| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] |
| { |
| func main<ios18>(tensor<fp32, [?, 64, 1, 4]> patches) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, dict<string, tensor<int32, [?]>>>>>((("DefaultShapes", {{"patches", [1, 64, 1, 4]}}), ("EnumeratedShapes", {{"35802335", {{"patches", [1, 64, 1, 4]}}}, {"e8a82450", {{"patches", [16, 64, 1, 4]}}}})))] { |
| string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> x_3_strides_0 = const()[name = string("x_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> x_3_dilations_0 = const()[name = string("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 x_3_groups_0 = const()[name = string("x_3_groups_0"), val = int32(1)]; |
| string patches_to_fp16_dtype_0 = const()[name = string("patches_to_fp16_dtype_0"), val = string("fp16")]; |
| tensor<fp16, [1024, 64, 1, 1]> var_61_to_fp16 = const()[name = string("op_61_to_fp16"), val = tensor<fp16, [1024, 64, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; |
| tensor<fp16, [1024]> var_60_to_fp16 = const()[name = string("op_60_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131200)))]; |
| tensor<fp16, [?, 64, 1, 4]> patches_to_fp16 = cast(dtype = patches_to_fp16_dtype_0, x = patches)[name = string("cast_74")]; |
| tensor<fp16, [?, 1024, 1, 4]> x_3_cast_fp16 = conv(bias = var_60_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = var_61_to_fp16, x = patches_to_fp16)[name = string("x_3_cast_fp16")]; |
| tensor<int32, [4]> var_71_begin_0 = const()[name = string("op_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> var_71_end_0 = const()[name = string("op_71_end_0"), val = tensor<int32, [4]>([0, 1024, 1, 1])]; |
| tensor<bool, [4]> var_71_end_mask_0 = const()[name = string("op_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])]; |
| tensor<fp16, [?, 1024, 1, 1]> var_71_cast_fp16 = slice_by_index(begin = var_71_begin_0, end = var_71_end_0, end_mask = var_71_end_mask_0, x = x_3_cast_fp16)[name = string("op_71_cast_fp16")]; |
| fp16 var_72_to_fp16 = const()[name = string("op_72_to_fp16"), val = fp16(0x0p+0)]; |
| tensor<fp16, [?, 1024, 1, 1]> var_73_cast_fp16 = mul(x = var_71_cast_fp16, y = var_72_to_fp16)[name = string("op_73_cast_fp16")]; |
| tensor<fp16, [1, 1024, 1, 1]> cls_column_to_fp16 = const()[name = string("cls_column_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133312)))]; |
| tensor<fp16, [?, 1024, 1, 1]> cls_column_1_cast_fp16 = add(x = var_73_cast_fp16, y = cls_column_to_fp16)[name = string("cls_column_1_cast_fp16")]; |
| int32 var_77 = const()[name = string("op_77"), val = int32(-1)]; |
| bool hidden_states_1_interleave_0 = const()[name = string("hidden_states_1_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_1_cast_fp16 = concat(axis = var_77, interleave = hidden_states_1_interleave_0, values = (cls_column_1_cast_fp16, x_3_cast_fp16))[name = string("hidden_states_1_cast_fp16")]; |
| int32 var_87 = const()[name = string("op_87"), val = int32(-2)]; |
| int32 var_90 = const()[name = string("op_90"), val = int32(1)]; |
| fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_101_cast_fp16 = mul(x = hidden_states_1_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_101_cast_fp16")]; |
| bool doubled_1_interleave_0 = const()[name = string("doubled_1_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_1_cast_fp16 = concat(axis = var_90, interleave = doubled_1_interleave_0, values = (hidden_states_1_cast_fp16, var_101_cast_fp16))[name = string("doubled_1_cast_fp16")]; |
| tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_1_gamma_0_to_fp16 = const()[name = string("out_1_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135424)))]; |
| fp16 var_111_to_fp16 = const()[name = string("op_111_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_111_to_fp16, gamma = out_1_gamma_0_to_fp16, x = doubled_1_cast_fp16)[name = string("out_1_cast_fp16")]; |
| tensor<int32, [2]> var_122_split_sizes_0 = const()[name = string("op_122_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_122_axis_0 = const()[name = string("op_122_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_122_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_122_cast_fp16_1 = split(axis = var_122_axis_0, split_sizes = var_122_split_sizes_0, x = out_1_cast_fp16)[name = string("op_122_cast_fp16")]; |
| string query_states_1_pad_type_0 = const()[name = string("query_states_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_1_strides_0 = const()[name = string("query_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_1_pad_0 = const()[name = string("query_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_1_dilations_0 = const()[name = string("query_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_1_groups_0 = const()[name = string("query_states_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_82_to_fp16 = const()[name = string("op_82_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139584)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_1_cast_fp16 = conv(dilations = query_states_1_dilations_0, groups = query_states_1_groups_0, pad = query_states_1_pad_0, pad_type = query_states_1_pad_type_0, strides = query_states_1_strides_0, weight = var_82_to_fp16, x = var_122_cast_fp16_0)[name = string("query_states_1_cast_fp16")]; |
| string key_states_1_pad_type_0 = const()[name = string("key_states_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_1_strides_0 = const()[name = string("key_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_1_pad_0 = const()[name = string("key_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_1_dilations_0 = const()[name = string("key_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_1_groups_0 = const()[name = string("key_states_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4333952)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_1_cast_fp16 = conv(dilations = key_states_1_dilations_0, groups = key_states_1_groups_0, pad = key_states_1_pad_0, pad_type = key_states_1_pad_type_0, strides = key_states_1_strides_0, weight = var_83_to_fp16, x = var_122_cast_fp16_0)[name = string("key_states_1_cast_fp16")]; |
| string value_states_1_pad_type_0 = const()[name = string("value_states_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_1_strides_0 = const()[name = string("value_states_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_1_pad_0 = const()[name = string("value_states_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_1_dilations_0 = const()[name = string("value_states_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_1_groups_0 = const()[name = string("value_states_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_84_to_fp16 = const()[name = string("op_84_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4858304)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_1_cast_fp16 = conv(dilations = value_states_1_dilations_0, groups = value_states_1_groups_0, pad = value_states_1_pad_0, pad_type = value_states_1_pad_type_0, strides = value_states_1_strides_0, weight = var_84_to_fp16, x = var_122_cast_fp16_0)[name = string("value_states_1_cast_fp16")]; |
| tensor<int32, [4]> concat_0x = const()[name = string("concat_0x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_5_cast_fp16 = reshape(shape = concat_0x, x = query_states_1_cast_fp16)[name = string("x_5_cast_fp16")]; |
| tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_7_cast_fp16 = reshape(shape = concat_1x, x = key_states_1_cast_fp16)[name = string("x_7_cast_fp16")]; |
| tensor<int32, [4]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_3_cast_fp16 = reshape(shape = concat_2x, x = value_states_1_cast_fp16)[name = string("value_states_3_cast_fp16")]; |
| tensor<fp16, [1, 1, 128, 5]> rope_cos_to_fp16 = const()[name = string("rope_cos_to_fp16"), val = tensor<fp16, [1, 1, 128, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5382656)))]; |
| tensor<fp16, [?, 16, 128, 5]> var_148_cast_fp16 = mul(x = x_5_cast_fp16, y = rope_cos_to_fp16)[name = string("op_148_cast_fp16")]; |
| tensor<int32, [2]> var_149_split_sizes_0 = const()[name = string("op_149_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_149_axis_0 = const()[name = string("op_149_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_149_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_149_cast_fp16_1 = split(axis = var_149_axis_0, split_sizes = var_149_split_sizes_0, x = x_5_cast_fp16)[name = string("op_149_cast_fp16")]; |
| bool var_152_interleave_0 = const()[name = string("op_152_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_152_cast_fp16 = concat(axis = var_87, interleave = var_152_interleave_0, values = (var_149_cast_fp16_1, var_149_cast_fp16_0))[name = string("op_152_cast_fp16")]; |
| tensor<fp16, [1, 1, 128, 5]> rope_sin_to_fp16 = const()[name = string("rope_sin_to_fp16"), val = tensor<fp16, [1, 1, 128, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5384000)))]; |
| tensor<fp16, [?, 16, 128, 5]> var_153_cast_fp16 = mul(x = var_152_cast_fp16, y = rope_sin_to_fp16)[name = string("op_153_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_3_cast_fp16 = add(x = var_148_cast_fp16, y = var_153_cast_fp16)[name = string("query_states_3_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_155_cast_fp16 = mul(x = x_7_cast_fp16, y = rope_cos_to_fp16)[name = string("op_155_cast_fp16")]; |
| tensor<int32, [2]> var_156_split_sizes_0 = const()[name = string("op_156_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_156_axis_0 = const()[name = string("op_156_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_156_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_156_cast_fp16_1 = split(axis = var_156_axis_0, split_sizes = var_156_split_sizes_0, x = x_7_cast_fp16)[name = string("op_156_cast_fp16")]; |
| bool var_159_interleave_0 = const()[name = string("op_159_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_159_cast_fp16 = concat(axis = var_87, interleave = var_159_interleave_0, values = (var_156_cast_fp16_1, var_156_cast_fp16_0))[name = string("op_159_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_160_cast_fp16 = mul(x = var_159_cast_fp16, y = rope_sin_to_fp16)[name = string("op_160_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_3_cast_fp16 = add(x = var_155_cast_fp16, y = var_160_cast_fp16)[name = string("key_states_3_cast_fp16")]; |
| tensor<int32, [2]> var_162_split_sizes_0 = const()[name = string("op_162_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_162_axis_0 = const()[name = string("op_162_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_162_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_162_cast_fp16_1 = split(axis = var_162_axis_0, split_sizes = var_162_split_sizes_0, x = query_states_3_cast_fp16)[name = string("op_162_cast_fp16")]; |
| tensor<int32, [2]> var_164_split_sizes_0 = const()[name = string("op_164_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_164_axis_0 = const()[name = string("op_164_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_164_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_164_cast_fp16_1 = split(axis = var_164_axis_0, split_sizes = var_164_split_sizes_0, x = key_states_3_cast_fp16)[name = string("op_164_cast_fp16")]; |
| tensor<int32, [2]> var_166_split_sizes_0 = const()[name = string("op_166_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_166_axis_0 = const()[name = string("op_166_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_166_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_166_cast_fp16_1 = split(axis = var_166_axis_0, split_sizes = var_166_split_sizes_0, x = value_states_3_cast_fp16)[name = string("op_166_cast_fp16")]; |
| bool attn_weights_1_transpose_x_1 = const()[name = string("attn_weights_1_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_1_transpose_y_1 = const()[name = string("attn_weights_1_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_1_cast_fp16 = matmul(transpose_x = attn_weights_1_transpose_x_1, transpose_y = attn_weights_1_transpose_y_1, x = var_164_cast_fp16_0, y = var_162_cast_fp16_0)[name = string("attn_weights_1_cast_fp16")]; |
| fp16 var_170_to_fp16 = const()[name = string("op_170_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_3_cast_fp16 = mul(x = attn_weights_1_cast_fp16, y = var_170_to_fp16)[name = string("attn_weights_3_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_5_cast_fp16 = softmax(axis = var_87, x = attn_weights_3_cast_fp16)[name = string("attn_weights_5_cast_fp16")]; |
| bool var_173_transpose_x_0 = const()[name = string("op_173_transpose_x_0"), val = bool(false)]; |
| bool var_173_transpose_y_0 = const()[name = string("op_173_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_173_cast_fp16 = matmul(transpose_x = var_173_transpose_x_0, transpose_y = var_173_transpose_y_0, x = var_166_cast_fp16_0, y = attn_weights_5_cast_fp16)[name = string("op_173_cast_fp16")]; |
| bool attn_weights_7_transpose_x_1 = const()[name = string("attn_weights_7_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_7_transpose_y_1 = const()[name = string("attn_weights_7_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_7_cast_fp16 = matmul(transpose_x = attn_weights_7_transpose_x_1, transpose_y = attn_weights_7_transpose_y_1, x = var_164_cast_fp16_1, y = var_162_cast_fp16_1)[name = string("attn_weights_7_cast_fp16")]; |
| fp16 var_176_to_fp16 = const()[name = string("op_176_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_9_cast_fp16 = mul(x = attn_weights_7_cast_fp16, y = var_176_to_fp16)[name = string("attn_weights_9_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_11_cast_fp16 = softmax(axis = var_87, x = attn_weights_9_cast_fp16)[name = string("attn_weights_11_cast_fp16")]; |
| bool attn_out_1_transpose_x_0 = const()[name = string("attn_out_1_transpose_x_0"), val = bool(false)]; |
| bool attn_out_1_transpose_y_0 = const()[name = string("attn_out_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_1_cast_fp16 = matmul(transpose_x = attn_out_1_transpose_x_0, transpose_y = attn_out_1_transpose_y_0, x = var_166_cast_fp16_1, y = attn_weights_11_cast_fp16)[name = string("attn_out_1_cast_fp16")]; |
| bool attn_output_1_interleave_0 = const()[name = string("attn_output_1_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_1_cast_fp16 = concat(axis = var_90, interleave = attn_output_1_interleave_0, values = (var_173_cast_fp16, attn_out_1_cast_fp16))[name = string("attn_output_1_cast_fp16")]; |
| tensor<int32, [4]> concat_3x = const()[name = string("concat_3x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_9_cast_fp16 = reshape(shape = concat_3x, x = attn_output_1_cast_fp16)[name = string("x_9_cast_fp16")]; |
| string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_89_to_fp16 = const()[name = string("op_89_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5385344)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_5_cast_fp16 = conv(dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = var_89_to_fp16, x = x_9_cast_fp16)[name = string("hidden_states_5_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_7_cast_fp16 = add(x = hidden_states_1_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("hidden_states_7_cast_fp16")]; |
| fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_191_cast_fp16 = mul(x = hidden_states_7_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_191_cast_fp16")]; |
| bool doubled_5_interleave_0 = const()[name = string("doubled_5_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_5_cast_fp16 = concat(axis = var_90, interleave = doubled_5_interleave_0, values = (hidden_states_7_cast_fp16, var_191_cast_fp16))[name = string("doubled_5_cast_fp16")]; |
| tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_3_gamma_0_to_fp16 = const()[name = string("out_3_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9579712)))]; |
| fp16 var_201_to_fp16 = const()[name = string("op_201_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_201_to_fp16, gamma = out_3_gamma_0_to_fp16, x = doubled_5_cast_fp16)[name = string("out_3_cast_fp16")]; |
| tensor<int32, [2]> var_212_split_sizes_0 = const()[name = string("op_212_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_212_axis_0 = const()[name = string("op_212_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_212_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_212_cast_fp16_1 = split(axis = var_212_axis_0, split_sizes = var_212_split_sizes_0, x = out_3_cast_fp16)[name = string("op_212_cast_fp16")]; |
| string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_79_to_fp16 = const()[name = string("op_79_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9583872)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = var_79_to_fp16, x = var_212_cast_fp16_0)[name = string("input_1_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_220_cast_fp16 = silu(x = input_1_cast_fp16)[name = string("op_220_cast_fp16")]; |
| string var_225_pad_type_0 = const()[name = string("op_225_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_225_strides_0 = const()[name = string("op_225_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_225_pad_0 = const()[name = string("op_225_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_225_dilations_0 = const()[name = string("op_225_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_225_groups_0 = const()[name = string("op_225_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_80_to_fp16 = const()[name = string("op_80_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17972544)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_225_cast_fp16 = conv(dilations = var_225_dilations_0, groups = var_225_groups_0, pad = var_225_pad_0, pad_type = var_225_pad_type_0, strides = var_225_strides_0, weight = var_80_to_fp16, x = var_212_cast_fp16_0)[name = string("op_225_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_13_cast_fp16 = mul(x = var_220_cast_fp16, y = var_225_cast_fp16)[name = string("x_13_cast_fp16")]; |
| string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_81_to_fp16 = const()[name = string("op_81_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26361216)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_9_cast_fp16 = conv(dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = var_81_to_fp16, x = x_13_cast_fp16)[name = string("hidden_states_9_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_11_cast_fp16 = add(x = hidden_states_7_cast_fp16, y = hidden_states_9_cast_fp16)[name = string("hidden_states_11_cast_fp16")]; |
| int32 var_241 = const()[name = string("op_241"), val = int32(-2)]; |
| int32 var_244 = const()[name = string("op_244"), val = int32(1)]; |
| fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_255_cast_fp16 = mul(x = hidden_states_11_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_255_cast_fp16")]; |
| bool doubled_9_interleave_0 = const()[name = string("doubled_9_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_9_cast_fp16 = concat(axis = var_244, interleave = doubled_9_interleave_0, values = (hidden_states_11_cast_fp16, var_255_cast_fp16))[name = string("doubled_9_cast_fp16")]; |
| tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_5_gamma_0_to_fp16 = const()[name = string("out_5_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34749888)))]; |
| fp16 var_265_to_fp16 = const()[name = string("op_265_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_265_to_fp16, gamma = out_5_gamma_0_to_fp16, x = doubled_9_cast_fp16)[name = string("out_5_cast_fp16")]; |
| tensor<int32, [2]> var_276_split_sizes_0 = const()[name = string("op_276_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_276_axis_0 = const()[name = string("op_276_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_276_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_276_cast_fp16_1 = split(axis = var_276_axis_0, split_sizes = var_276_split_sizes_0, x = out_5_cast_fp16)[name = string("op_276_cast_fp16")]; |
| string query_states_5_pad_type_0 = const()[name = string("query_states_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_5_strides_0 = const()[name = string("query_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_5_pad_0 = const()[name = string("query_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_5_dilations_0 = const()[name = string("query_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_5_groups_0 = const()[name = string("query_states_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_236_to_fp16 = const()[name = string("op_236_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34754048)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_5_cast_fp16 = conv(dilations = query_states_5_dilations_0, groups = query_states_5_groups_0, pad = query_states_5_pad_0, pad_type = query_states_5_pad_type_0, strides = query_states_5_strides_0, weight = var_236_to_fp16, x = var_276_cast_fp16_0)[name = string("query_states_5_cast_fp16")]; |
| string key_states_5_pad_type_0 = const()[name = string("key_states_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_5_strides_0 = const()[name = string("key_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_5_pad_0 = const()[name = string("key_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_5_dilations_0 = const()[name = string("key_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_5_groups_0 = const()[name = string("key_states_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_237_to_fp16 = const()[name = string("op_237_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38948416)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_5_cast_fp16 = conv(dilations = key_states_5_dilations_0, groups = key_states_5_groups_0, pad = key_states_5_pad_0, pad_type = key_states_5_pad_type_0, strides = key_states_5_strides_0, weight = var_237_to_fp16, x = var_276_cast_fp16_0)[name = string("key_states_5_cast_fp16")]; |
