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_64_perm_0 = const()[name = tensor("transpose_64_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_65_perm_0 = const()[name = tensor("transpose_65_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_157 = transpose(perm = transpose_65_perm_0, x = k_35_cast); + tensor transpose_158 = transpose(perm = transpose_64_perm_0, x = q_35_cast); + tensor qk_17_cast = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_158, y = transpose_157); + tensor var_1027_cast = softmax(axis = var_962, x = qk_17_cast); + tensor var_1029_transpose_x_0 = const()[name = tensor("op_1029_transpose_x_0"), val = tensor(false)]; + tensor var_1029_transpose_y_0 = const()[name = tensor("op_1029_transpose_y_0"), val = tensor(false)]; + tensor transpose_159 = transpose(perm = var_1023, x = var_1022_cast); + tensor var_1029_cast = matmul(transpose_x = var_1029_transpose_x_0, transpose_y = var_1029_transpose_y_0, x = var_1027_cast, y = transpose_159); + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 1024])]; + tensor transpose_156 = transpose(perm = var_1030, x = var_1029_cast); + tensor x_107_cast = reshape(shape = concat_8, x = transpose_156); + tensor var_1035_to_fp16 = const()[name = tensor("op_1035_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217709440)))]; + tensor var_1036_to_fp16 = const()[name = tensor("op_1036_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219806656)))]; + tensor var_1037_cast = linear(bias = var_1036_to_fp16, weight = var_1035_to_fp16, x = x_107_cast); + tensor x_109_cast = add(x = x_103_cast, y = var_1037_cast); + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219808768)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219810880)))]; + tensor var_1043_cast = layer_norm(axes = var_1043_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_968_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast); + tensor var_1052_to_fp16 = const()[name = tensor("op_1052_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219812992)))]; + tensor var_1053_to_fp16 = const()[name = tensor("op_1053_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228201664)))]; + tensor input_73_cast = linear(bias = var_1053_to_fp16, weight = var_1052_to_fp16, x = var_1043_cast); + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast = gelu(mode = x_113_mode_0, x = input_73_cast); + tensor var_1058_to_fp16 = const()[name = tensor("op_1058_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228209920)))]; + tensor var_1059_to_fp16 = const()[name = tensor("op_1059_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236598592)))]; + tensor var_1060_cast = linear(bias = var_1059_to_fp16, weight = var_1058_to_fp16, x = x_113_cast); + tensor x_115_cast = add(x = x_109_cast, y = var_1060_cast); + tensor var_1069 = const()[name = tensor("op_1069"), val = tensor(-1)]; + tensor var_1086_axes_0 = const()[name = tensor("op_1086_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236600704)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236602816)))]; + tensor var_1075_to_fp16 = const()[name = tensor("op_1075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1086_cast = layer_norm(axes = var_1086_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1075_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast); + tensor var_1097_to_fp16 = const()[name = tensor("op_1097_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236604928)))]; + tensor var_1098_to_fp16 = const()[name = tensor("op_1098_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238702144)))]; + tensor q_37_cast = linear(bias = var_1098_to_fp16, weight = var_1097_to_fp16, x = var_1086_cast); + tensor var_1101_to_fp16 = const()[name = tensor("op_1101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238704256)))]; + tensor k_37_bias_0_to_fp16 = const()[name = tensor("k_37_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240801472)))]; + tensor k_37_cast = linear(bias = k_37_bias_0_to_fp16, weight = var_1101_to_fp16, x = var_1086_cast); + tensor var_1105_to_fp16 = const()[name = tensor("op_1105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240803584)))]; + tensor var_1106_to_fp16 = const()[name = tensor("op_1106_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242900800)))]; + tensor v_37_cast = linear(bias = var_1106_to_fp16, weight = var_1105_to_fp16, x = var_1086_cast); + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([1, 1500, 16, -1])]; + tensor var_1115_cast = reshape(shape = var_1114, x = q_37_cast); + tensor const_186_to_fp16 = const()[name = tensor("const_186_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast = mul(x = var_1115_cast, y = const_186_to_fp16); + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor([1, 1500, 16, -1])]; + tensor var_1122_cast = reshape(shape = var_1121, x = k_37_cast); + tensor const_187_to_fp16 = const()[name = tensor("const_187_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast = mul(x = var_1122_cast, y = const_187_to_fp16); + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1500, 16, -1])]; + tensor var_1129_cast = reshape(shape = var_1128, x = v_37_cast); + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_66_perm_0 = const()[name = tensor("transpose_66_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_67_perm_0 = const()[name = tensor("transpose_67_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_153 = transpose(perm = transpose_67_perm_0, x = k_39_cast); + tensor transpose_154 = transpose(perm = transpose_66_perm_0, x = q_39_cast); + tensor qk_19_cast = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_154, y = transpose_153); + tensor var_1134_cast = softmax(axis = var_1069, x = qk_19_cast); + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_155 = transpose(perm = var_1130, x = var_1129_cast); + tensor var_1136_cast = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = var_1134_cast, y = transpose_155); + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 1024])]; + tensor transpose_152 = transpose(perm = var_1137, x = var_1136_cast); + tensor x_119_cast = reshape(shape = concat_9, x = transpose_152); + tensor var_1142_to_fp16 = const()[name = tensor("op_1142_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242902912)))]; + tensor var_1143_to_fp16 = const()[name = tensor("op_1143_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245000128)))]; + tensor var_1144_cast = linear(bias = var_1143_to_fp16, weight = var_1142_to_fp16, x = x_119_cast); + tensor x_121_cast = add(x = x_115_cast, y = var_1144_cast); + tensor var_1150_axes_0 = const()[name = tensor("op_1150_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245002240)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245004352)))]; + tensor var_1150_cast = layer_norm(axes = var_1150_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1075_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast); + tensor var_1159_to_fp16 = const()[name = tensor("op_1159_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245006464)))]; + tensor var_1160_to_fp16 = const()[name = tensor("op_1160_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253395136)))]; + tensor input_81_cast = linear(bias = var_1160_to_fp16, weight = var_1159_to_fp16, x = var_1150_cast); + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast = gelu(mode = x_125_mode_0, x = input_81_cast); + tensor var_1165_to_fp16 = const()[name = tensor("op_1165_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253403392)))]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261792064)))]; + tensor var_1167_cast = linear(bias = var_1166_to_fp16, weight = var_1165_to_fp16, x = x_125_cast); + tensor x_127_cast = add(x = x_121_cast, y = var_1167_cast); + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor(-1)]; + tensor var_1193_axes_0 = const()[name = tensor("op_1193_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261794176)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261796288)))]; + tensor var_1182_to_fp16 = const()[name = tensor("op_1182_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1193_cast = layer_norm(axes = var_1193_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1182_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast); + tensor var_1204_to_fp16 = const()[name = tensor("op_1204_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261798400)))]; + tensor var_1205_to_fp16 = const()[name = tensor("op_1205_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263895616)))]; + tensor q_41_cast = linear(bias = var_1205_to_fp16, weight = var_1204_to_fp16, x = var_1193_cast); + tensor var_1208_to_fp16 = const()[name = tensor("op_1208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263897728)))]; + tensor k_41_bias_0_to_fp16 = const()[name = tensor("k_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265994944)))]; + tensor k_41_cast = linear(bias = k_41_bias_0_to_fp16, weight = var_1208_to_fp16, x = var_1193_cast); + tensor var_1212_to_fp16 = const()[name = tensor("op_1212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265997056)))]; + tensor var_1213_to_fp16 = const()[name = tensor("op_1213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268094272)))]; + tensor v_41_cast = linear(bias = var_1213_to_fp16, weight = var_1212_to_fp16, x = var_1193_cast); + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 1500, 16, -1])]; + tensor var_1222_cast = reshape(shape = var_1221, x = q_41_cast); + tensor const_188_to_fp16 = const()[name = tensor("const_188_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast = mul(x = var_1222_cast, y = const_188_to_fp16); + tensor var_1228 = const()[name = tensor("op_1228"), val = tensor([1, 1500, 16, -1])]; + tensor var_1229_cast = reshape(shape = var_1228, x = k_41_cast); + tensor const_189_to_fp16 = const()[name = tensor("const_189_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast = mul(x = var_1229_cast, y = const_189_to_fp16); + tensor var_1235 = const()[name = tensor("op_1235"), val = tensor([1, 1500, 16, -1])]; + tensor var_1236_cast = reshape(shape = var_1235, x = v_41_cast); + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_68_perm_0 = const()[name = tensor("transpose_68_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_149 = transpose(perm = transpose_69_perm_0, x = k_43_cast); + tensor transpose_150 = transpose(perm = transpose_68_perm_0, x = q_43_cast); + tensor qk_21_cast = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_150, y = transpose_149); + tensor var_1241_cast = softmax(axis = var_1176, x = qk_21_cast); + tensor var_1243_transpose_x_0 = const()[name = tensor("op_1243_transpose_x_0"), val = tensor(false)]; + tensor var_1243_transpose_y_0 = const()[name = tensor("op_1243_transpose_y_0"), val = tensor(false)]; + tensor transpose_151 = transpose(perm = var_1237, x = var_1236_cast); + tensor var_1243_cast = matmul(transpose_x = var_1243_transpose_x_0, transpose_y = var_1243_transpose_y_0, x = var_1241_cast, y = transpose_151); + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 1024])]; + tensor transpose_148 = transpose(perm = var_1244, x = var_1243_cast); + tensor x_131_cast = reshape(shape = concat_10, x = transpose_148); + tensor var_1249_to_fp16 = const()[name = tensor("op_1249_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268096384)))]; + tensor var_1250_to_fp16 = const()[name = tensor("op_1250_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270193600)))]; + tensor var_1251_cast = linear(bias = var_1250_to_fp16, weight = var_1249_to_fp16, x = x_131_cast); + tensor x_133_cast = add(x = x_127_cast, y = var_1251_cast); + tensor var_1257_axes_0 = const()[name = tensor("op_1257_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270195712)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270197824)))]; + tensor var_1257_cast = layer_norm(axes = var_1257_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1182_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast); + tensor var_1266_to_fp16 = const()[name = tensor("op_1266_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270199936)))]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278588608)))]; + tensor input_89_cast = linear(bias = var_1267_to_fp16, weight = var_1266_to_fp16, x = var_1257_cast); + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast = gelu(mode = x_137_mode_0, x = input_89_cast); + tensor var_1272_to_fp16 = const()[name = tensor("op_1272_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278596864)))]; + tensor var_1273_to_fp16 = const()[name = tensor("op_1273_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286985536)))]; + tensor var_1274_cast = linear(bias = var_1273_to_fp16, weight = var_1272_to_fp16, x = x_137_cast); + tensor x_139_cast = add(x = x_133_cast, y = var_1274_cast); + tensor var_1283 = const()[name = tensor("op_1283"), val = tensor(-1)]; + tensor var_1300_axes_0 = const()[name = tensor("op_1300_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286987648)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286989760)))]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1300_cast = layer_norm(axes = var_1300_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1289_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast); + tensor var_1311_to_fp16 = const()[name = tensor("op_1311_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286991872)))]; + tensor var_1312_to_fp16 = const()[name = tensor("op_1312_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289089088)))]; + tensor q_45_cast = linear(bias = var_1312_to_fp16, weight = var_1311_to_fp16, x = var_1300_cast); + tensor var_1315_to_fp16 = const()[name = tensor("op_1315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289091200)))]; + tensor k_45_bias_0_to_fp16 = const()[name = tensor("k_45_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291188416)))]; + tensor k_45_cast = linear(bias = k_45_bias_0_to_fp16, weight = var_1315_to_fp16, x = var_1300_cast); + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291190528)))]; + tensor var_1320_to_fp16 = const()[name = tensor("op_1320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293287744)))]; + tensor v_45_cast = linear(bias = var_1320_to_fp16, weight = var_1319_to_fp16, x = var_1300_cast); + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([1, 1500, 16, -1])]; + tensor var_1329_cast = reshape(shape = var_1328, x = q_45_cast); + tensor const_190_to_fp16 = const()[name = tensor("const_190_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_47_cast = mul(x = var_1329_cast, y = const_190_to_fp16); + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 1500, 16, -1])]; + tensor var_1336_cast = reshape(shape = var_1335, x = k_45_cast); + tensor const_191_to_fp16 = const()[name = tensor("const_191_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_47_cast = mul(x = var_1336_cast, y = const_191_to_fp16); + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([1, 1500, 16, -1])]; + tensor var_1343_cast = reshape(shape = var_1342, x = v_45_cast); + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([0, 2, 1, 3])]; + tensor qk_23_transpose_x_0 = const()[name = tensor("qk_23_transpose_x_0"), val = tensor(false)]; + tensor qk_23_transpose_y_0 = const()[name = tensor("qk_23_transpose_y_0"), val = tensor(false)]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_145 = transpose(perm = transpose_71_perm_0, x = k_47_cast); + tensor transpose_146 = transpose(perm = transpose_70_perm_0, x = q_47_cast); + tensor qk_23_cast = matmul(transpose_x = qk_23_transpose_x_0, transpose_y = qk_23_transpose_y_0, x = transpose_146, y = transpose_145); + tensor var_1348_cast = softmax(axis = var_1283, x = qk_23_cast); + tensor var_1350_transpose_x_0 = const()[name = tensor("op_1350_transpose_x_0"), val = tensor(false)]; + tensor var_1350_transpose_y_0 = const()[name = tensor("op_1350_transpose_y_0"), val = tensor(false)]; + tensor transpose_147 = transpose(perm = var_1344, x = var_1343_cast); + tensor var_1350_cast = matmul(transpose_x = var_1350_transpose_x_0, transpose_y = var_1350_transpose_y_0, x = var_1348_cast, y = transpose_147); + tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 1024])]; + tensor transpose_144 = transpose(perm = var_1351, x = var_1350_cast); + tensor x_143_cast = reshape(shape = concat_11, x = transpose_144); + tensor var_1356_to_fp16 = const()[name = tensor("op_1356_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293289856)))]; + tensor var_1357_to_fp16 = const()[name = tensor("op_1357_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295387072)))]; + tensor var_1358_cast = linear(bias = var_1357_to_fp16, weight = var_1356_to_fp16, x = x_143_cast); + tensor x_145_cast = add(x = x_139_cast, y = var_1358_cast); + tensor var_1364_axes_0 = const()[name = tensor("op_1364_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295389184)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295391296)))]; + tensor var_1364_cast = layer_norm(axes = var_1364_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1289_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast); + tensor var_1373_to_fp16 = const()[name = tensor("op_1373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295393408)))]; + tensor var_1374_to_fp16 = const()[name = tensor("op_1374_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782080)))]; + tensor input_97_cast = linear(bias = var_1374_to_fp16, weight = var_1373_to_fp16, x = var_1364_cast); + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast = gelu(mode = x_149_mode_0, x = input_97_cast); + tensor var_1379_to_fp16 = const()[name = tensor("op_1379_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790336)))]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312179008)))]; + tensor var_1381_cast = linear(bias = var_1380_to_fp16, weight = var_1379_to_fp16, x = x_149_cast); + tensor x_151_cast = add(x = x_145_cast, y = var_1381_cast); + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor(-1)]; + tensor var_1407_axes_0 = const()[name = tensor("op_1407_axes_0"), val = tensor([-1])]; + tensor blocks_12_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_12_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312181120)))]; + tensor blocks_12_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_12_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312183232)))]; + tensor var_1396_to_fp16 = const()[name = tensor("op_1396_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1407_cast = layer_norm(axes = var_1407_axes_0, beta = blocks_12_attn_ln_bias_to_fp16, epsilon = var_1396_to_fp16, gamma = blocks_12_attn_ln_weight_to_fp16, x = x_151_cast); + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312185344)))]; + tensor var_1419_to_fp16 = const()[name = tensor("op_1419_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314282560)))]; + tensor q_49_cast = linear(bias = var_1419_to_fp16, weight = var_1418_to_fp16, x = var_1407_cast); + tensor var_1422_to_fp16 = const()[name = tensor("op_1422_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314284672)))]; + tensor k_49_bias_0_to_fp16 = const()[name = tensor("k_49_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316381888)))]; + tensor k_49_cast = linear(bias = k_49_bias_0_to_fp16, weight = var_1422_to_fp16, x = var_1407_cast); + tensor var_1426_to_fp16 = const()[name = tensor("op_1426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316384000)))]; + tensor var_1427_to_fp16 = const()[name = tensor("op_1427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318481216)))]; + tensor v_49_cast = linear(bias = var_1427_to_fp16, weight = var_1426_to_fp16, x = var_1407_cast); + tensor var_1435 = const()[name = tensor("op_1435"), val = tensor([1, 1500, 16, -1])]; + tensor var_1436_cast = reshape(shape = var_1435, x = q_49_cast); + tensor const_192_to_fp16 = const()[name = tensor("const_192_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_51_cast = mul(x = var_1436_cast, y = const_192_to_fp16); + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1500, 16, -1])]; + tensor var_1443_cast = reshape(shape = var_1442, x = k_49_cast); + tensor const_193_to_fp16 = const()[name = tensor("const_193_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_51_cast = mul(x = var_1443_cast, y = const_193_to_fp16); + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1500, 16, -1])]; + tensor var_1450_cast = reshape(shape = var_1449, x = v_49_cast); + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([0, 2, 1, 3])]; + tensor qk_25_transpose_x_0 = const()[name = tensor("qk_25_transpose_x_0"), val = tensor(false)]; + tensor qk_25_transpose_y_0 = const()[name = tensor("qk_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_141 = transpose(perm = transpose_73_perm_0, x = k_51_cast); + tensor transpose_142 = transpose(perm = transpose_72_perm_0, x = q_51_cast); + tensor qk_25_cast = matmul(transpose_x = qk_25_transpose_x_0, transpose_y = qk_25_transpose_y_0, x = transpose_142, y = transpose_141); + tensor var_1455_cast = softmax(axis = var_1390, x = qk_25_cast); + tensor var_1457_transpose_x_0 = const()[name = tensor("op_1457_transpose_x_0"), val = tensor(false)]; + tensor var_1457_transpose_y_0 = const()[name = tensor("op_1457_transpose_y_0"), val = tensor(false)]; + tensor transpose_143 = transpose(perm = var_1451, x = var_1450_cast); + tensor var_1457_cast = matmul(transpose_x = var_1457_transpose_x_0, transpose_y = var_1457_transpose_y_0, x = var_1455_cast, y = transpose_143); + tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([0, 2, 1, 3])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 1500, 1024])]; + tensor transpose_140 = transpose(perm = var_1458, x = var_1457_cast); + tensor x_155_cast = reshape(shape = concat_12, x = transpose_140); + tensor var_1463_to_fp16 = const()[name = tensor("op_1463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318483328)))]; + tensor var_1464_to_fp16 = const()[name = tensor("op_1464_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320580544)))]; + tensor var_1465_cast = linear(bias = var_1464_to_fp16, weight = var_1463_to_fp16, x = x_155_cast); + tensor x_157_cast = add(x = x_151_cast, y = var_1465_cast); + tensor var_1471_axes_0 = const()[name = tensor("op_1471_axes_0"), val = tensor([-1])]; + tensor blocks_12_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320582656)))]; + tensor blocks_12_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_12_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320584768)))]; + tensor var_1471_cast = layer_norm(axes = var_1471_axes_0, beta = blocks_12_mlp_ln_bias_to_fp16, epsilon = var_1396_to_fp16, gamma = blocks_12_mlp_ln_weight_to_fp16, x = x_157_cast); + tensor var_1480_to_fp16 = const()[name = tensor("op_1480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320586880)))]; + tensor var_1481_to_fp16 = const()[name = tensor("op_1481_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328975552)))]; + tensor input_105_cast = linear(bias = var_1481_to_fp16, weight = var_1480_to_fp16, x = var_1471_cast); + tensor x_161_mode_0 = const()[name = tensor("x_161_mode_0"), val = tensor("EXACT")]; + tensor x_161_cast = gelu(mode = x_161_mode_0, x = input_105_cast); + tensor var_1486_to_fp16 = const()[name = tensor("op_1486_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328983808)))]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337372480)))]; + tensor var_1488_cast = linear(bias = var_1487_to_fp16, weight = var_1486_to_fp16, x = x_161_cast); + tensor x_163_cast = add(x = x_157_cast, y = var_1488_cast); + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor(-1)]; + tensor var_1514_axes_0 = const()[name = tensor("op_1514_axes_0"), val = tensor([-1])]; + tensor blocks_13_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_13_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337374592)))]; + tensor blocks_13_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_13_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337376704)))]; + tensor var_1503_to_fp16 = const()[name = tensor("op_1503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1514_cast = layer_norm(axes = var_1514_axes_0, beta = blocks_13_attn_ln_bias_to_fp16, epsilon = var_1503_to_fp16, gamma = blocks_13_attn_ln_weight_to_fp16, x = x_163_cast); + tensor var_1525_to_fp16 = const()[name = tensor("op_1525_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337378816)))]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339476032)))]; + tensor q_53_cast = linear(bias = var_1526_to_fp16, weight = var_1525_to_fp16, x = var_1514_cast); + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339478144)))]; + tensor k_53_bias_0_to_fp16 = const()[name = tensor("k_53_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341575360)))]; + tensor k_53_cast = linear(bias = k_53_bias_0_to_fp16, weight = var_1529_to_fp16, x = var_1514_cast); + tensor var_1533_to_fp16 = const()[name = tensor("op_1533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341577472)))]; + tensor var_1534_to_fp16 = const()[name = tensor("op_1534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343674688)))]; + tensor v_53_cast = linear(bias = var_1534_to_fp16, weight = var_1533_to_fp16, x = var_1514_cast); + tensor var_1542 = const()[name = tensor("op_1542"), val = tensor([1, 1500, 16, -1])]; + tensor var_1543_cast = reshape(shape = var_1542, x = q_53_cast); + tensor const_194_to_fp16 = const()[name = tensor("const_194_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_55_cast = mul(x = var_1543_cast, y = const_194_to_fp16); + tensor var_1549 = const()[name = tensor("op_1549"), val = tensor([1, 1500, 16, -1])]; + tensor var_1550_cast = reshape(shape = var_1549, x = k_53_cast); + tensor const_195_to_fp16 = const()[name = tensor("const_195_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_55_cast = mul(x = var_1550_cast, y = const_195_to_fp16); + tensor var_1556 = const()[name = tensor("op_1556"), val = tensor([1, 1500, 16, -1])]; + tensor var_1557_cast = reshape(shape = var_1556, x = v_53_cast); + tensor var_1558 = const()[name = tensor("op_1558"), val = tensor([0, 2, 1, 3])]; + tensor qk_27_transpose_x_0 = const()[name = tensor("qk_27_transpose_x_0"), val = tensor(false)]; + tensor qk_27_transpose_y_0 = const()[name = tensor("qk_27_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_137 = transpose(perm = transpose_75_perm_0, x = k_55_cast); + tensor transpose_138 = transpose(perm = transpose_74_perm_0, x = q_55_cast); + tensor qk_27_cast = matmul(transpose_x = qk_27_transpose_x_0, transpose_y = qk_27_transpose_y_0, x = transpose_138, y = transpose_137); + tensor var_1562_cast = softmax(axis = var_1497, x = qk_27_cast); + tensor var_1564_transpose_x_0 = const()[name = tensor("op_1564_transpose_x_0"), val = tensor(false)]; + tensor var_1564_transpose_y_0 = const()[name = tensor("op_1564_transpose_y_0"), val = tensor(false)]; + tensor transpose_139 = transpose(perm = var_1558, x = var_1557_cast); + tensor var_1564_cast = matmul(transpose_x = var_1564_transpose_x_0, transpose_y = var_1564_transpose_y_0, x = var_1562_cast, y = transpose_139); + tensor var_1565 = const()[name = tensor("op_1565"), val = tensor([0, 2, 1, 3])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([1, 1500, 1024])]; + tensor transpose_136 = transpose(perm = var_1565, x = var_1564_cast); + tensor x_167_cast = reshape(shape = concat_13, x = transpose_136); + tensor var_1570_to_fp16 = const()[name = tensor("op_1570_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343676800)))]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345774016)))]; + tensor var_1572_cast = linear(bias = var_1571_to_fp16, weight = var_1570_to_fp16, x = x_167_cast); + tensor x_169_cast = add(x = x_163_cast, y = var_1572_cast); + tensor var_1578_axes_0 = const()[name = tensor("op_1578_axes_0"), val = tensor([-1])]; + tensor blocks_13_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345776128)))]; + tensor blocks_13_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_13_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345778240)))]; + tensor var_1578_cast = layer_norm(axes = var_1578_axes_0, beta = blocks_13_mlp_ln_bias_to_fp16, epsilon = var_1503_to_fp16, gamma = blocks_13_mlp_ln_weight_to_fp16, x = x_169_cast); + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345780352)))]; + tensor var_1588_to_fp16 = const()[name = tensor("op_1588_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354169024)))]; + tensor input_113_cast = linear(bias = var_1588_to_fp16, weight = var_1587_to_fp16, x = var_1578_cast); + tensor x_173_mode_0 = const()[name = tensor("x_173_mode_0"), val = tensor("EXACT")]; + tensor x_173_cast = gelu(mode = x_173_mode_0, x = input_113_cast); + tensor var_1593_to_fp16 = const()[name = tensor("op_1593_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354177280)))]; + tensor var_1594_to_fp16 = const()[name = tensor("op_1594_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362565952)))]; + tensor var_1595_cast = linear(bias = var_1594_to_fp16, weight = var_1593_to_fp16, x = x_173_cast); + tensor x_175_cast = add(x = x_169_cast, y = var_1595_cast); + tensor var_1604 = const()[name = tensor("op_1604"), val = tensor(-1)]; + tensor var_1621_axes_0 = const()[name = tensor("op_1621_axes_0"), val = tensor([-1])]; + tensor blocks_14_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_14_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362568064)))]; + tensor blocks_14_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_14_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362570176)))]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1621_cast = layer_norm(axes = var_1621_axes_0, beta = blocks_14_attn_ln_bias_to_fp16, epsilon = var_1610_to_fp16, gamma = blocks_14_attn_ln_weight_to_fp16, x = x_175_cast); + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362572288)))]; + tensor var_1633_to_fp16 = const()[name = tensor("op_1633_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364669504)))]; + tensor q_57_cast = linear(bias = var_1633_to_fp16, weight = var_1632_to_fp16, x = var_1621_cast); + tensor var_1636_to_fp16 = const()[name = tensor("op_1636_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364671616)))]; + tensor k_57_bias_0_to_fp16 = const()[name = tensor("k_57_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366768832)))]; + tensor k_57_cast = linear(bias = k_57_bias_0_to_fp16, weight = var_1636_to_fp16, x = var_1621_cast); + tensor var_1640_to_fp16 = const()[name = tensor("op_1640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366770944)))]; + tensor var_1641_to_fp16 = const()[name = tensor("op_1641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368868160)))]; + tensor v_57_cast = linear(bias = var_1641_to_fp16, weight = var_1640_to_fp16, x = var_1621_cast); + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor([1, 1500, 16, -1])]; + tensor var_1650_cast = reshape(shape = var_1649, x = q_57_cast); + tensor const_196_to_fp16 = const()[name = tensor("const_196_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_59_cast = mul(x = var_1650_cast, y = const_196_to_fp16); + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1500, 16, -1])]; + tensor var_1657_cast = reshape(shape = var_1656, x = k_57_cast); + tensor const_197_to_fp16 = const()[name = tensor("const_197_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_59_cast = mul(x = var_1657_cast, y = const_197_to_fp16); + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1500, 16, -1])]; + tensor var_1664_cast = reshape(shape = var_1663, x = v_57_cast); + tensor var_1665 = const()[name = tensor("op_1665"), val = tensor([0, 2, 1, 3])]; + tensor qk_29_transpose_x_0 = const()[name = tensor("qk_29_transpose_x_0"), val = tensor(false)]; + tensor qk_29_transpose_y_0 = const()[name = tensor("qk_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_133 = transpose(perm = transpose_77_perm_0, x = k_59_cast); + tensor transpose_134 = transpose(perm = transpose_76_perm_0, x = q_59_cast); + tensor qk_29_cast = matmul(transpose_x = qk_29_transpose_x_0, transpose_y = qk_29_transpose_y_0, x = transpose_134, y = transpose_133); + tensor var_1669_cast = softmax(axis = var_1604, x = qk_29_cast); + tensor var_1671_transpose_x_0 = const()[name = tensor("op_1671_transpose_x_0"), val = tensor(false)]; + tensor var_1671_transpose_y_0 = const()[name = tensor("op_1671_transpose_y_0"), val = tensor(false)]; + tensor transpose_135 = transpose(perm = var_1665, x = var_1664_cast); + tensor var_1671_cast = matmul(transpose_x = var_1671_transpose_x_0, transpose_y = var_1671_transpose_y_0, x = var_1669_cast, y = transpose_135); + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([0, 2, 1, 3])]; + tensor concat_14 = const()[name = tensor("concat_14"), val = tensor([1, 1500, 1024])]; + tensor transpose_132 = transpose(perm = var_1672, x = var_1671_cast); + tensor x_179_cast = reshape(shape = concat_14, x = transpose_132); + tensor var_1677_to_fp16 = const()[name = tensor("op_1677_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368870272)))]; + tensor var_1678_to_fp16 = const()[name = tensor("op_1678_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370967488)))]; + tensor var_1679_cast = linear(bias = var_1678_to_fp16, weight = var_1677_to_fp16, x = x_179_cast); + tensor x_181_cast = add(x = x_175_cast, y = var_1679_cast); + tensor var_1685_axes_0 = const()[name = tensor("op_1685_axes_0"), val = tensor([-1])]; + tensor blocks_14_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370969600)))]; + tensor blocks_14_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_14_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370971712)))]; + tensor var_1685_cast = layer_norm(axes = var_1685_axes_0, beta = blocks_14_mlp_ln_bias_to_fp16, epsilon = var_1610_to_fp16, gamma = blocks_14_mlp_ln_weight_to_fp16, x = x_181_cast); + tensor var_1694_to_fp16 = const()[name = tensor("op_1694_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370973824)))]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379362496)))]; + tensor input_121_cast = linear(bias = var_1695_to_fp16, weight = var_1694_to_fp16, x = var_1685_cast); + tensor x_185_mode_0 = const()[name = tensor("x_185_mode_0"), val = tensor("EXACT")]; + tensor x_185_cast = gelu(mode = x_185_mode_0, x = input_121_cast); + tensor var_1700_to_fp16 = const()[name = tensor("op_1700_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379370752)))]; + tensor var_1701_to_fp16 = const()[name = tensor("op_1701_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387759424)))]; + tensor var_1702_cast = linear(bias = var_1701_to_fp16, weight = var_1700_to_fp16, x = x_185_cast); + tensor x_187_cast = add(x = x_181_cast, y = var_1702_cast); + tensor var_1711 = const()[name = tensor("op_1711"), val = tensor(-1)]; + tensor var_1728_axes_0 = const()[name = tensor("op_1728_axes_0"), val = tensor([-1])]; + tensor blocks_15_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_15_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387761536)))]; + tensor blocks_15_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_15_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387763648)))]; + tensor var_1717_to_fp16 = const()[name = tensor("op_1717_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1728_cast = layer_norm(axes = var_1728_axes_0, beta = blocks_15_attn_ln_bias_to_fp16, epsilon = var_1717_to_fp16, gamma = blocks_15_attn_ln_weight_to_fp16, x = x_187_cast); + tensor var_1739_to_fp16 = const()[name = tensor("op_1739_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387765760)))]; + tensor var_1740_to_fp16 = const()[name = tensor("op_1740_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389862976)))]; + tensor q_61_cast = linear(bias = var_1740_to_fp16, weight = var_1739_to_fp16, x = var_1728_cast); + tensor var_1743_to_fp16 = const()[name = tensor("op_1743_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389865088)))]; + tensor k_61_bias_0_to_fp16 = const()[name = tensor("k_61_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391962304)))]; + tensor k_61_cast = linear(bias = k_61_bias_0_to_fp16, weight = var_1743_to_fp16, x = var_1728_cast); + tensor var_1747_to_fp16 = const()[name = tensor("op_1747_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391964416)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394061632)))]; + tensor v_61_cast = linear(bias = var_1748_to_fp16, weight = var_1747_to_fp16, x = var_1728_cast); + tensor var_1756 = const()[name = tensor("op_1756"), val = tensor([1, 1500, 16, -1])]; + tensor var_1757_cast = reshape(shape = var_1756, x = q_61_cast); + tensor const_198_to_fp16 = const()[name = tensor("const_198_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_63_cast = mul(x = var_1757_cast, y = const_198_to_fp16); + tensor var_1763 = const()[name = tensor("op_1763"), val = tensor([1, 1500, 16, -1])]; + tensor var_1764_cast = reshape(shape = var_1763, x = k_61_cast); + tensor const_199_to_fp16 = const()[name = tensor("const_199_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_63_cast = mul(x = var_1764_cast, y = const_199_to_fp16); + tensor var_1770 = const()[name = tensor("op_1770"), val = tensor([1, 1500, 16, -1])]; + tensor var_1771_cast = reshape(shape = var_1770, x = v_61_cast); + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([0, 2, 1, 3])]; + tensor qk_31_transpose_x_0 = const()[name = tensor("qk_31_transpose_x_0"), val = tensor(false)]; + tensor qk_31_transpose_y_0 = const()[name = tensor("qk_31_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_129 = transpose(perm = transpose_79_perm_0, x = k_63_cast); + tensor transpose_130 = transpose(perm = transpose_78_perm_0, x = q_63_cast); + tensor qk_31_cast = matmul(transpose_x = qk_31_transpose_x_0, transpose_y = qk_31_transpose_y_0, x = transpose_130, y = transpose_129); + tensor var_1776_cast = softmax(axis = var_1711, x = qk_31_cast); + tensor var_1778_transpose_x_0 = const()[name = tensor("op_1778_transpose_x_0"), val = tensor(false)]; + tensor var_1778_transpose_y_0 = const()[name = tensor("op_1778_transpose_y_0"), val = tensor(false)]; + tensor transpose_131 = transpose(perm = var_1772, x = var_1771_cast); + tensor var_1778_cast = matmul(transpose_x = var_1778_transpose_x_0, transpose_y = var_1778_transpose_y_0, x = var_1776_cast, y = transpose_131); + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([0, 2, 1, 3])]; + tensor concat_15 = const()[name = tensor("concat_15"), val = tensor([1, 1500, 1024])]; + tensor transpose_128 = transpose(perm = var_1779, x = var_1778_cast); + tensor x_191_cast = reshape(shape = concat_15, x = transpose_128); + tensor var_1784_to_fp16 = const()[name = tensor("op_1784_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394063744)))]; + tensor var_1785_to_fp16 = const()[name = tensor("op_1785_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396160960)))]; + tensor var_1786_cast = linear(bias = var_1785_to_fp16, weight = var_1784_to_fp16, x = x_191_cast); + tensor x_193_cast = add(x = x_187_cast, y = var_1786_cast); + tensor var_1792_axes_0 = const()[name = tensor("op_1792_axes_0"), val = tensor([-1])]; + tensor blocks_15_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396163072)))]; + tensor blocks_15_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_15_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396165184)))]; + tensor var_1792_cast = layer_norm(axes = var_1792_axes_0, beta = blocks_15_mlp_ln_bias_to_fp16, epsilon = var_1717_to_fp16, gamma = blocks_15_mlp_ln_weight_to_fp16, x = x_193_cast); + tensor var_1801_to_fp16 = const()[name = tensor("op_1801_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396167296)))]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404555968)))]; + tensor input_129_cast = linear(bias = var_1802_to_fp16, weight = var_1801_to_fp16, x = var_1792_cast); + tensor x_197_mode_0 = const()[name = tensor("x_197_mode_0"), val = tensor("EXACT")]; + tensor x_197_cast = gelu(mode = x_197_mode_0, x = input_129_cast); + tensor var_1807_to_fp16 = const()[name = tensor("op_1807_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404564224)))]; + tensor var_1808_to_fp16 = const()[name = tensor("op_1808_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412952896)))]; + tensor var_1809_cast = linear(bias = var_1808_to_fp16, weight = var_1807_to_fp16, x = x_197_cast); + tensor x_199_cast = add(x = x_193_cast, y = var_1809_cast); + tensor var_1818 = const()[name = tensor("op_1818"), val = tensor(-1)]; + tensor var_1835_axes_0 = const()[name = tensor("op_1835_axes_0"), val = tensor([-1])]; + tensor blocks_16_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_16_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412955008)))]; + tensor blocks_16_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_16_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412957120)))]; + tensor var_1824_to_fp16 = const()[name = tensor("op_1824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1835_cast = layer_norm(axes = var_1835_axes_0, beta = blocks_16_attn_ln_bias_to_fp16, epsilon = var_1824_to_fp16, gamma = blocks_16_attn_ln_weight_to_fp16, x = x_199_cast); + tensor var_1846_to_fp16 = const()[name = tensor("op_1846_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412959232)))]; + tensor var_1847_to_fp16 = const()[name = tensor("op_1847_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415056448)))]; + tensor q_65_cast = linear(bias = var_1847_to_fp16, weight = var_1846_to_fp16, x = var_1835_cast); + tensor var_1850_to_fp16 = const()[name = tensor("op_1850_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415058560)))]; + tensor k_65_bias_0_to_fp16 = const()[name = tensor("k_65_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417155776)))]; + tensor k_65_cast = linear(bias = k_65_bias_0_to_fp16, weight = var_1850_to_fp16, x = var_1835_cast); + tensor var_1854_to_fp16 = const()[name = tensor("op_1854_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417157888)))]; + tensor var_1855_to_fp16 = const()[name = tensor("op_1855_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419255104)))]; + tensor v_65_cast = linear(bias = var_1855_to_fp16, weight = var_1854_to_fp16, x = var_1835_cast); + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 1500, 16, -1])]; + tensor var_1864_cast = reshape(shape = var_1863, x = q_65_cast); + tensor const_200_to_fp16 = const()[name = tensor("const_200_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_67_cast = mul(x = var_1864_cast, y = const_200_to_fp16); + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 1500, 16, -1])]; + tensor var_1871_cast = reshape(shape = var_1870, x = k_65_cast); + tensor const_201_to_fp16 = const()[name = tensor("const_201_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_67_cast = mul(x = var_1871_cast, y = const_201_to_fp16); + tensor var_1877 = const()[name = tensor("op_1877"), val = tensor([1, 1500, 16, -1])]; + tensor var_1878_cast = reshape(shape = var_1877, x = v_65_cast); + tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([0, 2, 1, 3])]; + tensor qk_33_transpose_x_0 = const()[name = tensor("qk_33_transpose_x_0"), val = tensor(false)]; + tensor qk_33_transpose_y_0 = const()[name = tensor("qk_33_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_125 = transpose(perm = transpose_81_perm_0, x = k_67_cast); + tensor transpose_126 = transpose(perm = transpose_80_perm_0, x = q_67_cast); + tensor qk_33_cast = matmul(transpose_x = qk_33_transpose_x_0, transpose_y = qk_33_transpose_y_0, x = transpose_126, y = transpose_125); + tensor var_1883_cast = softmax(axis = var_1818, x = qk_33_cast); + tensor var_1885_transpose_x_0 = const()[name = tensor("op_1885_transpose_x_0"), val = tensor(false)]; + tensor var_1885_transpose_y_0 = const()[name = tensor("op_1885_transpose_y_0"), val = tensor(false)]; + tensor transpose_127 = transpose(perm = var_1879, x = var_1878_cast); + tensor var_1885_cast = matmul(transpose_x = var_1885_transpose_x_0, transpose_y = var_1885_transpose_y_0, x = var_1883_cast, y = transpose_127); + tensor var_1886 = const()[name = tensor("op_1886"), val = tensor([0, 2, 1, 3])]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([1, 1500, 1024])]; + tensor transpose_124 = transpose(perm = var_1886, x = var_1885_cast); + tensor x_203_cast = reshape(shape = concat_16, x = transpose_124); + tensor var_1891_to_fp16 = const()[name = tensor("op_1891_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419257216)))]; + tensor var_1892_to_fp16 = const()[name = tensor("op_1892_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421354432)))]; + tensor var_1893_cast = linear(bias = var_1892_to_fp16, weight = var_1891_to_fp16, x = x_203_cast); + tensor x_205_cast = add(x = x_199_cast, y = var_1893_cast); + tensor var_1899_axes_0 = const()[name = tensor("op_1899_axes_0"), val = tensor([-1])]; + tensor blocks_16_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421356544)))]; + tensor blocks_16_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_16_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421358656)))]; + tensor var_1899_cast = layer_norm(axes = var_1899_axes_0, beta = blocks_16_mlp_ln_bias_to_fp16, epsilon = var_1824_to_fp16, gamma = blocks_16_mlp_ln_weight_to_fp16, x = x_205_cast); + tensor var_1908_to_fp16 = const()[name = tensor("op_1908_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421360768)))]; + tensor var_1909_to_fp16 = const()[name = tensor("op_1909_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429749440)))]; + tensor input_137_cast = linear(bias = var_1909_to_fp16, weight = var_1908_to_fp16, x = var_1899_cast); + tensor x_209_mode_0 = const()[name = tensor("x_209_mode_0"), val = tensor("EXACT")]; + tensor x_209_cast = gelu(mode = x_209_mode_0, x = input_137_cast); + tensor var_1914_to_fp16 = const()[name = tensor("op_1914_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429757696)))]; + tensor var_1915_to_fp16 = const()[name = tensor("op_1915_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438146368)))]; + tensor var_1916_cast = linear(bias = var_1915_to_fp16, weight = var_1914_to_fp16, x = x_209_cast); + tensor x_211_cast = add(x = x_205_cast, y = var_1916_cast); + tensor var_1925 = const()[name = tensor("op_1925"), val = tensor(-1)]; + tensor var_1942_axes_0 = const()[name = tensor("op_1942_axes_0"), val = tensor([-1])]; + tensor blocks_17_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_17_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438148480)))]; + tensor blocks_17_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_17_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438150592)))]; + tensor var_1931_to_fp16 = const()[name = tensor("op_1931_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1942_cast = layer_norm(axes = var_1942_axes_0, beta = blocks_17_attn_ln_bias_to_fp16, epsilon = var_1931_to_fp16, gamma = blocks_17_attn_ln_weight_to_fp16, x = x_211_cast); + tensor var_1953_to_fp16 = const()[name = tensor("op_1953_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438152704)))]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440249920)))]; + tensor q_69_cast = linear(bias = var_1954_to_fp16, weight = var_1953_to_fp16, x = var_1942_cast); + tensor var_1957_to_fp16 = const()[name = tensor("op_1957_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440252032)))]; + tensor k_69_bias_0_to_fp16 = const()[name = tensor("k_69_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442349248)))]; + tensor k_69_cast = linear(bias = k_69_bias_0_to_fp16, weight = var_1957_to_fp16, x = var_1942_cast); + tensor var_1961_to_fp16 = const()[name = tensor("op_1961_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(442351360)))]; + tensor var_1962_to_fp16 = const()[name = tensor("op_1962_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444448576)))]; + tensor v_69_cast = linear(bias = var_1962_to_fp16, weight = var_1961_to_fp16, x = var_1942_cast); + tensor var_1970 = const()[name = tensor("op_1970"), val = tensor([1, 1500, 16, -1])]; + tensor var_1971_cast = reshape(shape = var_1970, x = q_69_cast); + tensor const_202_to_fp16 = const()[name = tensor("const_202_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_71_cast = mul(x = var_1971_cast, y = const_202_to_fp16); + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1500, 16, -1])]; + tensor var_1978_cast = reshape(shape = var_1977, x = k_69_cast); + tensor const_203_to_fp16 = const()[name = tensor("const_203_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_71_cast = mul(x = var_1978_cast, y = const_203_to_fp16); + tensor var_1984 = const()[name = tensor("op_1984"), val = tensor([1, 1500, 16, -1])]; + tensor var_1985_cast = reshape(shape = var_1984, x = v_69_cast); + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([0, 2, 1, 3])]; + tensor qk_35_transpose_x_0 = const()[name = tensor("qk_35_transpose_x_0"), val = tensor(false)]; + tensor qk_35_transpose_y_0 = const()[name = tensor("qk_35_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_121 = transpose(perm = transpose_83_perm_0, x = k_71_cast); + tensor transpose_122 = transpose(perm = transpose_82_perm_0, x = q_71_cast); + tensor qk_35_cast = matmul(transpose_x = qk_35_transpose_x_0, transpose_y = qk_35_transpose_y_0, x = transpose_122, y = transpose_121); + tensor var_1990_cast = softmax(axis = var_1925, x = qk_35_cast); + tensor var_1992_transpose_x_0 = const()[name = tensor("op_1992_transpose_x_0"), val = tensor(false)]; + tensor var_1992_transpose_y_0 = const()[name = tensor("op_1992_transpose_y_0"), val = tensor(false)]; + tensor transpose_123 = transpose(perm = var_1986, x = var_1985_cast); + tensor var_1992_cast = matmul(transpose_x = var_1992_transpose_x_0, transpose_y = var_1992_transpose_y_0, x = var_1990_cast, y = transpose_123); + tensor var_1993 = const()[name = tensor("op_1993"), val = tensor([0, 2, 1, 3])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([1, 1500, 1024])]; + tensor transpose_120 = transpose(perm = var_1993, x = var_1992_cast); + tensor x_215_cast = reshape(shape = concat_17, x = transpose_120); + tensor var_1998_to_fp16 = const()[name = tensor("op_1998_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444450688)))]; + tensor var_1999_to_fp16 = const()[name = tensor("op_1999_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446547904)))]; + tensor var_2000_cast = linear(bias = var_1999_to_fp16, weight = var_1998_to_fp16, x = x_215_cast); + tensor x_217_cast = add(x = x_211_cast, y = var_2000_cast); + tensor var_2006_axes_0 = const()[name = tensor("op_2006_axes_0"), val = tensor([-1])]; + tensor blocks_17_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446550016)))]; + tensor blocks_17_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_17_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446552128)))]; + tensor var_2006_cast = layer_norm(axes = var_2006_axes_0, beta = blocks_17_mlp_ln_bias_to_fp16, epsilon = var_1931_to_fp16, gamma = blocks_17_mlp_ln_weight_to_fp16, x = x_217_cast); + tensor var_2015_to_fp16 = const()[name = tensor("op_2015_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446554240)))]; + tensor var_2016_to_fp16 = const()[name = tensor("op_2016_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454942912)))]; + tensor input_145_cast = linear(bias = var_2016_to_fp16, weight = var_2015_to_fp16, x = var_2006_cast); + tensor x_221_mode_0 = const()[name = tensor("x_221_mode_0"), val = tensor("EXACT")]; + tensor x_221_cast = gelu(mode = x_221_mode_0, x = input_145_cast); + tensor var_2021_to_fp16 = const()[name = tensor("op_2021_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454951168)))]; + tensor var_2022_to_fp16 = const()[name = tensor("op_2022_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463339840)))]; + tensor var_2023_cast = linear(bias = var_2022_to_fp16, weight = var_2021_to_fp16, x = x_221_cast); + tensor x_223_cast = add(x = x_217_cast, y = var_2023_cast); + tensor var_2032 = const()[name = tensor("op_2032"), val = tensor(-1)]; + tensor var_2049_axes_0 = const()[name = tensor("op_2049_axes_0"), val = tensor([-1])]; + tensor blocks_18_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_18_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463341952)))]; + tensor blocks_18_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_18_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463344064)))]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2049_cast = layer_norm(axes = var_2049_axes_0, beta = blocks_18_attn_ln_bias_to_fp16, epsilon = var_2038_to_fp16, gamma = blocks_18_attn_ln_weight_to_fp16, x = x_223_cast); + tensor var_2060_to_fp16 = const()[name = tensor("op_2060_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463346176)))]; + tensor var_2061_to_fp16 = const()[name = tensor("op_2061_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465443392)))]; + tensor q_73_cast = linear(bias = var_2061_to_fp16, weight = var_2060_to_fp16, x = var_2049_cast); + tensor var_2064_to_fp16 = const()[name = tensor("op_2064_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465445504)))]; + tensor k_73_bias_0_to_fp16 = const()[name = tensor("k_73_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467542720)))]; + tensor k_73_cast = linear(bias = k_73_bias_0_to_fp16, weight = var_2064_to_fp16, x = var_2049_cast); + tensor var_2068_to_fp16 = const()[name = tensor("op_2068_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467544832)))]; + tensor var_2069_to_fp16 = const()[name = tensor("op_2069_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469642048)))]; + tensor v_73_cast = linear(bias = var_2069_to_fp16, weight = var_2068_to_fp16, x = var_2049_cast); + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1500, 16, -1])]; + tensor var_2078_cast = reshape(shape = var_2077, x = q_73_cast); + tensor const_204_to_fp16 = const()[name = tensor("const_204_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_75_cast = mul(x = var_2078_cast, y = const_204_to_fp16); + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1500, 16, -1])]; + tensor var_2085_cast = reshape(shape = var_2084, x = k_73_cast); + tensor const_205_to_fp16 = const()[name = tensor("const_205_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_75_cast = mul(x = var_2085_cast, y = const_205_to_fp16); + tensor var_2091 = const()[name = tensor("op_2091"), val = tensor([1, 1500, 16, -1])]; + tensor var_2092_cast = reshape(shape = var_2091, x = v_73_cast); + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([0, 2, 1, 3])]; + tensor qk_37_transpose_x_0 = const()[name = tensor("qk_37_transpose_x_0"), val = tensor(false)]; + tensor qk_37_transpose_y_0 = const()[name = tensor("qk_37_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_117 = transpose(perm = transpose_85_perm_0, x = k_75_cast); + tensor transpose_118 = transpose(perm = transpose_84_perm_0, x = q_75_cast); + tensor qk_37_cast = matmul(transpose_x = qk_37_transpose_x_0, transpose_y = qk_37_transpose_y_0, x = transpose_118, y = transpose_117); + tensor var_2097_cast = softmax(axis = var_2032, x = qk_37_cast); + tensor var_2099_transpose_x_0 = const()[name = tensor("op_2099_transpose_x_0"), val = tensor(false)]; + tensor var_2099_transpose_y_0 = const()[name = tensor("op_2099_transpose_y_0"), val = tensor(false)]; + tensor transpose_119 = transpose(perm = var_2093, x = var_2092_cast); + tensor var_2099_cast = matmul(transpose_x = var_2099_transpose_x_0, transpose_y = var_2099_transpose_y_0, x = var_2097_cast, y = transpose_119); + tensor var_2100 = const()[name = tensor("op_2100"), val = tensor([0, 2, 1, 3])]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 1500, 1024])]; + tensor transpose_116 = transpose(perm = var_2100, x = var_2099_cast); + tensor x_227_cast = reshape(shape = concat_18, x = transpose_116); + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469644160)))]; + tensor var_2106_to_fp16 = const()[name = tensor("op_2106_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471741376)))]; + tensor var_2107_cast = linear(bias = var_2106_to_fp16, weight = var_2105_to_fp16, x = x_227_cast); + tensor x_229_cast = add(x = x_223_cast, y = var_2107_cast); + tensor var_2113_axes_0 = const()[name = tensor("op_2113_axes_0"), val = tensor([-1])]; + tensor blocks_18_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471743488)))]; + tensor blocks_18_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_18_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471745600)))]; + tensor var_2113_cast = layer_norm(axes = var_2113_axes_0, beta = blocks_18_mlp_ln_bias_to_fp16, epsilon = var_2038_to_fp16, gamma = blocks_18_mlp_ln_weight_to_fp16, x = x_229_cast); + tensor var_2122_to_fp16 = const()[name = tensor("op_2122_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471747712)))]; + tensor var_2123_to_fp16 = const()[name = tensor("op_2123_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480136384)))]; + tensor input_153_cast = linear(bias = var_2123_to_fp16, weight = var_2122_to_fp16, x = var_2113_cast); + tensor x_233_mode_0 = const()[name = tensor("x_233_mode_0"), val = tensor("EXACT")]; + tensor x_233_cast = gelu(mode = x_233_mode_0, x = input_153_cast); + tensor var_2128_to_fp16 = const()[name = tensor("op_2128_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480144640)))]; + tensor var_2129_to_fp16 = const()[name = tensor("op_2129_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488533312)))]; + tensor var_2130_cast = linear(bias = var_2129_to_fp16, weight = var_2128_to_fp16, x = x_233_cast); + tensor x_235_cast = add(x = x_229_cast, y = var_2130_cast); + tensor var_2139 = const()[name = tensor("op_2139"), val = tensor(-1)]; + tensor var_2156_axes_0 = const()[name = tensor("op_2156_axes_0"), val = tensor([-1])]; + tensor blocks_19_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_19_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488535424)))]; + tensor blocks_19_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_19_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488537536)))]; + tensor var_2145_to_fp16 = const()[name = tensor("op_2145_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2156_cast = layer_norm(axes = var_2156_axes_0, beta = blocks_19_attn_ln_bias_to_fp16, epsilon = var_2145_to_fp16, gamma = blocks_19_attn_ln_weight_to_fp16, x = x_235_cast); + tensor var_2167_to_fp16 = const()[name = tensor("op_2167_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488539648)))]; + tensor var_2168_to_fp16 = const()[name = tensor("op_2168_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490636864)))]; + tensor q_77_cast = linear(bias = var_2168_to_fp16, weight = var_2167_to_fp16, x = var_2156_cast); + tensor var_2171_to_fp16 = const()[name = tensor("op_2171_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490638976)))]; + tensor k_77_bias_0_to_fp16 = const()[name = tensor("k_77_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492736192)))]; + tensor k_77_cast = linear(bias = k_77_bias_0_to_fp16, weight = var_2171_to_fp16, x = var_2156_cast); + tensor var_2175_to_fp16 = const()[name = tensor("op_2175_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492738304)))]; + tensor var_2176_to_fp16 = const()[name = tensor("op_2176_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494835520)))]; + tensor v_77_cast = linear(bias = var_2176_to_fp16, weight = var_2175_to_fp16, x = var_2156_cast); + tensor var_2184 = const()[name = tensor("op_2184"), val = tensor([1, 1500, 16, -1])]; + tensor var_2185_cast = reshape(shape = var_2184, x = q_77_cast); + tensor const_206_to_fp16 = const()[name = tensor("const_206_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_79_cast = mul(x = var_2185_cast, y = const_206_to_fp16); + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1500, 16, -1])]; + tensor var_2192_cast = reshape(shape = var_2191, x = k_77_cast); + tensor const_207_to_fp16 = const()[name = tensor("const_207_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_79_cast = mul(x = var_2192_cast, y = const_207_to_fp16); + tensor var_2198 = const()[name = tensor("op_2198"), val = tensor([1, 1500, 16, -1])]; + tensor var_2199_cast = reshape(shape = var_2198, x = v_77_cast); + tensor var_2200 = const()[name = tensor("op_2200"), val = tensor([0, 2, 1, 3])]; + tensor qk_39_transpose_x_0 = const()[name = tensor("qk_39_transpose_x_0"), val = tensor(false)]; + tensor qk_39_transpose_y_0 = const()[name = tensor("qk_39_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_113 = transpose(perm = transpose_87_perm_0, x = k_79_cast); + tensor transpose_114 = transpose(perm = transpose_86_perm_0, x = q_79_cast); + tensor qk_39_cast = matmul(transpose_x = qk_39_transpose_x_0, transpose_y = qk_39_transpose_y_0, x = transpose_114, y = transpose_113); + tensor var_2204_cast = softmax(axis = var_2139, x = qk_39_cast); + tensor var_2206_transpose_x_0 = const()[name = tensor("op_2206_transpose_x_0"), val = tensor(false)]; + tensor var_2206_transpose_y_0 = const()[name = tensor("op_2206_transpose_y_0"), val = tensor(false)]; + tensor transpose_115 = transpose(perm = var_2200, x = var_2199_cast); + tensor var_2206_cast = matmul(transpose_x = var_2206_transpose_x_0, transpose_y = var_2206_transpose_y_0, x = var_2204_cast, y = transpose_115); + tensor var_2207 = const()[name = tensor("op_2207"), val = tensor([0, 2, 1, 3])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([1, 1500, 1024])]; + tensor transpose_112 = transpose(perm = var_2207, x = var_2206_cast); + tensor x_239_cast = reshape(shape = concat_19, x = transpose_112); + tensor var_2212_to_fp16 = const()[name = tensor("op_2212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494837632)))]; + tensor var_2213_to_fp16 = const()[name = tensor("op_2213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496934848)))]; + tensor var_2214_cast = linear(bias = var_2213_to_fp16, weight = var_2212_to_fp16, x = x_239_cast); + tensor x_241_cast = add(x = x_235_cast, y = var_2214_cast); + tensor var_2220_axes_0 = const()[name = tensor("op_2220_axes_0"), val = tensor([-1])]; + tensor blocks_19_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496936960)))]; + tensor blocks_19_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_19_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496939072)))]; + tensor var_2220_cast = layer_norm(axes = var_2220_axes_0, beta = blocks_19_mlp_ln_bias_to_fp16, epsilon = var_2145_to_fp16, gamma = blocks_19_mlp_ln_weight_to_fp16, x = x_241_cast); + tensor var_2229_to_fp16 = const()[name = tensor("op_2229_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496941184)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505329856)))]; + tensor input_161_cast = linear(bias = var_2230_to_fp16, weight = var_2229_to_fp16, x = var_2220_cast); + tensor x_245_mode_0 = const()[name = tensor("x_245_mode_0"), val = tensor("EXACT")]; + tensor x_245_cast = gelu(mode = x_245_mode_0, x = input_161_cast); + tensor var_2235_to_fp16 = const()[name = tensor("op_2235_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(505338112)))]; + tensor var_2236_to_fp16 = const()[name = tensor("op_2236_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513726784)))]; + tensor var_2237_cast = linear(bias = var_2236_to_fp16, weight = var_2235_to_fp16, x = x_245_cast); + tensor x_247_cast = add(x = x_241_cast, y = var_2237_cast); + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor(-1)]; + tensor var_2263_axes_0 = const()[name = tensor("op_2263_axes_0"), val = tensor([-1])]; + tensor blocks_20_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_20_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513728896)))]; + tensor blocks_20_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_20_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513731008)))]; + tensor var_2252_to_fp16 = const()[name = tensor("op_2252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2263_cast = layer_norm(axes = var_2263_axes_0, beta = blocks_20_attn_ln_bias_to_fp16, epsilon = var_2252_to_fp16, gamma = blocks_20_attn_ln_weight_to_fp16, x = x_247_cast); + tensor var_2274_to_fp16 = const()[name = tensor("op_2274_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513733120)))]; + tensor var_2275_to_fp16 = const()[name = tensor("op_2275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515830336)))]; + tensor q_81_cast = linear(bias = var_2275_to_fp16, weight = var_2274_to_fp16, x = var_2263_cast); + tensor var_2278_to_fp16 = const()[name = tensor("op_2278_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(515832448)))]; + tensor k_81_bias_0_to_fp16 = const()[name = tensor("k_81_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517929664)))]; + tensor k_81_cast = linear(bias = k_81_bias_0_to_fp16, weight = var_2278_to_fp16, x = var_2263_cast); + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517931776)))]; + tensor var_2283_to_fp16 = const()[name = tensor("op_2283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520028992)))]; + tensor v_81_cast = linear(bias = var_2283_to_fp16, weight = var_2282_to_fp16, x = var_2263_cast); + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1500, 16, -1])]; + tensor var_2292_cast = reshape(shape = var_2291, x = q_81_cast); + tensor const_208_to_fp16 = const()[name = tensor("const_208_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_83_cast = mul(x = var_2292_cast, y = const_208_to_fp16); + tensor var_2298 = const()[name = tensor("op_2298"), val = tensor([1, 1500, 16, -1])]; + tensor var_2299_cast = reshape(shape = var_2298, x = k_81_cast); + tensor const_209_to_fp16 = const()[name = tensor("const_209_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_83_cast = mul(x = var_2299_cast, y = const_209_to_fp16); + tensor var_2305 = const()[name = tensor("op_2305"), val = tensor([1, 1500, 16, -1])]; + tensor var_2306_cast = reshape(shape = var_2305, x = v_81_cast); + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([0, 2, 1, 3])]; + tensor qk_41_transpose_x_0 = const()[name = tensor("qk_41_transpose_x_0"), val = tensor(false)]; + tensor qk_41_transpose_y_0 = const()[name = tensor("qk_41_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_109 = transpose(perm = transpose_89_perm_0, x = k_83_cast); + tensor transpose_110 = transpose(perm = transpose_88_perm_0, x = q_83_cast); + tensor qk_41_cast = matmul(transpose_x = qk_41_transpose_x_0, transpose_y = qk_41_transpose_y_0, x = transpose_110, y = transpose_109); + tensor var_2311_cast = softmax(axis = var_2246, x = qk_41_cast); + tensor var_2313_transpose_x_0 = const()[name = tensor("op_2313_transpose_x_0"), val = tensor(false)]; + tensor var_2313_transpose_y_0 = const()[name = tensor("op_2313_transpose_y_0"), val = tensor(false)]; + tensor transpose_111 = transpose(perm = var_2307, x = var_2306_cast); + tensor var_2313_cast = matmul(transpose_x = var_2313_transpose_x_0, transpose_y = var_2313_transpose_y_0, x = var_2311_cast, y = transpose_111); + tensor var_2314 = const()[name = tensor("op_2314"), val = tensor([0, 2, 1, 3])]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 1500, 1024])]; + tensor transpose_108 = transpose(perm = var_2314, x = var_2313_cast); + tensor x_251_cast = reshape(shape = concat_20, x = transpose_108); + tensor var_2319_to_fp16 = const()[name = tensor("op_2319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520031104)))]; + tensor var_2320_to_fp16 = const()[name = tensor("op_2320_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522128320)))]; + tensor var_2321_cast = linear(bias = var_2320_to_fp16, weight = var_2319_to_fp16, x = x_251_cast); + tensor x_253_cast = add(x = x_247_cast, y = var_2321_cast); + tensor var_2327_axes_0 = const()[name = tensor("op_2327_axes_0"), val = tensor([-1])]; + tensor blocks_20_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522130432)))]; + tensor blocks_20_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_20_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522132544)))]; + tensor var_2327_cast = layer_norm(axes = var_2327_axes_0, beta = blocks_20_mlp_ln_bias_to_fp16, epsilon = var_2252_to_fp16, gamma = blocks_20_mlp_ln_weight_to_fp16, x = x_253_cast); + tensor var_2336_to_fp16 = const()[name = tensor("op_2336_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522134656)))]; + tensor var_2337_to_fp16 = const()[name = tensor("op_2337_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530523328)))]; + tensor input_169_cast = linear(bias = var_2337_to_fp16, weight = var_2336_to_fp16, x = var_2327_cast); + tensor x_257_mode_0 = const()[name = tensor("x_257_mode_0"), val = tensor("EXACT")]; + tensor x_257_cast = gelu(mode = x_257_mode_0, x = input_169_cast); + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530531584)))]; + tensor var_2343_to_fp16 = const()[name = tensor("op_2343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538920256)))]; + tensor var_2344_cast = linear(bias = var_2343_to_fp16, weight = var_2342_to_fp16, x = x_257_cast); + tensor x_259_cast = add(x = x_253_cast, y = var_2344_cast); + tensor var_2353 = const()[name = tensor("op_2353"), val = tensor(-1)]; + tensor var_2370_axes_0 = const()[name = tensor("op_2370_axes_0"), val = tensor([-1])]; + tensor blocks_21_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_21_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538922368)))]; + tensor blocks_21_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_21_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538924480)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2370_cast = layer_norm(axes = var_2370_axes_0, beta = blocks_21_attn_ln_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = blocks_21_attn_ln_weight_to_fp16, x = x_259_cast); + tensor var_2381_to_fp16 = const()[name = tensor("op_2381_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538926592)))]; + tensor var_2382_to_fp16 = const()[name = tensor("op_2382_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541023808)))]; + tensor q_85_cast = linear(bias = var_2382_to_fp16, weight = var_2381_to_fp16, x = var_2370_cast); + tensor var_2385_to_fp16 = const()[name = tensor("op_2385_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541025920)))]; + tensor k_85_bias_0_to_fp16 = const()[name = tensor("k_85_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543123136)))]; + tensor k_85_cast = linear(bias = k_85_bias_0_to_fp16, weight = var_2385_to_fp16, x = var_2370_cast); + tensor var_2389_to_fp16 = const()[name = tensor("op_2389_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543125248)))]; + tensor var_2390_to_fp16 = const()[name = tensor("op_2390_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545222464)))]; + tensor v_85_cast = linear(bias = var_2390_to_fp16, weight = var_2389_to_fp16, x = var_2370_cast); + tensor var_2398 = const()[name = tensor("op_2398"), val = tensor([1, 1500, 16, -1])]; + tensor var_2399_cast = reshape(shape = var_2398, x = q_85_cast); + tensor const_210_to_fp16 = const()[name = tensor("const_210_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_87_cast = mul(x = var_2399_cast, y = const_210_to_fp16); + tensor var_2405 = const()[name = tensor("op_2405"), val = tensor([1, 1500, 16, -1])]; + tensor var_2406_cast = reshape(shape = var_2405, x = k_85_cast); + tensor const_211_to_fp16 = const()[name = tensor("const_211_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_87_cast = mul(x = var_2406_cast, y = const_211_to_fp16); + tensor var_2412 = const()[name = tensor("op_2412"), val = tensor([1, 1500, 16, -1])]; + tensor var_2413_cast = reshape(shape = var_2412, x = v_85_cast); + tensor var_2414 = const()[name = tensor("op_2414"), val = tensor([0, 2, 1, 3])]; + tensor qk_43_transpose_x_0 = const()[name = tensor("qk_43_transpose_x_0"), val = tensor(false)]; + tensor qk_43_transpose_y_0 = const()[name = tensor("qk_43_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_105 = transpose(perm = transpose_91_perm_0, x = k_87_cast); + tensor transpose_106 = transpose(perm = transpose_90_perm_0, x = q_87_cast); + tensor qk_43_cast = matmul(transpose_x = qk_43_transpose_x_0, transpose_y = qk_43_transpose_y_0, x = transpose_106, y = transpose_105); + tensor var_2418_cast = softmax(axis = var_2353, x = qk_43_cast); + tensor var_2420_transpose_x_0 = const()[name = tensor("op_2420_transpose_x_0"), val = tensor(false)]; + tensor var_2420_transpose_y_0 = const()[name = tensor("op_2420_transpose_y_0"), val = tensor(false)]; + tensor transpose_107 = transpose(perm = var_2414, x = var_2413_cast); + tensor var_2420_cast = matmul(transpose_x = var_2420_transpose_x_0, transpose_y = var_2420_transpose_y_0, x = var_2418_cast, y = transpose_107); + tensor var_2421 = const()[name = tensor("op_2421"), val = tensor([0, 2, 1, 3])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([1, 1500, 1024])]; + tensor transpose_104 = transpose(perm = var_2421, x = var_2420_cast); + tensor x_263_cast = reshape(shape = concat_21, x = transpose_104); + tensor var_2426_to_fp16 = const()[name = tensor("op_2426_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545224576)))]; + tensor var_2427_to_fp16 = const()[name = tensor("op_2427_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547321792)))]; + tensor var_2428_cast = linear(bias = var_2427_to_fp16, weight = var_2426_to_fp16, x = x_263_cast); + tensor x_265_cast = add(x = x_259_cast, y = var_2428_cast); + tensor var_2434_axes_0 = const()[name = tensor("op_2434_axes_0"), val = tensor([-1])]; + tensor blocks_21_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547323904)))]; + tensor blocks_21_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_21_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547326016)))]; + tensor var_2434_cast = layer_norm(axes = var_2434_axes_0, beta = blocks_21_mlp_ln_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = blocks_21_mlp_ln_weight_to_fp16, x = x_265_cast); + tensor var_2443_to_fp16 = const()[name = tensor("op_2443_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(547328128)))]; + tensor var_2444_to_fp16 = const()[name = tensor("op_2444_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555716800)))]; + tensor input_177_cast = linear(bias = var_2444_to_fp16, weight = var_2443_to_fp16, x = var_2434_cast); + tensor x_269_mode_0 = const()[name = tensor("x_269_mode_0"), val = tensor("EXACT")]; + tensor x_269_cast = gelu(mode = x_269_mode_0, x = input_177_cast); + tensor var_2449_to_fp16 = const()[name = tensor("op_2449_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555725056)))]; + tensor var_2450_to_fp16 = const()[name = tensor("op_2450_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564113728)))]; + tensor var_2451_cast = linear(bias = var_2450_to_fp16, weight = var_2449_to_fp16, x = x_269_cast); + tensor x_271_cast = add(x = x_265_cast, y = var_2451_cast); + tensor var_2460 = const()[name = tensor("op_2460"), val = tensor(-1)]; + tensor var_2477_axes_0 = const()[name = tensor("op_2477_axes_0"), val = tensor([-1])]; + tensor blocks_22_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_22_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564115840)))]; + tensor blocks_22_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_22_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564117952)))]; + tensor var_2466_to_fp16 = const()[name = tensor("op_2466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2477_cast = layer_norm(axes = var_2477_axes_0, beta = blocks_22_attn_ln_bias_to_fp16, epsilon = var_2466_to_fp16, gamma = blocks_22_attn_ln_weight_to_fp16, x = x_271_cast); + tensor var_2488_to_fp16 = const()[name = tensor("op_2488_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564120064)))]; + tensor var_2489_to_fp16 = const()[name = tensor("op_2489_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566217280)))]; + tensor q_89_cast = linear(bias = var_2489_to_fp16, weight = var_2488_to_fp16, x = var_2477_cast); + tensor var_2492_to_fp16 = const()[name = tensor("op_2492_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566219392)))]; + tensor k_89_bias_0_to_fp16 = const()[name = tensor("k_89_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568316608)))]; + tensor k_89_cast = linear(bias = k_89_bias_0_to_fp16, weight = var_2492_to_fp16, x = var_2477_cast); + tensor var_2496_to_fp16 = const()[name = tensor("op_2496_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568318720)))]; + tensor var_2497_to_fp16 = const()[name = tensor("op_2497_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570415936)))]; + tensor v_89_cast = linear(bias = var_2497_to_fp16, weight = var_2496_to_fp16, x = var_2477_cast); + tensor var_2505 = const()[name = tensor("op_2505"), val = tensor([1, 1500, 16, -1])]; + tensor var_2506_cast = reshape(shape = var_2505, x = q_89_cast); + tensor const_212_to_fp16 = const()[name = tensor("const_212_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_91_cast = mul(x = var_2506_cast, y = const_212_to_fp16); + tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 1500, 16, -1])]; + tensor var_2513_cast = reshape(shape = var_2512, x = k_89_cast); + tensor const_213_to_fp16 = const()[name = tensor("const_213_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_91_cast = mul(x = var_2513_cast, y = const_213_to_fp16); + tensor var_2519 = const()[name = tensor("op_2519"), val = tensor([1, 1500, 16, -1])]; + tensor var_2520_cast = reshape(shape = var_2519, x = v_89_cast); + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([0, 2, 1, 3])]; + tensor qk_45_transpose_x_0 = const()[name = tensor("qk_45_transpose_x_0"), val = tensor(false)]; + tensor qk_45_transpose_y_0 = const()[name = tensor("qk_45_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_101 = transpose(perm = transpose_93_perm_0, x = k_91_cast); + tensor transpose_102 = transpose(perm = transpose_92_perm_0, x = q_91_cast); + tensor qk_45_cast = matmul(transpose_x = qk_45_transpose_x_0, transpose_y = qk_45_transpose_y_0, x = transpose_102, y = transpose_101); + tensor var_2525_cast = softmax(axis = var_2460, x = qk_45_cast); + tensor var_2527_transpose_x_0 = const()[name = tensor("op_2527_transpose_x_0"), val = tensor(false)]; + tensor var_2527_transpose_y_0 = const()[name = tensor("op_2527_transpose_y_0"), val = tensor(false)]; + tensor transpose_103 = transpose(perm = var_2521, x = var_2520_cast); + tensor var_2527_cast = matmul(transpose_x = var_2527_transpose_x_0, transpose_y = var_2527_transpose_y_0, x = var_2525_cast, y = transpose_103); + tensor var_2528 = const()[name = tensor("op_2528"), val = tensor([0, 2, 1, 3])]; + tensor concat_22 = const()[name = tensor("concat_22"), val = tensor([1, 1500, 1024])]; + tensor transpose_100 = transpose(perm = var_2528, x = var_2527_cast); + tensor x_275_cast = reshape(shape = concat_22, x = transpose_100); + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570418048)))]; + tensor var_2534_to_fp16 = const()[name = tensor("op_2534_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572515264)))]; + tensor var_2535_cast = linear(bias = var_2534_to_fp16, weight = var_2533_to_fp16, x = x_275_cast); + tensor x_277_cast = add(x = x_271_cast, y = var_2535_cast); + tensor var_2541_axes_0 = const()[name = tensor("op_2541_axes_0"), val = tensor([-1])]; + tensor blocks_22_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572517376)))]; + tensor blocks_22_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_22_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572519488)))]; + tensor var_2541_cast = layer_norm(axes = var_2541_axes_0, beta = blocks_22_mlp_ln_bias_to_fp16, epsilon = var_2466_to_fp16, gamma = blocks_22_mlp_ln_weight_to_fp16, x = x_277_cast); + tensor var_2550_to_fp16 = const()[name = tensor("op_2550_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572521600)))]; + tensor var_2551_to_fp16 = const()[name = tensor("op_2551_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580910272)))]; + tensor input_185_cast = linear(bias = var_2551_to_fp16, weight = var_2550_to_fp16, x = var_2541_cast); + tensor x_281_mode_0 = const()[name = tensor("x_281_mode_0"), val = tensor("EXACT")]; + tensor x_281_cast = gelu(mode = x_281_mode_0, x = input_185_cast); + tensor var_2556_to_fp16 = const()[name = tensor("op_2556_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580918528)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589307200)))]; + tensor var_2558_cast = linear(bias = var_2557_to_fp16, weight = var_2556_to_fp16, x = x_281_cast); + tensor x_283_cast = add(x = x_277_cast, y = var_2558_cast); + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor(-1)]; + tensor var_2584_axes_0 = const()[name = tensor("op_2584_axes_0"), val = tensor([-1])]; + tensor blocks_23_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_23_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589309312)))]; + tensor blocks_23_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_23_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589311424)))]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2584_cast = layer_norm(axes = var_2584_axes_0, beta = blocks_23_attn_ln_bias_to_fp16, epsilon = var_2573_to_fp16, gamma = blocks_23_attn_ln_weight_to_fp16, x = x_283_cast); + tensor var_2595_to_fp16 = const()[name = tensor("op_2595_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(589313536)))]; + tensor var_2596_to_fp16 = const()[name = tensor("op_2596_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591410752)))]; + tensor q_93_cast = linear(bias = var_2596_to_fp16, weight = var_2595_to_fp16, x = var_2584_cast); + tensor var_2599_to_fp16 = const()[name = tensor("op_2599_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(591412864)))]; + tensor k_93_bias_0_to_fp16 = const()[name = tensor("k_93_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593510080)))]; + tensor k_93_cast = linear(bias = k_93_bias_0_to_fp16, weight = var_2599_to_fp16, x = var_2584_cast); + tensor var_2603_to_fp16 = const()[name = tensor("op_2603_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(593512192)))]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595609408)))]; + tensor v_93_cast = linear(bias = var_2604_to_fp16, weight = var_2603_to_fp16, x = var_2584_cast); + tensor var_2612 = const()[name = tensor("op_2612"), val = tensor([1, 1500, 16, -1])]; + tensor var_2613_cast = reshape(shape = var_2612, x = q_93_cast); + tensor const_214_to_fp16 = const()[name = tensor("const_214_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_2613_cast, y = const_214_to_fp16); + tensor var_2619 = const()[name = tensor("op_2619"), val = tensor([1, 1500, 16, -1])]; + tensor var_2620_cast = reshape(shape = var_2619, x = k_93_cast); + tensor const_215_to_fp16 = const()[name = tensor("const_215_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_2620_cast, y = const_215_to_fp16); + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor([1, 1500, 16, -1])]; + tensor var_2627_cast = reshape(shape = var_2626, x = v_93_cast); + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_97 = transpose(perm = transpose_95_perm_0, x = k_cast); + tensor transpose_98 = transpose(perm = transpose_94_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_98, y = transpose_97); + tensor var_2632_cast = softmax(axis = var_2567, x = qk_cast); + tensor var_2634_transpose_x_0 = const()[name = tensor("op_2634_transpose_x_0"), val = tensor(false)]; + tensor var_2634_transpose_y_0 = const()[name = tensor("op_2634_transpose_y_0"), val = tensor(false)]; + tensor transpose_99 = transpose(perm = var_2628, x = var_2627_cast); + tensor var_2634_cast = matmul(transpose_x = var_2634_transpose_x_0, transpose_y = var_2634_transpose_y_0, x = var_2632_cast, y = transpose_99); + tensor var_2635 = const()[name = tensor("op_2635"), val = tensor([0, 2, 1, 3])]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([1, 1500, 1024])]; + tensor transpose_96 = transpose(perm = var_2635, x = var_2634_cast); + tensor x_287_cast = reshape(shape = concat_23, x = transpose_96); + tensor var_2640_to_fp16 = const()[name = tensor("op_2640_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595611520)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597708736)))]; + tensor var_2642_cast = linear(bias = var_2641_to_fp16, weight = var_2640_to_fp16, x = x_287_cast); + tensor x_289_cast = add(x = x_283_cast, y = var_2642_cast); + tensor var_2648_axes_0 = const()[name = tensor("op_2648_axes_0"), val = tensor([-1])]; + tensor blocks_23_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597710848)))]; + tensor blocks_23_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_23_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597712960)))]; + tensor var_2648_cast = layer_norm(axes = var_2648_axes_0, beta = blocks_23_mlp_ln_bias_to_fp16, epsilon = var_2573_to_fp16, gamma = blocks_23_mlp_ln_weight_to_fp16, x = x_289_cast); + tensor var_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597715072)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606103744)))]; + tensor input_193_cast = linear(bias = var_2658_to_fp16, weight = var_2657_to_fp16, x = var_2648_cast); + tensor x_293_mode_0 = const()[name = tensor("x_293_mode_0"), val = tensor("EXACT")]; + tensor x_293_cast = gelu(mode = x_293_mode_0, x = input_193_cast); + tensor var_2663_to_fp16 = const()[name = tensor("op_2663_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(606112000)))]; + tensor var_2664_to_fp16 = const()[name = tensor("op_2664_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614500672)))]; + tensor var_2665_cast = linear(bias = var_2664_to_fp16, weight = var_2663_to_fp16, x = x_293_cast); + tensor x_cast = add(x = x_289_cast, y = var_2665_cast); + tensor var_2678_axes_0 = const()[name = tensor("op_2678_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614502784)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(614504896)))]; + tensor var_2669_to_fp16 = const()[name = tensor("op_2669_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2678_cast = layer_norm(axes = var_2678_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_2669_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_2678_cast_to_fp32_dtype_0 = const()[name = tensor("op_2678_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_2678_cast_to_fp32_dtype_0, x = var_2678_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..ef453169cca80cb0fb9fb7b6f6939e34b6acf521 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-medium.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63a6c6a2c6cb26aaca930ccc03fe96e2e5820d1ae2802010900eff36b226674d +size 614507008 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..e095a62f0c0e0274852b265955c63d22da9500d4 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0582ef7dc6601ff085fb4b40c286f3dc761f9ada0c450703766e9d1b505ca9ad +size 157879441 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9af8eb5beab7ec5764639dbc0fead5c8eb711493 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18ad2072ae82872c2ba8a187071e1e7d6c1105253685e7aa95138adcf07874e0 +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..9cd2f12c33b3b1a5fb4748833753f83b633e1628 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 72, + "Matmul" : 24, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 12, + "Add" : 25, + "LayerNorm" : 25, + "Mul" : 24, + "Transpose" : 49, + "Gelu" : 14, + "Reshape" : 48 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8149f70d09bf9f27147da1dfaed40fe475e486c2 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/model.mil @@ -0,0 +1,747 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_32 = const()[name = tensor("op_32"), val = tensor(1)]; + tensor var_40 = const()[name = tensor("op_40"), val = tensor([1])]; + tensor var_42 = const()[name = tensor("op_42"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368768)))]; + tensor cast_367 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_44_cast = conv(bias = bias_3_to_fp16, dilations = var_42, groups = var_32, pad = var_44_pad_0, pad_type = var_44_pad_type_0, strides = var_40, weight = weight_3_to_fp16, x = cast_367); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_44_cast); + tensor var_48 = const()[name = tensor("op_48"), val = tensor(1)]; + tensor var_57 = const()[name = tensor("op_57"), val = tensor([2])]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor([1])]; + tensor var_61_pad_type_0 = const()[name = tensor("op_61_pad_type_0"), val = tensor("custom")]; + tensor var_61_pad_0 = const()[name = tensor("op_61_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370368)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3909376)))]; + tensor var_61_cast = conv(bias = bias_7_to_fp16, dilations = var_59, groups = var_48, pad = var_61_pad_0, pad_type = var_61_pad_type_0, strides = var_57, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_61_cast); + tensor var_66 = const()[name = tensor("op_66"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3910976)))]; + tensor transpose_96 = transpose(perm = var_66, x = x_3_cast); + tensor var_69_cast = add(x = transpose_96, y = positional_embedding_to_fp16); + tensor var_82 = const()[name = tensor("op_82"), val = tensor(-1)]; + tensor var_99_axes_0 = const()[name = tensor("op_99_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6215040)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6216640)))]; + tensor var_88_to_fp16 = const()[name = tensor("op_88_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_99_cast = layer_norm(axes = var_99_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_88_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_69_cast); + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6218240)))]; + tensor var_111_to_fp16 = const()[name = tensor("op_111_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397952)))]; + tensor q_1_cast = linear(bias = var_111_to_fp16, weight = var_110_to_fp16, x = var_99_cast); + tensor var_114_to_fp16 = const()[name = tensor("op_114_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7399552)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8579264)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_114_to_fp16, x = var_99_cast); + tensor var_118_to_fp16 = const()[name = tensor("op_118_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8580864)))]; + tensor var_119_to_fp16 = const()[name = tensor("op_119_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9760576)))]; + tensor v_1_cast = linear(bias = var_119_to_fp16, weight = var_118_to_fp16, x = var_99_cast); + tensor var_127 = const()[name = tensor("op_127"), val = tensor([1, 1500, 12, -1])]; + tensor var_128_cast = reshape(shape = var_127, x = q_1_cast); + tensor const_84_to_fp16 = const()[name = tensor("const_84_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_128_cast, y = const_84_to_fp16); + tensor var_134 = const()[name = tensor("op_134"), val = tensor([1, 1500, 12, -1])]; + tensor var_135_cast = reshape(shape = var_134, x = k_1_cast); + tensor const_85_to_fp16 = const()[name = tensor("const_85_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_135_cast, y = const_85_to_fp16); + tensor var_141 = const()[name = tensor("op_141"), val = tensor([1, 1500, 12, -1])]; + tensor var_142_cast = reshape(shape = var_141, x = v_1_cast); + tensor var_143 = const()[name = tensor("op_143"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_93 = transpose(perm = transpose_25_perm_0, x = k_3_cast); + tensor transpose_94 = transpose(perm = transpose_24_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_94, y = transpose_93); + tensor var_147_cast = softmax(axis = var_82, x = qk_1_cast); + tensor var_149_transpose_x_0 = const()[name = tensor("op_149_transpose_x_0"), val = tensor(false)]; + tensor var_149_transpose_y_0 = const()[name = tensor("op_149_transpose_y_0"), val = tensor(false)]; + tensor transpose_95 = transpose(perm = var_143, x = var_142_cast); + tensor var_149_cast = matmul(transpose_x = var_149_transpose_x_0, transpose_y = var_149_transpose_y_0, x = var_147_cast, y = transpose_95); + tensor var_150 = const()[name = tensor("op_150"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 768])]; + tensor transpose_92 = transpose(perm = var_150, x = var_149_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_92); + tensor var_155_to_fp16 = const()[name = tensor("op_155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9762176)))]; + tensor var_156_to_fp16 = const()[name = tensor("op_156_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10941888)))]; + tensor var_157_cast = linear(bias = var_156_to_fp16, weight = var_155_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_69_cast, y = var_157_cast); + tensor var_163_axes_0 = const()[name = tensor("op_163_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10943488)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10945088)))]; + tensor var_163_cast = layer_norm(axes = var_163_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_88_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_172_to_fp16 = const()[name = tensor("op_172_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10946688)))]; + tensor var_173_to_fp16 = const()[name = tensor("op_173_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15665344)))]; + tensor input_9_cast = linear(bias = var_173_to_fp16, weight = var_172_to_fp16, x = var_163_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_178_to_fp16 = const()[name = tensor("op_178_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15671552)))]; + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20390208)))]; + tensor var_180_cast = linear(bias = var_179_to_fp16, weight = var_178_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_180_cast); + tensor var_189 = const()[name = tensor("op_189"), val = tensor(-1)]; + tensor var_206_axes_0 = const()[name = tensor("op_206_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20391808)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20393408)))]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_206_cast = layer_norm(axes = var_206_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_195_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_217_to_fp16 = const()[name = tensor("op_217_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395008)))]; + tensor var_218_to_fp16 = const()[name = tensor("op_218_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21574720)))]; + tensor q_5_cast = linear(bias = var_218_to_fp16, weight = var_217_to_fp16, x = var_206_cast); + tensor var_221_to_fp16 = const()[name = tensor("op_221_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21576320)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22756032)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_221_to_fp16, x = var_206_cast); + tensor var_225_to_fp16 = const()[name = tensor("op_225_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22757632)))]; + tensor var_226_to_fp16 = const()[name = tensor("op_226_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23937344)))]; + tensor v_5_cast = linear(bias = var_226_to_fp16, weight = var_225_to_fp16, x = var_206_cast); + tensor var_234 = const()[name = tensor("op_234"), val = tensor([1, 1500, 12, -1])]; + tensor var_235_cast = reshape(shape = var_234, x = q_5_cast); + tensor const_86_to_fp16 = const()[name = tensor("const_86_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_235_cast, y = const_86_to_fp16); + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1500, 12, -1])]; + tensor var_242_cast = reshape(shape = var_241, x = k_5_cast); + tensor const_87_to_fp16 = const()[name = tensor("const_87_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_242_cast, y = const_87_to_fp16); + tensor var_248 = const()[name = tensor("op_248"), val = tensor([1, 1500, 12, -1])]; + tensor var_249_cast = reshape(shape = var_248, x = v_5_cast); + tensor var_250 = const()[name = tensor("op_250"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_89 = transpose(perm = transpose_27_perm_0, x = k_7_cast); + tensor transpose_90 = transpose(perm = transpose_26_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_90, y = transpose_89); + tensor var_254_cast = softmax(axis = var_189, x = qk_3_cast); + tensor var_256_transpose_x_0 = const()[name = tensor("op_256_transpose_x_0"), val = tensor(false)]; + tensor var_256_transpose_y_0 = const()[name = tensor("op_256_transpose_y_0"), val = tensor(false)]; + tensor transpose_91 = transpose(perm = var_250, x = var_249_cast); + tensor var_256_cast = matmul(transpose_x = var_256_transpose_x_0, transpose_y = var_256_transpose_y_0, x = var_254_cast, y = transpose_91); + tensor var_257 = const()[name = tensor("op_257"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 768])]; + tensor transpose_88 = transpose(perm = var_257, x = var_256_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_88); + tensor var_262_to_fp16 = const()[name = tensor("op_262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23938944)))]; + tensor var_263_to_fp16 = const()[name = tensor("op_263_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25118656)))]; + tensor var_264_cast = linear(bias = var_263_to_fp16, weight = var_262_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_264_cast); + tensor var_270_axes_0 = const()[name = tensor("op_270_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25120256)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25121856)))]; + tensor var_270_cast = layer_norm(axes = var_270_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_195_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_279_to_fp16 = const()[name = tensor("op_279_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25123456)))]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29842112)))]; + tensor input_17_cast = linear(bias = var_280_to_fp16, weight = var_279_to_fp16, x = var_270_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_285_to_fp16 = const()[name = tensor("op_285_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29848320)))]; + tensor var_286_to_fp16 = const()[name = tensor("op_286_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34566976)))]; + tensor var_287_cast = linear(bias = var_286_to_fp16, weight = var_285_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_287_cast); + tensor var_296 = const()[name = tensor("op_296"), val = tensor(-1)]; + tensor var_313_axes_0 = const()[name = tensor("op_313_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34568576)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34570176)))]; + tensor var_302_to_fp16 = const()[name = tensor("op_302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_313_cast = layer_norm(axes = var_313_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_302_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_324_to_fp16 = const()[name = tensor("op_324_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34571776)))]; + tensor var_325_to_fp16 = const()[name = tensor("op_325_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35751488)))]; + tensor q_9_cast = linear(bias = var_325_to_fp16, weight = var_324_to_fp16, x = var_313_cast); + tensor var_328_to_fp16 = const()[name = tensor("op_328_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35753088)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36932800)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_328_to_fp16, x = var_313_cast); + tensor var_332_to_fp16 = const()[name = tensor("op_332_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36934400)))]; + tensor var_333_to_fp16 = const()[name = tensor("op_333_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38114112)))]; + tensor v_9_cast = linear(bias = var_333_to_fp16, weight = var_332_to_fp16, x = var_313_cast); + tensor var_341 = const()[name = tensor("op_341"), val = tensor([1, 1500, 12, -1])]; + tensor var_342_cast = reshape(shape = var_341, x = q_9_cast); + tensor const_88_to_fp16 = const()[name = tensor("const_88_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_342_cast, y = const_88_to_fp16); + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1500, 12, -1])]; + tensor var_349_cast = reshape(shape = var_348, x = k_9_cast); + tensor const_89_to_fp16 = const()[name = tensor("const_89_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_349_cast, y = const_89_to_fp16); + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 1500, 12, -1])]; + tensor var_356_cast = reshape(shape = var_355, x = v_9_cast); + tensor var_357 = const()[name = tensor("op_357"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_85 = transpose(perm = transpose_29_perm_0, x = k_11_cast); + tensor transpose_86 = transpose(perm = transpose_28_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_86, y = transpose_85); + tensor var_361_cast = softmax(axis = var_296, x = qk_5_cast); + tensor var_363_transpose_x_0 = const()[name = tensor("op_363_transpose_x_0"), val = tensor(false)]; + tensor var_363_transpose_y_0 = const()[name = tensor("op_363_transpose_y_0"), val = tensor(false)]; + tensor transpose_87 = transpose(perm = var_357, x = var_356_cast); + tensor var_363_cast = matmul(transpose_x = var_363_transpose_x_0, transpose_y = var_363_transpose_y_0, x = var_361_cast, y = transpose_87); + tensor var_364 = const()[name = tensor("op_364"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 768])]; + tensor transpose_84 = transpose(perm = var_364, x = var_363_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_84); + tensor var_369_to_fp16 = const()[name = tensor("op_369_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38115712)))]; + tensor var_370_to_fp16 = const()[name = tensor("op_370_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39295424)))]; + tensor var_371_cast = linear(bias = var_370_to_fp16, weight = var_369_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_371_cast); + tensor var_377_axes_0 = const()[name = tensor("op_377_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39297024)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39298624)))]; + tensor var_377_cast = layer_norm(axes = var_377_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_302_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_386_to_fp16 = const()[name = tensor("op_386_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39300224)))]; + tensor var_387_to_fp16 = const()[name = tensor("op_387_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44018880)))]; + tensor input_25_cast = linear(bias = var_387_to_fp16, weight = var_386_to_fp16, x = var_377_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_392_to_fp16 = const()[name = tensor("op_392_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44025088)))]; + tensor var_393_to_fp16 = const()[name = tensor("op_393_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48743744)))]; + tensor var_394_cast = linear(bias = var_393_to_fp16, weight = var_392_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_394_cast); + tensor var_403 = const()[name = tensor("op_403"), val = tensor(-1)]; + tensor var_420_axes_0 = const()[name = tensor("op_420_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48745344)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48746944)))]; + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_420_cast = layer_norm(axes = var_420_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_409_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_431_to_fp16 = const()[name = tensor("op_431_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48748544)))]; + tensor var_432_to_fp16 = const()[name = tensor("op_432_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49928256)))]; + tensor q_13_cast = linear(bias = var_432_to_fp16, weight = var_431_to_fp16, x = var_420_cast); + tensor var_435_to_fp16 = const()[name = tensor("op_435_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49929856)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51109568)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_435_to_fp16, x = var_420_cast); + tensor var_439_to_fp16 = const()[name = tensor("op_439_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51111168)))]; + tensor var_440_to_fp16 = const()[name = tensor("op_440_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52290880)))]; + tensor v_13_cast = linear(bias = var_440_to_fp16, weight = var_439_to_fp16, x = var_420_cast); + tensor var_448 = const()[name = tensor("op_448"), val = tensor([1, 1500, 12, -1])]; + tensor var_449_cast = reshape(shape = var_448, x = q_13_cast); + tensor const_90_to_fp16 = const()[name = tensor("const_90_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_449_cast, y = const_90_to_fp16); + tensor var_455 = const()[name = tensor("op_455"), val = tensor([1, 1500, 12, -1])]; + tensor var_456_cast = reshape(shape = var_455, x = k_13_cast); + tensor const_91_to_fp16 = const()[name = tensor("const_91_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_456_cast, y = const_91_to_fp16); + tensor var_462 = const()[name = tensor("op_462"), val = tensor([1, 1500, 12, -1])]; + tensor var_463_cast = reshape(shape = var_462, x = v_13_cast); + tensor var_464 = const()[name = tensor("op_464"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_30_perm_0 = const()[name = tensor("transpose_30_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_31_perm_0 = const()[name = tensor("transpose_31_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_81 = transpose(perm = transpose_31_perm_0, x = k_15_cast); + tensor transpose_82 = transpose(perm = transpose_30_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_82, y = transpose_81); + tensor var_468_cast = softmax(axis = var_403, x = qk_7_cast); + tensor var_470_transpose_x_0 = const()[name = tensor("op_470_transpose_x_0"), val = tensor(false)]; + tensor var_470_transpose_y_0 = const()[name = tensor("op_470_transpose_y_0"), val = tensor(false)]; + tensor transpose_83 = transpose(perm = var_464, x = var_463_cast); + tensor var_470_cast = matmul(transpose_x = var_470_transpose_x_0, transpose_y = var_470_transpose_y_0, x = var_468_cast, y = transpose_83); + tensor var_471 = const()[name = tensor("op_471"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 768])]; + tensor transpose_80 = transpose(perm = var_471, x = var_470_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_80); + tensor var_476_to_fp16 = const()[name = tensor("op_476_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52292480)))]; + tensor var_477_to_fp16 = const()[name = tensor("op_477_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53472192)))]; + tensor var_478_cast = linear(bias = var_477_to_fp16, weight = var_476_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_478_cast); + tensor var_484_axes_0 = const()[name = tensor("op_484_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53473792)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53475392)))]; + tensor var_484_cast = layer_norm(axes = var_484_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_409_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_493_to_fp16 = const()[name = tensor("op_493_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53476992)))]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58195648)))]; + tensor input_33_cast = linear(bias = var_494_to_fp16, weight = var_493_to_fp16, x = var_484_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_499_to_fp16 = const()[name = tensor("op_499_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58201856)))]; + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62920512)))]; + tensor var_501_cast = linear(bias = var_500_to_fp16, weight = var_499_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_501_cast); + tensor var_510 = const()[name = tensor("op_510"), val = tensor(-1)]; + tensor var_527_axes_0 = const()[name = tensor("op_527_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62922112)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62923712)))]; + tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_527_cast = layer_norm(axes = var_527_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_516_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62925312)))]; + tensor var_539_to_fp16 = const()[name = tensor("op_539_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64105024)))]; + tensor q_17_cast = linear(bias = var_539_to_fp16, weight = var_538_to_fp16, x = var_527_cast); + tensor var_542_to_fp16 = const()[name = tensor("op_542_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64106624)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65286336)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_542_to_fp16, x = var_527_cast); + tensor var_546_to_fp16 = const()[name = tensor("op_546_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65287936)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66467648)))]; + tensor v_17_cast = linear(bias = var_547_to_fp16, weight = var_546_to_fp16, x = var_527_cast); + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 1500, 12, -1])]; + tensor var_556_cast = reshape(shape = var_555, x = q_17_cast); + tensor const_92_to_fp16 = const()[name = tensor("const_92_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_556_cast, y = const_92_to_fp16); + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1500, 12, -1])]; + tensor var_563_cast = reshape(shape = var_562, x = k_17_cast); + tensor const_93_to_fp16 = const()[name = tensor("const_93_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_563_cast, y = const_93_to_fp16); + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1500, 12, -1])]; + tensor var_570_cast = reshape(shape = var_569, x = v_17_cast); + tensor var_571 = const()[name = tensor("op_571"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_32_perm_0 = const()[name = tensor("transpose_32_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_33_perm_0 = const()[name = tensor("transpose_33_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_77 = transpose(perm = transpose_33_perm_0, x = k_19_cast); + tensor transpose_78 = transpose(perm = transpose_32_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_78, y = transpose_77); + tensor var_575_cast = softmax(axis = var_510, x = qk_9_cast); + tensor var_577_transpose_x_0 = const()[name = tensor("op_577_transpose_x_0"), val = tensor(false)]; + tensor var_577_transpose_y_0 = const()[name = tensor("op_577_transpose_y_0"), val = tensor(false)]; + tensor transpose_79 = transpose(perm = var_571, x = var_570_cast); + tensor var_577_cast = matmul(transpose_x = var_577_transpose_x_0, transpose_y = var_577_transpose_y_0, x = var_575_cast, y = transpose_79); + tensor var_578 = const()[name = tensor("op_578"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 768])]; + tensor transpose_76 = transpose(perm = var_578, x = var_577_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_76); + tensor var_583_to_fp16 = const()[name = tensor("op_583_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66469248)))]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67648960)))]; + tensor var_585_cast = linear(bias = var_584_to_fp16, weight = var_583_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_585_cast); + tensor var_591_axes_0 = const()[name = tensor("op_591_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67650560)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67652160)))]; + tensor var_591_cast = layer_norm(axes = var_591_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_516_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_600_to_fp16 = const()[name = tensor("op_600_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67653760)))]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72372416)))]; + tensor input_41_cast = linear(bias = var_601_to_fp16, weight = var_600_to_fp16, x = var_591_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72378624)))]; + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77097280)))]; + tensor var_608_cast = linear(bias = var_607_to_fp16, weight = var_606_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_608_cast); + tensor var_617 = const()[name = tensor("op_617"), val = tensor(-1)]; + tensor var_634_axes_0 = const()[name = tensor("op_634_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77098880)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77100480)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_634_cast = layer_norm(axes = var_634_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_623_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_645_to_fp16 = const()[name = tensor("op_645_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77102080)))]; + tensor var_646_to_fp16 = const()[name = tensor("op_646_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78281792)))]; + tensor q_21_cast = linear(bias = var_646_to_fp16, weight = var_645_to_fp16, x = var_634_cast); + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78283392)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79463104)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_649_to_fp16, x = var_634_cast); + tensor var_653_to_fp16 = const()[name = tensor("op_653_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79464704)))]; + tensor var_654_to_fp16 = const()[name = tensor("op_654_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80644416)))]; + tensor v_21_cast = linear(bias = var_654_to_fp16, weight = var_653_to_fp16, x = var_634_cast); + tensor var_662 = const()[name = tensor("op_662"), val = tensor([1, 1500, 12, -1])]; + tensor var_663_cast = reshape(shape = var_662, x = q_21_cast); + tensor const_94_to_fp16 = const()[name = tensor("const_94_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast = mul(x = var_663_cast, y = const_94_to_fp16); + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 1500, 12, -1])]; + tensor var_670_cast = reshape(shape = var_669, x = k_21_cast); + tensor const_95_to_fp16 = const()[name = tensor("const_95_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast = mul(x = var_670_cast, y = const_95_to_fp16); + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 1500, 12, -1])]; + tensor var_677_cast = reshape(shape = var_676, x = v_21_cast); + tensor var_678 = const()[name = tensor("op_678"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_34_perm_0 = const()[name = tensor("transpose_34_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_35_perm_0 = const()[name = tensor("transpose_35_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_73 = transpose(perm = transpose_35_perm_0, x = k_23_cast); + tensor transpose_74 = transpose(perm = transpose_34_perm_0, x = q_23_cast); + tensor qk_11_cast = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_74, y = transpose_73); + tensor var_682_cast = softmax(axis = var_617, x = qk_11_cast); + tensor var_684_transpose_x_0 = const()[name = tensor("op_684_transpose_x_0"), val = tensor(false)]; + tensor var_684_transpose_y_0 = const()[name = tensor("op_684_transpose_y_0"), val = tensor(false)]; + tensor transpose_75 = transpose(perm = var_678, x = var_677_cast); + tensor var_684_cast = matmul(transpose_x = var_684_transpose_x_0, transpose_y = var_684_transpose_y_0, x = var_682_cast, y = transpose_75); + tensor var_685 = const()[name = tensor("op_685"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 768])]; + tensor transpose_72 = transpose(perm = var_685, x = var_684_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_72); + tensor var_690_to_fp16 = const()[name = tensor("op_690_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80646016)))]; + tensor var_691_to_fp16 = const()[name = tensor("op_691_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81825728)))]; + tensor var_692_cast = linear(bias = var_691_to_fp16, weight = var_690_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_692_cast); + tensor var_698_axes_0 = const()[name = tensor("op_698_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81827328)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81828928)))]; + tensor var_698_cast = layer_norm(axes = var_698_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_623_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_707_to_fp16 = const()[name = tensor("op_707_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81830528)))]; + tensor var_708_to_fp16 = const()[name = tensor("op_708_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86549184)))]; + tensor input_49_cast = linear(bias = var_708_to_fp16, weight = var_707_to_fp16, x = var_698_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_713_to_fp16 = const()[name = tensor("op_713_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86555392)))]; + tensor var_714_to_fp16 = const()[name = tensor("op_714_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91274048)))]; + tensor var_715_cast = linear(bias = var_714_to_fp16, weight = var_713_to_fp16, x = x_77_cast); + tensor x_79_cast = add(x = x_73_cast, y = var_715_cast); + tensor var_724 = const()[name = tensor("op_724"), val = tensor(-1)]; + tensor var_741_axes_0 = const()[name = tensor("op_741_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91275648)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91277248)))]; + tensor var_730_to_fp16 = const()[name = tensor("op_730_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_741_cast = layer_norm(axes = var_741_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_730_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast); + tensor var_752_to_fp16 = const()[name = tensor("op_752_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91278848)))]; + tensor var_753_to_fp16 = const()[name = tensor("op_753_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92458560)))]; + tensor q_25_cast = linear(bias = var_753_to_fp16, weight = var_752_to_fp16, x = var_741_cast); + tensor var_756_to_fp16 = const()[name = tensor("op_756_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92460160)))]; + tensor k_25_bias_0_to_fp16 = const()[name = tensor("k_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93639872)))]; + tensor k_25_cast = linear(bias = k_25_bias_0_to_fp16, weight = var_756_to_fp16, x = var_741_cast); + tensor var_760_to_fp16 = const()[name = tensor("op_760_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93641472)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94821184)))]; + tensor v_25_cast = linear(bias = var_761_to_fp16, weight = var_760_to_fp16, x = var_741_cast); + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1500, 12, -1])]; + tensor var_770_cast = reshape(shape = var_769, x = q_25_cast); + tensor const_96_to_fp16 = const()[name = tensor("const_96_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast = mul(x = var_770_cast, y = const_96_to_fp16); + tensor var_776 = const()[name = tensor("op_776"), val = tensor([1, 1500, 12, -1])]; + tensor var_777_cast = reshape(shape = var_776, x = k_25_cast); + tensor const_97_to_fp16 = const()[name = tensor("const_97_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast = mul(x = var_777_cast, y = const_97_to_fp16); + tensor var_783 = const()[name = tensor("op_783"), val = tensor([1, 1500, 12, -1])]; + tensor var_784_cast = reshape(shape = var_783, x = v_25_cast); + tensor var_785 = const()[name = tensor("op_785"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_36_perm_0 = const()[name = tensor("transpose_36_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_37_perm_0 = const()[name = tensor("transpose_37_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_69 = transpose(perm = transpose_37_perm_0, x = k_27_cast); + tensor transpose_70 = transpose(perm = transpose_36_perm_0, x = q_27_cast); + tensor qk_13_cast = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_70, y = transpose_69); + tensor var_789_cast = softmax(axis = var_724, x = qk_13_cast); + tensor var_791_transpose_x_0 = const()[name = tensor("op_791_transpose_x_0"), val = tensor(false)]; + tensor var_791_transpose_y_0 = const()[name = tensor("op_791_transpose_y_0"), val = tensor(false)]; + tensor transpose_71 = transpose(perm = var_785, x = var_784_cast); + tensor var_791_cast = matmul(transpose_x = var_791_transpose_x_0, transpose_y = var_791_transpose_y_0, x = var_789_cast, y = transpose_71); + tensor var_792 = const()[name = tensor("op_792"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 768])]; + tensor transpose_68 = transpose(perm = var_792, x = var_791_cast); + tensor x_83_cast = reshape(shape = concat_6, x = transpose_68); + tensor var_797_to_fp16 = const()[name = tensor("op_797_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94822784)))]; + tensor var_798_to_fp16 = const()[name = tensor("op_798_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96002496)))]; + tensor var_799_cast = linear(bias = var_798_to_fp16, weight = var_797_to_fp16, x = x_83_cast); + tensor x_85_cast = add(x = x_79_cast, y = var_799_cast); + tensor var_805_axes_0 = const()[name = tensor("op_805_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96004096)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96005696)))]; + tensor var_805_cast = layer_norm(axes = var_805_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_730_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast); + tensor var_814_to_fp16 = const()[name = tensor("op_814_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96007296)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100725952)))]; + tensor input_57_cast = linear(bias = var_815_to_fp16, weight = var_814_to_fp16, x = var_805_cast); + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast = gelu(mode = x_89_mode_0, x = input_57_cast); + tensor var_820_to_fp16 = const()[name = tensor("op_820_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100732160)))]; + tensor var_821_to_fp16 = const()[name = tensor("op_821_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105450816)))]; + tensor var_822_cast = linear(bias = var_821_to_fp16, weight = var_820_to_fp16, x = x_89_cast); + tensor x_91_cast = add(x = x_85_cast, y = var_822_cast); + tensor var_831 = const()[name = tensor("op_831"), val = tensor(-1)]; + tensor var_848_axes_0 = const()[name = tensor("op_848_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105452416)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105454016)))]; + tensor var_837_to_fp16 = const()[name = tensor("op_837_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_848_cast = layer_norm(axes = var_848_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_837_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast); + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105455616)))]; + tensor var_860_to_fp16 = const()[name = tensor("op_860_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106635328)))]; + tensor q_29_cast = linear(bias = var_860_to_fp16, weight = var_859_to_fp16, x = var_848_cast); + tensor var_863_to_fp16 = const()[name = tensor("op_863_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106636928)))]; + tensor k_29_bias_0_to_fp16 = const()[name = tensor("k_29_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107816640)))]; + tensor k_29_cast = linear(bias = k_29_bias_0_to_fp16, weight = var_863_to_fp16, x = var_848_cast); + tensor var_867_to_fp16 = const()[name = tensor("op_867_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107818240)))]; + tensor var_868_to_fp16 = const()[name = tensor("op_868_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108997952)))]; + tensor v_29_cast = linear(bias = var_868_to_fp16, weight = var_867_to_fp16, x = var_848_cast); + tensor var_876 = const()[name = tensor("op_876"), val = tensor([1, 1500, 12, -1])]; + tensor var_877_cast = reshape(shape = var_876, x = q_29_cast); + tensor const_98_to_fp16 = const()[name = tensor("const_98_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast = mul(x = var_877_cast, y = const_98_to_fp16); + tensor var_883 = const()[name = tensor("op_883"), val = tensor([1, 1500, 12, -1])]; + tensor var_884_cast = reshape(shape = var_883, x = k_29_cast); + tensor const_99_to_fp16 = const()[name = tensor("const_99_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast = mul(x = var_884_cast, y = const_99_to_fp16); + tensor var_890 = const()[name = tensor("op_890"), val = tensor([1, 1500, 12, -1])]; + tensor var_891_cast = reshape(shape = var_890, x = v_29_cast); + tensor var_892 = const()[name = tensor("op_892"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_38_perm_0 = const()[name = tensor("transpose_38_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_39_perm_0 = const()[name = tensor("transpose_39_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_65 = transpose(perm = transpose_39_perm_0, x = k_31_cast); + tensor transpose_66 = transpose(perm = transpose_38_perm_0, x = q_31_cast); + tensor qk_15_cast = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_66, y = transpose_65); + tensor var_896_cast = softmax(axis = var_831, x = qk_15_cast); + tensor var_898_transpose_x_0 = const()[name = tensor("op_898_transpose_x_0"), val = tensor(false)]; + tensor var_898_transpose_y_0 = const()[name = tensor("op_898_transpose_y_0"), val = tensor(false)]; + tensor transpose_67 = transpose(perm = var_892, x = var_891_cast); + tensor var_898_cast = matmul(transpose_x = var_898_transpose_x_0, transpose_y = var_898_transpose_y_0, x = var_896_cast, y = transpose_67); + tensor var_899 = const()[name = tensor("op_899"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 768])]; + tensor transpose_64 = transpose(perm = var_899, x = var_898_cast); + tensor x_95_cast = reshape(shape = concat_7, x = transpose_64); + tensor var_904_to_fp16 = const()[name = tensor("op_904_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108999552)))]; + tensor var_905_to_fp16 = const()[name = tensor("op_905_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110179264)))]; + tensor var_906_cast = linear(bias = var_905_to_fp16, weight = var_904_to_fp16, x = x_95_cast); + tensor x_97_cast = add(x = x_91_cast, y = var_906_cast); + tensor var_912_axes_0 = const()[name = tensor("op_912_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110180864)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110182464)))]; + tensor var_912_cast = layer_norm(axes = var_912_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_837_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast); + tensor var_921_to_fp16 = const()[name = tensor("op_921_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110184064)))]; + tensor var_922_to_fp16 = const()[name = tensor("op_922_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114902720)))]; + tensor input_65_cast = linear(bias = var_922_to_fp16, weight = var_921_to_fp16, x = var_912_cast); + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast = gelu(mode = x_101_mode_0, x = input_65_cast); + tensor var_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114908928)))]; + tensor var_928_to_fp16 = const()[name = tensor("op_928_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119627584)))]; + tensor var_929_cast = linear(bias = var_928_to_fp16, weight = var_927_to_fp16, x = x_101_cast); + tensor x_103_cast = add(x = x_97_cast, y = var_929_cast); + tensor var_938 = const()[name = tensor("op_938"), val = tensor(-1)]; + tensor var_955_axes_0 = const()[name = tensor("op_955_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119629184)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119630784)))]; + tensor var_944_to_fp16 = const()[name = tensor("op_944_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_955_cast = layer_norm(axes = var_955_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_944_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast); + tensor var_966_to_fp16 = const()[name = tensor("op_966_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119632384)))]; + tensor var_967_to_fp16 = const()[name = tensor("op_967_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120812096)))]; + tensor q_33_cast = linear(bias = var_967_to_fp16, weight = var_966_to_fp16, x = var_955_cast); + tensor var_970_to_fp16 = const()[name = tensor("op_970_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120813696)))]; + tensor k_33_bias_0_to_fp16 = const()[name = tensor("k_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121993408)))]; + tensor k_33_cast = linear(bias = k_33_bias_0_to_fp16, weight = var_970_to_fp16, x = var_955_cast); + tensor var_974_to_fp16 = const()[name = tensor("op_974_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121995008)))]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123174720)))]; + tensor v_33_cast = linear(bias = var_975_to_fp16, weight = var_974_to_fp16, x = var_955_cast); + tensor var_983 = const()[name = tensor("op_983"), val = tensor([1, 1500, 12, -1])]; + tensor var_984_cast = reshape(shape = var_983, x = q_33_cast); + tensor const_100_to_fp16 = const()[name = tensor("const_100_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast = mul(x = var_984_cast, y = const_100_to_fp16); + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1500, 12, -1])]; + tensor var_991_cast = reshape(shape = var_990, x = k_33_cast); + tensor const_101_to_fp16 = const()[name = tensor("const_101_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast = mul(x = var_991_cast, y = const_101_to_fp16); + tensor var_997 = const()[name = tensor("op_997"), val = tensor([1, 1500, 12, -1])]; + tensor var_998_cast = reshape(shape = var_997, x = v_33_cast); + tensor var_999 = const()[name = tensor("op_999"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_40_perm_0 = const()[name = tensor("transpose_40_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_41_perm_0 = const()[name = tensor("transpose_41_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_61 = transpose(perm = transpose_41_perm_0, x = k_35_cast); + tensor transpose_62 = transpose(perm = transpose_40_perm_0, x = q_35_cast); + tensor qk_17_cast = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_62, y = transpose_61); + tensor var_1003_cast = softmax(axis = var_938, x = qk_17_cast); + tensor var_1005_transpose_x_0 = const()[name = tensor("op_1005_transpose_x_0"), val = tensor(false)]; + tensor var_1005_transpose_y_0 = const()[name = tensor("op_1005_transpose_y_0"), val = tensor(false)]; + tensor transpose_63 = transpose(perm = var_999, x = var_998_cast); + tensor var_1005_cast = matmul(transpose_x = var_1005_transpose_x_0, transpose_y = var_1005_transpose_y_0, x = var_1003_cast, y = transpose_63); + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 768])]; + tensor transpose_60 = transpose(perm = var_1006, x = var_1005_cast); + tensor x_107_cast = reshape(shape = concat_8, x = transpose_60); + tensor var_1011_to_fp16 = const()[name = tensor("op_1011_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123176320)))]; + tensor var_1012_to_fp16 = const()[name = tensor("op_1012_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124356032)))]; + tensor var_1013_cast = linear(bias = var_1012_to_fp16, weight = var_1011_to_fp16, x = x_107_cast); + tensor x_109_cast = add(x = x_103_cast, y = var_1013_cast); + tensor var_1019_axes_0 = const()[name = tensor("op_1019_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124357632)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124359232)))]; + tensor var_1019_cast = layer_norm(axes = var_1019_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_944_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast); + tensor var_1028_to_fp16 = const()[name = tensor("op_1028_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124360832)))]; + tensor var_1029_to_fp16 = const()[name = tensor("op_1029_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129079488)))]; + tensor input_73_cast = linear(bias = var_1029_to_fp16, weight = var_1028_to_fp16, x = var_1019_cast); + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast = gelu(mode = x_113_mode_0, x = input_73_cast); + tensor var_1034_to_fp16 = const()[name = tensor("op_1034_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129085696)))]; + tensor var_1035_to_fp16 = const()[name = tensor("op_1035_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133804352)))]; + tensor var_1036_cast = linear(bias = var_1035_to_fp16, weight = var_1034_to_fp16, x = x_113_cast); + tensor x_115_cast = add(x = x_109_cast, y = var_1036_cast); + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(-1)]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133805952)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133807552)))]; + tensor var_1051_to_fp16 = const()[name = tensor("op_1051_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1062_cast = layer_norm(axes = var_1062_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1051_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast); + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133809152)))]; + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134988864)))]; + tensor q_37_cast = linear(bias = var_1074_to_fp16, weight = var_1073_to_fp16, x = var_1062_cast); + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134990464)))]; + tensor k_37_bias_0_to_fp16 = const()[name = tensor("k_37_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136170176)))]; + tensor k_37_cast = linear(bias = k_37_bias_0_to_fp16, weight = var_1077_to_fp16, x = var_1062_cast); + tensor var_1081_to_fp16 = const()[name = tensor("op_1081_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136171776)))]; + tensor var_1082_to_fp16 = const()[name = tensor("op_1082_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137351488)))]; + tensor v_37_cast = linear(bias = var_1082_to_fp16, weight = var_1081_to_fp16, x = var_1062_cast); + tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 1500, 12, -1])]; + tensor var_1091_cast = reshape(shape = var_1090, x = q_37_cast); + tensor const_102_to_fp16 = const()[name = tensor("const_102_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast = mul(x = var_1091_cast, y = const_102_to_fp16); + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1500, 12, -1])]; + tensor var_1098_cast = reshape(shape = var_1097, x = k_37_cast); + tensor const_103_to_fp16 = const()[name = tensor("const_103_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast = mul(x = var_1098_cast, y = const_103_to_fp16); + tensor var_1104 = const()[name = tensor("op_1104"), val = tensor([1, 1500, 12, -1])]; + tensor var_1105_cast = reshape(shape = var_1104, x = v_37_cast); + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_42_perm_0 = const()[name = tensor("transpose_42_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_43_perm_0 = const()[name = tensor("transpose_43_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_57 = transpose(perm = transpose_43_perm_0, x = k_39_cast); + tensor transpose_58 = transpose(perm = transpose_42_perm_0, x = q_39_cast); + tensor qk_19_cast = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_58, y = transpose_57); + tensor var_1110_cast = softmax(axis = var_1045, x = qk_19_cast); + tensor var_1112_transpose_x_0 = const()[name = tensor("op_1112_transpose_x_0"), val = tensor(false)]; + tensor var_1112_transpose_y_0 = const()[name = tensor("op_1112_transpose_y_0"), val = tensor(false)]; + tensor transpose_59 = transpose(perm = var_1106, x = var_1105_cast); + tensor var_1112_cast = matmul(transpose_x = var_1112_transpose_x_0, transpose_y = var_1112_transpose_y_0, x = var_1110_cast, y = transpose_59); + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 768])]; + tensor transpose_56 = transpose(perm = var_1113, x = var_1112_cast); + tensor x_119_cast = reshape(shape = concat_9, x = transpose_56); + tensor var_1118_to_fp16 = const()[name = tensor("op_1118_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137353088)))]; + tensor var_1119_to_fp16 = const()[name = tensor("op_1119_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138532800)))]; + tensor var_1120_cast = linear(bias = var_1119_to_fp16, weight = var_1118_to_fp16, x = x_119_cast); + tensor x_121_cast = add(x = x_115_cast, y = var_1120_cast); + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138534400)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138536000)))]; + tensor var_1126_cast = layer_norm(axes = var_1126_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1051_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast); + tensor var_1135_to_fp16 = const()[name = tensor("op_1135_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138537600)))]; + tensor var_1136_to_fp16 = const()[name = tensor("op_1136_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143256256)))]; + tensor input_81_cast = linear(bias = var_1136_to_fp16, weight = var_1135_to_fp16, x = var_1126_cast); + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast = gelu(mode = x_125_mode_0, x = input_81_cast); + tensor var_1141_to_fp16 = const()[name = tensor("op_1141_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143262464)))]; + tensor var_1142_to_fp16 = const()[name = tensor("op_1142_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147981120)))]; + tensor var_1143_cast = linear(bias = var_1142_to_fp16, weight = var_1141_to_fp16, x = x_125_cast); + tensor x_127_cast = add(x = x_121_cast, y = var_1143_cast); + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor(-1)]; + tensor var_1169_axes_0 = const()[name = tensor("op_1169_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147982720)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147984320)))]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1169_cast = layer_norm(axes = var_1169_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1158_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast); + tensor var_1180_to_fp16 = const()[name = tensor("op_1180_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147985920)))]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149165632)))]; + tensor q_41_cast = linear(bias = var_1181_to_fp16, weight = var_1180_to_fp16, x = var_1169_cast); + tensor var_1184_to_fp16 = const()[name = tensor("op_1184_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149167232)))]; + tensor k_41_bias_0_to_fp16 = const()[name = tensor("k_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150346944)))]; + tensor