| string value_states_5_pad_type_0 = const()[name = string("value_states_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_5_strides_0 = const()[name = string("value_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_5_pad_0 = const()[name = string("value_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_5_dilations_0 = const()[name = string("value_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_5_groups_0 = const()[name = string("value_states_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_238_to_fp16 = const()[name = string("op_238_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39472768)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_5_cast_fp16 = conv(dilations = value_states_5_dilations_0, groups = value_states_5_groups_0, pad = value_states_5_pad_0, pad_type = value_states_5_pad_type_0, strides = value_states_5_strides_0, weight = var_238_to_fp16, x = var_276_cast_fp16_0)[name = string("value_states_5_cast_fp16")]; |
| tensor<int32, [4]> concat_4x = const()[name = string("concat_4x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_15_cast_fp16 = reshape(shape = concat_4x, x = query_states_5_cast_fp16)[name = string("x_15_cast_fp16")]; |
| tensor<int32, [4]> concat_5x = const()[name = string("concat_5x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_17_cast_fp16 = reshape(shape = concat_5x, x = key_states_5_cast_fp16)[name = string("x_17_cast_fp16")]; |
| tensor<int32, [4]> concat_6x = const()[name = string("concat_6x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_7_cast_fp16 = reshape(shape = concat_6x, x = value_states_5_cast_fp16)[name = string("value_states_7_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_302_cast_fp16 = mul(x = x_15_cast_fp16, y = rope_cos_to_fp16)[name = string("op_302_cast_fp16")]; |
| tensor<int32, [2]> var_303_split_sizes_0 = const()[name = string("op_303_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_303_axis_0 = const()[name = string("op_303_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_303_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_303_cast_fp16_1 = split(axis = var_303_axis_0, split_sizes = var_303_split_sizes_0, x = x_15_cast_fp16)[name = string("op_303_cast_fp16")]; |
| bool var_306_interleave_0 = const()[name = string("op_306_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_306_cast_fp16 = concat(axis = var_241, interleave = var_306_interleave_0, values = (var_303_cast_fp16_1, var_303_cast_fp16_0))[name = string("op_306_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_307_cast_fp16 = mul(x = var_306_cast_fp16, y = rope_sin_to_fp16)[name = string("op_307_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_7_cast_fp16 = add(x = var_302_cast_fp16, y = var_307_cast_fp16)[name = string("query_states_7_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_309_cast_fp16 = mul(x = x_17_cast_fp16, y = rope_cos_to_fp16)[name = string("op_309_cast_fp16")]; |
| tensor<int32, [2]> var_310_split_sizes_0 = const()[name = string("op_310_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_310_axis_0 = const()[name = string("op_310_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_310_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_310_cast_fp16_1 = split(axis = var_310_axis_0, split_sizes = var_310_split_sizes_0, x = x_17_cast_fp16)[name = string("op_310_cast_fp16")]; |
| bool var_313_interleave_0 = const()[name = string("op_313_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_313_cast_fp16 = concat(axis = var_241, interleave = var_313_interleave_0, values = (var_310_cast_fp16_1, var_310_cast_fp16_0))[name = string("op_313_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_314_cast_fp16 = mul(x = var_313_cast_fp16, y = rope_sin_to_fp16)[name = string("op_314_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_7_cast_fp16 = add(x = var_309_cast_fp16, y = var_314_cast_fp16)[name = string("key_states_7_cast_fp16")]; |
| tensor<int32, [2]> var_316_split_sizes_0 = const()[name = string("op_316_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_316_axis_0 = const()[name = string("op_316_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_316_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_316_cast_fp16_1 = split(axis = var_316_axis_0, split_sizes = var_316_split_sizes_0, x = query_states_7_cast_fp16)[name = string("op_316_cast_fp16")]; |
| tensor<int32, [2]> var_318_split_sizes_0 = const()[name = string("op_318_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_318_axis_0 = const()[name = string("op_318_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_318_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_318_cast_fp16_1 = split(axis = var_318_axis_0, split_sizes = var_318_split_sizes_0, x = key_states_7_cast_fp16)[name = string("op_318_cast_fp16")]; |
| tensor<int32, [2]> var_320_split_sizes_0 = const()[name = string("op_320_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_320_axis_0 = const()[name = string("op_320_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_320_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_320_cast_fp16_1 = split(axis = var_320_axis_0, split_sizes = var_320_split_sizes_0, x = value_states_7_cast_fp16)[name = string("op_320_cast_fp16")]; |
| bool attn_weights_13_transpose_x_1 = const()[name = string("attn_weights_13_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_13_transpose_y_1 = const()[name = string("attn_weights_13_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_13_cast_fp16 = matmul(transpose_x = attn_weights_13_transpose_x_1, transpose_y = attn_weights_13_transpose_y_1, x = var_318_cast_fp16_0, y = var_316_cast_fp16_0)[name = string("attn_weights_13_cast_fp16")]; |
| fp16 var_324_to_fp16 = const()[name = string("op_324_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_15_cast_fp16 = mul(x = attn_weights_13_cast_fp16, y = var_324_to_fp16)[name = string("attn_weights_15_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_17_cast_fp16 = softmax(axis = var_241, x = attn_weights_15_cast_fp16)[name = string("attn_weights_17_cast_fp16")]; |
| bool var_327_transpose_x_0 = const()[name = string("op_327_transpose_x_0"), val = bool(false)]; |
| bool var_327_transpose_y_0 = const()[name = string("op_327_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_327_cast_fp16 = matmul(transpose_x = var_327_transpose_x_0, transpose_y = var_327_transpose_y_0, x = var_320_cast_fp16_0, y = attn_weights_17_cast_fp16)[name = string("op_327_cast_fp16")]; |
| bool attn_weights_19_transpose_x_1 = const()[name = string("attn_weights_19_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_19_transpose_y_1 = const()[name = string("attn_weights_19_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_19_cast_fp16 = matmul(transpose_x = attn_weights_19_transpose_x_1, transpose_y = attn_weights_19_transpose_y_1, x = var_318_cast_fp16_1, y = var_316_cast_fp16_1)[name = string("attn_weights_19_cast_fp16")]; |
| fp16 var_330_to_fp16 = const()[name = string("op_330_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_21_cast_fp16 = mul(x = attn_weights_19_cast_fp16, y = var_330_to_fp16)[name = string("attn_weights_21_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_23_cast_fp16 = softmax(axis = var_241, x = attn_weights_21_cast_fp16)[name = string("attn_weights_23_cast_fp16")]; |
| bool attn_out_3_transpose_x_0 = const()[name = string("attn_out_3_transpose_x_0"), val = bool(false)]; |
| bool attn_out_3_transpose_y_0 = const()[name = string("attn_out_3_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_3_cast_fp16 = matmul(transpose_x = attn_out_3_transpose_x_0, transpose_y = attn_out_3_transpose_y_0, x = var_320_cast_fp16_1, y = attn_weights_23_cast_fp16)[name = string("attn_out_3_cast_fp16")]; |
| bool attn_output_3_interleave_0 = const()[name = string("attn_output_3_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_3_cast_fp16 = concat(axis = var_244, interleave = attn_output_3_interleave_0, values = (var_327_cast_fp16, attn_out_3_cast_fp16))[name = string("attn_output_3_cast_fp16")]; |
| tensor<int32, [4]> concat_7x = const()[name = string("concat_7x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_19_cast_fp16 = reshape(shape = concat_7x, x = attn_output_3_cast_fp16)[name = string("x_19_cast_fp16")]; |
| string hidden_states_15_pad_type_0 = const()[name = string("hidden_states_15_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_15_strides_0 = const()[name = string("hidden_states_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_15_pad_0 = const()[name = string("hidden_states_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_15_dilations_0 = const()[name = string("hidden_states_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_15_groups_0 = const()[name = string("hidden_states_15_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_243_to_fp16 = const()[name = string("op_243_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39997120)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_15_cast_fp16 = conv(dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = var_243_to_fp16, x = x_19_cast_fp16)[name = string("hidden_states_15_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_17_cast_fp16 = add(x = hidden_states_11_cast_fp16, y = hidden_states_15_cast_fp16)[name = string("hidden_states_17_cast_fp16")]; |
| fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_345_cast_fp16 = mul(x = hidden_states_17_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_345_cast_fp16")]; |
| bool doubled_13_interleave_0 = const()[name = string("doubled_13_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_13_cast_fp16 = concat(axis = var_244, interleave = doubled_13_interleave_0, values = (hidden_states_17_cast_fp16, var_345_cast_fp16))[name = string("doubled_13_cast_fp16")]; |
| tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_7_gamma_0_to_fp16 = const()[name = string("out_7_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44191488)))]; |
| fp16 var_355_to_fp16 = const()[name = string("op_355_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_355_to_fp16, gamma = out_7_gamma_0_to_fp16, x = doubled_13_cast_fp16)[name = string("out_7_cast_fp16")]; |
| tensor<int32, [2]> var_366_split_sizes_0 = const()[name = string("op_366_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_366_axis_0 = const()[name = string("op_366_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_366_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_366_cast_fp16_1 = split(axis = var_366_axis_0, split_sizes = var_366_split_sizes_0, x = out_7_cast_fp16)[name = string("op_366_cast_fp16")]; |
| string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_233_to_fp16 = const()[name = string("op_233_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44195648)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = var_233_to_fp16, x = var_366_cast_fp16_0)[name = string("input_3_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_374_cast_fp16 = silu(x = input_3_cast_fp16)[name = string("op_374_cast_fp16")]; |
| string var_379_pad_type_0 = const()[name = string("op_379_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_379_strides_0 = const()[name = string("op_379_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_379_pad_0 = const()[name = string("op_379_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_379_dilations_0 = const()[name = string("op_379_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_379_groups_0 = const()[name = string("op_379_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_234_to_fp16 = const()[name = string("op_234_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52584320)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_379_cast_fp16 = conv(dilations = var_379_dilations_0, groups = var_379_groups_0, pad = var_379_pad_0, pad_type = var_379_pad_type_0, strides = var_379_strides_0, weight = var_234_to_fp16, x = var_366_cast_fp16_0)[name = string("op_379_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_23_cast_fp16 = mul(x = var_374_cast_fp16, y = var_379_cast_fp16)[name = string("x_23_cast_fp16")]; |
| string hidden_states_19_pad_type_0 = const()[name = string("hidden_states_19_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_19_strides_0 = const()[name = string("hidden_states_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_19_pad_0 = const()[name = string("hidden_states_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_19_dilations_0 = const()[name = string("hidden_states_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_19_groups_0 = const()[name = string("hidden_states_19_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_235_to_fp16 = const()[name = string("op_235_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60972992)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_19_cast_fp16 = conv(dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = var_235_to_fp16, x = x_23_cast_fp16)[name = string("hidden_states_19_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_21_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = hidden_states_19_cast_fp16)[name = string("hidden_states_21_cast_fp16")]; |
| int32 var_395 = const()[name = string("op_395"), val = int32(-2)]; |
| int32 var_398 = const()[name = string("op_398"), val = int32(1)]; |
| fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_409_cast_fp16 = mul(x = hidden_states_21_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_409_cast_fp16")]; |
| bool doubled_17_interleave_0 = const()[name = string("doubled_17_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_17_cast_fp16 = concat(axis = var_398, interleave = doubled_17_interleave_0, values = (hidden_states_21_cast_fp16, var_409_cast_fp16))[name = string("doubled_17_cast_fp16")]; |
| tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_9_gamma_0_to_fp16 = const()[name = string("out_9_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69361664)))]; |
| fp16 var_419_to_fp16 = const()[name = string("op_419_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_419_to_fp16, gamma = out_9_gamma_0_to_fp16, x = doubled_17_cast_fp16)[name = string("out_9_cast_fp16")]; |
| tensor<int32, [2]> var_430_split_sizes_0 = const()[name = string("op_430_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_430_axis_0 = const()[name = string("op_430_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_430_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_430_cast_fp16_1 = split(axis = var_430_axis_0, split_sizes = var_430_split_sizes_0, x = out_9_cast_fp16)[name = string("op_430_cast_fp16")]; |
| string query_states_9_pad_type_0 = const()[name = string("query_states_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_9_strides_0 = const()[name = string("query_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_9_pad_0 = const()[name = string("query_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_9_dilations_0 = const()[name = string("query_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_9_groups_0 = const()[name = string("query_states_9_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_390_to_fp16 = const()[name = string("op_390_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69365824)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_9_cast_fp16 = conv(dilations = query_states_9_dilations_0, groups = query_states_9_groups_0, pad = query_states_9_pad_0, pad_type = query_states_9_pad_type_0, strides = query_states_9_strides_0, weight = var_390_to_fp16, x = var_430_cast_fp16_0)[name = string("query_states_9_cast_fp16")]; |
| string key_states_9_pad_type_0 = const()[name = string("key_states_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_9_strides_0 = const()[name = string("key_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_9_pad_0 = const()[name = string("key_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_9_dilations_0 = const()[name = string("key_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_9_groups_0 = const()[name = string("key_states_9_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_391_to_fp16 = const()[name = string("op_391_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73560192)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_9_cast_fp16 = conv(dilations = key_states_9_dilations_0, groups = key_states_9_groups_0, pad = key_states_9_pad_0, pad_type = key_states_9_pad_type_0, strides = key_states_9_strides_0, weight = var_391_to_fp16, x = var_430_cast_fp16_0)[name = string("key_states_9_cast_fp16")]; |
| string value_states_9_pad_type_0 = const()[name = string("value_states_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_9_strides_0 = const()[name = string("value_states_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_9_pad_0 = const()[name = string("value_states_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_9_dilations_0 = const()[name = string("value_states_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_9_groups_0 = const()[name = string("value_states_9_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_392_to_fp16 = const()[name = string("op_392_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74084544)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_9_cast_fp16 = conv(dilations = value_states_9_dilations_0, groups = value_states_9_groups_0, pad = value_states_9_pad_0, pad_type = value_states_9_pad_type_0, strides = value_states_9_strides_0, weight = var_392_to_fp16, x = var_430_cast_fp16_0)[name = string("value_states_9_cast_fp16")]; |
| tensor<int32, [4]> concat_8x = const()[name = string("concat_8x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_25_cast_fp16 = reshape(shape = concat_8x, x = query_states_9_cast_fp16)[name = string("x_25_cast_fp16")]; |
| tensor<int32, [4]> concat_9x = const()[name = string("concat_9x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_27_cast_fp16 = reshape(shape = concat_9x, x = key_states_9_cast_fp16)[name = string("x_27_cast_fp16")]; |
| tensor<int32, [4]> concat_10x = const()[name = string("concat_10x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_11_cast_fp16 = reshape(shape = concat_10x, x = value_states_9_cast_fp16)[name = string("value_states_11_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_456_cast_fp16 = mul(x = x_25_cast_fp16, y = rope_cos_to_fp16)[name = string("op_456_cast_fp16")]; |
| tensor<int32, [2]> var_457_split_sizes_0 = const()[name = string("op_457_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_457_axis_0 = const()[name = string("op_457_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_457_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_457_cast_fp16_1 = split(axis = var_457_axis_0, split_sizes = var_457_split_sizes_0, x = x_25_cast_fp16)[name = string("op_457_cast_fp16")]; |
| bool var_460_interleave_0 = const()[name = string("op_460_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_460_cast_fp16 = concat(axis = var_395, interleave = var_460_interleave_0, values = (var_457_cast_fp16_1, var_457_cast_fp16_0))[name = string("op_460_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_461_cast_fp16 = mul(x = var_460_cast_fp16, y = rope_sin_to_fp16)[name = string("op_461_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_11_cast_fp16 = add(x = var_456_cast_fp16, y = var_461_cast_fp16)[name = string("query_states_11_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_463_cast_fp16 = mul(x = x_27_cast_fp16, y = rope_cos_to_fp16)[name = string("op_463_cast_fp16")]; |
| tensor<int32, [2]> var_464_split_sizes_0 = const()[name = string("op_464_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_464_axis_0 = const()[name = string("op_464_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_464_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_464_cast_fp16_1 = split(axis = var_464_axis_0, split_sizes = var_464_split_sizes_0, x = x_27_cast_fp16)[name = string("op_464_cast_fp16")]; |
| bool var_467_interleave_0 = const()[name = string("op_467_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_467_cast_fp16 = concat(axis = var_395, interleave = var_467_interleave_0, values = (var_464_cast_fp16_1, var_464_cast_fp16_0))[name = string("op_467_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_468_cast_fp16 = mul(x = var_467_cast_fp16, y = rope_sin_to_fp16)[name = string("op_468_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_11_cast_fp16 = add(x = var_463_cast_fp16, y = var_468_cast_fp16)[name = string("key_states_11_cast_fp16")]; |
| tensor<int32, [2]> var_470_split_sizes_0 = const()[name = string("op_470_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_470_axis_0 = const()[name = string("op_470_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_470_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_470_cast_fp16_1 = split(axis = var_470_axis_0, split_sizes = var_470_split_sizes_0, x = query_states_11_cast_fp16)[name = string("op_470_cast_fp16")]; |
| tensor<int32, [2]> var_472_split_sizes_0 = const()[name = string("op_472_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_472_axis_0 = const()[name = string("op_472_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_472_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_472_cast_fp16_1 = split(axis = var_472_axis_0, split_sizes = var_472_split_sizes_0, x = key_states_11_cast_fp16)[name = string("op_472_cast_fp16")]; |
| tensor<int32, [2]> var_474_split_sizes_0 = const()[name = string("op_474_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_474_axis_0 = const()[name = string("op_474_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_474_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_474_cast_fp16_1 = split(axis = var_474_axis_0, split_sizes = var_474_split_sizes_0, x = value_states_11_cast_fp16)[name = string("op_474_cast_fp16")]; |
| bool attn_weights_25_transpose_x_1 = const()[name = string("attn_weights_25_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_25_transpose_y_1 = const()[name = string("attn_weights_25_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_25_cast_fp16 = matmul(transpose_x = attn_weights_25_transpose_x_1, transpose_y = attn_weights_25_transpose_y_1, x = var_472_cast_fp16_0, y = var_470_cast_fp16_0)[name = string("attn_weights_25_cast_fp16")]; |
| fp16 var_478_to_fp16 = const()[name = string("op_478_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_27_cast_fp16 = mul(x = attn_weights_25_cast_fp16, y = var_478_to_fp16)[name = string("attn_weights_27_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_29_cast_fp16 = softmax(axis = var_395, x = attn_weights_27_cast_fp16)[name = string("attn_weights_29_cast_fp16")]; |
| bool var_481_transpose_x_0 = const()[name = string("op_481_transpose_x_0"), val = bool(false)]; |
| bool var_481_transpose_y_0 = const()[name = string("op_481_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_481_cast_fp16 = matmul(transpose_x = var_481_transpose_x_0, transpose_y = var_481_transpose_y_0, x = var_474_cast_fp16_0, y = attn_weights_29_cast_fp16)[name = string("op_481_cast_fp16")]; |
| bool attn_weights_31_transpose_x_1 = const()[name = string("attn_weights_31_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_31_transpose_y_1 = const()[name = string("attn_weights_31_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_31_cast_fp16 = matmul(transpose_x = attn_weights_31_transpose_x_1, transpose_y = attn_weights_31_transpose_y_1, x = var_472_cast_fp16_1, y = var_470_cast_fp16_1)[name = string("attn_weights_31_cast_fp16")]; |
| fp16 var_484_to_fp16 = const()[name = string("op_484_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_33_cast_fp16 = mul(x = attn_weights_31_cast_fp16, y = var_484_to_fp16)[name = string("attn_weights_33_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_35_cast_fp16 = softmax(axis = var_395, x = attn_weights_33_cast_fp16)[name = string("attn_weights_35_cast_fp16")]; |
| bool attn_out_5_transpose_x_0 = const()[name = string("attn_out_5_transpose_x_0"), val = bool(false)]; |
| bool attn_out_5_transpose_y_0 = const()[name = string("attn_out_5_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_5_cast_fp16 = matmul(transpose_x = attn_out_5_transpose_x_0, transpose_y = attn_out_5_transpose_y_0, x = var_474_cast_fp16_1, y = attn_weights_35_cast_fp16)[name = string("attn_out_5_cast_fp16")]; |
| bool attn_output_5_interleave_0 = const()[name = string("attn_output_5_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_5_cast_fp16 = concat(axis = var_398, interleave = attn_output_5_interleave_0, values = (var_481_cast_fp16, attn_out_5_cast_fp16))[name = string("attn_output_5_cast_fp16")]; |
| tensor<int32, [4]> concat_11x = const()[name = string("concat_11x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_29_cast_fp16 = reshape(shape = concat_11x, x = attn_output_5_cast_fp16)[name = string("x_29_cast_fp16")]; |
| string hidden_states_25_pad_type_0 = const()[name = string("hidden_states_25_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_25_strides_0 = const()[name = string("hidden_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_25_pad_0 = const()[name = string("hidden_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_25_dilations_0 = const()[name = string("hidden_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_25_groups_0 = const()[name = string("hidden_states_25_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_397_to_fp16 = const()[name = string("op_397_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74608896)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_25_cast_fp16 = conv(dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = var_397_to_fp16, x = x_29_cast_fp16)[name = string("hidden_states_25_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_27_cast_fp16 = add(x = hidden_states_21_cast_fp16, y = hidden_states_25_cast_fp16)[name = string("hidden_states_27_cast_fp16")]; |
| fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_499_cast_fp16 = mul(x = hidden_states_27_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_499_cast_fp16")]; |
| bool doubled_21_interleave_0 = const()[name = string("doubled_21_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_21_cast_fp16 = concat(axis = var_398, interleave = doubled_21_interleave_0, values = (hidden_states_27_cast_fp16, var_499_cast_fp16))[name = string("doubled_21_cast_fp16")]; |
| tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_11_gamma_0_to_fp16 = const()[name = string("out_11_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78803264)))]; |
| fp16 var_509_to_fp16 = const()[name = string("op_509_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_509_to_fp16, gamma = out_11_gamma_0_to_fp16, x = doubled_21_cast_fp16)[name = string("out_11_cast_fp16")]; |
| tensor<int32, [2]> var_520_split_sizes_0 = const()[name = string("op_520_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_520_axis_0 = const()[name = string("op_520_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_520_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_520_cast_fp16_1 = split(axis = var_520_axis_0, split_sizes = var_520_split_sizes_0, x = out_11_cast_fp16)[name = string("op_520_cast_fp16")]; |
| string input_5_pad_type_0 = const()[name = string("input_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_5_strides_0 = const()[name = string("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_5_dilations_0 = const()[name = string("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_5_groups_0 = const()[name = string("input_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_387_to_fp16 = const()[name = string("op_387_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78807424)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = var_387_to_fp16, x = var_520_cast_fp16_0)[name = string("input_5_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_528_cast_fp16 = silu(x = input_5_cast_fp16)[name = string("op_528_cast_fp16")]; |
| string var_533_pad_type_0 = const()[name = string("op_533_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_533_strides_0 = const()[name = string("op_533_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_533_pad_0 = const()[name = string("op_533_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_533_dilations_0 = const()[name = string("op_533_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_533_groups_0 = const()[name = string("op_533_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_388_to_fp16 = const()[name = string("op_388_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87196096)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_533_cast_fp16 = conv(dilations = var_533_dilations_0, groups = var_533_groups_0, pad = var_533_pad_0, pad_type = var_533_pad_type_0, strides = var_533_strides_0, weight = var_388_to_fp16, x = var_520_cast_fp16_0)[name = string("op_533_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_33_cast_fp16 = mul(x = var_528_cast_fp16, y = var_533_cast_fp16)[name = string("x_33_cast_fp16")]; |
| string hidden_states_29_pad_type_0 = const()[name = string("hidden_states_29_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_29_strides_0 = const()[name = string("hidden_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_29_pad_0 = const()[name = string("hidden_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_29_dilations_0 = const()[name = string("hidden_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_29_groups_0 = const()[name = string("hidden_states_29_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_389_to_fp16 = const()[name = string("op_389_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95584768)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_29_cast_fp16 = conv(dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = var_389_to_fp16, x = x_33_cast_fp16)[name = string("hidden_states_29_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_31_cast_fp16 = add(x = hidden_states_27_cast_fp16, y = hidden_states_29_cast_fp16)[name = string("hidden_states_31_cast_fp16")]; |
| int32 var_549 = const()[name = string("op_549"), val = int32(-2)]; |
| int32 var_552 = const()[name = string("op_552"), val = int32(1)]; |
| fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_563_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_563_cast_fp16")]; |
| bool doubled_25_interleave_0 = const()[name = string("doubled_25_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_25_cast_fp16 = concat(axis = var_552, interleave = doubled_25_interleave_0, values = (hidden_states_31_cast_fp16, var_563_cast_fp16))[name = string("doubled_25_cast_fp16")]; |
| tensor<int32, [1]> out_13_axes_0 = const()[name = string("out_13_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_13_gamma_0_to_fp16 = const()[name = string("out_13_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103973440)))]; |
| fp16 var_573_to_fp16 = const()[name = string("op_573_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_573_to_fp16, gamma = out_13_gamma_0_to_fp16, x = doubled_25_cast_fp16)[name = string("out_13_cast_fp16")]; |
| tensor<int32, [2]> var_584_split_sizes_0 = const()[name = string("op_584_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_584_axis_0 = const()[name = string("op_584_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_584_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_584_cast_fp16_1 = split(axis = var_584_axis_0, split_sizes = var_584_split_sizes_0, x = out_13_cast_fp16)[name = string("op_584_cast_fp16")]; |
| string query_states_13_pad_type_0 = const()[name = string("query_states_13_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_13_strides_0 = const()[name = string("query_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_13_pad_0 = const()[name = string("query_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_13_dilations_0 = const()[name = string("query_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_13_groups_0 = const()[name = string("query_states_13_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_544_to_fp16 = const()[name = string("op_544_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103977600)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_13_cast_fp16 = conv(dilations = query_states_13_dilations_0, groups = query_states_13_groups_0, pad = query_states_13_pad_0, pad_type = query_states_13_pad_type_0, strides = query_states_13_strides_0, weight = var_544_to_fp16, x = var_584_cast_fp16_0)[name = string("query_states_13_cast_fp16")]; |
| string key_states_13_pad_type_0 = const()[name = string("key_states_13_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_13_strides_0 = const()[name = string("key_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_13_pad_0 = const()[name = string("key_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_13_dilations_0 = const()[name = string("key_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_13_groups_0 = const()[name = string("key_states_13_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_545_to_fp16 = const()[name = string("op_545_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108171968)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_13_cast_fp16 = conv(dilations = key_states_13_dilations_0, groups = key_states_13_groups_0, pad = key_states_13_pad_0, pad_type = key_states_13_pad_type_0, strides = key_states_13_strides_0, weight = var_545_to_fp16, x = var_584_cast_fp16_0)[name = string("key_states_13_cast_fp16")]; |
| string value_states_13_pad_type_0 = const()[name = string("value_states_13_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_13_strides_0 = const()[name = string("value_states_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_13_pad_0 = const()[name = string("value_states_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_13_dilations_0 = const()[name = string("value_states_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_13_groups_0 = const()[name = string("value_states_13_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_546_to_fp16 = const()[name = string("op_546_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108696320)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_13_cast_fp16 = conv(dilations = value_states_13_dilations_0, groups = value_states_13_groups_0, pad = value_states_13_pad_0, pad_type = value_states_13_pad_type_0, strides = value_states_13_strides_0, weight = var_546_to_fp16, x = var_584_cast_fp16_0)[name = string("value_states_13_cast_fp16")]; |
| tensor<int32, [4]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_35_cast_fp16 = reshape(shape = concat_12x, x = query_states_13_cast_fp16)[name = string("x_35_cast_fp16")]; |
| tensor<int32, [4]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_37_cast_fp16 = reshape(shape = concat_13x, x = key_states_13_cast_fp16)[name = string("x_37_cast_fp16")]; |
| tensor<int32, [4]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_15_cast_fp16 = reshape(shape = concat_14x, x = value_states_13_cast_fp16)[name = string("value_states_15_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_610_cast_fp16 = mul(x = x_35_cast_fp16, y = rope_cos_to_fp16)[name = string("op_610_cast_fp16")]; |
| tensor<int32, [2]> var_611_split_sizes_0 = const()[name = string("op_611_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_611_axis_0 = const()[name = string("op_611_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_611_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_611_cast_fp16_1 = split(axis = var_611_axis_0, split_sizes = var_611_split_sizes_0, x = x_35_cast_fp16)[name = string("op_611_cast_fp16")]; |
| bool var_614_interleave_0 = const()[name = string("op_614_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_614_cast_fp16 = concat(axis = var_549, interleave = var_614_interleave_0, values = (var_611_cast_fp16_1, var_611_cast_fp16_0))[name = string("op_614_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_615_cast_fp16 = mul(x = var_614_cast_fp16, y = rope_sin_to_fp16)[name = string("op_615_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_15_cast_fp16 = add(x = var_610_cast_fp16, y = var_615_cast_fp16)[name = string("query_states_15_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_617_cast_fp16 = mul(x = x_37_cast_fp16, y = rope_cos_to_fp16)[name = string("op_617_cast_fp16")]; |
| tensor<int32, [2]> var_618_split_sizes_0 = const()[name = string("op_618_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_618_axis_0 = const()[name = string("op_618_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_618_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_618_cast_fp16_1 = split(axis = var_618_axis_0, split_sizes = var_618_split_sizes_0, x = x_37_cast_fp16)[name = string("op_618_cast_fp16")]; |
| bool var_621_interleave_0 = const()[name = string("op_621_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_621_cast_fp16 = concat(axis = var_549, interleave = var_621_interleave_0, values = (var_618_cast_fp16_1, var_618_cast_fp16_0))[name = string("op_621_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_622_cast_fp16 = mul(x = var_621_cast_fp16, y = rope_sin_to_fp16)[name = string("op_622_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_15_cast_fp16 = add(x = var_617_cast_fp16, y = var_622_cast_fp16)[name = string("key_states_15_cast_fp16")]; |
| tensor<int32, [2]> var_624_split_sizes_0 = const()[name = string("op_624_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_624_axis_0 = const()[name = string("op_624_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_624_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_624_cast_fp16_1 = split(axis = var_624_axis_0, split_sizes = var_624_split_sizes_0, x = query_states_15_cast_fp16)[name = string("op_624_cast_fp16")]; |
| tensor<int32, [2]> var_626_split_sizes_0 = const()[name = string("op_626_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_626_axis_0 = const()[name = string("op_626_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_626_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_626_cast_fp16_1 = split(axis = var_626_axis_0, split_sizes = var_626_split_sizes_0, x = key_states_15_cast_fp16)[name = string("op_626_cast_fp16")]; |
| tensor<int32, [2]> var_628_split_sizes_0 = const()[name = string("op_628_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_628_axis_0 = const()[name = string("op_628_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_628_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_628_cast_fp16_1 = split(axis = var_628_axis_0, split_sizes = var_628_split_sizes_0, x = value_states_15_cast_fp16)[name = string("op_628_cast_fp16")]; |
| bool attn_weights_37_transpose_x_1 = const()[name = string("attn_weights_37_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_37_transpose_y_1 = const()[name = string("attn_weights_37_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_37_cast_fp16 = matmul(transpose_x = attn_weights_37_transpose_x_1, transpose_y = attn_weights_37_transpose_y_1, x = var_626_cast_fp16_0, y = var_624_cast_fp16_0)[name = string("attn_weights_37_cast_fp16")]; |
| fp16 var_632_to_fp16 = const()[name = string("op_632_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_39_cast_fp16 = mul(x = attn_weights_37_cast_fp16, y = var_632_to_fp16)[name = string("attn_weights_39_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_41_cast_fp16 = softmax(axis = var_549, x = attn_weights_39_cast_fp16)[name = string("attn_weights_41_cast_fp16")]; |
| bool var_635_transpose_x_0 = const()[name = string("op_635_transpose_x_0"), val = bool(false)]; |
| bool var_635_transpose_y_0 = const()[name = string("op_635_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_635_cast_fp16 = matmul(transpose_x = var_635_transpose_x_0, transpose_y = var_635_transpose_y_0, x = var_628_cast_fp16_0, y = attn_weights_41_cast_fp16)[name = string("op_635_cast_fp16")]; |
| bool attn_weights_43_transpose_x_1 = const()[name = string("attn_weights_43_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_43_transpose_y_1 = const()[name = string("attn_weights_43_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_43_cast_fp16 = matmul(transpose_x = attn_weights_43_transpose_x_1, transpose_y = attn_weights_43_transpose_y_1, x = var_626_cast_fp16_1, y = var_624_cast_fp16_1)[name = string("attn_weights_43_cast_fp16")]; |
| fp16 var_638_to_fp16 = const()[name = string("op_638_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_45_cast_fp16 = mul(x = attn_weights_43_cast_fp16, y = var_638_to_fp16)[name = string("attn_weights_45_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_47_cast_fp16 = softmax(axis = var_549, x = attn_weights_45_cast_fp16)[name = string("attn_weights_47_cast_fp16")]; |
| bool attn_out_7_transpose_x_0 = const()[name = string("attn_out_7_transpose_x_0"), val = bool(false)]; |
| bool attn_out_7_transpose_y_0 = const()[name = string("attn_out_7_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_7_cast_fp16 = matmul(transpose_x = attn_out_7_transpose_x_0, transpose_y = attn_out_7_transpose_y_0, x = var_628_cast_fp16_1, y = attn_weights_47_cast_fp16)[name = string("attn_out_7_cast_fp16")]; |
| bool attn_output_7_interleave_0 = const()[name = string("attn_output_7_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_7_cast_fp16 = concat(axis = var_552, interleave = attn_output_7_interleave_0, values = (var_635_cast_fp16, attn_out_7_cast_fp16))[name = string("attn_output_7_cast_fp16")]; |
| tensor<int32, [4]> concat_15x = const()[name = string("concat_15x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_39_cast_fp16 = reshape(shape = concat_15x, x = attn_output_7_cast_fp16)[name = string("x_39_cast_fp16")]; |
| string hidden_states_35_pad_type_0 = const()[name = string("hidden_states_35_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_35_strides_0 = const()[name = string("hidden_states_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_35_pad_0 = const()[name = string("hidden_states_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_35_dilations_0 = const()[name = string("hidden_states_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_35_groups_0 = const()[name = string("hidden_states_35_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_551_to_fp16 = const()[name = string("op_551_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109220672)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_35_cast_fp16 = conv(dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = var_551_to_fp16, x = x_39_cast_fp16)[name = string("hidden_states_35_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_37_cast_fp16 = add(x = hidden_states_31_cast_fp16, y = hidden_states_35_cast_fp16)[name = string("hidden_states_37_cast_fp16")]; |
| fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_653_cast_fp16 = mul(x = hidden_states_37_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_653_cast_fp16")]; |
| bool doubled_29_interleave_0 = const()[name = string("doubled_29_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_29_cast_fp16 = concat(axis = var_552, interleave = doubled_29_interleave_0, values = (hidden_states_37_cast_fp16, var_653_cast_fp16))[name = string("doubled_29_cast_fp16")]; |
| tensor<int32, [1]> out_15_axes_0 = const()[name = string("out_15_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_15_gamma_0_to_fp16 = const()[name = string("out_15_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113415040)))]; |