k_41_cast = linear(bias = k_41_bias_0_to_fp16, weight = var_1184_to_fp16, x = var_1169_cast); + tensor var_1188_to_fp16 = const()[name = tensor("op_1188_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150348544)))]; + tensor var_1189_to_fp16 = const()[name = tensor("op_1189_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151528256)))]; + tensor v_41_cast = linear(bias = var_1189_to_fp16, weight = var_1188_to_fp16, x = var_1169_cast); + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1, 1500, 12, -1])]; + tensor var_1198_cast = reshape(shape = var_1197, x = q_41_cast); + tensor const_104_to_fp16 = const()[name = tensor("const_104_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast = mul(x = var_1198_cast, y = const_104_to_fp16); + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([1, 1500, 12, -1])]; + tensor var_1205_cast = reshape(shape = var_1204, x = k_41_cast); + tensor const_105_to_fp16 = const()[name = tensor("const_105_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast = mul(x = var_1205_cast, y = const_105_to_fp16); + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([1, 1500, 12, -1])]; + tensor var_1212_cast = reshape(shape = var_1211, x = v_41_cast); + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_44_perm_0 = const()[name = tensor("transpose_44_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_45_perm_0 = const()[name = tensor("transpose_45_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_53 = transpose(perm = transpose_45_perm_0, x = k_43_cast); + tensor transpose_54 = transpose(perm = transpose_44_perm_0, x = q_43_cast); + tensor qk_21_cast = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_54, y = transpose_53); + tensor var_1217_cast = softmax(axis = var_1152, x = qk_21_cast); + tensor var_1219_transpose_x_0 = const()[name = tensor("op_1219_transpose_x_0"), val = tensor(false)]; + tensor var_1219_transpose_y_0 = const()[name = tensor("op_1219_transpose_y_0"), val = tensor(false)]; + tensor transpose_55 = transpose(perm = var_1213, x = var_1212_cast); + tensor var_1219_cast = matmul(transpose_x = var_1219_transpose_x_0, transpose_y = var_1219_transpose_y_0, x = var_1217_cast, y = transpose_55); + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 768])]; + tensor transpose_52 = transpose(perm = var_1220, x = var_1219_cast); + tensor x_131_cast = reshape(shape = concat_10, x = transpose_52); + tensor var_1225_to_fp16 = const()[name = tensor("op_1225_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151529856)))]; + tensor var_1226_to_fp16 = const()[name = tensor("op_1226_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152709568)))]; + tensor var_1227_cast = linear(bias = var_1226_to_fp16, weight = var_1225_to_fp16, x = x_131_cast); + tensor x_133_cast = add(x = x_127_cast, y = var_1227_cast); + tensor var_1233_axes_0 = const()[name = tensor("op_1233_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152711168)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152712768)))]; + tensor var_1233_cast = layer_norm(axes = var_1233_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1158_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast); + tensor var_1242_to_fp16 = const()[name = tensor("op_1242_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152714368)))]; + tensor var_1243_to_fp16 = const()[name = tensor("op_1243_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157433024)))]; + tensor input_89_cast = linear(bias = var_1243_to_fp16, weight = var_1242_to_fp16, x = var_1233_cast); + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast = gelu(mode = x_137_mode_0, x = input_89_cast); + tensor var_1248_to_fp16 = const()[name = tensor("op_1248_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157439232)))]; + tensor var_1249_to_fp16 = const()[name = tensor("op_1249_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162157888)))]; + tensor var_1250_cast = linear(bias = var_1249_to_fp16, weight = var_1248_to_fp16, x = x_137_cast); + tensor x_139_cast = add(x = x_133_cast, y = var_1250_cast); + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor(-1)]; + tensor var_1276_axes_0 = const()[name = tensor("op_1276_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162159488)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162161088)))]; + tensor var_1265_to_fp16 = const()[name = tensor("op_1265_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1276_cast = layer_norm(axes = var_1276_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1265_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast); + tensor var_1287_to_fp16 = const()[name = tensor("op_1287_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162162688)))]; + tensor var_1288_to_fp16 = const()[name = tensor("op_1288_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163342400)))]; + tensor q_45_cast = linear(bias = var_1288_to_fp16, weight = var_1287_to_fp16, x = var_1276_cast); + tensor var_1291_to_fp16 = const()[name = tensor("op_1291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163344000)))]; + tensor k_45_bias_0_to_fp16 = const()[name = tensor("k_45_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164523712)))]; + tensor k_45_cast = linear(bias = k_45_bias_0_to_fp16, weight = var_1291_to_fp16, x = var_1276_cast); + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164525312)))]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165705024)))]; + tensor v_45_cast = linear(bias = var_1296_to_fp16, weight = var_1295_to_fp16, x = var_1276_cast); + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([1, 1500, 12, -1])]; + tensor var_1305_cast = reshape(shape = var_1304, x = q_45_cast); + tensor const_106_to_fp16 = const()[name = tensor("const_106_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_1305_cast, y = const_106_to_fp16); + tensor var_1311 = const()[name = tensor("op_1311"), val = tensor([1, 1500, 12, -1])]; + tensor var_1312_cast = reshape(shape = var_1311, x = k_45_cast); + tensor const_107_to_fp16 = const()[name = tensor("const_107_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_1312_cast, y = const_107_to_fp16); + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([1, 1500, 12, -1])]; + tensor var_1319_cast = reshape(shape = var_1318, x = v_45_cast); + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_46_perm_0 = const()[name = tensor("transpose_46_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_47_perm_0 = const()[name = tensor("transpose_47_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_49 = transpose(perm = transpose_47_perm_0, x = k_cast); + tensor transpose_50 = transpose(perm = transpose_46_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_50, y = transpose_49); + tensor var_1324_cast = softmax(axis = var_1259, x = qk_cast); + tensor var_1326_transpose_x_0 = const()[name = tensor("op_1326_transpose_x_0"), val = tensor(false)]; + tensor var_1326_transpose_y_0 = const()[name = tensor("op_1326_transpose_y_0"), val = tensor(false)]; + tensor transpose_51 = transpose(perm = var_1320, x = var_1319_cast); + tensor var_1326_cast = matmul(transpose_x = var_1326_transpose_x_0, transpose_y = var_1326_transpose_y_0, x = var_1324_cast, y = transpose_51); + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 768])]; + tensor transpose_48 = transpose(perm = var_1327, x = var_1326_cast); + tensor x_143_cast = reshape(shape = concat_11, x = transpose_48); + tensor var_1332_to_fp16 = const()[name = tensor("op_1332_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165706624)))]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166886336)))]; + tensor var_1334_cast = linear(bias = var_1333_to_fp16, weight = var_1332_to_fp16, x = x_143_cast); + tensor x_145_cast = add(x = x_139_cast, y = var_1334_cast); + tensor var_1340_axes_0 = const()[name = tensor("op_1340_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166887936)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166889536)))]; + tensor var_1340_cast = layer_norm(axes = var_1340_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1265_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast); + tensor var_1349_to_fp16 = const()[name = tensor("op_1349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166891136)))]; + tensor var_1350_to_fp16 = const()[name = tensor("op_1350_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171609792)))]; + tensor input_97_cast = linear(bias = var_1350_to_fp16, weight = var_1349_to_fp16, x = var_1340_cast); + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast = gelu(mode = x_149_mode_0, x = input_97_cast); + tensor var_1355_to_fp16 = const()[name = tensor("op_1355_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171616000)))]; + tensor var_1356_to_fp16 = const()[name = tensor("op_1356_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176334656)))]; + tensor var_1357_cast = linear(bias = var_1356_to_fp16, weight = var_1355_to_fp16, x = x_149_cast); + tensor x_cast = add(x = x_145_cast, y = var_1357_cast); + tensor var_1370_axes_0 = const()[name = tensor("op_1370_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176336256)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176337856)))]; + tensor var_1361_to_fp16 = const()[name = tensor("op_1361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1370_cast = layer_norm(axes = var_1370_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_1361_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_1370_cast_to_fp32_dtype_0 = const()[name = tensor("op_1370_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_1370_cast_to_fp32_dtype_0, x = var_1370_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..07e9e74a32beb311ae36f6e021b4a0cafb02b44c --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87eed4ae76f11a2d4a50786bc7423d4b45c2d0d9ca05577a3bd2557452072eaf +size 176339456 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..fbfb0c57c2ce04a7828b26a6b1cfd3e019326989 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f114bb1ae10b45b5ccbd6ffcdf018733b37f4b9a20290803d4e3df708035eee1 +size 155920470 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9af8eb5beab7ec5764639dbc0fead5c8eb711493 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18ad2072ae82872c2ba8a187071e1e7d6c1105253685e7aa95138adcf07874e0 +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..19154636e803e3816e6f39e22ba5a864da93204e --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 72, + "Matmul" : 24, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 12, + "Add" : 25, + "LayerNorm" : 25, + "Mul" : 24, + "Transpose" : 49, + "Gelu" : 14, + "Reshape" : 48 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..8149f70d09bf9f27147da1dfaed40fe475e486c2 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/model.mil @@ -0,0 +1,747 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_32 = const()[name = tensor("op_32"), val = tensor(1)]; + tensor var_40 = const()[name = tensor("op_40"), val = tensor([1])]; + tensor var_42 = const()[name = tensor("op_42"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368768)))]; + tensor cast_367 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_44_cast = conv(bias = bias_3_to_fp16, dilations = var_42, groups = var_32, pad = var_44_pad_0, pad_type = var_44_pad_type_0, strides = var_40, weight = weight_3_to_fp16, x = cast_367); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_44_cast); + tensor var_48 = const()[name = tensor("op_48"), val = tensor(1)]; + tensor var_57 = const()[name = tensor("op_57"), val = tensor([2])]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor([1])]; + tensor var_61_pad_type_0 = const()[name = tensor("op_61_pad_type_0"), val = tensor("custom")]; + tensor var_61_pad_0 = const()[name = tensor("op_61_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370368)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3909376)))]; + tensor var_61_cast = conv(bias = bias_7_to_fp16, dilations = var_59, groups = var_48, pad = var_61_pad_0, pad_type = var_61_pad_type_0, strides = var_57, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_61_cast); + tensor var_66 = const()[name = tensor("op_66"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3910976)))]; + tensor transpose_96 = transpose(perm = var_66, x = x_3_cast); + tensor var_69_cast = add(x = transpose_96, y = positional_embedding_to_fp16); + tensor var_82 = const()[name = tensor("op_82"), val = tensor(-1)]; + tensor var_99_axes_0 = const()[name = tensor("op_99_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6215040)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6216640)))]; + tensor var_88_to_fp16 = const()[name = tensor("op_88_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_99_cast = layer_norm(axes = var_99_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_88_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_69_cast); + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6218240)))]; + tensor var_111_to_fp16 = const()[name = tensor("op_111_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397952)))]; + tensor q_1_cast = linear(bias = var_111_to_fp16, weight = var_110_to_fp16, x = var_99_cast); + tensor var_114_to_fp16 = const()[name = tensor("op_114_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7399552)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8579264)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_114_to_fp16, x = var_99_cast); + tensor var_118_to_fp16 = const()[name = tensor("op_118_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8580864)))]; + tensor var_119_to_fp16 = const()[name = tensor("op_119_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9760576)))]; + tensor v_1_cast = linear(bias = var_119_to_fp16, weight = var_118_to_fp16, x = var_99_cast); + tensor var_127 = const()[name = tensor("op_127"), val = tensor([1, 1500, 12, -1])]; + tensor var_128_cast = reshape(shape = var_127, x = q_1_cast); + tensor const_84_to_fp16 = const()[name = tensor("const_84_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_128_cast, y = const_84_to_fp16); + tensor var_134 = const()[name = tensor("op_134"), val = tensor([1, 1500, 12, -1])]; + tensor var_135_cast = reshape(shape = var_134, x = k_1_cast); + tensor const_85_to_fp16 = const()[name = tensor("const_85_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_135_cast, y = const_85_to_fp16); + tensor var_141 = const()[name = tensor("op_141"), val = tensor([1, 1500, 12, -1])]; + tensor var_142_cast = reshape(shape = var_141, x = v_1_cast); + tensor var_143 = const()[name = tensor("op_143"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_93 = transpose(perm = transpose_25_perm_0, x = k_3_cast); + tensor transpose_94 = transpose(perm = transpose_24_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_94, y = transpose_93); + tensor var_147_cast = softmax(axis = var_82, x = qk_1_cast); + tensor var_149_transpose_x_0 = const()[name = tensor("op_149_transpose_x_0"), val = tensor(false)]; + tensor var_149_transpose_y_0 = const()[name = tensor("op_149_transpose_y_0"), val = tensor(false)]; + tensor transpose_95 = transpose(perm = var_143, x = var_142_cast); + tensor var_149_cast = matmul(transpose_x = var_149_transpose_x_0, transpose_y = var_149_transpose_y_0, x = var_147_cast, y = transpose_95); + tensor var_150 = const()[name = tensor("op_150"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 768])]; + tensor transpose_92 = transpose(perm = var_150, x = var_149_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_92); + tensor var_155_to_fp16 = const()[name = tensor("op_155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9762176)))]; + tensor var_156_to_fp16 = const()[name = tensor("op_156_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10941888)))]; + tensor var_157_cast = linear(bias = var_156_to_fp16, weight = var_155_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_69_cast, y = var_157_cast); + tensor var_163_axes_0 = const()[name = tensor("op_163_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10943488)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10945088)))]; + tensor var_163_cast = layer_norm(axes = var_163_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_88_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_172_to_fp16 = const()[name = tensor("op_172_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10946688)))]; + tensor var_173_to_fp16 = const()[name = tensor("op_173_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15665344)))]; + tensor input_9_cast = linear(bias = var_173_to_fp16, weight = var_172_to_fp16, x = var_163_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_178_to_fp16 = const()[name = tensor("op_178_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15671552)))]; + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20390208)))]; + tensor var_180_cast = linear(bias = var_179_to_fp16, weight = var_178_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_180_cast); + tensor var_189 = const()[name = tensor("op_189"), val = tensor(-1)]; + tensor var_206_axes_0 = const()[name = tensor("op_206_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20391808)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20393408)))]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_206_cast = layer_norm(axes = var_206_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_195_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_217_to_fp16 = const()[name = tensor("op_217_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395008)))]; + tensor var_218_to_fp16 = const()[name = tensor("op_218_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21574720)))]; + tensor q_5_cast = linear(bias = var_218_to_fp16, weight = var_217_to_fp16, x = var_206_cast); + tensor var_221_to_fp16 = const()[name = tensor("op_221_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21576320)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22756032)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_221_to_fp16, x = var_206_cast); + tensor var_225_to_fp16 = const()[name = tensor("op_225_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22757632)))]; + tensor var_226_to_fp16 = const()[name = tensor("op_226_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23937344)))]; + tensor v_5_cast = linear(bias = var_226_to_fp16, weight = var_225_to_fp16, x = var_206_cast); + tensor var_234 = const()[name = tensor("op_234"), val = tensor([1, 1500, 12, -1])]; + tensor var_235_cast = reshape(shape = var_234, x = q_5_cast); + tensor const_86_to_fp16 = const()[name = tensor("const_86_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_235_cast, y = const_86_to_fp16); + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1500, 12, -1])]; + tensor var_242_cast = reshape(shape = var_241, x = k_5_cast); + tensor const_87_to_fp16 = const()[name = tensor("const_87_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_242_cast, y = const_87_to_fp16); + tensor var_248 = const()[name = tensor("op_248"), val = tensor([1, 1500, 12, -1])]; + tensor var_249_cast = reshape(shape = var_248, x = v_5_cast); + tensor var_250 = const()[name = tensor("op_250"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_89 = transpose(perm = transpose_27_perm_0, x = k_7_cast); + tensor transpose_90 = transpose(perm = transpose_26_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_90, y = transpose_89); + tensor var_254_cast = softmax(axis = var_189, x = qk_3_cast); + tensor var_256_transpose_x_0 = const()[name = tensor("op_256_transpose_x_0"), val = tensor(false)]; + tensor var_256_transpose_y_0 = const()[name = tensor("op_256_transpose_y_0"), val = tensor(false)]; + tensor transpose_91 = transpose(perm = var_250, x = var_249_cast); + tensor var_256_cast = matmul(transpose_x = var_256_transpose_x_0, transpose_y = var_256_transpose_y_0, x = var_254_cast, y = transpose_91); + tensor var_257 = const()[name = tensor("op_257"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 768])]; + tensor transpose_88 = transpose(perm = var_257, x = var_256_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_88); + tensor var_262_to_fp16 = const()[name = tensor("op_262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23938944)))]; + tensor var_263_to_fp16 = const()[name = tensor("op_263_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25118656)))]; + tensor var_264_cast = linear(bias = var_263_to_fp16, weight = var_262_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_264_cast); + tensor var_270_axes_0 = const()[name = tensor("op_270_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25120256)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25121856)))]; + tensor var_270_cast = layer_norm(axes = var_270_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_195_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_279_to_fp16 = const()[name = tensor("op_279_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25123456)))]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29842112)))]; + tensor input_17_cast = linear(bias = var_280_to_fp16, weight = var_279_to_fp16, x = var_270_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_285_to_fp16 = const()[name = tensor("op_285_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29848320)))]; + tensor var_286_to_fp16 = const()[name = tensor("op_286_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34566976)))]; + tensor var_287_cast = linear(bias = var_286_to_fp16, weight = var_285_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_287_cast); + tensor var_296 = const()[name = tensor("op_296"), val = tensor(-1)]; + tensor var_313_axes_0 = const()[name = tensor("op_313_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34568576)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34570176)))]; + tensor var_302_to_fp16 = const()[name = tensor("op_302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_313_cast = layer_norm(axes = var_313_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_302_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_324_to_fp16 = const()[name = tensor("op_324_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34571776)))]; + tensor var_325_to_fp16 = const()[name = tensor("op_325_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35751488)))]; + tensor q_9_cast = linear(bias = var_325_to_fp16, weight = var_324_to_fp16, x = var_313_cast); + tensor var_328_to_fp16 = const()[name = tensor("op_328_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35753088)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36932800)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_328_to_fp16, x = var_313_cast); + tensor var_332_to_fp16 = const()[name = tensor("op_332_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36934400)))]; + tensor var_333_to_fp16 = const()[name = tensor("op_333_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38114112)))]; + tensor v_9_cast = linear(bias = var_333_to_fp16, weight = var_332_to_fp16, x = var_313_cast); + tensor var_341 = const()[name = tensor("op_341"), val = tensor([1, 1500, 12, -1])]; + tensor var_342_cast = reshape(shape = var_341, x = q_9_cast); + tensor const_88_to_fp16 = const()[name = tensor("const_88_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_342_cast, y = const_88_to_fp16); + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1500, 12, -1])]; + tensor var_349_cast = reshape(shape = var_348, x = k_9_cast); + tensor const_89_to_fp16 = const()[name = tensor("const_89_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_349_cast, y = const_89_to_fp16); + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 1500, 12, -1])]; + tensor var_356_cast = reshape(shape = var_355, x = v_9_cast); + tensor var_357 = const()[name = tensor("op_357"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_85 = transpose(perm = transpose_29_perm_0, x = k_11_cast); + tensor transpose_86 = transpose(perm = transpose_28_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_86, y = transpose_85); + tensor var_361_cast = softmax(axis = var_296, x = qk_5_cast); + tensor var_363_transpose_x_0 = const()[name = tensor("op_363_transpose_x_0"), val = tensor(false)]; + tensor var_363_transpose_y_0 = const()[name = tensor("op_363_transpose_y_0"), val = tensor(false)]; + tensor transpose_87 = transpose(perm = var_357, x = var_356_cast); + tensor var_363_cast = matmul(transpose_x = var_363_transpose_x_0, transpose_y = var_363_transpose_y_0, x = var_361_cast, y = transpose_87); + tensor var_364 = const()[name = tensor("op_364"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 768])]; + tensor transpose_84 = transpose(perm = var_364, x = var_363_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_84); + tensor var_369_to_fp16 = const()[name = tensor("op_369_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38115712)))]; + tensor var_370_to_fp16 = const()[name = tensor("op_370_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39295424)))]; + tensor var_371_cast = linear(bias = var_370_to_fp16, weight = var_369_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_371_cast); + tensor var_377_axes_0 = const()[name = tensor("op_377_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39297024)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39298624)))]; + tensor var_377_cast = layer_norm(axes = var_377_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_302_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_386_to_fp16 = const()[name = tensor("op_386_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39300224)))]; + tensor var_387_to_fp16 = const()[name = tensor("op_387_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44018880)))]; + tensor input_25_cast = linear(bias = var_387_to_fp16, weight = var_386_to_fp16, x = var_377_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_392_to_fp16 = const()[name = tensor("op_392_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44025088)))]; + tensor var_393_to_fp16 = const()[name = tensor("op_393_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48743744)))]; + tensor var_394_cast = linear(bias = var_393_to_fp16, weight = var_392_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_394_cast); + tensor var_403 = const()[name = tensor("op_403"), val = tensor(-1)]; + tensor var_420_axes_0 = const()[name = tensor("op_420_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48745344)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48746944)))]; + tensor var_409_to_fp16 = const()[name = tensor("op_409_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_420_cast = layer_norm(axes = var_420_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_409_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_431_to_fp16 = const()[name = tensor("op_431_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48748544)))]; + tensor var_432_to_fp16 = const()[name = tensor("op_432_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49928256)))]; + tensor q_13_cast = linear(bias = var_432_to_fp16, weight = var_431_to_fp16, x = var_420_cast); + tensor var_435_to_fp16 = const()[name = tensor("op_435_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49929856)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51109568)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_435_to_fp16, x = var_420_cast); + tensor var_439_to_fp16 = const()[name = tensor("op_439_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51111168)))]; + tensor var_440_to_fp16 = const()[name = tensor("op_440_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52290880)))]; + tensor v_13_cast = linear(bias = var_440_to_fp16, weight = var_439_to_fp16, x = var_420_cast); + tensor var_448 = const()[name = tensor("op_448"), val = tensor([1, 1500, 12, -1])]; + tensor var_449_cast = reshape(shape = var_448, x = q_13_cast); + tensor const_90_to_fp16 = const()[name = tensor("const_90_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_15_cast = mul(x = var_449_cast, y = const_90_to_fp16); + tensor var_455 = const()[name = tensor("op_455"), val = tensor([1, 1500, 12, -1])]; + tensor var_456_cast = reshape(shape = var_455, x = k_13_cast); + tensor const_91_to_fp16 = const()[name = tensor("const_91_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_15_cast = mul(x = var_456_cast, y = const_91_to_fp16); + tensor var_462 = const()[name = tensor("op_462"), val = tensor([1, 1500, 12, -1])]; + tensor var_463_cast = reshape(shape = var_462, x = v_13_cast); + tensor var_464 = const()[name = tensor("op_464"), val = tensor([0, 2, 1, 3])]; + tensor qk_7_transpose_x_0 = const()[name = tensor("qk_7_transpose_x_0"), val = tensor(false)]; + tensor qk_7_transpose_y_0 = const()[name = tensor("qk_7_transpose_y_0"), val = tensor(false)]; + tensor transpose_30_perm_0 = const()[name = tensor("transpose_30_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_31_perm_0 = const()[name = tensor("transpose_31_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_81 = transpose(perm = transpose_31_perm_0, x = k_15_cast); + tensor transpose_82 = transpose(perm = transpose_30_perm_0, x = q_15_cast); + tensor qk_7_cast = matmul(transpose_x = qk_7_transpose_x_0, transpose_y = qk_7_transpose_y_0, x = transpose_82, y = transpose_81); + tensor var_468_cast = softmax(axis = var_403, x = qk_7_cast); + tensor var_470_transpose_x_0 = const()[name = tensor("op_470_transpose_x_0"), val = tensor(false)]; + tensor var_470_transpose_y_0 = const()[name = tensor("op_470_transpose_y_0"), val = tensor(false)]; + tensor transpose_83 = transpose(perm = var_464, x = var_463_cast); + tensor var_470_cast = matmul(transpose_x = var_470_transpose_x_0, transpose_y = var_470_transpose_y_0, x = var_468_cast, y = transpose_83); + tensor var_471 = const()[name = tensor("op_471"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 768])]; + tensor transpose_80 = transpose(perm = var_471, x = var_470_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_80); + tensor var_476_to_fp16 = const()[name = tensor("op_476_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52292480)))]; + tensor var_477_to_fp16 = const()[name = tensor("op_477_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53472192)))]; + tensor var_478_cast = linear(bias = var_477_to_fp16, weight = var_476_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_478_cast); + tensor var_484_axes_0 = const()[name = tensor("op_484_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53473792)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53475392)))]; + tensor var_484_cast = layer_norm(axes = var_484_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_409_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_493_to_fp16 = const()[name = tensor("op_493_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53476992)))]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58195648)))]; + tensor input_33_cast = linear(bias = var_494_to_fp16, weight = var_493_to_fp16, x = var_484_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_499_to_fp16 = const()[name = tensor("op_499_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58201856)))]; + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62920512)))]; + tensor var_501_cast = linear(bias = var_500_to_fp16, weight = var_499_to_fp16, x = x_53_cast); + tensor x_55_cast = add(x = x_49_cast, y = var_501_cast); + tensor var_510 = const()[name = tensor("op_510"), val = tensor(-1)]; + tensor var_527_axes_0 = const()[name = tensor("op_527_axes_0"), val = tensor([-1])]; + tensor blocks_4_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_4_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62922112)))]; + tensor blocks_4_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_4_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62923712)))]; + tensor var_516_to_fp16 = const()[name = tensor("op_516_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_527_cast = layer_norm(axes = var_527_axes_0, beta = blocks_4_attn_ln_bias_to_fp16, epsilon = var_516_to_fp16, gamma = blocks_4_attn_ln_weight_to_fp16, x = x_55_cast); + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62925312)))]; + tensor var_539_to_fp16 = const()[name = tensor("op_539_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64105024)))]; + tensor q_17_cast = linear(bias = var_539_to_fp16, weight = var_538_to_fp16, x = var_527_cast); + tensor var_542_to_fp16 = const()[name = tensor("op_542_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64106624)))]; + tensor k_17_bias_0_to_fp16 = const()[name = tensor("k_17_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65286336)))]; + tensor k_17_cast = linear(bias = k_17_bias_0_to_fp16, weight = var_542_to_fp16, x = var_527_cast); + tensor var_546_to_fp16 = const()[name = tensor("op_546_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65287936)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66467648)))]; + tensor v_17_cast = linear(bias = var_547_to_fp16, weight = var_546_to_fp16, x = var_527_cast); + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 1500, 12, -1])]; + tensor var_556_cast = reshape(shape = var_555, x = q_17_cast); + tensor const_92_to_fp16 = const()[name = tensor("const_92_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_19_cast = mul(x = var_556_cast, y = const_92_to_fp16); + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1500, 12, -1])]; + tensor var_563_cast = reshape(shape = var_562, x = k_17_cast); + tensor const_93_to_fp16 = const()[name = tensor("const_93_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_19_cast = mul(x = var_563_cast, y = const_93_to_fp16); + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1500, 12, -1])]; + tensor var_570_cast = reshape(shape = var_569, x = v_17_cast); + tensor var_571 = const()[name = tensor("op_571"), val = tensor([0, 2, 1, 3])]; + tensor qk_9_transpose_x_0 = const()[name = tensor("qk_9_transpose_x_0"), val = tensor(false)]; + tensor qk_9_transpose_y_0 = const()[name = tensor("qk_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_32_perm_0 = const()[name = tensor("transpose_32_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_33_perm_0 = const()[name = tensor("transpose_33_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_77 = transpose(perm = transpose_33_perm_0, x = k_19_cast); + tensor transpose_78 = transpose(perm = transpose_32_perm_0, x = q_19_cast); + tensor qk_9_cast = matmul(transpose_x = qk_9_transpose_x_0, transpose_y = qk_9_transpose_y_0, x = transpose_78, y = transpose_77); + tensor var_575_cast = softmax(axis = var_510, x = qk_9_cast); + tensor