| fp16 var_663_to_fp16 = const()[name = string("op_663_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_663_to_fp16, gamma = out_15_gamma_0_to_fp16, x = doubled_29_cast_fp16)[name = string("out_15_cast_fp16")]; |
| tensor<int32, [2]> var_674_split_sizes_0 = const()[name = string("op_674_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_674_axis_0 = const()[name = string("op_674_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_674_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_674_cast_fp16_1 = split(axis = var_674_axis_0, split_sizes = var_674_split_sizes_0, x = out_15_cast_fp16)[name = string("op_674_cast_fp16")]; |
| string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_541_to_fp16 = const()[name = string("op_541_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113419200)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_7_cast_fp16 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = var_541_to_fp16, x = var_674_cast_fp16_0)[name = string("input_7_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_682_cast_fp16 = silu(x = input_7_cast_fp16)[name = string("op_682_cast_fp16")]; |
| string var_687_pad_type_0 = const()[name = string("op_687_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_687_strides_0 = const()[name = string("op_687_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_687_pad_0 = const()[name = string("op_687_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_687_dilations_0 = const()[name = string("op_687_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_687_groups_0 = const()[name = string("op_687_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_542_to_fp16 = const()[name = string("op_542_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121807872)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_687_cast_fp16 = conv(dilations = var_687_dilations_0, groups = var_687_groups_0, pad = var_687_pad_0, pad_type = var_687_pad_type_0, strides = var_687_strides_0, weight = var_542_to_fp16, x = var_674_cast_fp16_0)[name = string("op_687_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_43_cast_fp16 = mul(x = var_682_cast_fp16, y = var_687_cast_fp16)[name = string("x_43_cast_fp16")]; |
| string hidden_states_39_pad_type_0 = const()[name = string("hidden_states_39_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_39_strides_0 = const()[name = string("hidden_states_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_39_pad_0 = const()[name = string("hidden_states_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_39_dilations_0 = const()[name = string("hidden_states_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_39_groups_0 = const()[name = string("hidden_states_39_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_543_to_fp16 = const()[name = string("op_543_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130196544)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_39_cast_fp16 = conv(dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = var_543_to_fp16, x = x_43_cast_fp16)[name = string("hidden_states_39_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_41_cast_fp16 = add(x = hidden_states_37_cast_fp16, y = hidden_states_39_cast_fp16)[name = string("hidden_states_41_cast_fp16")]; |
| int32 var_703 = const()[name = string("op_703"), val = int32(-2)]; |
| int32 var_706 = const()[name = string("op_706"), val = int32(1)]; |
| fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_717_cast_fp16 = mul(x = hidden_states_41_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_717_cast_fp16")]; |
| bool doubled_33_interleave_0 = const()[name = string("doubled_33_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_33_cast_fp16 = concat(axis = var_706, interleave = doubled_33_interleave_0, values = (hidden_states_41_cast_fp16, var_717_cast_fp16))[name = string("doubled_33_cast_fp16")]; |
| tensor<int32, [1]> out_17_axes_0 = const()[name = string("out_17_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_17_gamma_0_to_fp16 = const()[name = string("out_17_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138585216)))]; |
| fp16 var_727_to_fp16 = const()[name = string("op_727_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_727_to_fp16, gamma = out_17_gamma_0_to_fp16, x = doubled_33_cast_fp16)[name = string("out_17_cast_fp16")]; |
| tensor<int32, [2]> var_738_split_sizes_0 = const()[name = string("op_738_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_738_axis_0 = const()[name = string("op_738_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_738_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_738_cast_fp16_1 = split(axis = var_738_axis_0, split_sizes = var_738_split_sizes_0, x = out_17_cast_fp16)[name = string("op_738_cast_fp16")]; |
| string query_states_17_pad_type_0 = const()[name = string("query_states_17_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_17_strides_0 = const()[name = string("query_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_17_pad_0 = const()[name = string("query_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_17_dilations_0 = const()[name = string("query_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_17_groups_0 = const()[name = string("query_states_17_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_698_to_fp16 = const()[name = string("op_698_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138589376)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_17_cast_fp16 = conv(dilations = query_states_17_dilations_0, groups = query_states_17_groups_0, pad = query_states_17_pad_0, pad_type = query_states_17_pad_type_0, strides = query_states_17_strides_0, weight = var_698_to_fp16, x = var_738_cast_fp16_0)[name = string("query_states_17_cast_fp16")]; |
| string key_states_17_pad_type_0 = const()[name = string("key_states_17_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_17_strides_0 = const()[name = string("key_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_17_pad_0 = const()[name = string("key_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_17_dilations_0 = const()[name = string("key_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_17_groups_0 = const()[name = string("key_states_17_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_699_to_fp16 = const()[name = string("op_699_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142783744)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_17_cast_fp16 = conv(dilations = key_states_17_dilations_0, groups = key_states_17_groups_0, pad = key_states_17_pad_0, pad_type = key_states_17_pad_type_0, strides = key_states_17_strides_0, weight = var_699_to_fp16, x = var_738_cast_fp16_0)[name = string("key_states_17_cast_fp16")]; |
| string value_states_17_pad_type_0 = const()[name = string("value_states_17_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_17_strides_0 = const()[name = string("value_states_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_17_pad_0 = const()[name = string("value_states_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_17_dilations_0 = const()[name = string("value_states_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_17_groups_0 = const()[name = string("value_states_17_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_700_to_fp16 = const()[name = string("op_700_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143308096)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_17_cast_fp16 = conv(dilations = value_states_17_dilations_0, groups = value_states_17_groups_0, pad = value_states_17_pad_0, pad_type = value_states_17_pad_type_0, strides = value_states_17_strides_0, weight = var_700_to_fp16, x = var_738_cast_fp16_0)[name = string("value_states_17_cast_fp16")]; |
| tensor<int32, [4]> concat_16x = const()[name = string("concat_16x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_45_cast_fp16 = reshape(shape = concat_16x, x = query_states_17_cast_fp16)[name = string("x_45_cast_fp16")]; |
| tensor<int32, [4]> concat_17x = const()[name = string("concat_17x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_47_cast_fp16 = reshape(shape = concat_17x, x = key_states_17_cast_fp16)[name = string("x_47_cast_fp16")]; |
| tensor<int32, [4]> concat_18x = const()[name = string("concat_18x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_19_cast_fp16 = reshape(shape = concat_18x, x = value_states_17_cast_fp16)[name = string("value_states_19_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_764_cast_fp16 = mul(x = x_45_cast_fp16, y = rope_cos_to_fp16)[name = string("op_764_cast_fp16")]; |
| tensor<int32, [2]> var_765_split_sizes_0 = const()[name = string("op_765_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_765_axis_0 = const()[name = string("op_765_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_765_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_765_cast_fp16_1 = split(axis = var_765_axis_0, split_sizes = var_765_split_sizes_0, x = x_45_cast_fp16)[name = string("op_765_cast_fp16")]; |
| bool var_768_interleave_0 = const()[name = string("op_768_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_768_cast_fp16 = concat(axis = var_703, interleave = var_768_interleave_0, values = (var_765_cast_fp16_1, var_765_cast_fp16_0))[name = string("op_768_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_769_cast_fp16 = mul(x = var_768_cast_fp16, y = rope_sin_to_fp16)[name = string("op_769_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_19_cast_fp16 = add(x = var_764_cast_fp16, y = var_769_cast_fp16)[name = string("query_states_19_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_771_cast_fp16 = mul(x = x_47_cast_fp16, y = rope_cos_to_fp16)[name = string("op_771_cast_fp16")]; |
| tensor<int32, [2]> var_772_split_sizes_0 = const()[name = string("op_772_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_772_axis_0 = const()[name = string("op_772_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_772_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_772_cast_fp16_1 = split(axis = var_772_axis_0, split_sizes = var_772_split_sizes_0, x = x_47_cast_fp16)[name = string("op_772_cast_fp16")]; |
| bool var_775_interleave_0 = const()[name = string("op_775_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_775_cast_fp16 = concat(axis = var_703, interleave = var_775_interleave_0, values = (var_772_cast_fp16_1, var_772_cast_fp16_0))[name = string("op_775_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_776_cast_fp16 = mul(x = var_775_cast_fp16, y = rope_sin_to_fp16)[name = string("op_776_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_19_cast_fp16 = add(x = var_771_cast_fp16, y = var_776_cast_fp16)[name = string("key_states_19_cast_fp16")]; |
| tensor<int32, [2]> var_778_split_sizes_0 = const()[name = string("op_778_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_778_axis_0 = const()[name = string("op_778_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_778_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_778_cast_fp16_1 = split(axis = var_778_axis_0, split_sizes = var_778_split_sizes_0, x = query_states_19_cast_fp16)[name = string("op_778_cast_fp16")]; |
| tensor<int32, [2]> var_780_split_sizes_0 = const()[name = string("op_780_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_780_axis_0 = const()[name = string("op_780_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_780_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_780_cast_fp16_1 = split(axis = var_780_axis_0, split_sizes = var_780_split_sizes_0, x = key_states_19_cast_fp16)[name = string("op_780_cast_fp16")]; |
| tensor<int32, [2]> var_782_split_sizes_0 = const()[name = string("op_782_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_782_axis_0 = const()[name = string("op_782_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_782_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_782_cast_fp16_1 = split(axis = var_782_axis_0, split_sizes = var_782_split_sizes_0, x = value_states_19_cast_fp16)[name = string("op_782_cast_fp16")]; |
| bool attn_weights_49_transpose_x_1 = const()[name = string("attn_weights_49_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_49_transpose_y_1 = const()[name = string("attn_weights_49_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_49_cast_fp16 = matmul(transpose_x = attn_weights_49_transpose_x_1, transpose_y = attn_weights_49_transpose_y_1, x = var_780_cast_fp16_0, y = var_778_cast_fp16_0)[name = string("attn_weights_49_cast_fp16")]; |
| fp16 var_786_to_fp16 = const()[name = string("op_786_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_51_cast_fp16 = mul(x = attn_weights_49_cast_fp16, y = var_786_to_fp16)[name = string("attn_weights_51_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_53_cast_fp16 = softmax(axis = var_703, x = attn_weights_51_cast_fp16)[name = string("attn_weights_53_cast_fp16")]; |
| bool var_789_transpose_x_0 = const()[name = string("op_789_transpose_x_0"), val = bool(false)]; |
| bool var_789_transpose_y_0 = const()[name = string("op_789_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_789_cast_fp16 = matmul(transpose_x = var_789_transpose_x_0, transpose_y = var_789_transpose_y_0, x = var_782_cast_fp16_0, y = attn_weights_53_cast_fp16)[name = string("op_789_cast_fp16")]; |
| bool attn_weights_55_transpose_x_1 = const()[name = string("attn_weights_55_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_55_transpose_y_1 = const()[name = string("attn_weights_55_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_55_cast_fp16 = matmul(transpose_x = attn_weights_55_transpose_x_1, transpose_y = attn_weights_55_transpose_y_1, x = var_780_cast_fp16_1, y = var_778_cast_fp16_1)[name = string("attn_weights_55_cast_fp16")]; |
| fp16 var_792_to_fp16 = const()[name = string("op_792_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_57_cast_fp16 = mul(x = attn_weights_55_cast_fp16, y = var_792_to_fp16)[name = string("attn_weights_57_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_59_cast_fp16 = softmax(axis = var_703, x = attn_weights_57_cast_fp16)[name = string("attn_weights_59_cast_fp16")]; |
| bool attn_out_9_transpose_x_0 = const()[name = string("attn_out_9_transpose_x_0"), val = bool(false)]; |
| bool attn_out_9_transpose_y_0 = const()[name = string("attn_out_9_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_9_cast_fp16 = matmul(transpose_x = attn_out_9_transpose_x_0, transpose_y = attn_out_9_transpose_y_0, x = var_782_cast_fp16_1, y = attn_weights_59_cast_fp16)[name = string("attn_out_9_cast_fp16")]; |
| bool attn_output_9_interleave_0 = const()[name = string("attn_output_9_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_9_cast_fp16 = concat(axis = var_706, interleave = attn_output_9_interleave_0, values = (var_789_cast_fp16, attn_out_9_cast_fp16))[name = string("attn_output_9_cast_fp16")]; |
| tensor<int32, [4]> concat_19x = const()[name = string("concat_19x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_49_cast_fp16 = reshape(shape = concat_19x, x = attn_output_9_cast_fp16)[name = string("x_49_cast_fp16")]; |
| string hidden_states_45_pad_type_0 = const()[name = string("hidden_states_45_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_45_strides_0 = const()[name = string("hidden_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_45_pad_0 = const()[name = string("hidden_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_45_dilations_0 = const()[name = string("hidden_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_45_groups_0 = const()[name = string("hidden_states_45_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_705_to_fp16 = const()[name = string("op_705_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143832448)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_45_cast_fp16 = conv(dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = var_705_to_fp16, x = x_49_cast_fp16)[name = string("hidden_states_45_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_47_cast_fp16 = add(x = hidden_states_41_cast_fp16, y = hidden_states_45_cast_fp16)[name = string("hidden_states_47_cast_fp16")]; |
| fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_807_cast_fp16 = mul(x = hidden_states_47_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_807_cast_fp16")]; |
| bool doubled_37_interleave_0 = const()[name = string("doubled_37_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_37_cast_fp16 = concat(axis = var_706, interleave = doubled_37_interleave_0, values = (hidden_states_47_cast_fp16, var_807_cast_fp16))[name = string("doubled_37_cast_fp16")]; |
| tensor<int32, [1]> out_19_axes_0 = const()[name = string("out_19_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_19_gamma_0_to_fp16 = const()[name = string("out_19_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148026816)))]; |
| fp16 var_817_to_fp16 = const()[name = string("op_817_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_817_to_fp16, gamma = out_19_gamma_0_to_fp16, x = doubled_37_cast_fp16)[name = string("out_19_cast_fp16")]; |
| tensor<int32, [2]> var_828_split_sizes_0 = const()[name = string("op_828_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_828_axis_0 = const()[name = string("op_828_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_828_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_828_cast_fp16_1 = split(axis = var_828_axis_0, split_sizes = var_828_split_sizes_0, x = out_19_cast_fp16)[name = string("op_828_cast_fp16")]; |
| string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_695_to_fp16 = const()[name = string("op_695_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148030976)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_9_cast_fp16 = conv(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 = var_695_to_fp16, x = var_828_cast_fp16_0)[name = string("input_9_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_836_cast_fp16 = silu(x = input_9_cast_fp16)[name = string("op_836_cast_fp16")]; |
| string var_841_pad_type_0 = const()[name = string("op_841_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_841_strides_0 = const()[name = string("op_841_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_841_pad_0 = const()[name = string("op_841_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_841_dilations_0 = const()[name = string("op_841_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_841_groups_0 = const()[name = string("op_841_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_696_to_fp16 = const()[name = string("op_696_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156419648)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_841_cast_fp16 = conv(dilations = var_841_dilations_0, groups = var_841_groups_0, pad = var_841_pad_0, pad_type = var_841_pad_type_0, strides = var_841_strides_0, weight = var_696_to_fp16, x = var_828_cast_fp16_0)[name = string("op_841_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_53_cast_fp16 = mul(x = var_836_cast_fp16, y = var_841_cast_fp16)[name = string("x_53_cast_fp16")]; |
| string hidden_states_49_pad_type_0 = const()[name = string("hidden_states_49_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_49_strides_0 = const()[name = string("hidden_states_49_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_49_pad_0 = const()[name = string("hidden_states_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_49_dilations_0 = const()[name = string("hidden_states_49_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_49_groups_0 = const()[name = string("hidden_states_49_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_697_to_fp16 = const()[name = string("op_697_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164808320)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_49_cast_fp16 = conv(dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = var_697_to_fp16, x = x_53_cast_fp16)[name = string("hidden_states_49_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_51_cast_fp16 = add(x = hidden_states_47_cast_fp16, y = hidden_states_49_cast_fp16)[name = string("hidden_states_51_cast_fp16")]; |
| int32 var_857 = const()[name = string("op_857"), val = int32(-2)]; |
| int32 var_860 = const()[name = string("op_860"), val = int32(1)]; |
| fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_871_cast_fp16 = mul(x = hidden_states_51_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_871_cast_fp16")]; |
| bool doubled_41_interleave_0 = const()[name = string("doubled_41_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_41_cast_fp16 = concat(axis = var_860, interleave = doubled_41_interleave_0, values = (hidden_states_51_cast_fp16, var_871_cast_fp16))[name = string("doubled_41_cast_fp16")]; |
| tensor<int32, [1]> out_21_axes_0 = const()[name = string("out_21_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_21_gamma_0_to_fp16 = const()[name = string("out_21_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173196992)))]; |
| fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_881_to_fp16, gamma = out_21_gamma_0_to_fp16, x = doubled_41_cast_fp16)[name = string("out_21_cast_fp16")]; |
| tensor<int32, [2]> var_892_split_sizes_0 = const()[name = string("op_892_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_892_axis_0 = const()[name = string("op_892_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_892_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_892_cast_fp16_1 = split(axis = var_892_axis_0, split_sizes = var_892_split_sizes_0, x = out_21_cast_fp16)[name = string("op_892_cast_fp16")]; |
| string query_states_21_pad_type_0 = const()[name = string("query_states_21_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_21_strides_0 = const()[name = string("query_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_21_pad_0 = const()[name = string("query_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_21_dilations_0 = const()[name = string("query_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_21_groups_0 = const()[name = string("query_states_21_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_852_to_fp16 = const()[name = string("op_852_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173201152)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_21_cast_fp16 = conv(dilations = query_states_21_dilations_0, groups = query_states_21_groups_0, pad = query_states_21_pad_0, pad_type = query_states_21_pad_type_0, strides = query_states_21_strides_0, weight = var_852_to_fp16, x = var_892_cast_fp16_0)[name = string("query_states_21_cast_fp16")]; |
| string key_states_21_pad_type_0 = const()[name = string("key_states_21_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_21_strides_0 = const()[name = string("key_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_21_pad_0 = const()[name = string("key_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_21_dilations_0 = const()[name = string("key_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_21_groups_0 = const()[name = string("key_states_21_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_853_to_fp16 = const()[name = string("op_853_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177395520)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_21_cast_fp16 = conv(dilations = key_states_21_dilations_0, groups = key_states_21_groups_0, pad = key_states_21_pad_0, pad_type = key_states_21_pad_type_0, strides = key_states_21_strides_0, weight = var_853_to_fp16, x = var_892_cast_fp16_0)[name = string("key_states_21_cast_fp16")]; |
| string value_states_21_pad_type_0 = const()[name = string("value_states_21_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_21_strides_0 = const()[name = string("value_states_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_21_pad_0 = const()[name = string("value_states_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_21_dilations_0 = const()[name = string("value_states_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_21_groups_0 = const()[name = string("value_states_21_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_854_to_fp16 = const()[name = string("op_854_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177919872)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_21_cast_fp16 = conv(dilations = value_states_21_dilations_0, groups = value_states_21_groups_0, pad = value_states_21_pad_0, pad_type = value_states_21_pad_type_0, strides = value_states_21_strides_0, weight = var_854_to_fp16, x = var_892_cast_fp16_0)[name = string("value_states_21_cast_fp16")]; |
| tensor<int32, [4]> concat_20x = const()[name = string("concat_20x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_55_cast_fp16 = reshape(shape = concat_20x, x = query_states_21_cast_fp16)[name = string("x_55_cast_fp16")]; |
| tensor<int32, [4]> concat_21x = const()[name = string("concat_21x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_57_cast_fp16 = reshape(shape = concat_21x, x = key_states_21_cast_fp16)[name = string("x_57_cast_fp16")]; |
| tensor<int32, [4]> concat_22x = const()[name = string("concat_22x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_23_cast_fp16 = reshape(shape = concat_22x, x = value_states_21_cast_fp16)[name = string("value_states_23_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_918_cast_fp16 = mul(x = x_55_cast_fp16, y = rope_cos_to_fp16)[name = string("op_918_cast_fp16")]; |
| tensor<int32, [2]> var_919_split_sizes_0 = const()[name = string("op_919_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_919_axis_0 = const()[name = string("op_919_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_919_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_919_cast_fp16_1 = split(axis = var_919_axis_0, split_sizes = var_919_split_sizes_0, x = x_55_cast_fp16)[name = string("op_919_cast_fp16")]; |
| bool var_922_interleave_0 = const()[name = string("op_922_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_922_cast_fp16 = concat(axis = var_857, interleave = var_922_interleave_0, values = (var_919_cast_fp16_1, var_919_cast_fp16_0))[name = string("op_922_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_923_cast_fp16 = mul(x = var_922_cast_fp16, y = rope_sin_to_fp16)[name = string("op_923_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_23_cast_fp16 = add(x = var_918_cast_fp16, y = var_923_cast_fp16)[name = string("query_states_23_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_925_cast_fp16 = mul(x = x_57_cast_fp16, y = rope_cos_to_fp16)[name = string("op_925_cast_fp16")]; |
| tensor<int32, [2]> var_926_split_sizes_0 = const()[name = string("op_926_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_926_axis_0 = const()[name = string("op_926_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_926_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_926_cast_fp16_1 = split(axis = var_926_axis_0, split_sizes = var_926_split_sizes_0, x = x_57_cast_fp16)[name = string("op_926_cast_fp16")]; |
| bool var_929_interleave_0 = const()[name = string("op_929_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_929_cast_fp16 = concat(axis = var_857, interleave = var_929_interleave_0, values = (var_926_cast_fp16_1, var_926_cast_fp16_0))[name = string("op_929_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_930_cast_fp16 = mul(x = var_929_cast_fp16, y = rope_sin_to_fp16)[name = string("op_930_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_23_cast_fp16 = add(x = var_925_cast_fp16, y = var_930_cast_fp16)[name = string("key_states_23_cast_fp16")]; |
| tensor<int32, [2]> var_932_split_sizes_0 = const()[name = string("op_932_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_932_axis_0 = const()[name = string("op_932_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_932_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_932_cast_fp16_1 = split(axis = var_932_axis_0, split_sizes = var_932_split_sizes_0, x = query_states_23_cast_fp16)[name = string("op_932_cast_fp16")]; |
| tensor<int32, [2]> var_934_split_sizes_0 = const()[name = string("op_934_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_934_axis_0 = const()[name = string("op_934_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_934_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_934_cast_fp16_1 = split(axis = var_934_axis_0, split_sizes = var_934_split_sizes_0, x = key_states_23_cast_fp16)[name = string("op_934_cast_fp16")]; |
| tensor<int32, [2]> var_936_split_sizes_0 = const()[name = string("op_936_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_936_axis_0 = const()[name = string("op_936_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_936_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_936_cast_fp16_1 = split(axis = var_936_axis_0, split_sizes = var_936_split_sizes_0, x = value_states_23_cast_fp16)[name = string("op_936_cast_fp16")]; |
| bool attn_weights_61_transpose_x_1 = const()[name = string("attn_weights_61_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_61_transpose_y_1 = const()[name = string("attn_weights_61_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_61_cast_fp16 = matmul(transpose_x = attn_weights_61_transpose_x_1, transpose_y = attn_weights_61_transpose_y_1, x = var_934_cast_fp16_0, y = var_932_cast_fp16_0)[name = string("attn_weights_61_cast_fp16")]; |
| fp16 var_940_to_fp16 = const()[name = string("op_940_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_63_cast_fp16 = mul(x = attn_weights_61_cast_fp16, y = var_940_to_fp16)[name = string("attn_weights_63_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_65_cast_fp16 = softmax(axis = var_857, x = attn_weights_63_cast_fp16)[name = string("attn_weights_65_cast_fp16")]; |
| bool var_943_transpose_x_0 = const()[name = string("op_943_transpose_x_0"), val = bool(false)]; |
| bool var_943_transpose_y_0 = const()[name = string("op_943_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_943_cast_fp16 = matmul(transpose_x = var_943_transpose_x_0, transpose_y = var_943_transpose_y_0, x = var_936_cast_fp16_0, y = attn_weights_65_cast_fp16)[name = string("op_943_cast_fp16")]; |
| bool attn_weights_67_transpose_x_1 = const()[name = string("attn_weights_67_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_67_transpose_y_1 = const()[name = string("attn_weights_67_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_67_cast_fp16 = matmul(transpose_x = attn_weights_67_transpose_x_1, transpose_y = attn_weights_67_transpose_y_1, x = var_934_cast_fp16_1, y = var_932_cast_fp16_1)[name = string("attn_weights_67_cast_fp16")]; |
| fp16 var_946_to_fp16 = const()[name = string("op_946_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_69_cast_fp16 = mul(x = attn_weights_67_cast_fp16, y = var_946_to_fp16)[name = string("attn_weights_69_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_71_cast_fp16 = softmax(axis = var_857, x = attn_weights_69_cast_fp16)[name = string("attn_weights_71_cast_fp16")]; |
| bool attn_out_11_transpose_x_0 = const()[name = string("attn_out_11_transpose_x_0"), val = bool(false)]; |
| bool attn_out_11_transpose_y_0 = const()[name = string("attn_out_11_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_11_cast_fp16 = matmul(transpose_x = attn_out_11_transpose_x_0, transpose_y = attn_out_11_transpose_y_0, x = var_936_cast_fp16_1, y = attn_weights_71_cast_fp16)[name = string("attn_out_11_cast_fp16")]; |
| bool attn_output_11_interleave_0 = const()[name = string("attn_output_11_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_11_cast_fp16 = concat(axis = var_860, interleave = attn_output_11_interleave_0, values = (var_943_cast_fp16, attn_out_11_cast_fp16))[name = string("attn_output_11_cast_fp16")]; |
| tensor<int32, [4]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_59_cast_fp16 = reshape(shape = concat_23x, x = attn_output_11_cast_fp16)[name = string("x_59_cast_fp16")]; |
| string hidden_states_55_pad_type_0 = const()[name = string("hidden_states_55_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_55_strides_0 = const()[name = string("hidden_states_55_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_55_pad_0 = const()[name = string("hidden_states_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_55_dilations_0 = const()[name = string("hidden_states_55_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_55_groups_0 = const()[name = string("hidden_states_55_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_859_to_fp16 = const()[name = string("op_859_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178444224)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_55_cast_fp16 = conv(dilations = hidden_states_55_dilations_0, groups = hidden_states_55_groups_0, pad = hidden_states_55_pad_0, pad_type = hidden_states_55_pad_type_0, strides = hidden_states_55_strides_0, weight = var_859_to_fp16, x = x_59_cast_fp16)[name = string("hidden_states_55_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_57_cast_fp16 = add(x = hidden_states_51_cast_fp16, y = hidden_states_55_cast_fp16)[name = string("hidden_states_57_cast_fp16")]; |
| fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_961_cast_fp16 = mul(x = hidden_states_57_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_961_cast_fp16")]; |
| bool doubled_45_interleave_0 = const()[name = string("doubled_45_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_45_cast_fp16 = concat(axis = var_860, interleave = doubled_45_interleave_0, values = (hidden_states_57_cast_fp16, var_961_cast_fp16))[name = string("doubled_45_cast_fp16")]; |
| tensor<int32, [1]> out_23_axes_0 = const()[name = string("out_23_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_23_gamma_0_to_fp16 = const()[name = string("out_23_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182638592)))]; |
| fp16 var_971_to_fp16 = const()[name = string("op_971_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_971_to_fp16, gamma = out_23_gamma_0_to_fp16, x = doubled_45_cast_fp16)[name = string("out_23_cast_fp16")]; |
| tensor<int32, [2]> var_982_split_sizes_0 = const()[name = string("op_982_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_982_axis_0 = const()[name = string("op_982_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_982_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_982_cast_fp16_1 = split(axis = var_982_axis_0, split_sizes = var_982_split_sizes_0, x = out_23_cast_fp16)[name = string("op_982_cast_fp16")]; |
| string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_849_to_fp16 = const()[name = string("op_849_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182642752)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = var_849_to_fp16, x = var_982_cast_fp16_0)[name = string("input_11_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_990_cast_fp16 = silu(x = input_11_cast_fp16)[name = string("op_990_cast_fp16")]; |
| string var_995_pad_type_0 = const()[name = string("op_995_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_995_strides_0 = const()[name = string("op_995_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_995_pad_0 = const()[name = string("op_995_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_995_dilations_0 = const()[name = string("op_995_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_995_groups_0 = const()[name = string("op_995_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_850_to_fp16 = const()[name = string("op_850_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191031424)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_995_cast_fp16 = conv(dilations = var_995_dilations_0, groups = var_995_groups_0, pad = var_995_pad_0, pad_type = var_995_pad_type_0, strides = var_995_strides_0, weight = var_850_to_fp16, x = var_982_cast_fp16_0)[name = string("op_995_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_63_cast_fp16 = mul(x = var_990_cast_fp16, y = var_995_cast_fp16)[name = string("x_63_cast_fp16")]; |
| string hidden_states_59_pad_type_0 = const()[name = string("hidden_states_59_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_59_strides_0 = const()[name = string("hidden_states_59_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_59_pad_0 = const()[name = string("hidden_states_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_59_dilations_0 = const()[name = string("hidden_states_59_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_59_groups_0 = const()[name = string("hidden_states_59_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_851_to_fp16 = const()[name = string("op_851_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199420096)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_59_cast_fp16 = conv(dilations = hidden_states_59_dilations_0, groups = hidden_states_59_groups_0, pad = hidden_states_59_pad_0, pad_type = hidden_states_59_pad_type_0, strides = hidden_states_59_strides_0, weight = var_851_to_fp16, x = x_63_cast_fp16)[name = string("hidden_states_59_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_61_cast_fp16 = add(x = hidden_states_57_cast_fp16, y = hidden_states_59_cast_fp16)[name = string("hidden_states_61_cast_fp16")]; |
| int32 var_1011 = const()[name = string("op_1011"), val = int32(-2)]; |
| int32 var_1014 = const()[name = string("op_1014"), val = int32(1)]; |
| fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1025_cast_fp16 = mul(x = hidden_states_61_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_1025_cast_fp16")]; |
| bool doubled_49_interleave_0 = const()[name = string("doubled_49_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_49_cast_fp16 = concat(axis = var_1014, interleave = doubled_49_interleave_0, values = (hidden_states_61_cast_fp16, var_1025_cast_fp16))[name = string("doubled_49_cast_fp16")]; |
| tensor<int32, [1]> out_25_axes_0 = const()[name = string("out_25_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_25_gamma_0_to_fp16 = const()[name = string("out_25_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207808768)))]; |
| fp16 var_1035_to_fp16 = const()[name = string("op_1035_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1035_to_fp16, gamma = out_25_gamma_0_to_fp16, x = doubled_49_cast_fp16)[name = string("out_25_cast_fp16")]; |
| tensor<int32, [2]> var_1046_split_sizes_0 = const()[name = string("op_1046_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1046_axis_0 = const()[name = string("op_1046_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1046_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1046_cast_fp16_1 = split(axis = var_1046_axis_0, split_sizes = var_1046_split_sizes_0, x = out_25_cast_fp16)[name = string("op_1046_cast_fp16")]; |
| string query_states_25_pad_type_0 = const()[name = string("query_states_25_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_25_strides_0 = const()[name = string("query_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_25_pad_0 = const()[name = string("query_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_25_dilations_0 = const()[name = string("query_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_25_groups_0 = const()[name = string("query_states_25_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1006_to_fp16 = const()[name = string("op_1006_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207812928)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_25_cast_fp16 = conv(dilations = query_states_25_dilations_0, groups = query_states_25_groups_0, pad = query_states_25_pad_0, pad_type = query_states_25_pad_type_0, strides = query_states_25_strides_0, weight = var_1006_to_fp16, x = var_1046_cast_fp16_0)[name = string("query_states_25_cast_fp16")]; |
| string key_states_25_pad_type_0 = const()[name = string("key_states_25_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_25_strides_0 = const()[name = string("key_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_25_pad_0 = const()[name = string("key_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_25_dilations_0 = const()[name = string("key_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_25_groups_0 = const()[name = string("key_states_25_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1007_to_fp16 = const()[name = string("op_1007_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212007296)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_25_cast_fp16 = conv(dilations = key_states_25_dilations_0, groups = key_states_25_groups_0, pad = key_states_25_pad_0, pad_type = key_states_25_pad_type_0, strides = key_states_25_strides_0, weight = var_1007_to_fp16, x = var_1046_cast_fp16_0)[name = string("key_states_25_cast_fp16")]; |
| string value_states_25_pad_type_0 = const()[name = string("value_states_25_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_25_strides_0 = const()[name = string("value_states_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_25_pad_0 = const()[name = string("value_states_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_25_dilations_0 = const()[name = string("value_states_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_25_groups_0 = const()[name = string("value_states_25_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1008_to_fp16 = const()[name = string("op_1008_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212531648)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_25_cast_fp16 = conv(dilations = value_states_25_dilations_0, groups = value_states_25_groups_0, pad = value_states_25_pad_0, pad_type = value_states_25_pad_type_0, strides = value_states_25_strides_0, weight = var_1008_to_fp16, x = var_1046_cast_fp16_0)[name = string("value_states_25_cast_fp16")]; |
| tensor<int32, [4]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_65_cast_fp16 = reshape(shape = concat_24x, x = query_states_25_cast_fp16)[name = string("x_65_cast_fp16")]; |
| tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_67_cast_fp16 = reshape(shape = concat_25x, x = key_states_25_cast_fp16)[name = string("x_67_cast_fp16")]; |
| tensor<int32, [4]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_27_cast_fp16 = reshape(shape = concat_26x, x = value_states_25_cast_fp16)[name = string("value_states_27_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1072_cast_fp16 = mul(x = x_65_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1072_cast_fp16")]; |
| tensor<int32, [2]> var_1073_split_sizes_0 = const()[name = string("op_1073_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1073_axis_0 = const()[name = string("op_1073_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_1073_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_1073_cast_fp16_1 = split(axis = var_1073_axis_0, split_sizes = var_1073_split_sizes_0, x = x_65_cast_fp16)[name = string("op_1073_cast_fp16")]; |
| bool var_1076_interleave_0 = const()[name = string("op_1076_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_1076_cast_fp16 = concat(axis = var_1011, interleave = var_1076_interleave_0, values = (var_1073_cast_fp16_1, var_1073_cast_fp16_0))[name = string("op_1076_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1077_cast_fp16 = mul(x = var_1076_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1077_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_27_cast_fp16 = add(x = var_1072_cast_fp16, y = var_1077_cast_fp16)[name = string("query_states_27_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1079_cast_fp16 = mul(x = x_67_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1079_cast_fp16")]; |