var_577_transpose_x_0 = const()[name = tensor("op_577_transpose_x_0"), val = tensor(false)]; + tensor var_577_transpose_y_0 = const()[name = tensor("op_577_transpose_y_0"), val = tensor(false)]; + tensor transpose_79 = transpose(perm = var_571, x = var_570_cast); + tensor var_577_cast = matmul(transpose_x = var_577_transpose_x_0, transpose_y = var_577_transpose_y_0, x = var_575_cast, y = transpose_79); + tensor var_578 = const()[name = tensor("op_578"), val = tensor([0, 2, 1, 3])]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 1500, 768])]; + tensor transpose_76 = transpose(perm = var_578, x = var_577_cast); + tensor x_59_cast = reshape(shape = concat_4, x = transpose_76); + tensor var_583_to_fp16 = const()[name = tensor("op_583_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66469248)))]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67648960)))]; + tensor var_585_cast = linear(bias = var_584_to_fp16, weight = var_583_to_fp16, x = x_59_cast); + tensor x_61_cast = add(x = x_55_cast, y = var_585_cast); + tensor var_591_axes_0 = const()[name = tensor("op_591_axes_0"), val = tensor([-1])]; + tensor blocks_4_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67650560)))]; + tensor blocks_4_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_4_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67652160)))]; + tensor var_591_cast = layer_norm(axes = var_591_axes_0, beta = blocks_4_mlp_ln_bias_to_fp16, epsilon = var_516_to_fp16, gamma = blocks_4_mlp_ln_weight_to_fp16, x = x_61_cast); + tensor var_600_to_fp16 = const()[name = tensor("op_600_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67653760)))]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72372416)))]; + tensor input_41_cast = linear(bias = var_601_to_fp16, weight = var_600_to_fp16, x = var_591_cast); + tensor x_65_mode_0 = const()[name = tensor("x_65_mode_0"), val = tensor("EXACT")]; + tensor x_65_cast = gelu(mode = x_65_mode_0, x = input_41_cast); + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72378624)))]; + tensor var_607_to_fp16 = const()[name = tensor("op_607_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77097280)))]; + tensor var_608_cast = linear(bias = var_607_to_fp16, weight = var_606_to_fp16, x = x_65_cast); + tensor x_67_cast = add(x = x_61_cast, y = var_608_cast); + tensor var_617 = const()[name = tensor("op_617"), val = tensor(-1)]; + tensor var_634_axes_0 = const()[name = tensor("op_634_axes_0"), val = tensor([-1])]; + tensor blocks_5_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_5_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77098880)))]; + tensor blocks_5_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_5_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77100480)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_634_cast = layer_norm(axes = var_634_axes_0, beta = blocks_5_attn_ln_bias_to_fp16, epsilon = var_623_to_fp16, gamma = blocks_5_attn_ln_weight_to_fp16, x = x_67_cast); + tensor var_645_to_fp16 = const()[name = tensor("op_645_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77102080)))]; + tensor var_646_to_fp16 = const()[name = tensor("op_646_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78281792)))]; + tensor q_21_cast = linear(bias = var_646_to_fp16, weight = var_645_to_fp16, x = var_634_cast); + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78283392)))]; + tensor k_21_bias_0_to_fp16 = const()[name = tensor("k_21_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79463104)))]; + tensor k_21_cast = linear(bias = k_21_bias_0_to_fp16, weight = var_649_to_fp16, x = var_634_cast); + tensor var_653_to_fp16 = const()[name = tensor("op_653_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79464704)))]; + tensor var_654_to_fp16 = const()[name = tensor("op_654_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80644416)))]; + tensor v_21_cast = linear(bias = var_654_to_fp16, weight = var_653_to_fp16, x = var_634_cast); + tensor var_662 = const()[name = tensor("op_662"), val = tensor([1, 1500, 12, -1])]; + tensor var_663_cast = reshape(shape = var_662, x = q_21_cast); + tensor const_94_to_fp16 = const()[name = tensor("const_94_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_23_cast = mul(x = var_663_cast, y = const_94_to_fp16); + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 1500, 12, -1])]; + tensor var_670_cast = reshape(shape = var_669, x = k_21_cast); + tensor const_95_to_fp16 = const()[name = tensor("const_95_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_23_cast = mul(x = var_670_cast, y = const_95_to_fp16); + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 1500, 12, -1])]; + tensor var_677_cast = reshape(shape = var_676, x = v_21_cast); + tensor var_678 = const()[name = tensor("op_678"), val = tensor([0, 2, 1, 3])]; + tensor qk_11_transpose_x_0 = const()[name = tensor("qk_11_transpose_x_0"), val = tensor(false)]; + tensor qk_11_transpose_y_0 = const()[name = tensor("qk_11_transpose_y_0"), val = tensor(false)]; + tensor transpose_34_perm_0 = const()[name = tensor("transpose_34_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_35_perm_0 = const()[name = tensor("transpose_35_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_73 = transpose(perm = transpose_35_perm_0, x = k_23_cast); + tensor transpose_74 = transpose(perm = transpose_34_perm_0, x = q_23_cast); + tensor qk_11_cast = matmul(transpose_x = qk_11_transpose_x_0, transpose_y = qk_11_transpose_y_0, x = transpose_74, y = transpose_73); + tensor var_682_cast = softmax(axis = var_617, x = qk_11_cast); + tensor var_684_transpose_x_0 = const()[name = tensor("op_684_transpose_x_0"), val = tensor(false)]; + tensor var_684_transpose_y_0 = const()[name = tensor("op_684_transpose_y_0"), val = tensor(false)]; + tensor transpose_75 = transpose(perm = var_678, x = var_677_cast); + tensor var_684_cast = matmul(transpose_x = var_684_transpose_x_0, transpose_y = var_684_transpose_y_0, x = var_682_cast, y = transpose_75); + tensor var_685 = const()[name = tensor("op_685"), val = tensor([0, 2, 1, 3])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([1, 1500, 768])]; + tensor transpose_72 = transpose(perm = var_685, x = var_684_cast); + tensor x_71_cast = reshape(shape = concat_5, x = transpose_72); + tensor var_690_to_fp16 = const()[name = tensor("op_690_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80646016)))]; + tensor var_691_to_fp16 = const()[name = tensor("op_691_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81825728)))]; + tensor var_692_cast = linear(bias = var_691_to_fp16, weight = var_690_to_fp16, x = x_71_cast); + tensor x_73_cast = add(x = x_67_cast, y = var_692_cast); + tensor var_698_axes_0 = const()[name = tensor("op_698_axes_0"), val = tensor([-1])]; + tensor blocks_5_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81827328)))]; + tensor blocks_5_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_5_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81828928)))]; + tensor var_698_cast = layer_norm(axes = var_698_axes_0, beta = blocks_5_mlp_ln_bias_to_fp16, epsilon = var_623_to_fp16, gamma = blocks_5_mlp_ln_weight_to_fp16, x = x_73_cast); + tensor var_707_to_fp16 = const()[name = tensor("op_707_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81830528)))]; + tensor var_708_to_fp16 = const()[name = tensor("op_708_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86549184)))]; + tensor input_49_cast = linear(bias = var_708_to_fp16, weight = var_707_to_fp16, x = var_698_cast); + tensor x_77_mode_0 = const()[name = tensor("x_77_mode_0"), val = tensor("EXACT")]; + tensor x_77_cast = gelu(mode = x_77_mode_0, x = input_49_cast); + tensor var_713_to_fp16 = const()[name = tensor("op_713_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86555392)))]; + tensor var_714_to_fp16 = const()[name = tensor("op_714_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91274048)))]; + tensor var_715_cast = linear(bias = var_714_to_fp16, weight = var_713_to_fp16, x = x_77_cast); + tensor x_79_cast = add(x = x_73_cast, y = var_715_cast); + tensor var_724 = const()[name = tensor("op_724"), val = tensor(-1)]; + tensor var_741_axes_0 = const()[name = tensor("op_741_axes_0"), val = tensor([-1])]; + tensor blocks_6_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_6_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91275648)))]; + tensor blocks_6_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_6_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91277248)))]; + tensor var_730_to_fp16 = const()[name = tensor("op_730_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_741_cast = layer_norm(axes = var_741_axes_0, beta = blocks_6_attn_ln_bias_to_fp16, epsilon = var_730_to_fp16, gamma = blocks_6_attn_ln_weight_to_fp16, x = x_79_cast); + tensor var_752_to_fp16 = const()[name = tensor("op_752_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91278848)))]; + tensor var_753_to_fp16 = const()[name = tensor("op_753_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92458560)))]; + tensor q_25_cast = linear(bias = var_753_to_fp16, weight = var_752_to_fp16, x = var_741_cast); + tensor var_756_to_fp16 = const()[name = tensor("op_756_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92460160)))]; + tensor k_25_bias_0_to_fp16 = const()[name = tensor("k_25_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93639872)))]; + tensor k_25_cast = linear(bias = k_25_bias_0_to_fp16, weight = var_756_to_fp16, x = var_741_cast); + tensor var_760_to_fp16 = const()[name = tensor("op_760_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93641472)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94821184)))]; + tensor v_25_cast = linear(bias = var_761_to_fp16, weight = var_760_to_fp16, x = var_741_cast); + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1500, 12, -1])]; + tensor var_770_cast = reshape(shape = var_769, x = q_25_cast); + tensor const_96_to_fp16 = const()[name = tensor("const_96_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_27_cast = mul(x = var_770_cast, y = const_96_to_fp16); + tensor var_776 = const()[name = tensor("op_776"), val = tensor([1, 1500, 12, -1])]; + tensor var_777_cast = reshape(shape = var_776, x = k_25_cast); + tensor const_97_to_fp16 = const()[name = tensor("const_97_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_27_cast = mul(x = var_777_cast, y = const_97_to_fp16); + tensor var_783 = const()[name = tensor("op_783"), val = tensor([1, 1500, 12, -1])]; + tensor var_784_cast = reshape(shape = var_783, x = v_25_cast); + tensor var_785 = const()[name = tensor("op_785"), val = tensor([0, 2, 1, 3])]; + tensor qk_13_transpose_x_0 = const()[name = tensor("qk_13_transpose_x_0"), val = tensor(false)]; + tensor qk_13_transpose_y_0 = const()[name = tensor("qk_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_36_perm_0 = const()[name = tensor("transpose_36_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_37_perm_0 = const()[name = tensor("transpose_37_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_69 = transpose(perm = transpose_37_perm_0, x = k_27_cast); + tensor transpose_70 = transpose(perm = transpose_36_perm_0, x = q_27_cast); + tensor qk_13_cast = matmul(transpose_x = qk_13_transpose_x_0, transpose_y = qk_13_transpose_y_0, x = transpose_70, y = transpose_69); + tensor var_789_cast = softmax(axis = var_724, x = qk_13_cast); + tensor var_791_transpose_x_0 = const()[name = tensor("op_791_transpose_x_0"), val = tensor(false)]; + tensor var_791_transpose_y_0 = const()[name = tensor("op_791_transpose_y_0"), val = tensor(false)]; + tensor transpose_71 = transpose(perm = var_785, x = var_784_cast); + tensor var_791_cast = matmul(transpose_x = var_791_transpose_x_0, transpose_y = var_791_transpose_y_0, x = var_789_cast, y = transpose_71); + tensor var_792 = const()[name = tensor("op_792"), val = tensor([0, 2, 1, 3])]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([1, 1500, 768])]; + tensor transpose_68 = transpose(perm = var_792, x = var_791_cast); + tensor x_83_cast = reshape(shape = concat_6, x = transpose_68); + tensor var_797_to_fp16 = const()[name = tensor("op_797_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94822784)))]; + tensor var_798_to_fp16 = const()[name = tensor("op_798_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96002496)))]; + tensor var_799_cast = linear(bias = var_798_to_fp16, weight = var_797_to_fp16, x = x_83_cast); + tensor x_85_cast = add(x = x_79_cast, y = var_799_cast); + tensor var_805_axes_0 = const()[name = tensor("op_805_axes_0"), val = tensor([-1])]; + tensor blocks_6_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96004096)))]; + tensor blocks_6_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_6_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96005696)))]; + tensor var_805_cast = layer_norm(axes = var_805_axes_0, beta = blocks_6_mlp_ln_bias_to_fp16, epsilon = var_730_to_fp16, gamma = blocks_6_mlp_ln_weight_to_fp16, x = x_85_cast); + tensor var_814_to_fp16 = const()[name = tensor("op_814_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96007296)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100725952)))]; + tensor input_57_cast = linear(bias = var_815_to_fp16, weight = var_814_to_fp16, x = var_805_cast); + tensor x_89_mode_0 = const()[name = tensor("x_89_mode_0"), val = tensor("EXACT")]; + tensor x_89_cast = gelu(mode = x_89_mode_0, x = input_57_cast); + tensor var_820_to_fp16 = const()[name = tensor("op_820_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100732160)))]; + tensor var_821_to_fp16 = const()[name = tensor("op_821_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105450816)))]; + tensor var_822_cast = linear(bias = var_821_to_fp16, weight = var_820_to_fp16, x = x_89_cast); + tensor x_91_cast = add(x = x_85_cast, y = var_822_cast); + tensor var_831 = const()[name = tensor("op_831"), val = tensor(-1)]; + tensor var_848_axes_0 = const()[name = tensor("op_848_axes_0"), val = tensor([-1])]; + tensor blocks_7_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_7_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105452416)))]; + tensor blocks_7_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_7_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105454016)))]; + tensor var_837_to_fp16 = const()[name = tensor("op_837_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_848_cast = layer_norm(axes = var_848_axes_0, beta = blocks_7_attn_ln_bias_to_fp16, epsilon = var_837_to_fp16, gamma = blocks_7_attn_ln_weight_to_fp16, x = x_91_cast); + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105455616)))]; + tensor var_860_to_fp16 = const()[name = tensor("op_860_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106635328)))]; + tensor q_29_cast = linear(bias = var_860_to_fp16, weight = var_859_to_fp16, x = var_848_cast); + tensor var_863_to_fp16 = const()[name = tensor("op_863_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106636928)))]; + tensor k_29_bias_0_to_fp16 = const()[name = tensor("k_29_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107816640)))]; + tensor k_29_cast = linear(bias = k_29_bias_0_to_fp16, weight = var_863_to_fp16, x = var_848_cast); + tensor var_867_to_fp16 = const()[name = tensor("op_867_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107818240)))]; + tensor var_868_to_fp16 = const()[name = tensor("op_868_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108997952)))]; + tensor v_29_cast = linear(bias = var_868_to_fp16, weight = var_867_to_fp16, x = var_848_cast); + tensor var_876 = const()[name = tensor("op_876"), val = tensor([1, 1500, 12, -1])]; + tensor var_877_cast = reshape(shape = var_876, x = q_29_cast); + tensor const_98_to_fp16 = const()[name = tensor("const_98_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_31_cast = mul(x = var_877_cast, y = const_98_to_fp16); + tensor var_883 = const()[name = tensor("op_883"), val = tensor([1, 1500, 12, -1])]; + tensor var_884_cast = reshape(shape = var_883, x = k_29_cast); + tensor const_99_to_fp16 = const()[name = tensor("const_99_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_31_cast = mul(x = var_884_cast, y = const_99_to_fp16); + tensor var_890 = const()[name = tensor("op_890"), val = tensor([1, 1500, 12, -1])]; + tensor var_891_cast = reshape(shape = var_890, x = v_29_cast); + tensor var_892 = const()[name = tensor("op_892"), val = tensor([0, 2, 1, 3])]; + tensor qk_15_transpose_x_0 = const()[name = tensor("qk_15_transpose_x_0"), val = tensor(false)]; + tensor qk_15_transpose_y_0 = const()[name = tensor("qk_15_transpose_y_0"), val = tensor(false)]; + tensor transpose_38_perm_0 = const()[name = tensor("transpose_38_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_39_perm_0 = const()[name = tensor("transpose_39_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_65 = transpose(perm = transpose_39_perm_0, x = k_31_cast); + tensor transpose_66 = transpose(perm = transpose_38_perm_0, x = q_31_cast); + tensor qk_15_cast = matmul(transpose_x = qk_15_transpose_x_0, transpose_y = qk_15_transpose_y_0, x = transpose_66, y = transpose_65); + tensor var_896_cast = softmax(axis = var_831, x = qk_15_cast); + tensor var_898_transpose_x_0 = const()[name = tensor("op_898_transpose_x_0"), val = tensor(false)]; + tensor var_898_transpose_y_0 = const()[name = tensor("op_898_transpose_y_0"), val = tensor(false)]; + tensor transpose_67 = transpose(perm = var_892, x = var_891_cast); + tensor var_898_cast = matmul(transpose_x = var_898_transpose_x_0, transpose_y = var_898_transpose_y_0, x = var_896_cast, y = transpose_67); + tensor var_899 = const()[name = tensor("op_899"), val = tensor([0, 2, 1, 3])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 1500, 768])]; + tensor transpose_64 = transpose(perm = var_899, x = var_898_cast); + tensor x_95_cast = reshape(shape = concat_7, x = transpose_64); + tensor var_904_to_fp16 = const()[name = tensor("op_904_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108999552)))]; + tensor var_905_to_fp16 = const()[name = tensor("op_905_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110179264)))]; + tensor var_906_cast = linear(bias = var_905_to_fp16, weight = var_904_to_fp16, x = x_95_cast); + tensor x_97_cast = add(x = x_91_cast, y = var_906_cast); + tensor var_912_axes_0 = const()[name = tensor("op_912_axes_0"), val = tensor([-1])]; + tensor blocks_7_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110180864)))]; + tensor blocks_7_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_7_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110182464)))]; + tensor var_912_cast = layer_norm(axes = var_912_axes_0, beta = blocks_7_mlp_ln_bias_to_fp16, epsilon = var_837_to_fp16, gamma = blocks_7_mlp_ln_weight_to_fp16, x = x_97_cast); + tensor var_921_to_fp16 = const()[name = tensor("op_921_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110184064)))]; + tensor var_922_to_fp16 = const()[name = tensor("op_922_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114902720)))]; + tensor input_65_cast = linear(bias = var_922_to_fp16, weight = var_921_to_fp16, x = var_912_cast); + tensor x_101_mode_0 = const()[name = tensor("x_101_mode_0"), val = tensor("EXACT")]; + tensor x_101_cast = gelu(mode = x_101_mode_0, x = input_65_cast); + tensor var_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(114908928)))]; + tensor var_928_to_fp16 = const()[name = tensor("op_928_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119627584)))]; + tensor var_929_cast = linear(bias = var_928_to_fp16, weight = var_927_to_fp16, x = x_101_cast); + tensor x_103_cast = add(x = x_97_cast, y = var_929_cast); + tensor var_938 = const()[name = tensor("op_938"), val = tensor(-1)]; + tensor var_955_axes_0 = const()[name = tensor("op_955_axes_0"), val = tensor([-1])]; + tensor blocks_8_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_8_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119629184)))]; + tensor blocks_8_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_8_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119630784)))]; + tensor var_944_to_fp16 = const()[name = tensor("op_944_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_955_cast = layer_norm(axes = var_955_axes_0, beta = blocks_8_attn_ln_bias_to_fp16, epsilon = var_944_to_fp16, gamma = blocks_8_attn_ln_weight_to_fp16, x = x_103_cast); + tensor var_966_to_fp16 = const()[name = tensor("op_966_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119632384)))]; + tensor var_967_to_fp16 = const()[name = tensor("op_967_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120812096)))]; + tensor q_33_cast = linear(bias = var_967_to_fp16, weight = var_966_to_fp16, x = var_955_cast); + tensor var_970_to_fp16 = const()[name = tensor("op_970_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120813696)))]; + tensor k_33_bias_0_to_fp16 = const()[name = tensor("k_33_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121993408)))]; + tensor k_33_cast = linear(bias = k_33_bias_0_to_fp16, weight = var_970_to_fp16, x = var_955_cast); + tensor var_974_to_fp16 = const()[name = tensor("op_974_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121995008)))]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123174720)))]; + tensor v_33_cast = linear(bias = var_975_to_fp16, weight = var_974_to_fp16, x = var_955_cast); + tensor var_983 = const()[name = tensor("op_983"), val = tensor([1, 1500, 12, -1])]; + tensor var_984_cast = reshape(shape = var_983, x = q_33_cast); + tensor const_100_to_fp16 = const()[name = tensor("const_100_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_35_cast = mul(x = var_984_cast, y = const_100_to_fp16); + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1500, 12, -1])]; + tensor var_991_cast = reshape(shape = var_990, x = k_33_cast); + tensor const_101_to_fp16 = const()[name = tensor("const_101_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_35_cast = mul(x = var_991_cast, y = const_101_to_fp16); + tensor var_997 = const()[name = tensor("op_997"), val = tensor([1, 1500, 12, -1])]; + tensor var_998_cast = reshape(shape = var_997, x = v_33_cast); + tensor var_999 = const()[name = tensor("op_999"), val = tensor([0, 2, 1, 3])]; + tensor qk_17_transpose_x_0 = const()[name = tensor("qk_17_transpose_x_0"), val = tensor(false)]; + tensor qk_17_transpose_y_0 = const()[name = tensor("qk_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_40_perm_0 = const()[name = tensor("transpose_40_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_41_perm_0 = const()[name = tensor("transpose_41_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_61 = transpose(perm = transpose_41_perm_0, x = k_35_cast); + tensor transpose_62 = transpose(perm = transpose_40_perm_0, x = q_35_cast); + tensor qk_17_cast = matmul(transpose_x = qk_17_transpose_x_0, transpose_y = qk_17_transpose_y_0, x = transpose_62, y = transpose_61); + tensor var_1003_cast = softmax(axis = var_938, x = qk_17_cast); + tensor var_1005_transpose_x_0 = const()[name = tensor("op_1005_transpose_x_0"), val = tensor(false)]; + tensor var_1005_transpose_y_0 = const()[name = tensor("op_1005_transpose_y_0"), val = tensor(false)]; + tensor transpose_63 = transpose(perm = var_999, x = var_998_cast); + tensor var_1005_cast = matmul(transpose_x = var_1005_transpose_x_0, transpose_y = var_1005_transpose_y_0, x = var_1003_cast, y = transpose_63); + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor([0, 2, 1, 3])]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([1, 1500, 768])]; + tensor transpose_60 = transpose(perm = var_1006, x = var_1005_cast); + tensor x_107_cast = reshape(shape = concat_8, x = transpose_60); + tensor var_1011_to_fp16 = const()[name = tensor("op_1011_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123176320)))]; + tensor var_1012_to_fp16 = const()[name = tensor("op_1012_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124356032)))]; + tensor var_1013_cast = linear(bias = var_1012_to_fp16, weight = var_1011_to_fp16, x = x_107_cast); + tensor x_109_cast = add(x = x_103_cast, y = var_1013_cast); + tensor var_1019_axes_0 = const()[name = tensor("op_1019_axes_0"), val = tensor([-1])]; + tensor blocks_8_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124357632)))]; + tensor blocks_8_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_8_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124359232)))]; + tensor var_1019_cast = layer_norm(axes = var_1019_axes_0, beta = blocks_8_mlp_ln_bias_to_fp16, epsilon = var_944_to_fp16, gamma = blocks_8_mlp_ln_weight_to_fp16, x = x_109_cast); + tensor var_1028_to_fp16 = const()[name = tensor("op_1028_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124360832)))]; + tensor var_1029_to_fp16 = const()[name = tensor("op_1029_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129079488)))]; + tensor input_73_cast = linear(bias = var_1029_to_fp16, weight = var_1028_to_fp16, x = var_1019_cast); + tensor x_113_mode_0 = const()[name = tensor("x_113_mode_0"), val = tensor("EXACT")]; + tensor x_113_cast = gelu(mode = x_113_mode_0, x = input_73_cast); + tensor var_1034_to_fp16 = const()[name = tensor("op_1034_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129085696)))]; + tensor var_1035_to_fp16 = const()[name = tensor("op_1035_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133804352)))]; + tensor var_1036_cast = linear(bias = var_1035_to_fp16, weight = var_1034_to_fp16, x = x_113_cast); + tensor x_115_cast = add(x = x_109_cast, y = var_1036_cast); + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(-1)]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([-1])]; + tensor blocks_9_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_9_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133805952)))]; + tensor blocks_9_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_9_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133807552)))]; + tensor var_1051_to_fp16 = const()[name = tensor("op_1051_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1062_cast = layer_norm(axes = var_1062_axes_0, beta = blocks_9_attn_ln_bias_to_fp16, epsilon = var_1051_to_fp16, gamma = blocks_9_attn_ln_weight_to_fp16, x = x_115_cast); + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133809152)))]; + tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134988864)))]; + tensor q_37_cast = linear(bias = var_1074_to_fp16, weight = var_1073_to_fp16, x = var_1062_cast); + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134990464)))]; + tensor k_37_bias_0_to_fp16 = const()[name = tensor("k_37_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136170176)))]; + tensor k_37_cast = linear(bias = k_37_bias_0_to_fp16, weight = var_1077_to_fp16, x = var_1062_cast); + tensor var_1081_to_fp16 = const()[name = tensor("op_1081_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136171776)))]; + tensor var_1082_to_fp16 = const()[name = tensor("op_1082_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137351488)))]; + tensor v_37_cast = linear(bias = var_1082_to_fp16, weight = var_1081_to_fp16, x = var_1062_cast); + tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 1500, 12, -1])]; + tensor var_1091_cast = reshape(shape = var_1090, x = q_37_cast); + tensor const_102_to_fp16 = const()[name = tensor("const_102_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_39_cast = mul(x = var_1091_cast, y = const_102_to_fp16); + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1500, 12, -1])]; + tensor var_1098_cast = reshape(shape = var_1097, x = k_37_cast); + tensor const_103_to_fp16 = const()[name = tensor("const_103_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_39_cast = mul(x = var_1098_cast, y = const_103_to_fp16); + tensor var_1104 = const()[name = tensor("op_1104"), val = tensor([1, 1500, 12, -1])]; + tensor var_1105_cast = reshape(shape = var_1104, x = v_37_cast); + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([0, 2, 1, 3])]; + tensor qk_19_transpose_x_0 = const()[name = tensor("qk_19_transpose_x_0"), val = tensor(false)]; + tensor qk_19_transpose_y_0 = const()[name = tensor("qk_19_transpose_y_0"), val = tensor(false)]; + tensor transpose_42_perm_0 = const()[name = tensor("transpose_42_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_43_perm_0 = const()[name = tensor("transpose_43_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_57 = transpose(perm = transpose_43_perm_0, x = k_39_cast); + tensor transpose_58 = transpose(perm = transpose_42_perm_0, x = q_39_cast); + tensor qk_19_cast = matmul(transpose_x = qk_19_transpose_x_0, transpose_y = qk_19_transpose_y_0, x = transpose_58, y = transpose_57); + tensor var_1110_cast = softmax(axis = var_1045, x = qk_19_cast); + tensor var_1112_transpose_x_0 = const()[name = tensor("op_1112_transpose_x_0"), val = tensor(false)]; + tensor var_1112_transpose_y_0 = const()[name = tensor("op_1112_transpose_y_0"), val = tensor(false)]; + tensor transpose_59 = transpose(perm = var_1106, x = var_1105_cast); + tensor var_1112_cast = matmul(transpose_x = var_1112_transpose_x_0, transpose_y = var_1112_transpose_y_0, x = var_1110_cast, y = transpose_59); + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([0, 2, 1, 3])]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([1, 1500, 768])]; + tensor transpose_56 = transpose(perm = var_1113, x = var_1112_cast); + tensor x_119_cast = reshape(shape = concat_9, x = transpose_56); + tensor var_1118_to_fp16 = const()[name = tensor("op_1118_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137353088)))]; + tensor var_1119_to_fp16 = const()[name = tensor("op_1119_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138532800)))]; + tensor var_1120_cast = linear(bias = var_1119_to_fp16, weight = var_1118_to_fp16, x = x_119_cast); + tensor x_121_cast = add(x = x_115_cast, y = var_1120_cast); + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-1])]; + tensor blocks_9_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138534400)))]; + tensor blocks_9_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_9_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138536000)))]; + tensor var_1126_cast = layer_norm(axes = var_1126_axes_0, beta = blocks_9_mlp_ln_bias_to_fp16, epsilon = var_1051_to_fp16, gamma = blocks_9_mlp_ln_weight_to_fp16, x = x_121_cast); + tensor var_1135_to_fp16 = const()[name = tensor("op_1135_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138537600)))]; + tensor var_1136_to_fp16 = const()[name = tensor("op_1136_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143256256)))]; + tensor input_81_cast = linear(bias = var_1136_to_fp16, weight = var_1135_to_fp16, x = var_1126_cast); + tensor x_125_mode_0 = const()[name = tensor("x_125_mode_0"), val = tensor("EXACT")]; + tensor x_125_cast = gelu(mode = x_125_mode_0, x = input_81_cast); + tensor var_1141_to_fp16 = const()[name = tensor("op_1141_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143262464)))]; + tensor var_1142_to_fp16 = const()[name = tensor("op_1142_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147981120)))]; + tensor var_1143_cast = linear(bias = var_1142_to_fp16, weight = var_1141_to_fp16, x = x_125_cast); + tensor x_127_cast = add(x = x_121_cast, y = var_1143_cast); + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor(-1)]; + tensor var_1169_axes_0 = const()[name = tensor("op_1169_axes_0"), val = tensor([-1])]; + tensor blocks_10_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_10_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147982720)))]; + tensor blocks_10_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_10_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147984320)))]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1169_cast = layer_norm(axes = var_1169_axes_0, beta = blocks_10_attn_ln_bias_to_fp16, epsilon = var_1158_to_fp16, gamma = blocks_10_attn_ln_weight_to_fp16, x = x_127_cast); + tensor var_1180_to_fp16 = const()[name = tensor("op_1180_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147985920)))]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149165632)))]; + tensor q_41_cast = linear(bias = var_1181_to_fp16, weight = var_1180_to_fp16, x = var_1169_cast); + tensor var_1184_to_fp16 = const()[name = tensor("op_1184_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149167232)))]; + tensor k_41_bias_0_to_fp16 = const()[name = tensor("k_41_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150346944)))]; + tensor k_41_cast = linear(bias = k_41_bias_0_to_fp16, weight = var_1184_to_fp16, x = var_1169_cast); + tensor var_1188_to_fp16 = const()[name = tensor("op_1188_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150348544)))]; + tensor var_1189_to_fp16 = const()[name = tensor("op_1189_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151528256)))]; + tensor v_41_cast = linear(bias = var_1189_to_fp16, weight = var_1188_to_fp16, x = var_1169_cast); + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1, 1500, 12, -1])]; + tensor var_1198_cast = reshape(shape = var_1197, x = q_41_cast); + tensor const_104_to_fp16 = const()[name = tensor("const_104_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_43_cast = mul(x = var_1198_cast, y = const_104_to_fp16); + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([1, 1500, 12, -1])]; + tensor var_1205_cast = reshape(shape = var_1204, x = k_41_cast); + tensor const_105_to_fp16 = const()[name = tensor("const_105_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_43_cast = mul(x = var_1205_cast, y = const_105_to_fp16); + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([1, 1500, 12, -1])]; + tensor var_1212_cast = reshape(shape = var_1211, x = v_41_cast); + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([0, 2, 1, 3])]; + tensor qk_21_transpose_x_0 = const()[name = tensor("qk_21_transpose_x_0"), val = tensor(false)]; + tensor qk_21_transpose_y_0 = const()[name = tensor("qk_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_44_perm_0 = const()[name = tensor("transpose_44_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_45_perm_0 = const()[name = tensor("transpose_45_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_53 = transpose(perm = transpose_45_perm_0, x = k_43_cast); + tensor transpose_54 = transpose(perm = transpose_44_perm_0, x = q_43_cast); + tensor qk_21_cast = matmul(transpose_x = qk_21_transpose_x_0, transpose_y = qk_21_transpose_y_0, x = transpose_54, y = transpose_53); + tensor var_1217_cast = softmax(axis = var_1152, x = qk_21_cast); + tensor var_1219_transpose_x_0 = const()[name = tensor("op_1219_transpose_x_0"), val = tensor(false)]; + tensor var_1219_transpose_y_0 = const()[name = tensor("op_1219_transpose_y_0"), val = tensor(false)]; + tensor transpose_55 = transpose(perm = var_1213, x = var_1212_cast); + tensor var_1219_cast = matmul(transpose_x = var_1219_transpose_x_0, transpose_y = var_1219_transpose_y_0, x = var_1217_cast, y = transpose_55); + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([0, 2, 1, 3])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 1500, 768])]; + tensor transpose_52 = transpose(perm = var_1220, x = var_1219_cast); + tensor x_131_cast = reshape(shape = concat_10, x = transpose_52); + tensor var_1225_to_fp16 = const()[name = tensor("op_1225_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151529856)))]; + tensor var_1226_to_fp16 = const()[name = tensor("op_1226_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152709568)))]; + tensor var_1227_cast = linear(bias = var_1226_to_fp16, weight = var_1225_to_fp16, x = x_131_cast); + tensor x_133_cast = add(x = x_127_cast, y = var_1227_cast); + tensor var_1233_axes_0 = const()[name = tensor("op_1233_axes_0"), val = tensor([-1])]; + tensor blocks_10_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152711168)))]; + tensor blocks_10_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_10_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152712768)))]; + tensor var_1233_cast = layer_norm(axes = var_1233_axes_0, beta = blocks_10_mlp_ln_bias_to_fp16, epsilon = var_1158_to_fp16, gamma = blocks_10_mlp_ln_weight_to_fp16, x = x_133_cast); + tensor var_1242_to_fp16 = const()[name = tensor("op_1242_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152714368)))]; + tensor var_1243_to_fp16 = const()[name = tensor("op_1243_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157433024)))]; + tensor input_89_cast = linear(bias = var_1243_to_fp16, weight = var_1242_to_fp16, x = var_1233_cast); + tensor x_137_mode_0 = const()[name = tensor("x_137_mode_0"), val = tensor("EXACT")]; + tensor x_137_cast = gelu(mode = x_137_mode_0, x = input_89_cast); + tensor var_1248_to_fp16 = const()[name = tensor("op_1248_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157439232)))]; + tensor var_1249_to_fp16 = const()[name = tensor("op_1249_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162157888)))]; + tensor var_1250_cast = linear(bias = var_1249_to_fp16, weight = var_1248_to_fp16, x = x_137_cast); + tensor x_139_cast = add(x = x_133_cast, y = var_1250_cast); + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor(-1)]; + tensor var_1276_axes_0 = const()[name = tensor("op_1276_axes_0"), val = tensor([-1])]; + tensor blocks_11_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_11_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162159488)))]; + tensor blocks_11_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_11_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162161088)))]; + tensor var_1265_to_fp16 = const()[name = tensor("op_1265_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1276_cast = layer_norm(axes = var_1276_axes_0, beta = blocks_11_attn_ln_bias_to_fp16, epsilon = var_1265_to_fp16, gamma = blocks_11_attn_ln_weight_to_fp16, x = x_139_cast); + tensor var_1287_to_fp16 = const()[name = tensor("op_1287_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162162688)))]; + tensor var_1288_to_fp16 = const()[name = tensor("op_1288_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163342400)))]; + tensor q_45_cast = linear(bias = var_1288_to_fp16, weight = var_1287_to_fp16, x = var_1276_cast); + tensor var_1291_to_fp16 = const()[name = tensor("op_1291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163344000)))]; + tensor k_45_bias_0_to_fp16 = const()[name = tensor("k_45_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164523712)))]; + tensor k_45_cast = linear(bias = k_45_bias_0_to_fp16, weight = var_1291_to_fp16, x = var_1276_cast); + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164525312)))]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165705024)))]; + tensor v_45_cast = linear(bias = var_1296_to_fp16, weight = var_1295_to_fp16, x = var_1276_cast); + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([1, 1500, 12, -1])]; + tensor var_1305_cast = reshape(shape = var_1304, x = q_45_cast); + tensor const_106_to_fp16 = const()[name = tensor("const_106_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_1305_cast, y = const_106_to_fp16); + tensor var_1311 = const()[name = tensor("op_1311"), val = tensor([1, 1500, 12, -1])]; + tensor var_1312_cast = reshape(shape = var_1311, x = k_45_cast); + tensor const_107_to_fp16 = const()[name = tensor("const_107_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_1312_cast, y = const_107_to_fp16); + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([1, 1500, 12, -1])]; + tensor var_1319_cast = reshape(shape = var_1318, x = v_45_cast); + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_46_perm_0 = const()[name = tensor("transpose_46_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_47_perm_0 = const()[name = tensor("transpose_47_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_49 = transpose(perm = transpose_47_perm_0, x = k_cast); + tensor transpose_50 = transpose(perm = transpose_46_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_50, y = transpose_49); + tensor var_1324_cast = softmax(axis = var_1259, x = qk_cast); + tensor var_1326_transpose_x_0 = const()[name = tensor("op_1326_transpose_x_0"), val = tensor(false)]; + tensor var_1326_transpose_y_0 = const()[name = tensor("op_1326_transpose_y_0"), val = tensor(false)]; + tensor transpose_51 = transpose(perm = var_1320, x = var_1319_cast); + tensor var_1326_cast = matmul(transpose_x = var_1326_transpose_x_0, transpose_y = var_1326_transpose_y_0, x = var_1324_cast, y = transpose_51); + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([0, 2, 1, 3])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 1500, 768])]; + tensor transpose_48 = transpose(perm = var_1327, x = var_1326_cast); + tensor x_143_cast = reshape(shape = concat_11, x = transpose_48); + tensor var_1332_to_fp16 = const()[name = tensor("op_1332_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165706624)))]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166886336)))]; + tensor var_1334_cast = linear(bias = var_1333_to_fp16, weight = var_1332_to_fp16, x = x_143_cast); + tensor x_145_cast = add(x = x_139_cast, y = var_1334_cast); + tensor var_1340_axes_0 = const()[name = tensor("op_1340_axes_0"), val = tensor([-1])]; + tensor blocks_11_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166887936)))]; + tensor blocks_11_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_11_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166889536)))]; + tensor var_1340_cast = layer_norm(axes = var_1340_axes_0, beta = blocks_11_mlp_ln_bias_to_fp16, epsilon = var_1265_to_fp16, gamma = blocks_11_mlp_ln_weight_to_fp16, x = x_145_cast); + tensor var_1349_to_fp16 = const()[name = tensor("op_1349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166891136)))]; + tensor var_1350_to_fp16 = const()[name = tensor("op_1350_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171609792)))]; + tensor input_97_cast = linear(bias = var_1350_to_fp16, weight = var_1349_to_fp16, x = var_1340_cast); + tensor x_149_mode_0 = const()[name = tensor("x_149_mode_0"), val = tensor("EXACT")]; + tensor x_149_cast = gelu(mode = x_149_mode_0, x = input_97_cast); + tensor var_1355_to_fp16 = const()[name = tensor("op_1355_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171616000)))]; + tensor var_1356_to_fp16 = const()[name = tensor("op_1356_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176334656)))]; + tensor var_1357_cast = linear(bias = var_1356_to_fp16, weight = var_1355_to_fp16, x = x_149_cast); + tensor x_cast = add(x = x_145_cast, y = var_1357_cast); + tensor var_1370_axes_0 = const()[name = tensor("op_1370_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176336256)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176337856)))]; + tensor var_1361_to_fp16 = const()[name = tensor("op_1361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1370_cast = layer_norm(axes = var_1370_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_1361_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_1370_cast_to_fp32_dtype_0 = const()[name = tensor("op_1370_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_1370_cast_to_fp32_dtype_0, x = var_1370_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d1229df486ac8a7a1e9667bd28651b80e7b116ea --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-small.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d66ece4975aeea37c628f6b59d85de85d34beb2eb9b5f61f7d3dc7c1aca18143 +size 176339456 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..0966691f36b0cf00cb253c089587746cc35ba7cb --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:369a4cca64a069c07d48f6aa3cc4fc3f081ae825e44d3c349c0d3ef147d85acc +size 14292994 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..16ba4c4c65c22b09941cb787a9302b5fc0130c55 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2f746e204d04779629b4959dad86d4ada040edd54f0ed1c80aa6f482a0fa72e +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..01f1ff45fbbba4fa68d9d81e145c8dd405299f11 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 24, + "Matmul" : 8, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 4, + "Add" : 9, + "LayerNorm" : 9, + "Mul" : 8, + "Transpose" : 17, + "Gelu" : 6, + "Reshape" : 16 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..ccedc27eec99a20a310efef71e5bbc4d874439cc --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/model.mil @@ -0,0 +1,275 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_16 = const()[name = tensor("op_16"), val = tensor(1)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor([1])]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184448)))]; + tensor cast_127 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_28_cast = conv(bias = bias_3_to_fp16, dilations = var_26, groups = var_16, pad = var_28_pad_0, pad_type = var_28_pad_type_0, strides = var_24, weight = weight_3_to_fp16, x = cast_127); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_28_cast); + tensor var_32 = const()[name = tensor("op_32"), val = tensor(1)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor([2])]; + tensor var_43 = const()[name = tensor("op_43"), val = tensor([1])]; + tensor var_45_pad_type_0 = const()[name = tensor("op_45_pad_type_0"), val = tensor("custom")]; + tensor var_45_pad_0 = const()[name = tensor("op_45_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185280)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070080)))]; + tensor var_45_cast = conv(bias = bias_7_to_fp16, dilations = var_43, groups = var_32, pad = var_45_pad_0, pad_type = var_45_pad_type_0, strides = var_41, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_45_cast); + tensor var_50 = const()[name = tensor("op_50"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070912)))]; + tensor transpose_32 = transpose(perm = var_50, x = x_3_cast); + tensor var_53_cast = add(x = transpose_32, y = positional_embedding_to_fp16); + tensor var_65 = const()[name = tensor("op_65"), val = tensor(-1)]; + tensor var_82_axes_0 = const()[name = tensor("op_82_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2222976)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2223808)))]; + tensor var_71_to_fp16 = const()[name = tensor("op_71_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_82_cast = layer_norm(axes = var_82_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_71_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_53_cast); + tensor var_93_to_fp16 = const()[name = tensor("op_93_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2224640)))]; + tensor var_94_to_fp16 = const()[name = tensor("op_94_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2519616)))]; + tensor q_1_cast = linear(bias = var_94_to_fp16, weight = var_93_to_fp16, x = var_82_cast); + tensor var_97_to_fp16 = const()[name = tensor("op_97_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2520448)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2815424)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_97_to_fp16, x = var_82_cast); + tensor var_101_to_fp16 = const()[name = tensor("op_101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2816256)))]; + tensor var_102_to_fp16 = const()[name = tensor("op_102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3111232)))]; + tensor v_1_cast = linear(bias = var_102_to_fp16, weight = var_101_to_fp16, x = var_82_cast); + tensor var_110 = const()[name = tensor("op_110"), val = tensor([1, 1500, 6, -1])]; + tensor var_111_cast = reshape(shape = var_110, x = q_1_cast); + tensor const_28_to_fp16 = const()[name = tensor("const_28_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_111_cast, y = const_28_to_fp16); + tensor var_117 = const()[name = tensor("op_117"), val = tensor([1, 1500, 6, -1])]; + tensor var_118_cast = reshape(shape = var_117, x = k_1_cast); + tensor const_29_to_fp16 = const()[name = tensor("const_29_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_118_cast, y = const_29_to_fp16); + tensor var_124 = const()[name = tensor("op_124"), val = tensor([1, 1500, 6, -1])]; + tensor var_125_cast = reshape(shape = var_124, x = v_1_cast); + tensor var_126 = const()[name = tensor("op_126"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_29 = transpose(perm = transpose_9_perm_0, x = k_3_cast); + tensor transpose_30 = transpose(perm = transpose_8_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_30, y = transpose_29); + tensor var_130_cast = softmax(axis = var_65, x = qk_1_cast); + tensor var_132_transpose_x_0 = const()[name = tensor("op_132_transpose_x_0"), val = tensor(false)]; + tensor var_132_transpose_y_0 = const()[name = tensor("op_132_transpose_y_0"), val = tensor(false)]; + tensor transpose_31 = transpose(perm = var_126, x = var_125_cast); + tensor var_132_cast = matmul(transpose_x = var_132_transpose_x_0, transpose_y = var_132_transpose_y_0, x = var_130_cast, y = transpose_31); + tensor var_133 = const()[name = tensor("op_133"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 384])]; + tensor transpose_28 = transpose(perm = var_133, x = var_132_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_28); + tensor var_138_to_fp16 = const()[name = tensor("op_138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3112064)))]; + tensor var_139_to_fp16 = const()[name = tensor("op_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407040)))]; + tensor var_140_cast = linear(bias = var_139_to_fp16, weight = var_138_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_53_cast, y = var_140_cast); + tensor var_146_axes_0 = const()[name = tensor("op_146_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407872)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3408704)))]; + tensor var_146_cast = layer_norm(axes = var_146_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_71_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_155_to_fp16 = const()[name = tensor("op_155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3409536)))]; + tensor var_156_to_fp16 = const()[name = tensor("op_156_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4589248)))]; + tensor input_9_cast = linear(bias = var_156_to_fp16, weight = var_155_to_fp16, x = var_146_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_161_to_fp16 = const()[name = tensor("op_161_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4592384)))]; + tensor var_162_to_fp16 = const()[name = tensor("op_162_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772096)))]; + tensor var_163_cast = linear(bias = var_162_to_fp16, weight = var_161_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_163_cast); + tensor var_171 = const()[name = tensor("op_171"), val = tensor(-1)]; + tensor var_188_axes_0 = const()[name = tensor("op_188_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772928)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5773760)))]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_188_cast = layer_norm(axes = var_188_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_177_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_199_to_fp16 = const()[name = tensor("op_199_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5774592)))]; + tensor var_200_to_fp16 = const()[name = tensor("op_200_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6069568)))]; + tensor q_5_cast = linear(bias = var_200_to_fp16, weight = var_199_to_fp16, x = var_188_cast); + tensor var_203_to_fp16 = const()[name = tensor("op_203_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6070400)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6365376)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_203_to_fp16, x = var_188_cast); + tensor var_207_to_fp16 = const()[name = tensor("op_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6366208)))]; + tensor var_208_to_fp16 = const()[name = tensor("op_208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6661184)))]; + tensor v_5_cast = linear(bias = var_208_to_fp16, weight = var_207_to_fp16, x = var_188_cast); + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1500, 6, -1])]; + tensor var_217_cast = reshape(shape = var_216, x = q_5_cast); + tensor const_30_to_fp16 = const()[name = tensor("const_30_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_217_cast, y = const_30_to_fp16); + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 1500, 6, -1])]; + tensor var_224_cast = reshape(shape = var_223, x = k_5_cast); + tensor const_31_to_fp16 = const()[name = tensor("const_31_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_224_cast, y = const_31_to_fp16); + tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 1500, 6, -1])]; + tensor var_231_cast = reshape(shape = var_230, x = v_5_cast); + tensor var_232 = const()[name = tensor("op_232"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_25 = transpose(perm = transpose_11_perm_0, x = k_7_cast); + tensor transpose_26 = transpose(perm = transpose_10_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_26, y = transpose_25); + tensor var_236_cast = softmax(axis = var_171, x = qk_3_cast); + tensor var_238_transpose_x_0 = const()[name = tensor("op_238_transpose_x_0"), val = tensor(false)]; + tensor var_238_transpose_y_0 = const()[name = tensor("op_238_transpose_y_0"), val = tensor(false)]; + tensor transpose_27 = transpose(perm = var_232, x = var_231_cast); + tensor var_238_cast = matmul(transpose_x = var_238_transpose_x_0, transpose_y = var_238_transpose_y_0, x = var_236_cast, y = transpose_27); + tensor var_239 = const()[name = tensor("op_239"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 384])]; + tensor transpose_24 = transpose(perm = var_239, x = var_238_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_24); + tensor var_244_to_fp16 = const()[name = tensor("op_244_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6662016)))]; + tensor var_245_to_fp16 = const()[name = tensor("op_245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956992)))]; + tensor var_246_cast = linear(bias = var_245_to_fp16, weight = var_244_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_246_cast); + tensor var_252_axes_0 = const()[name = tensor("op_252_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6957824)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6958656)))]; + tensor var_252_cast = layer_norm(axes = var_252_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_177_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6959488)))]; + tensor var_262_to_fp16 = const()[name = tensor("op_262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8139200)))]; + tensor input_17_cast = linear(bias = var_262_to_fp16, weight = var_261_to_fp16, x = var_252_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8142336)))]; + tensor var_268_to_fp16 = const()[name = tensor("op_268_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322048)))]; + tensor var_269_cast = linear(bias = var_268_to_fp16, weight = var_267_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_269_cast); + tensor var_277 = const()[name = tensor("op_277"), val = tensor(-1)]; + tensor var_294_axes_0 = const()[name = tensor("op_294_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322880)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9323712)))]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_294_cast = layer_norm(axes = var_294_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_305_to_fp16 = const()[name = tensor("op_305_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9324544)))]; + tensor var_306_to_fp16 = const()[name = tensor("op_306_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9619520)))]; + tensor q_9_cast = linear(bias = var_306_to_fp16, weight = var_305_to_fp16, x = var_294_cast); + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9620352)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9915328)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_309_to_fp16, x = var_294_cast); + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9916160)))]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10211136)))]; + tensor v_9_cast = linear(bias = var_314_to_fp16, weight = var_313_to_fp16, x = var_294_cast); + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 1500, 6, -1])]; + tensor var_323_cast = reshape(shape = var_322, x = q_9_cast); + tensor const_32_to_fp16 = const()[name = tensor("const_32_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_323_cast, y = const_32_to_fp16); + tensor var_329 = const()[name = tensor("op_329"), val = tensor([1, 1500, 6, -1])]; + tensor var_330_cast = reshape(shape = var_329, x = k_9_cast); + tensor const_33_to_fp16 = const()[name = tensor("const_33_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_330_cast, y = const_33_to_fp16); + tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1500, 6, -1])]; + tensor var_337_cast = reshape(shape = var_336, x = v_9_cast); + tensor var_338 = const()[name = tensor("op_338"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_21 = transpose(perm = transpose_13_perm_0, x = k_11_cast); + tensor transpose_22 = transpose(perm = transpose_12_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_22, y = transpose_21); + tensor var_342_cast = softmax(axis = var_277, x = qk_5_cast); + tensor var_344_transpose_x_0 = const()[name = tensor("op_344_transpose_x_0"), val = tensor(false)]; + tensor var_344_transpose_y_0 = const()[name = tensor("op_344_transpose_y_0"), val = tensor(false)]; + tensor transpose_23 = transpose(perm = var_338, x = var_337_cast); + tensor var_344_cast = matmul(transpose_x = var_344_transpose_x_0, transpose_y = var_344_transpose_y_0, x = var_342_cast, y = transpose_23); + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 384])]; + tensor transpose_20 = transpose(perm = var_345, x = var_344_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_20); + tensor var_350_to_fp16 = const()[name = tensor("op_350_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10211968)))]; + tensor var_351_to_fp16 = const()[name = tensor("op_351_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506944)))]; + tensor var_352_cast = linear(bias = var_351_to_fp16, weight = var_350_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_352_cast); + tensor var_358_axes_0 = const()[name = tensor("op_358_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10507776)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10508608)))]; + tensor var_358_cast = layer_norm(axes = var_358_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_367_to_fp16 = const()[name = tensor("op_367_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10509440)))]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11689152)))]; + tensor input_25_cast = linear(bias = var_368_to_fp16, weight = var_367_to_fp16, x = var_358_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_373_to_fp16 = const()[name = tensor("op_373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11692288)))]; + tensor var_374_to_fp16 = const()[name = tensor("op_374_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12872000)))]; + tensor var_375_cast = linear(bias = var_374_to_fp16, weight = var_373_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_375_cast); + tensor var_383 = const()[name = tensor("op_383"), val = tensor(-1)]; + tensor var_400_axes_0 = const()[name = tensor("op_400_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12872832)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12873664)))]; + tensor var_389_to_fp16 = const()[name = tensor("op_389_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_400_cast = layer_norm(axes = var_400_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_389_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_411_to_fp16 = const()[name = tensor("op_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12874496)))]; + tensor var_412_to_fp16 = const()[name = tensor("op_412_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13169472)))]; + tensor q_13_cast = linear(bias = var_412_to_fp16, weight = var_411_to_fp16, x = var_400_cast); + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13170304)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13465280)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_415_to_fp16, x = var_400_cast); + tensor var_419_to_fp16 = const()[name = tensor("op_419_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13466112)))]; + tensor var_420_to_fp16 = const()[name = tensor("op_420_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13761088)))]; + tensor v_13_cast = linear(bias = var_420_to_fp16, weight = var_419_to_fp16, x = var_400_cast); + tensor var_428 = const()[name = tensor("op_428"), val = tensor([1, 1500, 6, -1])]; + tensor var_429_cast = reshape(shape = var_428, x = q_13_cast); + tensor const_34_to_fp16 = const()[name = tensor("const_34_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_429_cast, y = const_34_to_fp16); + tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 1500, 6, -1])]; + tensor var_436_cast = reshape(shape = var_435, x = k_13_cast); + tensor const_35_to_fp16 = const()[name = tensor("const_35_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_436_cast, y = const_35_to_fp16); + tensor var_442 = const()[name = tensor("op_442"), val = tensor([1, 1500, 6, -1])]; + tensor var_443_cast = reshape(shape = var_442, x = v_13_cast); + tensor var_444 = const()[name = tensor("op_444"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_17 = transpose(perm = transpose_15_perm_0, x = k_cast); + tensor transpose_18 = transpose(perm = transpose_14_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_18, y = transpose_17); + tensor var_448_cast = softmax(axis = var_383, x = qk_cast); + tensor var_450_transpose_x_0 = const()[name = tensor("op_450_transpose_x_0"), val = tensor(false)]; + tensor var_450_transpose_y_0 = const()[name = tensor("op_450_transpose_y_0"), val = tensor(false)]; + tensor transpose_19 = transpose(perm = var_444, x = var_443_cast); + tensor var_450_cast = matmul(transpose_x = var_450_transpose_x_0, transpose_y = var_450_transpose_y_0, x = var_448_cast, y = transpose_19); + tensor var_451 = const()[name = tensor("op_451"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 384])]; + tensor transpose_16 = transpose(perm = var_451, x = var_450_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_16); + tensor var_456_to_fp16 = const()[name = tensor("op_456_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13761920)))]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14056896)))]; + tensor var_458_cast = linear(bias = var_457_to_fp16, weight = var_456_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_458_cast); + tensor var_464_axes_0 = const()[name = tensor("op_464_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14057728)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14058560)))]; + tensor var_464_cast = layer_norm(axes = var_464_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_389_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_473_to_fp16 = const()[name = tensor("op_473_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14059392)))]; + tensor var_474_to_fp16 = const()[name = tensor("op_474_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15239104)))]; + tensor input_33_cast = linear(bias = var_474_to_fp16, weight = var_473_to_fp16, x = var_464_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15242240)))]; + tensor var_480_to_fp16 = const()[name = tensor("op_480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16421952)))]; + tensor var_481_cast = linear(bias = var_480_to_fp16, weight = var_479_to_fp16, x = x_53_cast); + tensor x_cast = add(x = x_49_cast, y = var_481_cast); + tensor var_494_axes_0 = const()[name = tensor("op_494_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16422784)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16423616)))]; + tensor var_485_to_fp16 = const()[name = tensor("op_485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_494_cast = layer_norm(axes = var_494_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_485_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_494_cast_to_fp32_dtype_0 = const()[name = tensor("op_494_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_494_cast_to_fp32_dtype_0, x = var_494_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1e6c109974ed71f6326eca3e75c025a1c6234b26 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d3e4113a57206004a0ca73454ac9fcc7aed518a1729a83402629ccec54a022c +size 16424448 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc.7z b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc.7z new file mode 100644 index 0000000000000000000000000000000000000000..ae0b1243cbbfbb30b3c7c6f54ff8e5c81309fb85 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc.7z @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f7313a3a35f564f528ecb6a4f0259ed0d53f486b60a6f492906a15d63f848c0 +size 14290462 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..16ba4c4c65c22b09941cb787a9302b5fc0130c55 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2f746e204d04779629b4959dad86d4ada040edd54f0ed1c80aa6f482a0fa72e +size 207 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/coremldata.bin b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c732ed7b8caeac3c15acda1c2836f3e1f525a5a8 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05fe28591b40616fa0c34ad7b853133623f5300923ec812acb11459c411acf3b +size 149 diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/metadata.json b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..de3f1ce87c6a39a44201cfdcb1bdd865eb81e4d3 --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/metadata.json @@ -0,0 +1,64 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float16", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32)", + "shortDescription" : "", + "shape" : "[]", + "name" : "output", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 6, + "mlProgramOperationTypeHistogram" : { + "Linear" : 24, + "Matmul" : 8, + "Cast" : 2, + "Conv" : 2, + "Softmax" : 4, + "Add" : 9, + "LayerNorm" : 9, + "Mul" : 8, + "Transpose" : 17, + "Gelu" : 6, + "Reshape" : 16 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "availability" : { + "macOS" : "12.0", + "tvOS" : "15.0", + "watchOS" : "8.0", + "iOS" : "15.0", + "macCatalyst" : "15.