| tensor<int32, [2]> var_1080_split_sizes_0 = const()[name = string("op_1080_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1080_axis_0 = const()[name = string("op_1080_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_1080_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_1080_cast_fp16_1 = split(axis = var_1080_axis_0, split_sizes = var_1080_split_sizes_0, x = x_67_cast_fp16)[name = string("op_1080_cast_fp16")]; |
| bool var_1083_interleave_0 = const()[name = string("op_1083_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_1083_cast_fp16 = concat(axis = var_1011, interleave = var_1083_interleave_0, values = (var_1080_cast_fp16_1, var_1080_cast_fp16_0))[name = string("op_1083_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1084_cast_fp16 = mul(x = var_1083_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1084_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_27_cast_fp16 = add(x = var_1079_cast_fp16, y = var_1084_cast_fp16)[name = string("key_states_27_cast_fp16")]; |
| tensor<int32, [2]> var_1086_split_sizes_0 = const()[name = string("op_1086_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_1086_axis_0 = const()[name = string("op_1086_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1086_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_1086_cast_fp16_1 = split(axis = var_1086_axis_0, split_sizes = var_1086_split_sizes_0, x = query_states_27_cast_fp16)[name = string("op_1086_cast_fp16")]; |
| tensor<int32, [2]> var_1088_split_sizes_0 = const()[name = string("op_1088_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1088_axis_0 = const()[name = string("op_1088_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1088_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1088_cast_fp16_1 = split(axis = var_1088_axis_0, split_sizes = var_1088_split_sizes_0, x = key_states_27_cast_fp16)[name = string("op_1088_cast_fp16")]; |
| tensor<int32, [2]> var_1090_split_sizes_0 = const()[name = string("op_1090_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1090_axis_0 = const()[name = string("op_1090_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1090_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1090_cast_fp16_1 = split(axis = var_1090_axis_0, split_sizes = var_1090_split_sizes_0, x = value_states_27_cast_fp16)[name = string("op_1090_cast_fp16")]; |
| bool attn_weights_73_transpose_x_1 = const()[name = string("attn_weights_73_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_73_transpose_y_1 = const()[name = string("attn_weights_73_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_73_cast_fp16 = matmul(transpose_x = attn_weights_73_transpose_x_1, transpose_y = attn_weights_73_transpose_y_1, x = var_1088_cast_fp16_0, y = var_1086_cast_fp16_0)[name = string("attn_weights_73_cast_fp16")]; |
| fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_75_cast_fp16 = mul(x = attn_weights_73_cast_fp16, y = var_1094_to_fp16)[name = string("attn_weights_75_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_77_cast_fp16 = softmax(axis = var_1011, x = attn_weights_75_cast_fp16)[name = string("attn_weights_77_cast_fp16")]; |
| bool var_1097_transpose_x_0 = const()[name = string("op_1097_transpose_x_0"), val = bool(false)]; |
| bool var_1097_transpose_y_0 = const()[name = string("op_1097_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1097_cast_fp16 = matmul(transpose_x = var_1097_transpose_x_0, transpose_y = var_1097_transpose_y_0, x = var_1090_cast_fp16_0, y = attn_weights_77_cast_fp16)[name = string("op_1097_cast_fp16")]; |
| bool attn_weights_79_transpose_x_1 = const()[name = string("attn_weights_79_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_79_transpose_y_1 = const()[name = string("attn_weights_79_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_79_cast_fp16 = matmul(transpose_x = attn_weights_79_transpose_x_1, transpose_y = attn_weights_79_transpose_y_1, x = var_1088_cast_fp16_1, y = var_1086_cast_fp16_1)[name = string("attn_weights_79_cast_fp16")]; |
| fp16 var_1100_to_fp16 = const()[name = string("op_1100_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_81_cast_fp16 = mul(x = attn_weights_79_cast_fp16, y = var_1100_to_fp16)[name = string("attn_weights_81_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_83_cast_fp16 = softmax(axis = var_1011, x = attn_weights_81_cast_fp16)[name = string("attn_weights_83_cast_fp16")]; |
| bool attn_out_13_transpose_x_0 = const()[name = string("attn_out_13_transpose_x_0"), val = bool(false)]; |
| bool attn_out_13_transpose_y_0 = const()[name = string("attn_out_13_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_13_cast_fp16 = matmul(transpose_x = attn_out_13_transpose_x_0, transpose_y = attn_out_13_transpose_y_0, x = var_1090_cast_fp16_1, y = attn_weights_83_cast_fp16)[name = string("attn_out_13_cast_fp16")]; |
| bool attn_output_13_interleave_0 = const()[name = string("attn_output_13_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_13_cast_fp16 = concat(axis = var_1014, interleave = attn_output_13_interleave_0, values = (var_1097_cast_fp16, attn_out_13_cast_fp16))[name = string("attn_output_13_cast_fp16")]; |
| tensor<int32, [4]> concat_27x = const()[name = string("concat_27x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_69_cast_fp16 = reshape(shape = concat_27x, x = attn_output_13_cast_fp16)[name = string("x_69_cast_fp16")]; |
| string hidden_states_65_pad_type_0 = const()[name = string("hidden_states_65_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_65_strides_0 = const()[name = string("hidden_states_65_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_65_pad_0 = const()[name = string("hidden_states_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_65_dilations_0 = const()[name = string("hidden_states_65_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_65_groups_0 = const()[name = string("hidden_states_65_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_1013_to_fp16 = const()[name = string("op_1013_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213056000)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_65_cast_fp16 = conv(dilations = hidden_states_65_dilations_0, groups = hidden_states_65_groups_0, pad = hidden_states_65_pad_0, pad_type = hidden_states_65_pad_type_0, strides = hidden_states_65_strides_0, weight = var_1013_to_fp16, x = x_69_cast_fp16)[name = string("hidden_states_65_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_67_cast_fp16 = add(x = hidden_states_61_cast_fp16, y = hidden_states_65_cast_fp16)[name = string("hidden_states_67_cast_fp16")]; |
| fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1115_cast_fp16 = mul(x = hidden_states_67_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_1115_cast_fp16")]; |
| bool doubled_53_interleave_0 = const()[name = string("doubled_53_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_53_cast_fp16 = concat(axis = var_1014, interleave = doubled_53_interleave_0, values = (hidden_states_67_cast_fp16, var_1115_cast_fp16))[name = string("doubled_53_cast_fp16")]; |
| tensor<int32, [1]> out_27_axes_0 = const()[name = string("out_27_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_27_gamma_0_to_fp16 = const()[name = string("out_27_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217250368)))]; |
| fp16 var_1125_to_fp16 = const()[name = string("op_1125_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1125_to_fp16, gamma = out_27_gamma_0_to_fp16, x = doubled_53_cast_fp16)[name = string("out_27_cast_fp16")]; |
| tensor<int32, [2]> var_1136_split_sizes_0 = const()[name = string("op_1136_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1136_axis_0 = const()[name = string("op_1136_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1136_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1136_cast_fp16_1 = split(axis = var_1136_axis_0, split_sizes = var_1136_split_sizes_0, x = out_27_cast_fp16)[name = string("op_1136_cast_fp16")]; |
| string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1003_to_fp16 = const()[name = string("op_1003_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217254528)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = var_1003_to_fp16, x = var_1136_cast_fp16_0)[name = string("input_13_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1144_cast_fp16 = silu(x = input_13_cast_fp16)[name = string("op_1144_cast_fp16")]; |
| string var_1149_pad_type_0 = const()[name = string("op_1149_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1149_strides_0 = const()[name = string("op_1149_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1149_pad_0 = const()[name = string("op_1149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1149_dilations_0 = const()[name = string("op_1149_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1149_groups_0 = const()[name = string("op_1149_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1004_to_fp16 = const()[name = string("op_1004_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225643200)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1149_cast_fp16 = conv(dilations = var_1149_dilations_0, groups = var_1149_groups_0, pad = var_1149_pad_0, pad_type = var_1149_pad_type_0, strides = var_1149_strides_0, weight = var_1004_to_fp16, x = var_1136_cast_fp16_0)[name = string("op_1149_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_73_cast_fp16 = mul(x = var_1144_cast_fp16, y = var_1149_cast_fp16)[name = string("x_73_cast_fp16")]; |
| string hidden_states_69_pad_type_0 = const()[name = string("hidden_states_69_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_69_strides_0 = const()[name = string("hidden_states_69_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_69_pad_0 = const()[name = string("hidden_states_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_69_dilations_0 = const()[name = string("hidden_states_69_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_69_groups_0 = const()[name = string("hidden_states_69_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_1005_to_fp16 = const()[name = string("op_1005_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234031872)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_69_cast_fp16 = conv(dilations = hidden_states_69_dilations_0, groups = hidden_states_69_groups_0, pad = hidden_states_69_pad_0, pad_type = hidden_states_69_pad_type_0, strides = hidden_states_69_strides_0, weight = var_1005_to_fp16, x = x_73_cast_fp16)[name = string("hidden_states_69_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_71_cast_fp16 = add(x = hidden_states_67_cast_fp16, y = hidden_states_69_cast_fp16)[name = string("hidden_states_71_cast_fp16")]; |
| int32 var_1165 = const()[name = string("op_1165"), val = int32(-2)]; |
| int32 var_1168 = const()[name = string("op_1168"), val = int32(1)]; |
| fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1179_cast_fp16 = mul(x = hidden_states_71_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_1179_cast_fp16")]; |
| bool doubled_57_interleave_0 = const()[name = string("doubled_57_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_57_cast_fp16 = concat(axis = var_1168, interleave = doubled_57_interleave_0, values = (hidden_states_71_cast_fp16, var_1179_cast_fp16))[name = string("doubled_57_cast_fp16")]; |
| tensor<int32, [1]> out_29_axes_0 = const()[name = string("out_29_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_29_gamma_0_to_fp16 = const()[name = string("out_29_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242420544)))]; |
| fp16 var_1189_to_fp16 = const()[name = string("op_1189_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1189_to_fp16, gamma = out_29_gamma_0_to_fp16, x = doubled_57_cast_fp16)[name = string("out_29_cast_fp16")]; |
| tensor<int32, [2]> var_1200_split_sizes_0 = const()[name = string("op_1200_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1200_axis_0 = const()[name = string("op_1200_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1200_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1200_cast_fp16_1 = split(axis = var_1200_axis_0, split_sizes = var_1200_split_sizes_0, x = out_29_cast_fp16)[name = string("op_1200_cast_fp16")]; |
| string query_states_29_pad_type_0 = const()[name = string("query_states_29_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_29_strides_0 = const()[name = string("query_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_29_pad_0 = const()[name = string("query_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_29_dilations_0 = const()[name = string("query_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_29_groups_0 = const()[name = string("query_states_29_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1160_to_fp16 = const()[name = string("op_1160_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242424704)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_29_cast_fp16 = conv(dilations = query_states_29_dilations_0, groups = query_states_29_groups_0, pad = query_states_29_pad_0, pad_type = query_states_29_pad_type_0, strides = query_states_29_strides_0, weight = var_1160_to_fp16, x = var_1200_cast_fp16_0)[name = string("query_states_29_cast_fp16")]; |
| string key_states_29_pad_type_0 = const()[name = string("key_states_29_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_29_strides_0 = const()[name = string("key_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_29_pad_0 = const()[name = string("key_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_29_dilations_0 = const()[name = string("key_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_29_groups_0 = const()[name = string("key_states_29_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1161_to_fp16 = const()[name = string("op_1161_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246619072)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_29_cast_fp16 = conv(dilations = key_states_29_dilations_0, groups = key_states_29_groups_0, pad = key_states_29_pad_0, pad_type = key_states_29_pad_type_0, strides = key_states_29_strides_0, weight = var_1161_to_fp16, x = var_1200_cast_fp16_0)[name = string("key_states_29_cast_fp16")]; |
| string value_states_29_pad_type_0 = const()[name = string("value_states_29_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_29_strides_0 = const()[name = string("value_states_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_29_pad_0 = const()[name = string("value_states_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_29_dilations_0 = const()[name = string("value_states_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_29_groups_0 = const()[name = string("value_states_29_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1162_to_fp16 = const()[name = string("op_1162_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247143424)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_29_cast_fp16 = conv(dilations = value_states_29_dilations_0, groups = value_states_29_groups_0, pad = value_states_29_pad_0, pad_type = value_states_29_pad_type_0, strides = value_states_29_strides_0, weight = var_1162_to_fp16, x = var_1200_cast_fp16_0)[name = string("value_states_29_cast_fp16")]; |
| tensor<int32, [4]> concat_28x = const()[name = string("concat_28x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_75_cast_fp16 = reshape(shape = concat_28x, x = query_states_29_cast_fp16)[name = string("x_75_cast_fp16")]; |
| tensor<int32, [4]> concat_29x = const()[name = string("concat_29x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_77_cast_fp16 = reshape(shape = concat_29x, x = key_states_29_cast_fp16)[name = string("x_77_cast_fp16")]; |
| tensor<int32, [4]> concat_30x = const()[name = string("concat_30x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_31_cast_fp16 = reshape(shape = concat_30x, x = value_states_29_cast_fp16)[name = string("value_states_31_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1226_cast_fp16 = mul(x = x_75_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1226_cast_fp16")]; |
| tensor<int32, [2]> var_1227_split_sizes_0 = const()[name = string("op_1227_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1227_axis_0 = const()[name = string("op_1227_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_1227_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_1227_cast_fp16_1 = split(axis = var_1227_axis_0, split_sizes = var_1227_split_sizes_0, x = x_75_cast_fp16)[name = string("op_1227_cast_fp16")]; |
| bool var_1230_interleave_0 = const()[name = string("op_1230_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_1230_cast_fp16 = concat(axis = var_1165, interleave = var_1230_interleave_0, values = (var_1227_cast_fp16_1, var_1227_cast_fp16_0))[name = string("op_1230_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1231_cast_fp16 = mul(x = var_1230_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1231_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_31_cast_fp16 = add(x = var_1226_cast_fp16, y = var_1231_cast_fp16)[name = string("query_states_31_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1233_cast_fp16 = mul(x = x_77_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1233_cast_fp16")]; |
| tensor<int32, [2]> var_1234_split_sizes_0 = const()[name = string("op_1234_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1234_axis_0 = const()[name = string("op_1234_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_1234_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_1234_cast_fp16_1 = split(axis = var_1234_axis_0, split_sizes = var_1234_split_sizes_0, x = x_77_cast_fp16)[name = string("op_1234_cast_fp16")]; |
| bool var_1237_interleave_0 = const()[name = string("op_1237_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_1237_cast_fp16 = concat(axis = var_1165, interleave = var_1237_interleave_0, values = (var_1234_cast_fp16_1, var_1234_cast_fp16_0))[name = string("op_1237_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1238_cast_fp16 = mul(x = var_1237_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1238_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_31_cast_fp16 = add(x = var_1233_cast_fp16, y = var_1238_cast_fp16)[name = string("key_states_31_cast_fp16")]; |
| tensor<int32, [2]> var_1240_split_sizes_0 = const()[name = string("op_1240_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_1240_axis_0 = const()[name = string("op_1240_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1240_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_1240_cast_fp16_1 = split(axis = var_1240_axis_0, split_sizes = var_1240_split_sizes_0, x = query_states_31_cast_fp16)[name = string("op_1240_cast_fp16")]; |
| tensor<int32, [2]> var_1242_split_sizes_0 = const()[name = string("op_1242_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1242_axis_0 = const()[name = string("op_1242_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1242_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1242_cast_fp16_1 = split(axis = var_1242_axis_0, split_sizes = var_1242_split_sizes_0, x = key_states_31_cast_fp16)[name = string("op_1242_cast_fp16")]; |
| tensor<int32, [2]> var_1244_split_sizes_0 = const()[name = string("op_1244_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1244_axis_0 = const()[name = string("op_1244_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1244_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1244_cast_fp16_1 = split(axis = var_1244_axis_0, split_sizes = var_1244_split_sizes_0, x = value_states_31_cast_fp16)[name = string("op_1244_cast_fp16")]; |
| bool attn_weights_85_transpose_x_1 = const()[name = string("attn_weights_85_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_85_transpose_y_1 = const()[name = string("attn_weights_85_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_85_cast_fp16 = matmul(transpose_x = attn_weights_85_transpose_x_1, transpose_y = attn_weights_85_transpose_y_1, x = var_1242_cast_fp16_0, y = var_1240_cast_fp16_0)[name = string("attn_weights_85_cast_fp16")]; |
| fp16 var_1248_to_fp16 = const()[name = string("op_1248_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_87_cast_fp16 = mul(x = attn_weights_85_cast_fp16, y = var_1248_to_fp16)[name = string("attn_weights_87_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_89_cast_fp16 = softmax(axis = var_1165, x = attn_weights_87_cast_fp16)[name = string("attn_weights_89_cast_fp16")]; |
| bool var_1251_transpose_x_0 = const()[name = string("op_1251_transpose_x_0"), val = bool(false)]; |
| bool var_1251_transpose_y_0 = const()[name = string("op_1251_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1251_cast_fp16 = matmul(transpose_x = var_1251_transpose_x_0, transpose_y = var_1251_transpose_y_0, x = var_1244_cast_fp16_0, y = attn_weights_89_cast_fp16)[name = string("op_1251_cast_fp16")]; |
| bool attn_weights_91_transpose_x_1 = const()[name = string("attn_weights_91_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_91_transpose_y_1 = const()[name = string("attn_weights_91_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_91_cast_fp16 = matmul(transpose_x = attn_weights_91_transpose_x_1, transpose_y = attn_weights_91_transpose_y_1, x = var_1242_cast_fp16_1, y = var_1240_cast_fp16_1)[name = string("attn_weights_91_cast_fp16")]; |
| fp16 var_1254_to_fp16 = const()[name = string("op_1254_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_93_cast_fp16 = mul(x = attn_weights_91_cast_fp16, y = var_1254_to_fp16)[name = string("attn_weights_93_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_95_cast_fp16 = softmax(axis = var_1165, x = attn_weights_93_cast_fp16)[name = string("attn_weights_95_cast_fp16")]; |
| bool attn_out_15_transpose_x_0 = const()[name = string("attn_out_15_transpose_x_0"), val = bool(false)]; |
| bool attn_out_15_transpose_y_0 = const()[name = string("attn_out_15_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_15_cast_fp16 = matmul(transpose_x = attn_out_15_transpose_x_0, transpose_y = attn_out_15_transpose_y_0, x = var_1244_cast_fp16_1, y = attn_weights_95_cast_fp16)[name = string("attn_out_15_cast_fp16")]; |
| bool attn_output_15_interleave_0 = const()[name = string("attn_output_15_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_15_cast_fp16 = concat(axis = var_1168, interleave = attn_output_15_interleave_0, values = (var_1251_cast_fp16, attn_out_15_cast_fp16))[name = string("attn_output_15_cast_fp16")]; |
| tensor<int32, [4]> concat_31x = const()[name = string("concat_31x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_79_cast_fp16 = reshape(shape = concat_31x, x = attn_output_15_cast_fp16)[name = string("x_79_cast_fp16")]; |
| string hidden_states_75_pad_type_0 = const()[name = string("hidden_states_75_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_75_strides_0 = const()[name = string("hidden_states_75_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_75_pad_0 = const()[name = string("hidden_states_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_75_dilations_0 = const()[name = string("hidden_states_75_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_75_groups_0 = const()[name = string("hidden_states_75_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247667776)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_75_cast_fp16 = conv(dilations = hidden_states_75_dilations_0, groups = hidden_states_75_groups_0, pad = hidden_states_75_pad_0, pad_type = hidden_states_75_pad_type_0, strides = hidden_states_75_strides_0, weight = var_1167_to_fp16, x = x_79_cast_fp16)[name = string("hidden_states_75_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_77_cast_fp16 = add(x = hidden_states_71_cast_fp16, y = hidden_states_75_cast_fp16)[name = string("hidden_states_77_cast_fp16")]; |
| fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1269_cast_fp16 = mul(x = hidden_states_77_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_1269_cast_fp16")]; |
| bool doubled_61_interleave_0 = const()[name = string("doubled_61_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_61_cast_fp16 = concat(axis = var_1168, interleave = doubled_61_interleave_0, values = (hidden_states_77_cast_fp16, var_1269_cast_fp16))[name = string("doubled_61_cast_fp16")]; |
| tensor<int32, [1]> out_31_axes_0 = const()[name = string("out_31_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_31_gamma_0_to_fp16 = const()[name = string("out_31_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251862144)))]; |
| fp16 var_1279_to_fp16 = const()[name = string("op_1279_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1279_to_fp16, gamma = out_31_gamma_0_to_fp16, x = doubled_61_cast_fp16)[name = string("out_31_cast_fp16")]; |
| tensor<int32, [2]> var_1290_split_sizes_0 = const()[name = string("op_1290_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1290_axis_0 = const()[name = string("op_1290_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1290_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1290_cast_fp16_1 = split(axis = var_1290_axis_0, split_sizes = var_1290_split_sizes_0, x = out_31_cast_fp16)[name = string("op_1290_cast_fp16")]; |
| string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1157_to_fp16 = const()[name = string("op_1157_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251866304)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = var_1157_to_fp16, x = var_1290_cast_fp16_0)[name = string("input_15_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1298_cast_fp16 = silu(x = input_15_cast_fp16)[name = string("op_1298_cast_fp16")]; |
| string var_1303_pad_type_0 = const()[name = string("op_1303_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1303_strides_0 = const()[name = string("op_1303_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1303_pad_0 = const()[name = string("op_1303_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1303_dilations_0 = const()[name = string("op_1303_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1303_groups_0 = const()[name = string("op_1303_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1158_to_fp16 = const()[name = string("op_1158_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260254976)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1303_cast_fp16 = conv(dilations = var_1303_dilations_0, groups = var_1303_groups_0, pad = var_1303_pad_0, pad_type = var_1303_pad_type_0, strides = var_1303_strides_0, weight = var_1158_to_fp16, x = var_1290_cast_fp16_0)[name = string("op_1303_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_83_cast_fp16 = mul(x = var_1298_cast_fp16, y = var_1303_cast_fp16)[name = string("x_83_cast_fp16")]; |
| string hidden_states_79_pad_type_0 = const()[name = string("hidden_states_79_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_79_strides_0 = const()[name = string("hidden_states_79_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_79_pad_0 = const()[name = string("hidden_states_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_79_dilations_0 = const()[name = string("hidden_states_79_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_79_groups_0 = const()[name = string("hidden_states_79_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_1159_to_fp16 = const()[name = string("op_1159_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268643648)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_79_cast_fp16 = conv(dilations = hidden_states_79_dilations_0, groups = hidden_states_79_groups_0, pad = hidden_states_79_pad_0, pad_type = hidden_states_79_pad_type_0, strides = hidden_states_79_strides_0, weight = var_1159_to_fp16, x = x_83_cast_fp16)[name = string("hidden_states_79_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_81_cast_fp16 = add(x = hidden_states_77_cast_fp16, y = hidden_states_79_cast_fp16)[name = string("hidden_states_81_cast_fp16")]; |
| int32 var_1319 = const()[name = string("op_1319"), val = int32(-2)]; |
| int32 var_1322 = const()[name = string("op_1322"), val = int32(1)]; |
| fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1333_cast_fp16 = mul(x = hidden_states_81_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_1333_cast_fp16")]; |
| bool doubled_65_interleave_0 = const()[name = string("doubled_65_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_65_cast_fp16 = concat(axis = var_1322, interleave = doubled_65_interleave_0, values = (hidden_states_81_cast_fp16, var_1333_cast_fp16))[name = string("doubled_65_cast_fp16")]; |
| tensor<int32, [1]> out_33_axes_0 = const()[name = string("out_33_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_33_gamma_0_to_fp16 = const()[name = string("out_33_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277032320)))]; |
| fp16 var_1343_to_fp16 = const()[name = string("op_1343_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1343_to_fp16, gamma = out_33_gamma_0_to_fp16, x = doubled_65_cast_fp16)[name = string("out_33_cast_fp16")]; |
| tensor<int32, [2]> var_1354_split_sizes_0 = const()[name = string("op_1354_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1354_axis_0 = const()[name = string("op_1354_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1354_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1354_cast_fp16_1 = split(axis = var_1354_axis_0, split_sizes = var_1354_split_sizes_0, x = out_33_cast_fp16)[name = string("op_1354_cast_fp16")]; |
| string query_states_33_pad_type_0 = const()[name = string("query_states_33_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_33_strides_0 = const()[name = string("query_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_33_pad_0 = const()[name = string("query_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_33_dilations_0 = const()[name = string("query_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_33_groups_0 = const()[name = string("query_states_33_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1314_to_fp16 = const()[name = string("op_1314_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277036480)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_33_cast_fp16 = conv(dilations = query_states_33_dilations_0, groups = query_states_33_groups_0, pad = query_states_33_pad_0, pad_type = query_states_33_pad_type_0, strides = query_states_33_strides_0, weight = var_1314_to_fp16, x = var_1354_cast_fp16_0)[name = string("query_states_33_cast_fp16")]; |
| string key_states_33_pad_type_0 = const()[name = string("key_states_33_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_33_strides_0 = const()[name = string("key_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_33_pad_0 = const()[name = string("key_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_33_dilations_0 = const()[name = string("key_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_33_groups_0 = const()[name = string("key_states_33_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1315_to_fp16 = const()[name = string("op_1315_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281230848)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_33_cast_fp16 = conv(dilations = key_states_33_dilations_0, groups = key_states_33_groups_0, pad = key_states_33_pad_0, pad_type = key_states_33_pad_type_0, strides = key_states_33_strides_0, weight = var_1315_to_fp16, x = var_1354_cast_fp16_0)[name = string("key_states_33_cast_fp16")]; |
| string value_states_33_pad_type_0 = const()[name = string("value_states_33_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_33_strides_0 = const()[name = string("value_states_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_33_pad_0 = const()[name = string("value_states_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_33_dilations_0 = const()[name = string("value_states_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_33_groups_0 = const()[name = string("value_states_33_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1316_to_fp16 = const()[name = string("op_1316_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281755200)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_33_cast_fp16 = conv(dilations = value_states_33_dilations_0, groups = value_states_33_groups_0, pad = value_states_33_pad_0, pad_type = value_states_33_pad_type_0, strides = value_states_33_strides_0, weight = var_1316_to_fp16, x = var_1354_cast_fp16_0)[name = string("value_states_33_cast_fp16")]; |
| tensor<int32, [4]> concat_32x = const()[name = string("concat_32x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_85_cast_fp16 = reshape(shape = concat_32x, x = query_states_33_cast_fp16)[name = string("x_85_cast_fp16")]; |
| tensor<int32, [4]> concat_33x = const()[name = string("concat_33x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_87_cast_fp16 = reshape(shape = concat_33x, x = key_states_33_cast_fp16)[name = string("x_87_cast_fp16")]; |
| tensor<int32, [4]> concat_34x = const()[name = string("concat_34x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_35_cast_fp16 = reshape(shape = concat_34x, x = value_states_33_cast_fp16)[name = string("value_states_35_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1380_cast_fp16 = mul(x = x_85_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1380_cast_fp16")]; |
| tensor<int32, [2]> var_1381_split_sizes_0 = const()[name = string("op_1381_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1381_axis_0 = const()[name = string("op_1381_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_1381_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_1381_cast_fp16_1 = split(axis = var_1381_axis_0, split_sizes = var_1381_split_sizes_0, x = x_85_cast_fp16)[name = string("op_1381_cast_fp16")]; |
| bool var_1384_interleave_0 = const()[name = string("op_1384_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_1384_cast_fp16 = concat(axis = var_1319, interleave = var_1384_interleave_0, values = (var_1381_cast_fp16_1, var_1381_cast_fp16_0))[name = string("op_1384_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1385_cast_fp16 = mul(x = var_1384_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1385_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_35_cast_fp16 = add(x = var_1380_cast_fp16, y = var_1385_cast_fp16)[name = string("query_states_35_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1387_cast_fp16 = mul(x = x_87_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1387_cast_fp16")]; |
| tensor<int32, [2]> var_1388_split_sizes_0 = const()[name = string("op_1388_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1388_axis_0 = const()[name = string("op_1388_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_1388_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_1388_cast_fp16_1 = split(axis = var_1388_axis_0, split_sizes = var_1388_split_sizes_0, x = x_87_cast_fp16)[name = string("op_1388_cast_fp16")]; |
| bool var_1391_interleave_0 = const()[name = string("op_1391_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_1391_cast_fp16 = concat(axis = var_1319, interleave = var_1391_interleave_0, values = (var_1388_cast_fp16_1, var_1388_cast_fp16_0))[name = string("op_1391_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1392_cast_fp16 = mul(x = var_1391_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1392_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_35_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1392_cast_fp16)[name = string("key_states_35_cast_fp16")]; |
| tensor<int32, [2]> var_1394_split_sizes_0 = const()[name = string("op_1394_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_1394_axis_0 = const()[name = string("op_1394_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1394_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_1394_cast_fp16_1 = split(axis = var_1394_axis_0, split_sizes = var_1394_split_sizes_0, x = query_states_35_cast_fp16)[name = string("op_1394_cast_fp16")]; |
| tensor<int32, [2]> var_1396_split_sizes_0 = const()[name = string("op_1396_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1396_axis_0 = const()[name = string("op_1396_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1396_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1396_cast_fp16_1 = split(axis = var_1396_axis_0, split_sizes = var_1396_split_sizes_0, x = key_states_35_cast_fp16)[name = string("op_1396_cast_fp16")]; |
| tensor<int32, [2]> var_1398_split_sizes_0 = const()[name = string("op_1398_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1398_axis_0 = const()[name = string("op_1398_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1398_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1398_cast_fp16_1 = split(axis = var_1398_axis_0, split_sizes = var_1398_split_sizes_0, x = value_states_35_cast_fp16)[name = string("op_1398_cast_fp16")]; |
| bool attn_weights_97_transpose_x_1 = const()[name = string("attn_weights_97_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_97_transpose_y_1 = const()[name = string("attn_weights_97_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_97_cast_fp16 = matmul(transpose_x = attn_weights_97_transpose_x_1, transpose_y = attn_weights_97_transpose_y_1, x = var_1396_cast_fp16_0, y = var_1394_cast_fp16_0)[name = string("attn_weights_97_cast_fp16")]; |
| fp16 var_1402_to_fp16 = const()[name = string("op_1402_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_99_cast_fp16 = mul(x = attn_weights_97_cast_fp16, y = var_1402_to_fp16)[name = string("attn_weights_99_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_101_cast_fp16 = softmax(axis = var_1319, x = attn_weights_99_cast_fp16)[name = string("attn_weights_101_cast_fp16")]; |
| bool var_1405_transpose_x_0 = const()[name = string("op_1405_transpose_x_0"), val = bool(false)]; |
| bool var_1405_transpose_y_0 = const()[name = string("op_1405_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1405_cast_fp16 = matmul(transpose_x = var_1405_transpose_x_0, transpose_y = var_1405_transpose_y_0, x = var_1398_cast_fp16_0, y = attn_weights_101_cast_fp16)[name = string("op_1405_cast_fp16")]; |
| bool attn_weights_103_transpose_x_1 = const()[name = string("attn_weights_103_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_103_transpose_y_1 = const()[name = string("attn_weights_103_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_103_cast_fp16 = matmul(transpose_x = attn_weights_103_transpose_x_1, transpose_y = attn_weights_103_transpose_y_1, x = var_1396_cast_fp16_1, y = var_1394_cast_fp16_1)[name = string("attn_weights_103_cast_fp16")]; |
| fp16 var_1408_to_fp16 = const()[name = string("op_1408_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_105_cast_fp16 = mul(x = attn_weights_103_cast_fp16, y = var_1408_to_fp16)[name = string("attn_weights_105_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_107_cast_fp16 = softmax(axis = var_1319, x = attn_weights_105_cast_fp16)[name = string("attn_weights_107_cast_fp16")]; |
| bool attn_out_17_transpose_x_0 = const()[name = string("attn_out_17_transpose_x_0"), val = bool(false)]; |
| bool attn_out_17_transpose_y_0 = const()[name = string("attn_out_17_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_17_cast_fp16 = matmul(transpose_x = attn_out_17_transpose_x_0, transpose_y = attn_out_17_transpose_y_0, x = var_1398_cast_fp16_1, y = attn_weights_107_cast_fp16)[name = string("attn_out_17_cast_fp16")]; |
| bool attn_output_17_interleave_0 = const()[name = string("attn_output_17_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_17_cast_fp16 = concat(axis = var_1322, interleave = attn_output_17_interleave_0, values = (var_1405_cast_fp16, attn_out_17_cast_fp16))[name = string("attn_output_17_cast_fp16")]; |
| tensor<int32, [4]> concat_35x = const()[name = string("concat_35x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_89_cast_fp16 = reshape(shape = concat_35x, x = attn_output_17_cast_fp16)[name = string("x_89_cast_fp16")]; |
| string hidden_states_85_pad_type_0 = const()[name = string("hidden_states_85_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_85_strides_0 = const()[name = string("hidden_states_85_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_85_pad_0 = const()[name = string("hidden_states_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_85_dilations_0 = const()[name = string("hidden_states_85_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_85_groups_0 = const()[name = string("hidden_states_85_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_1321_to_fp16 = const()[name = string("op_1321_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282279552)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_85_cast_fp16 = conv(dilations = hidden_states_85_dilations_0, groups = hidden_states_85_groups_0, pad = hidden_states_85_pad_0, pad_type = hidden_states_85_pad_type_0, strides = hidden_states_85_strides_0, weight = var_1321_to_fp16, x = x_89_cast_fp16)[name = string("hidden_states_85_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_87_cast_fp16 = add(x = hidden_states_81_cast_fp16, y = hidden_states_85_cast_fp16)[name = string("hidden_states_87_cast_fp16")]; |
| fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1423_cast_fp16 = mul(x = hidden_states_87_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_1423_cast_fp16")]; |
| bool doubled_69_interleave_0 = const()[name = string("doubled_69_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_69_cast_fp16 = concat(axis = var_1322, interleave = doubled_69_interleave_0, values = (hidden_states_87_cast_fp16, var_1423_cast_fp16))[name = string("doubled_69_cast_fp16")]; |
| tensor<int32, [1]> out_35_axes_0 = const()[name = string("out_35_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_35_gamma_0_to_fp16 = const()[name = string("out_35_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286473920)))]; |
| fp16 var_1433_to_fp16 = const()[name = string("op_1433_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1433_to_fp16, gamma = out_35_gamma_0_to_fp16, x = doubled_69_cast_fp16)[name = string("out_35_cast_fp16")]; |
| tensor<int32, [2]> var_1444_split_sizes_0 = const()[name = string("op_1444_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1444_axis_0 = const()[name = string("op_1444_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1444_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1444_cast_fp16_1 = split(axis = var_1444_axis_0, split_sizes = var_1444_split_sizes_0, x = out_35_cast_fp16)[name = string("op_1444_cast_fp16")]; |
| string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1311_to_fp16 = const()[name = string("op_1311_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286478080)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_17_cast_fp16 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = var_1311_to_fp16, x = var_1444_cast_fp16_0)[name = string("input_17_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1452_cast_fp16 = silu(x = input_17_cast_fp16)[name = string("op_1452_cast_fp16")]; |
| string var_1457_pad_type_0 = const()[name = string("op_1457_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1457_strides_0 = const()[name = string("op_1457_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1457_pad_0 = const()[name = string("op_1457_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1457_dilations_0 = const()[name = string("op_1457_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1457_groups_0 = const()[name = string("op_1457_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1312_to_fp16 = const()[name = string("op_1312_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294866752)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1457_cast_fp16 = conv(dilations = var_1457_dilations_0, groups = var_1457_groups_0, pad = var_1457_pad_0, pad_type = var_1457_pad_type_0, strides = var_1457_strides_0, weight = var_1312_to_fp16, x = var_1444_cast_fp16_0)[name = string("op_1457_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_93_cast_fp16 = mul(x = var_1452_cast_fp16, y = var_1457_cast_fp16)[name = string("x_93_cast_fp16")]; |