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + + }, + "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/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/model.mil b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..ccedc27eec99a20a310efef71e5bbc4d874439cc --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/model.mil @@ -0,0 +1,275 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "4.28.4"}, {"coremlc-version", "1436.100.10"}})] +{ + func main(tensor logmel_data) { + tensor var_16 = const()[name = tensor("op_16"), val = tensor(1)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor([1])]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1])]; + 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 logmel_data_to_fp16_dtype_0 = const()[name = tensor("logmel_data_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor weight_3_to_fp16 = const()[name = tensor("weight_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor bias_3_to_fp16 = const()[name = tensor("bias_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184448)))]; + tensor cast_127 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data); + tensor var_28_cast = conv(bias = bias_3_to_fp16, dilations = var_26, groups = var_16, pad = var_28_pad_0, pad_type = var_28_pad_type_0, strides = var_24, weight = weight_3_to_fp16, x = cast_127); + tensor input_1_mode_0 = const()[name = tensor("input_1_mode_0"), val = tensor("EXACT")]; + tensor input_1_cast = gelu(mode = input_1_mode_0, x = var_28_cast); + tensor var_32 = const()[name = tensor("op_32"), val = tensor(1)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor([2])]; + tensor var_43 = const()[name = tensor("op_43"), val = tensor([1])]; + tensor var_45_pad_type_0 = const()[name = tensor("op_45_pad_type_0"), val = tensor("custom")]; + tensor var_45_pad_0 = const()[name = tensor("op_45_pad_0"), val = tensor([1, 1])]; + tensor weight_7_to_fp16 = const()[name = tensor("weight_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185280)))]; + tensor bias_7_to_fp16 = const()[name = tensor("bias_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070080)))]; + tensor var_45_cast = conv(bias = bias_7_to_fp16, dilations = var_43, groups = var_32, pad = var_45_pad_0, pad_type = var_45_pad_type_0, strides = var_41, weight = weight_7_to_fp16, x = input_1_cast); + tensor x_3_mode_0 = const()[name = tensor("x_3_mode_0"), val = tensor("EXACT")]; + tensor x_3_cast = gelu(mode = x_3_mode_0, x = var_45_cast); + tensor var_50 = const()[name = tensor("op_50"), val = tensor([0, 2, 1])]; + tensor positional_embedding_to_fp16 = const()[name = tensor("positional_embedding_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1070912)))]; + tensor transpose_32 = transpose(perm = var_50, x = x_3_cast); + tensor var_53_cast = add(x = transpose_32, y = positional_embedding_to_fp16); + tensor var_65 = const()[name = tensor("op_65"), val = tensor(-1)]; + tensor var_82_axes_0 = const()[name = tensor("op_82_axes_0"), val = tensor([-1])]; + tensor blocks_0_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_0_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2222976)))]; + tensor blocks_0_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_0_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2223808)))]; + tensor var_71_to_fp16 = const()[name = tensor("op_71_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_82_cast = layer_norm(axes = var_82_axes_0, beta = blocks_0_attn_ln_bias_to_fp16, epsilon = var_71_to_fp16, gamma = blocks_0_attn_ln_weight_to_fp16, x = var_53_cast); + tensor var_93_to_fp16 = const()[name = tensor("op_93_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2224640)))]; + tensor var_94_to_fp16 = const()[name = tensor("op_94_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2519616)))]; + tensor q_1_cast = linear(bias = var_94_to_fp16, weight = var_93_to_fp16, x = var_82_cast); + tensor var_97_to_fp16 = const()[name = tensor("op_97_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2520448)))]; + tensor k_1_bias_0_to_fp16 = const()[name = tensor("k_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2815424)))]; + tensor k_1_cast = linear(bias = k_1_bias_0_to_fp16, weight = var_97_to_fp16, x = var_82_cast); + tensor var_101_to_fp16 = const()[name = tensor("op_101_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2816256)))]; + tensor var_102_to_fp16 = const()[name = tensor("op_102_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3111232)))]; + tensor v_1_cast = linear(bias = var_102_to_fp16, weight = var_101_to_fp16, x = var_82_cast); + tensor var_110 = const()[name = tensor("op_110"), val = tensor([1, 1500, 6, -1])]; + tensor var_111_cast = reshape(shape = var_110, x = q_1_cast); + tensor const_28_to_fp16 = const()[name = tensor("const_28_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_3_cast = mul(x = var_111_cast, y = const_28_to_fp16); + tensor var_117 = const()[name = tensor("op_117"), val = tensor([1, 1500, 6, -1])]; + tensor var_118_cast = reshape(shape = var_117, x = k_1_cast); + tensor const_29_to_fp16 = const()[name = tensor("const_29_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_3_cast = mul(x = var_118_cast, y = const_29_to_fp16); + tensor var_124 = const()[name = tensor("op_124"), val = tensor([1, 1500, 6, -1])]; + tensor var_125_cast = reshape(shape = var_124, x = v_1_cast); + tensor var_126 = const()[name = tensor("op_126"), val = tensor([0, 2, 1, 3])]; + tensor qk_1_transpose_x_0 = const()[name = tensor("qk_1_transpose_x_0"), val = tensor(false)]; + tensor qk_1_transpose_y_0 = const()[name = tensor("qk_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_29 = transpose(perm = transpose_9_perm_0, x = k_3_cast); + tensor transpose_30 = transpose(perm = transpose_8_perm_0, x = q_3_cast); + tensor qk_1_cast = matmul(transpose_x = qk_1_transpose_x_0, transpose_y = qk_1_transpose_y_0, x = transpose_30, y = transpose_29); + tensor var_130_cast = softmax(axis = var_65, x = qk_1_cast); + tensor var_132_transpose_x_0 = const()[name = tensor("op_132_transpose_x_0"), val = tensor(false)]; + tensor var_132_transpose_y_0 = const()[name = tensor("op_132_transpose_y_0"), val = tensor(false)]; + tensor transpose_31 = transpose(perm = var_126, x = var_125_cast); + tensor var_132_cast = matmul(transpose_x = var_132_transpose_x_0, transpose_y = var_132_transpose_y_0, x = var_130_cast, y = transpose_31); + tensor var_133 = const()[name = tensor("op_133"), val = tensor([0, 2, 1, 3])]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([1, 1500, 384])]; + tensor transpose_28 = transpose(perm = var_133, x = var_132_cast); + tensor x_11_cast = reshape(shape = concat_0, x = transpose_28); + tensor var_138_to_fp16 = const()[name = tensor("op_138_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3112064)))]; + tensor var_139_to_fp16 = const()[name = tensor("op_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407040)))]; + tensor var_140_cast = linear(bias = var_139_to_fp16, weight = var_138_to_fp16, x = x_11_cast); + tensor x_13_cast = add(x = var_53_cast, y = var_140_cast); + tensor var_146_axes_0 = const()[name = tensor("op_146_axes_0"), val = tensor([-1])]; + tensor blocks_0_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3407872)))]; + tensor blocks_0_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_0_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3408704)))]; + tensor var_146_cast = layer_norm(axes = var_146_axes_0, beta = blocks_0_mlp_ln_bias_to_fp16, epsilon = var_71_to_fp16, gamma = blocks_0_mlp_ln_weight_to_fp16, x = x_13_cast); + tensor var_155_to_fp16 = const()[name = tensor("op_155_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3409536)))]; + tensor var_156_to_fp16 = const()[name = tensor("op_156_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4589248)))]; + tensor input_9_cast = linear(bias = var_156_to_fp16, weight = var_155_to_fp16, x = var_146_cast); + tensor x_17_mode_0 = const()[name = tensor("x_17_mode_0"), val = tensor("EXACT")]; + tensor x_17_cast = gelu(mode = x_17_mode_0, x = input_9_cast); + tensor var_161_to_fp16 = const()[name = tensor("op_161_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4592384)))]; + tensor var_162_to_fp16 = const()[name = tensor("op_162_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772096)))]; + tensor var_163_cast = linear(bias = var_162_to_fp16, weight = var_161_to_fp16, x = x_17_cast); + tensor x_19_cast = add(x = x_13_cast, y = var_163_cast); + tensor var_171 = const()[name = tensor("op_171"), val = tensor(-1)]; + tensor var_188_axes_0 = const()[name = tensor("op_188_axes_0"), val = tensor([-1])]; + tensor blocks_1_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_1_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5772928)))]; + tensor blocks_1_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_1_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5773760)))]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_188_cast = layer_norm(axes = var_188_axes_0, beta = blocks_1_attn_ln_bias_to_fp16, epsilon = var_177_to_fp16, gamma = blocks_1_attn_ln_weight_to_fp16, x = x_19_cast); + tensor var_199_to_fp16 = const()[name = tensor("op_199_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5774592)))]; + tensor var_200_to_fp16 = const()[name = tensor("op_200_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6069568)))]; + tensor q_5_cast = linear(bias = var_200_to_fp16, weight = var_199_to_fp16, x = var_188_cast); + tensor var_203_to_fp16 = const()[name = tensor("op_203_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6070400)))]; + tensor k_5_bias_0_to_fp16 = const()[name = tensor("k_5_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6365376)))]; + tensor k_5_cast = linear(bias = k_5_bias_0_to_fp16, weight = var_203_to_fp16, x = var_188_cast); + tensor var_207_to_fp16 = const()[name = tensor("op_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6366208)))]; + tensor var_208_to_fp16 = const()[name = tensor("op_208_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6661184)))]; + tensor v_5_cast = linear(bias = var_208_to_fp16, weight = var_207_to_fp16, x = var_188_cast); + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1500, 6, -1])]; + tensor var_217_cast = reshape(shape = var_216, x = q_5_cast); + tensor const_30_to_fp16 = const()[name = tensor("const_30_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_7_cast = mul(x = var_217_cast, y = const_30_to_fp16); + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 1500, 6, -1])]; + tensor var_224_cast = reshape(shape = var_223, x = k_5_cast); + tensor const_31_to_fp16 = const()[name = tensor("const_31_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_7_cast = mul(x = var_224_cast, y = const_31_to_fp16); + tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 1500, 6, -1])]; + tensor var_231_cast = reshape(shape = var_230, x = v_5_cast); + tensor var_232 = const()[name = tensor("op_232"), val = tensor([0, 2, 1, 3])]; + tensor qk_3_transpose_x_0 = const()[name = tensor("qk_3_transpose_x_0"), val = tensor(false)]; + tensor qk_3_transpose_y_0 = const()[name = tensor("qk_3_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_25 = transpose(perm = transpose_11_perm_0, x = k_7_cast); + tensor transpose_26 = transpose(perm = transpose_10_perm_0, x = q_7_cast); + tensor qk_3_cast = matmul(transpose_x = qk_3_transpose_x_0, transpose_y = qk_3_transpose_y_0, x = transpose_26, y = transpose_25); + tensor var_236_cast = softmax(axis = var_171, x = qk_3_cast); + tensor var_238_transpose_x_0 = const()[name = tensor("op_238_transpose_x_0"), val = tensor(false)]; + tensor var_238_transpose_y_0 = const()[name = tensor("op_238_transpose_y_0"), val = tensor(false)]; + tensor transpose_27 = transpose(perm = var_232, x = var_231_cast); + tensor var_238_cast = matmul(transpose_x = var_238_transpose_x_0, transpose_y = var_238_transpose_y_0, x = var_236_cast, y = transpose_27); + tensor var_239 = const()[name = tensor("op_239"), val = tensor([0, 2, 1, 3])]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([1, 1500, 384])]; + tensor transpose_24 = transpose(perm = var_239, x = var_238_cast); + tensor x_23_cast = reshape(shape = concat_1, x = transpose_24); + tensor var_244_to_fp16 = const()[name = tensor("op_244_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6662016)))]; + tensor var_245_to_fp16 = const()[name = tensor("op_245_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6956992)))]; + tensor var_246_cast = linear(bias = var_245_to_fp16, weight = var_244_to_fp16, x = x_23_cast); + tensor x_25_cast = add(x = x_19_cast, y = var_246_cast); + tensor var_252_axes_0 = const()[name = tensor("op_252_axes_0"), val = tensor([-1])]; + tensor blocks_1_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6957824)))]; + tensor blocks_1_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_1_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6958656)))]; + tensor var_252_cast = layer_norm(axes = var_252_axes_0, beta = blocks_1_mlp_ln_bias_to_fp16, epsilon = var_177_to_fp16, gamma = blocks_1_mlp_ln_weight_to_fp16, x = x_25_cast); + tensor var_261_to_fp16 = const()[name = tensor("op_261_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6959488)))]; + tensor var_262_to_fp16 = const()[name = tensor("op_262_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8139200)))]; + tensor input_17_cast = linear(bias = var_262_to_fp16, weight = var_261_to_fp16, x = var_252_cast); + tensor x_29_mode_0 = const()[name = tensor("x_29_mode_0"), val = tensor("EXACT")]; + tensor x_29_cast = gelu(mode = x_29_mode_0, x = input_17_cast); + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8142336)))]; + tensor var_268_to_fp16 = const()[name = tensor("op_268_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322048)))]; + tensor var_269_cast = linear(bias = var_268_to_fp16, weight = var_267_to_fp16, x = x_29_cast); + tensor x_31_cast = add(x = x_25_cast, y = var_269_cast); + tensor var_277 = const()[name = tensor("op_277"), val = tensor(-1)]; + tensor var_294_axes_0 = const()[name = tensor("op_294_axes_0"), val = tensor([-1])]; + tensor blocks_2_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_2_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9322880)))]; + tensor blocks_2_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_2_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9323712)))]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_294_cast = layer_norm(axes = var_294_axes_0, beta = blocks_2_attn_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_2_attn_ln_weight_to_fp16, x = x_31_cast); + tensor var_305_to_fp16 = const()[name = tensor("op_305_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9324544)))]; + tensor var_306_to_fp16 = const()[name = tensor("op_306_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9619520)))]; + tensor q_9_cast = linear(bias = var_306_to_fp16, weight = var_305_to_fp16, x = var_294_cast); + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9620352)))]; + tensor k_9_bias_0_to_fp16 = const()[name = tensor("k_9_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9915328)))]; + tensor k_9_cast = linear(bias = k_9_bias_0_to_fp16, weight = var_309_to_fp16, x = var_294_cast); + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9916160)))]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10211136)))]; + tensor v_9_cast = linear(bias = var_314_to_fp16, weight = var_313_to_fp16, x = var_294_cast); + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 1500, 6, -1])]; + tensor var_323_cast = reshape(shape = var_322, x = q_9_cast); + tensor const_32_to_fp16 = const()[name = tensor("const_32_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_11_cast = mul(x = var_323_cast, y = const_32_to_fp16); + tensor var_329 = const()[name = tensor("op_329"), val = tensor([1, 1500, 6, -1])]; + tensor var_330_cast = reshape(shape = var_329, x = k_9_cast); + tensor const_33_to_fp16 = const()[name = tensor("const_33_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_11_cast = mul(x = var_330_cast, y = const_33_to_fp16); + tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1500, 6, -1])]; + tensor var_337_cast = reshape(shape = var_336, x = v_9_cast); + tensor var_338 = const()[name = tensor("op_338"), val = tensor([0, 2, 1, 3])]; + tensor qk_5_transpose_x_0 = const()[name = tensor("qk_5_transpose_x_0"), val = tensor(false)]; + tensor qk_5_transpose_y_0 = const()[name = tensor("qk_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_21 = transpose(perm = transpose_13_perm_0, x = k_11_cast); + tensor transpose_22 = transpose(perm = transpose_12_perm_0, x = q_11_cast); + tensor qk_5_cast = matmul(transpose_x = qk_5_transpose_x_0, transpose_y = qk_5_transpose_y_0, x = transpose_22, y = transpose_21); + tensor var_342_cast = softmax(axis = var_277, x = qk_5_cast); + tensor var_344_transpose_x_0 = const()[name = tensor("op_344_transpose_x_0"), val = tensor(false)]; + tensor var_344_transpose_y_0 = const()[name = tensor("op_344_transpose_y_0"), val = tensor(false)]; + tensor transpose_23 = transpose(perm = var_338, x = var_337_cast); + tensor var_344_cast = matmul(transpose_x = var_344_transpose_x_0, transpose_y = var_344_transpose_y_0, x = var_342_cast, y = transpose_23); + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([1, 1500, 384])]; + tensor transpose_20 = transpose(perm = var_345, x = var_344_cast); + tensor x_35_cast = reshape(shape = concat_2, x = transpose_20); + tensor var_350_to_fp16 = const()[name = tensor("op_350_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10211968)))]; + tensor var_351_to_fp16 = const()[name = tensor("op_351_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10506944)))]; + tensor var_352_cast = linear(bias = var_351_to_fp16, weight = var_350_to_fp16, x = x_35_cast); + tensor x_37_cast = add(x = x_31_cast, y = var_352_cast); + tensor var_358_axes_0 = const()[name = tensor("op_358_axes_0"), val = tensor([-1])]; + tensor blocks_2_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10507776)))]; + tensor blocks_2_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_2_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10508608)))]; + tensor var_358_cast = layer_norm(axes = var_358_axes_0, beta = blocks_2_mlp_ln_bias_to_fp16, epsilon = var_283_to_fp16, gamma = blocks_2_mlp_ln_weight_to_fp16, x = x_37_cast); + tensor var_367_to_fp16 = const()[name = tensor("op_367_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10509440)))]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11689152)))]; + tensor input_25_cast = linear(bias = var_368_to_fp16, weight = var_367_to_fp16, x = var_358_cast); + tensor x_41_mode_0 = const()[name = tensor("x_41_mode_0"), val = tensor("EXACT")]; + tensor x_41_cast = gelu(mode = x_41_mode_0, x = input_25_cast); + tensor var_373_to_fp16 = const()[name = tensor("op_373_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11692288)))]; + tensor var_374_to_fp16 = const()[name = tensor("op_374_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12872000)))]; + tensor var_375_cast = linear(bias = var_374_to_fp16, weight = var_373_to_fp16, x = x_41_cast); + tensor x_43_cast = add(x = x_37_cast, y = var_375_cast); + tensor var_383 = const()[name = tensor("op_383"), val = tensor(-1)]; + tensor var_400_axes_0 = const()[name = tensor("op_400_axes_0"), val = tensor([-1])]; + tensor blocks_3_attn_ln_weight_to_fp16 = const()[name = tensor("blocks_3_attn_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12872832)))]; + tensor blocks_3_attn_ln_bias_to_fp16 = const()[name = tensor("blocks_3_attn_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12873664)))]; + tensor var_389_to_fp16 = const()[name = tensor("op_389_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_400_cast = layer_norm(axes = var_400_axes_0, beta = blocks_3_attn_ln_bias_to_fp16, epsilon = var_389_to_fp16, gamma = blocks_3_attn_ln_weight_to_fp16, x = x_43_cast); + tensor var_411_to_fp16 = const()[name = tensor("op_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12874496)))]; + tensor var_412_to_fp16 = const()[name = tensor("op_412_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13169472)))]; + tensor q_13_cast = linear(bias = var_412_to_fp16, weight = var_411_to_fp16, x = var_400_cast); + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13170304)))]; + tensor k_13_bias_0_to_fp16 = const()[name = tensor("k_13_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13465280)))]; + tensor k_13_cast = linear(bias = k_13_bias_0_to_fp16, weight = var_415_to_fp16, x = var_400_cast); + tensor var_419_to_fp16 = const()[name = tensor("op_419_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13466112)))]; + tensor var_420_to_fp16 = const()[name = tensor("op_420_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13761088)))]; + tensor v_13_cast = linear(bias = var_420_to_fp16, weight = var_419_to_fp16, x = var_400_cast); + tensor var_428 = const()[name = tensor("op_428"), val = tensor([1, 1500, 6, -1])]; + tensor var_429_cast = reshape(shape = var_428, x = q_13_cast); + tensor const_34_to_fp16 = const()[name = tensor("const_34_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor q_cast = mul(x = var_429_cast, y = const_34_to_fp16); + tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 1500, 6, -1])]; + tensor var_436_cast = reshape(shape = var_435, x = k_13_cast); + tensor const_35_to_fp16 = const()[name = tensor("const_35_to_fp16"), val = tensor([[[[0x1.6ap-2]]]])]; + tensor k_cast = mul(x = var_436_cast, y = const_35_to_fp16); + tensor var_442 = const()[name = tensor("op_442"), val = tensor([1, 1500, 6, -1])]; + tensor var_443_cast = reshape(shape = var_442, x = v_13_cast); + tensor var_444 = const()[name = tensor("op_444"), val = tensor([0, 2, 1, 3])]; + tensor qk_transpose_x_0 = const()[name = tensor("qk_transpose_x_0"), val = tensor(false)]; + tensor qk_transpose_y_0 = const()[name = tensor("qk_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor transpose_17 = transpose(perm = transpose_15_perm_0, x = k_cast); + tensor transpose_18 = transpose(perm = transpose_14_perm_0, x = q_cast); + tensor qk_cast = matmul(transpose_x = qk_transpose_x_0, transpose_y = qk_transpose_y_0, x = transpose_18, y = transpose_17); + tensor var_448_cast = softmax(axis = var_383, x = qk_cast); + tensor var_450_transpose_x_0 = const()[name = tensor("op_450_transpose_x_0"), val = tensor(false)]; + tensor var_450_transpose_y_0 = const()[name = tensor("op_450_transpose_y_0"), val = tensor(false)]; + tensor transpose_19 = transpose(perm = var_444, x = var_443_cast); + tensor var_450_cast = matmul(transpose_x = var_450_transpose_x_0, transpose_y = var_450_transpose_y_0, x = var_448_cast, y = transpose_19); + tensor var_451 = const()[name = tensor("op_451"), val = tensor([0, 2, 1, 3])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([1, 1500, 384])]; + tensor transpose_16 = transpose(perm = var_451, x = var_450_cast); + tensor x_47_cast = reshape(shape = concat_3, x = transpose_16); + tensor var_456_to_fp16 = const()[name = tensor("op_456_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13761920)))]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14056896)))]; + tensor var_458_cast = linear(bias = var_457_to_fp16, weight = var_456_to_fp16, x = x_47_cast); + tensor x_49_cast = add(x = x_43_cast, y = var_458_cast); + tensor var_464_axes_0 = const()[name = tensor("op_464_axes_0"), val = tensor([-1])]; + tensor blocks_3_mlp_ln_weight_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14057728)))]; + tensor blocks_3_mlp_ln_bias_to_fp16 = const()[name = tensor("blocks_3_mlp_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14058560)))]; + tensor var_464_cast = layer_norm(axes = var_464_axes_0, beta = blocks_3_mlp_ln_bias_to_fp16, epsilon = var_389_to_fp16, gamma = blocks_3_mlp_ln_weight_to_fp16, x = x_49_cast); + tensor var_473_to_fp16 = const()[name = tensor("op_473_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14059392)))]; + tensor var_474_to_fp16 = const()[name = tensor("op_474_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15239104)))]; + tensor input_33_cast = linear(bias = var_474_to_fp16, weight = var_473_to_fp16, x = var_464_cast); + tensor x_53_mode_0 = const()[name = tensor("x_53_mode_0"), val = tensor("EXACT")]; + tensor x_53_cast = gelu(mode = x_53_mode_0, x = input_33_cast); + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15242240)))]; + tensor var_480_to_fp16 = const()[name = tensor("op_480_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16421952)))]; + tensor var_481_cast = linear(bias = var_480_to_fp16, weight = var_479_to_fp16, x = x_53_cast); + tensor x_cast = add(x = x_49_cast, y = var_481_cast); + tensor var_494_axes_0 = const()[name = tensor("op_494_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16422784)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16423616)))]; + tensor var_485_to_fp16 = const()[name = tensor("op_485_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_494_cast = layer_norm(axes = var_494_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_485_to_fp16, gamma = ln_post_weight_to_fp16, x = x_cast); + tensor var_494_cast_to_fp32_dtype_0 = const()[name = tensor("op_494_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor output = cast(dtype = var_494_cast_to_fp32_dtype_0, x = var_494_cast); + } -> (output); +} \ No newline at end of file diff --git a/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..6c5a335a3ef0654b49359ba603cc1449c90a7e1e --- /dev/null +++ b/whisper.cpp/encoder.mlmodelc/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c59faa13544fde24098d9594a3f955d0515b4aa8f3e8a60cd1c022e76bba8d31 +size 16424448 diff --git a/whisper.cpp/ggml-base-q5_1.bin b/whisper.cpp/ggml-base-q5_1.bin new file mode 100644 index 0000000000000000000000000000000000000000..4947e22c7365d941a9864b2d0f9f474a505cf39b --- /dev/null +++ b/whisper.cpp/ggml-base-q5_1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422f1ae452ade6f30a004d7e5c6a43195e4433bc370bf23fac9cc591f01a8898 +size 59707625 diff --git a/whisper.cpp/ggml-base-q8_0.bin b/whisper.cpp/ggml-base-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..1a710c5bff0c9b764ee332dd9a8422bc978bfb1e --- /dev/null +++ b/whisper.cpp/ggml-base-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c577b9a86e7e048a0b7eada054f4dd79a56bbfa911fbdacf900ac5b567cbb7d9 +size 81768585 diff --git a/whisper.cpp/ggml-base.bin b/whisper.cpp/ggml-base.bin new file mode 100644 index 0000000000000000000000000000000000000000..17993a254ac4e9db7642a261cf53af3dc446145c --- /dev/null +++ b/whisper.cpp/ggml-base.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe +size 147951465 diff --git a/whisper.cpp/ggml-base.en-q5_1.bin b/whisper.cpp/ggml-base.en-q5_1.bin new file mode 100644 index 0000000000000000000000000000000000000000..5c89fe2c0d47f0d16d536610dc5f7c4a6fe81bc7 --- /dev/null +++ b/whisper.cpp/ggml-base.en-q5_1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4baf70dd0d7c4247ba2b81fafd9c01005ac77c2f9ef064e00dcf195d0e2fdd2f +size 59721011 diff --git a/whisper.cpp/ggml-base.en-q8_0.bin b/whisper.cpp/ggml-base.en-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..f459c0ad21ffd7b3661f62914745c59bb63f2422 --- /dev/null +++ b/whisper.cpp/ggml-base.en-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4d4a0768075e13cfd7e19df3ae2dbc4a68d37d36a7dad45e8410c9a34f8c87e +size 81781811 diff --git a/whisper.cpp/ggml-base.en.bin b/whisper.cpp/ggml-base.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..87c664c563ef3ff52424dd4fa925cf95b306dba6 --- /dev/null +++ b/whisper.cpp/ggml-base.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a03779c86df3323075f5e796cb2ce5029f00ec8869eee3fdfb897afe36c6d002 +size 147964211 diff --git a/whisper.cpp/ggml-large-v1.bin b/whisper.cpp/ggml-large-v1.bin new file mode 100644 index 0000000000000000000000000000000000000000..047db73ddfd7789113dfc94c20b22bc1d044586f --- /dev/null +++ b/whisper.cpp/ggml-large-v1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d99f41a10525d0206bddadd86760181fa920438b6b33237e3118ff6c83bb53d +size 3094623691 diff --git a/whisper.cpp/ggml-large-v2-q5_0.bin b/whisper.cpp/ggml-large-v2-q5_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..8ea090d899ab4099f57a60abcba4d4008553b675 --- /dev/null +++ b/whisper.cpp/ggml-large-v2-q5_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a214837221e4530dbc1fe8d734f302af393eb30bd0ed046042ebf4baf70f6f2 +size 1080732091 diff --git a/whisper.cpp/ggml-large-v2-q8_0.bin b/whisper.cpp/ggml-large-v2-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..423bfb2f2932572249bc17a16139883d9aaed65f --- /dev/null +++ b/whisper.cpp/ggml-large-v2-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fef54e6d898246a65c8285bfa83bd1807e27fadf54d5d4e81754c47634737e8c +size 1656129691 diff --git a/whisper.cpp/ggml-large-v2.bin b/whisper.cpp/ggml-large-v2.bin new file mode 100644 index 0000000000000000000000000000000000000000..649aafd67e30021d0140c24342ee2ffb947f4bde --- /dev/null +++ b/whisper.cpp/ggml-large-v2.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a423fe4d40c82774b6af34115b8b935f34152246eb19e80e376071d3f999487 +size 3094623691 diff --git a/whisper.cpp/ggml-large-v3-q5_0.bin b/whisper.cpp/ggml-large-v3-q5_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..14150f8463725bde9626c0656683ad5735e01ac2 --- /dev/null +++ b/whisper.cpp/ggml-large-v3-q5_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d75795ecff3f83b5faa89d1900604ad8c780abd5739fae406de19f23ecd98ad1 +size 1081140203 diff --git a/whisper.cpp/ggml-large-v3-turbo-q5_0.bin b/whisper.cpp/ggml-large-v3-turbo-q5_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..0e2474d5ec0361bb1726829aa83317ed4cbc3f18 --- /dev/null +++ b/whisper.cpp/ggml-large-v3-turbo-q5_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:394221709cd5ad1f40c46e6031ca61bce88931e6e088c188294c6d5a55ffa7e2 +size 574041195 diff --git a/whisper.cpp/ggml-large-v3-turbo-q8_0.bin b/whisper.cpp/ggml-large-v3-turbo-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..dd827281ffb9642e91f152f2a90de53907d4603a --- /dev/null +++ b/whisper.cpp/ggml-large-v3-turbo-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:317eb69c11673c9de1e1f0d459b253999804ec71ac4c23c17ecf5fbe24e259a1 +size 874188075 diff --git a/whisper.cpp/ggml-large-v3-turbo.bin b/whisper.cpp/ggml-large-v3-turbo.bin new file mode 100644 index 0000000000000000000000000000000000000000..819841c70bdf4488c4ff778f8becdcb37df43969 --- /dev/null +++ b/whisper.cpp/ggml-large-v3-turbo.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fc70f774d38eb169993ac391eea357ef47c88757ef72ee5943879b7e8e2bc69 +size 1624555275 diff --git a/whisper.cpp/ggml-large-v3.bin b/whisper.cpp/ggml-large-v3.bin new file mode 100644 index 0000000000000000000000000000000000000000..30488f6b9eeae93e026c978ac7a3190274732ea2 --- /dev/null +++ b/whisper.cpp/ggml-large-v3.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64d182b440b98d5203c4f9bd541544d84c605196c4f7b845dfa11fb23594d1e2 +size 3095033483 diff --git a/whisper.cpp/ggml-medium-q5_0.bin b/whisper.cpp/ggml-medium-q5_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..9cc19f38fec2314f6c00f666ca708e8cf22d189d --- /dev/null +++ b/whisper.cpp/ggml-medium-q5_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19fea4b380c3a618ec4723c3eef2eb785ffba0d0538cf43f8f235e7b3b34220f +size 539212467 diff --git a/whisper.cpp/ggml-medium-q8_0.bin b/whisper.cpp/ggml-medium-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..6c783bed539147853779c28702e59d170ab6a5bb --- /dev/null +++ b/whisper.cpp/ggml-medium-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42a1ffcbe4167d224232443396968db4d02d4e8e87e213d3ee2e03095dea6502 +size 823369779 diff --git a/whisper.cpp/ggml-medium.bin b/whisper.cpp/ggml-medium.bin new file mode 100644 index 0000000000000000000000000000000000000000..be775c464024bdcd9e834c2f9ed42aae6b708d2b --- /dev/null +++ b/whisper.cpp/ggml-medium.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c14d5adee5f86394037b4e4e8b59f1673b6cee10e3cf0b11bbdbee79c156208 +size 1533763059 diff --git a/whisper.cpp/ggml-medium.en-q5_0.bin b/whisper.cpp/ggml-medium.en-q5_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..8722db4993ba67f49c4f20cdd9880202c8cc3f2b --- /dev/null +++ b/whisper.cpp/ggml-medium.en-q5_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76733e26ad8fe1c7a5bf7531a9d41917b2adc0f20f2e4f5531688a8c6cd88eb0 +size 539225533 diff --git a/whisper.cpp/ggml-medium.en-q8_0.bin b/whisper.cpp/ggml-medium.en-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..c4ca4e1562dffe8db2381e32299fe64110228156 --- /dev/null +++ b/whisper.cpp/ggml-medium.en-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43fa2cd084de5a04399a896a9a7a786064e221365c01700cea4666005218f11c +size 823382461 diff --git a/whisper.cpp/ggml-medium.en.bin b/whisper.cpp/ggml-medium.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..f8d7f988b60916d7f7e7feee9897c037a09b2f85 --- /dev/null +++ b/whisper.cpp/ggml-medium.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc37e93478338ec7700281a7ac30a10128929eb8f427dda2e865faa8f6da4356 +size 1533774781 diff --git a/whisper.cpp/ggml-small-q5_1.bin b/whisper.cpp/ggml-small-q5_1.bin new file mode 100644 index 0000000000000000000000000000000000000000..87631c52c89d37ed2337a46ec4a009976db1f3f4 --- /dev/null +++ b/whisper.cpp/ggml-small-q5_1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae85e4a935d7a567bd102fe55afc16bb595bdb618e11b2fc7591bc08120411bb +size 190085487 diff --git a/whisper.cpp/ggml-small-q8_0.bin b/whisper.cpp/ggml-small-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..8a9937d2be60bae6b1efbb01b9ca1612ab2a617c --- /dev/null +++ b/whisper.cpp/ggml-small-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49c8fb02b65e6049d5fa6c04f81f53b867b5ec9540406812c643f177317f779f +size 264464607 diff --git a/whisper.cpp/ggml-small.bin b/whisper.cpp/ggml-small.bin new file mode 100644 index 0000000000000000000000000000000000000000..bc4c4d96b528654fb067e2b243bdc438cfaf0072 --- /dev/null +++ b/whisper.cpp/ggml-small.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1be3a9b2063867b937e64e2ec7483364a79917e157fa98c5d94b5c1fffea987b +size 487601967 diff --git a/whisper.cpp/ggml-small.en-q5_1.bin b/whisper.cpp/ggml-small.en-q5_1.bin new file mode 100644 index 0000000000000000000000000000000000000000..bf61baaf463626fe45fc9157b8e9c69de8943c87 --- /dev/null +++ b/whisper.cpp/ggml-small.en-q5_1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfdff4894dcb76bbf647d56263ea2a96645423f1669176f4844a1bf8e478ad30 +size 190098681 diff --git a/whisper.cpp/ggml-small.en-q8_0.bin b/whisper.cpp/ggml-small.en-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..9c826215654cf46674a38d02f19e123e17c5b2d3 --- /dev/null +++ b/whisper.cpp/ggml-small.en-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67a179f608ea6114bd3fdb9060e762b588a3fb3bd00c4387971be4d177958067 +size 264477561 diff --git a/whisper.cpp/ggml-small.en.bin b/whisper.cpp/ggml-small.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..eaeeb6d63378cf6515ff2c1cb4e33486ae6bcc2f --- /dev/null +++ b/whisper.cpp/ggml-small.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6138d6d58ecc8322097e0f987c32f1be8bb0a18532a3f88f734d1bbf9c41e5d +size 487614201 diff --git a/whisper.cpp/ggml-tiny-q5_1.bin b/whisper.cpp/ggml-tiny-q5_1.bin new file mode 100644 index 0000000000000000000000000000000000000000..42701a9c1d8a9f4f0b96dbcefcc049075814bcf3 --- /dev/null +++ b/whisper.cpp/ggml-tiny-q5_1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:818710568da3ca15689e31a743197b520007872ff9576237bda97bd1b469c3d7 +size 32152673 diff --git a/whisper.cpp/ggml-tiny-q8_0.bin b/whisper.cpp/ggml-tiny-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..fc48fd17b13f7eb147bab1c66dee10c97f471b27 --- /dev/null +++ b/whisper.cpp/ggml-tiny-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2085835d3f50733e2ff6e4b41ae8a2b8d8110461e18821b09a15c40c42d1cca +size 43537433 diff --git a/whisper.cpp/ggml-tiny.bin b/whisper.cpp/ggml-tiny.bin new file mode 100644 index 0000000000000000000000000000000000000000..d144f735b005ae8cbfa04a49e22fe40faa24dbec --- /dev/null +++ b/whisper.cpp/ggml-tiny.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be07e048e1e599ad46341c8d2a135645097a538221678b7acdd1b1919c6e1b21 +size 77691713 diff --git a/whisper.cpp/ggml-tiny.en-q5_1.bin b/whisper.cpp/ggml-tiny.en-q5_1.bin new file mode 100644 index 0000000000000000000000000000000000000000..46aeed3b01f3fc86f05fa3e8ca711cc1773ead08 --- /dev/null +++ b/whisper.cpp/ggml-tiny.en-q5_1.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c77c5766f1cef09b6b7d47f21b546cbddd4157886b3b5d6d4f709e91e66c7c2b +size 32166155 diff --git a/whisper.cpp/ggml-tiny.en-q8_0.bin b/whisper.cpp/ggml-tiny.en-q8_0.bin new file mode 100644 index 0000000000000000000000000000000000000000..959468eb9f72c2100ab5e4a9efc102a45cc8b313 --- /dev/null +++ b/whisper.cpp/ggml-tiny.en-q8_0.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bc2b3860aa151a4c6e7bb095e1fcce7cf12c7b020ca08dcec0c6d018bb7dd94 +size 43550795 diff --git a/whisper.cpp/ggml-tiny.en.bin b/whisper.cpp/ggml-tiny.en.bin new file mode 100644 index 0000000000000000000000000000000000000000..17ad750438d1d42162fe06ab4b21aef2389d2137 --- /dev/null +++ b/whisper.cpp/ggml-tiny.en.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:921e4cf8686fdd993dcd081a5da5b6c365bfde1162e72b08d75ac75289920b1f +size 77704715