| string hidden_states_89_pad_type_0 = const()[name = string("hidden_states_89_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_89_strides_0 = const()[name = string("hidden_states_89_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_89_pad_0 = const()[name = string("hidden_states_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_89_dilations_0 = const()[name = string("hidden_states_89_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_89_groups_0 = const()[name = string("hidden_states_89_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_1313_to_fp16 = const()[name = string("op_1313_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303255424)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_89_cast_fp16 = conv(dilations = hidden_states_89_dilations_0, groups = hidden_states_89_groups_0, pad = hidden_states_89_pad_0, pad_type = hidden_states_89_pad_type_0, strides = hidden_states_89_strides_0, weight = var_1313_to_fp16, x = x_93_cast_fp16)[name = string("hidden_states_89_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_91_cast_fp16 = add(x = hidden_states_87_cast_fp16, y = hidden_states_89_cast_fp16)[name = string("hidden_states_91_cast_fp16")]; |
| int32 var_1473 = const()[name = string("op_1473"), val = int32(-2)]; |
| int32 var_1476 = const()[name = string("op_1476"), val = int32(1)]; |
| fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1487_cast_fp16 = mul(x = hidden_states_91_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_1487_cast_fp16")]; |
| bool doubled_73_interleave_0 = const()[name = string("doubled_73_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_73_cast_fp16 = concat(axis = var_1476, interleave = doubled_73_interleave_0, values = (hidden_states_91_cast_fp16, var_1487_cast_fp16))[name = string("doubled_73_cast_fp16")]; |
| tensor<int32, [1]> out_37_axes_0 = const()[name = string("out_37_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_37_gamma_0_to_fp16 = const()[name = string("out_37_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311644096)))]; |
| fp16 var_1497_to_fp16 = const()[name = string("op_1497_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1497_to_fp16, gamma = out_37_gamma_0_to_fp16, x = doubled_73_cast_fp16)[name = string("out_37_cast_fp16")]; |
| tensor<int32, [2]> var_1508_split_sizes_0 = const()[name = string("op_1508_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1508_axis_0 = const()[name = string("op_1508_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1508_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1508_cast_fp16_1 = split(axis = var_1508_axis_0, split_sizes = var_1508_split_sizes_0, x = out_37_cast_fp16)[name = string("op_1508_cast_fp16")]; |
| string query_states_37_pad_type_0 = const()[name = string("query_states_37_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_37_strides_0 = const()[name = string("query_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_37_pad_0 = const()[name = string("query_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_37_dilations_0 = const()[name = string("query_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_37_groups_0 = const()[name = string("query_states_37_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1468_to_fp16 = const()[name = string("op_1468_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311648256)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_37_cast_fp16 = conv(dilations = query_states_37_dilations_0, groups = query_states_37_groups_0, pad = query_states_37_pad_0, pad_type = query_states_37_pad_type_0, strides = query_states_37_strides_0, weight = var_1468_to_fp16, x = var_1508_cast_fp16_0)[name = string("query_states_37_cast_fp16")]; |
| string key_states_37_pad_type_0 = const()[name = string("key_states_37_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_37_strides_0 = const()[name = string("key_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_37_pad_0 = const()[name = string("key_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_37_dilations_0 = const()[name = string("key_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_37_groups_0 = const()[name = string("key_states_37_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315842624)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_37_cast_fp16 = conv(dilations = key_states_37_dilations_0, groups = key_states_37_groups_0, pad = key_states_37_pad_0, pad_type = key_states_37_pad_type_0, strides = key_states_37_strides_0, weight = var_1469_to_fp16, x = var_1508_cast_fp16_0)[name = string("key_states_37_cast_fp16")]; |
| string value_states_37_pad_type_0 = const()[name = string("value_states_37_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_37_strides_0 = const()[name = string("value_states_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_37_pad_0 = const()[name = string("value_states_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_37_dilations_0 = const()[name = string("value_states_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_37_groups_0 = const()[name = string("value_states_37_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1470_to_fp16 = const()[name = string("op_1470_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316366976)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_37_cast_fp16 = conv(dilations = value_states_37_dilations_0, groups = value_states_37_groups_0, pad = value_states_37_pad_0, pad_type = value_states_37_pad_type_0, strides = value_states_37_strides_0, weight = var_1470_to_fp16, x = var_1508_cast_fp16_0)[name = string("value_states_37_cast_fp16")]; |
| tensor<int32, [4]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_95_cast_fp16 = reshape(shape = concat_36x, x = query_states_37_cast_fp16)[name = string("x_95_cast_fp16")]; |
| tensor<int32, [4]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_97_cast_fp16 = reshape(shape = concat_37x, x = key_states_37_cast_fp16)[name = string("x_97_cast_fp16")]; |
| tensor<int32, [4]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_39_cast_fp16 = reshape(shape = concat_38x, x = value_states_37_cast_fp16)[name = string("value_states_39_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1534_cast_fp16 = mul(x = x_95_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1534_cast_fp16")]; |
| tensor<int32, [2]> var_1535_split_sizes_0 = const()[name = string("op_1535_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1535_axis_0 = const()[name = string("op_1535_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_1535_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_1535_cast_fp16_1 = split(axis = var_1535_axis_0, split_sizes = var_1535_split_sizes_0, x = x_95_cast_fp16)[name = string("op_1535_cast_fp16")]; |
| bool var_1538_interleave_0 = const()[name = string("op_1538_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_1538_cast_fp16 = concat(axis = var_1473, interleave = var_1538_interleave_0, values = (var_1535_cast_fp16_1, var_1535_cast_fp16_0))[name = string("op_1538_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1539_cast_fp16 = mul(x = var_1538_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1539_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_39_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1539_cast_fp16)[name = string("query_states_39_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1541_cast_fp16 = mul(x = x_97_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1541_cast_fp16")]; |
| tensor<int32, [2]> var_1542_split_sizes_0 = const()[name = string("op_1542_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1542_axis_0 = const()[name = string("op_1542_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_1542_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_1542_cast_fp16_1 = split(axis = var_1542_axis_0, split_sizes = var_1542_split_sizes_0, x = x_97_cast_fp16)[name = string("op_1542_cast_fp16")]; |
| bool var_1545_interleave_0 = const()[name = string("op_1545_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_1545_cast_fp16 = concat(axis = var_1473, interleave = var_1545_interleave_0, values = (var_1542_cast_fp16_1, var_1542_cast_fp16_0))[name = string("op_1545_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1546_cast_fp16 = mul(x = var_1545_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1546_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_39_cast_fp16 = add(x = var_1541_cast_fp16, y = var_1546_cast_fp16)[name = string("key_states_39_cast_fp16")]; |
| tensor<int32, [2]> var_1548_split_sizes_0 = const()[name = string("op_1548_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_1548_axis_0 = const()[name = string("op_1548_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1548_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_1548_cast_fp16_1 = split(axis = var_1548_axis_0, split_sizes = var_1548_split_sizes_0, x = query_states_39_cast_fp16)[name = string("op_1548_cast_fp16")]; |
| tensor<int32, [2]> var_1550_split_sizes_0 = const()[name = string("op_1550_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1550_axis_0 = const()[name = string("op_1550_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1550_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1550_cast_fp16_1 = split(axis = var_1550_axis_0, split_sizes = var_1550_split_sizes_0, x = key_states_39_cast_fp16)[name = string("op_1550_cast_fp16")]; |
| tensor<int32, [2]> var_1552_split_sizes_0 = const()[name = string("op_1552_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1552_axis_0 = const()[name = string("op_1552_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1552_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1552_cast_fp16_1 = split(axis = var_1552_axis_0, split_sizes = var_1552_split_sizes_0, x = value_states_39_cast_fp16)[name = string("op_1552_cast_fp16")]; |
| bool attn_weights_109_transpose_x_1 = const()[name = string("attn_weights_109_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_109_transpose_y_1 = const()[name = string("attn_weights_109_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_109_cast_fp16 = matmul(transpose_x = attn_weights_109_transpose_x_1, transpose_y = attn_weights_109_transpose_y_1, x = var_1550_cast_fp16_0, y = var_1548_cast_fp16_0)[name = string("attn_weights_109_cast_fp16")]; |
| fp16 var_1556_to_fp16 = const()[name = string("op_1556_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_111_cast_fp16 = mul(x = attn_weights_109_cast_fp16, y = var_1556_to_fp16)[name = string("attn_weights_111_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_113_cast_fp16 = softmax(axis = var_1473, x = attn_weights_111_cast_fp16)[name = string("attn_weights_113_cast_fp16")]; |
| bool var_1559_transpose_x_0 = const()[name = string("op_1559_transpose_x_0"), val = bool(false)]; |
| bool var_1559_transpose_y_0 = const()[name = string("op_1559_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1559_cast_fp16 = matmul(transpose_x = var_1559_transpose_x_0, transpose_y = var_1559_transpose_y_0, x = var_1552_cast_fp16_0, y = attn_weights_113_cast_fp16)[name = string("op_1559_cast_fp16")]; |
| bool attn_weights_115_transpose_x_1 = const()[name = string("attn_weights_115_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_115_transpose_y_1 = const()[name = string("attn_weights_115_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_115_cast_fp16 = matmul(transpose_x = attn_weights_115_transpose_x_1, transpose_y = attn_weights_115_transpose_y_1, x = var_1550_cast_fp16_1, y = var_1548_cast_fp16_1)[name = string("attn_weights_115_cast_fp16")]; |
| fp16 var_1562_to_fp16 = const()[name = string("op_1562_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_117_cast_fp16 = mul(x = attn_weights_115_cast_fp16, y = var_1562_to_fp16)[name = string("attn_weights_117_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_119_cast_fp16 = softmax(axis = var_1473, x = attn_weights_117_cast_fp16)[name = string("attn_weights_119_cast_fp16")]; |
| bool attn_out_19_transpose_x_0 = const()[name = string("attn_out_19_transpose_x_0"), val = bool(false)]; |
| bool attn_out_19_transpose_y_0 = const()[name = string("attn_out_19_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_19_cast_fp16 = matmul(transpose_x = attn_out_19_transpose_x_0, transpose_y = attn_out_19_transpose_y_0, x = var_1552_cast_fp16_1, y = attn_weights_119_cast_fp16)[name = string("attn_out_19_cast_fp16")]; |
| bool attn_output_19_interleave_0 = const()[name = string("attn_output_19_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_19_cast_fp16 = concat(axis = var_1476, interleave = attn_output_19_interleave_0, values = (var_1559_cast_fp16, attn_out_19_cast_fp16))[name = string("attn_output_19_cast_fp16")]; |
| tensor<int32, [4]> concat_39x = const()[name = string("concat_39x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_99_cast_fp16 = reshape(shape = concat_39x, x = attn_output_19_cast_fp16)[name = string("x_99_cast_fp16")]; |
| string hidden_states_95_pad_type_0 = const()[name = string("hidden_states_95_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_95_strides_0 = const()[name = string("hidden_states_95_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_95_pad_0 = const()[name = string("hidden_states_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_95_dilations_0 = const()[name = string("hidden_states_95_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_95_groups_0 = const()[name = string("hidden_states_95_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_1475_to_fp16 = const()[name = string("op_1475_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316891328)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_95_cast_fp16 = conv(dilations = hidden_states_95_dilations_0, groups = hidden_states_95_groups_0, pad = hidden_states_95_pad_0, pad_type = hidden_states_95_pad_type_0, strides = hidden_states_95_strides_0, weight = var_1475_to_fp16, x = x_99_cast_fp16)[name = string("hidden_states_95_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_97_cast_fp16 = add(x = hidden_states_91_cast_fp16, y = hidden_states_95_cast_fp16)[name = string("hidden_states_97_cast_fp16")]; |
| fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1577_cast_fp16 = mul(x = hidden_states_97_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_1577_cast_fp16")]; |
| bool doubled_77_interleave_0 = const()[name = string("doubled_77_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_77_cast_fp16 = concat(axis = var_1476, interleave = doubled_77_interleave_0, values = (hidden_states_97_cast_fp16, var_1577_cast_fp16))[name = string("doubled_77_cast_fp16")]; |
| tensor<int32, [1]> out_39_axes_0 = const()[name = string("out_39_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_39_gamma_0_to_fp16 = const()[name = string("out_39_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321085696)))]; |
| fp16 var_1587_to_fp16 = const()[name = string("op_1587_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1587_to_fp16, gamma = out_39_gamma_0_to_fp16, x = doubled_77_cast_fp16)[name = string("out_39_cast_fp16")]; |
| tensor<int32, [2]> var_1598_split_sizes_0 = const()[name = string("op_1598_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1598_axis_0 = const()[name = string("op_1598_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1598_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1598_cast_fp16_1 = split(axis = var_1598_axis_0, split_sizes = var_1598_split_sizes_0, x = out_39_cast_fp16)[name = string("op_1598_cast_fp16")]; |
| string input_19_pad_type_0 = const()[name = string("input_19_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_19_strides_0 = const()[name = string("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_19_pad_0 = const()[name = string("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_19_dilations_0 = const()[name = string("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_19_groups_0 = const()[name = string("input_19_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1465_to_fp16 = const()[name = string("op_1465_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321089856)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_19_cast_fp16 = conv(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 = var_1465_to_fp16, x = var_1598_cast_fp16_0)[name = string("input_19_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1606_cast_fp16 = silu(x = input_19_cast_fp16)[name = string("op_1606_cast_fp16")]; |
| string var_1611_pad_type_0 = const()[name = string("op_1611_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1611_strides_0 = const()[name = string("op_1611_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1611_pad_0 = const()[name = string("op_1611_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1611_dilations_0 = const()[name = string("op_1611_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1611_groups_0 = const()[name = string("op_1611_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1466_to_fp16 = const()[name = string("op_1466_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329478528)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1611_cast_fp16 = conv(dilations = var_1611_dilations_0, groups = var_1611_groups_0, pad = var_1611_pad_0, pad_type = var_1611_pad_type_0, strides = var_1611_strides_0, weight = var_1466_to_fp16, x = var_1598_cast_fp16_0)[name = string("op_1611_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_103_cast_fp16 = mul(x = var_1606_cast_fp16, y = var_1611_cast_fp16)[name = string("x_103_cast_fp16")]; |
| string hidden_states_99_pad_type_0 = const()[name = string("hidden_states_99_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_99_strides_0 = const()[name = string("hidden_states_99_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_99_pad_0 = const()[name = string("hidden_states_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_99_dilations_0 = const()[name = string("hidden_states_99_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_99_groups_0 = const()[name = string("hidden_states_99_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_1467_to_fp16 = const()[name = string("op_1467_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337867200)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_99_cast_fp16 = conv(dilations = hidden_states_99_dilations_0, groups = hidden_states_99_groups_0, pad = hidden_states_99_pad_0, pad_type = hidden_states_99_pad_type_0, strides = hidden_states_99_strides_0, weight = var_1467_to_fp16, x = x_103_cast_fp16)[name = string("hidden_states_99_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_101_cast_fp16 = add(x = hidden_states_97_cast_fp16, y = hidden_states_99_cast_fp16)[name = string("hidden_states_101_cast_fp16")]; |
| int32 var_1627 = const()[name = string("op_1627"), val = int32(-2)]; |
| int32 var_1630 = const()[name = string("op_1630"), val = int32(1)]; |
| fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1641_cast_fp16 = mul(x = hidden_states_101_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_1641_cast_fp16")]; |
| bool doubled_81_interleave_0 = const()[name = string("doubled_81_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_81_cast_fp16 = concat(axis = var_1630, interleave = doubled_81_interleave_0, values = (hidden_states_101_cast_fp16, var_1641_cast_fp16))[name = string("doubled_81_cast_fp16")]; |
| tensor<int32, [1]> out_41_axes_0 = const()[name = string("out_41_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_41_gamma_0_to_fp16 = const()[name = string("out_41_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346255872)))]; |
| fp16 var_1651_to_fp16 = const()[name = string("op_1651_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1651_to_fp16, gamma = out_41_gamma_0_to_fp16, x = doubled_81_cast_fp16)[name = string("out_41_cast_fp16")]; |
| tensor<int32, [2]> var_1662_split_sizes_0 = const()[name = string("op_1662_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1662_axis_0 = const()[name = string("op_1662_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1662_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1662_cast_fp16_1 = split(axis = var_1662_axis_0, split_sizes = var_1662_split_sizes_0, x = out_41_cast_fp16)[name = string("op_1662_cast_fp16")]; |
| string query_states_41_pad_type_0 = const()[name = string("query_states_41_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_41_strides_0 = const()[name = string("query_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_41_pad_0 = const()[name = string("query_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_41_dilations_0 = const()[name = string("query_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_41_groups_0 = const()[name = string("query_states_41_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1622_to_fp16 = const()[name = string("op_1622_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346260032)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_41_cast_fp16 = conv(dilations = query_states_41_dilations_0, groups = query_states_41_groups_0, pad = query_states_41_pad_0, pad_type = query_states_41_pad_type_0, strides = query_states_41_strides_0, weight = var_1622_to_fp16, x = var_1662_cast_fp16_0)[name = string("query_states_41_cast_fp16")]; |
| string key_states_41_pad_type_0 = const()[name = string("key_states_41_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_41_strides_0 = const()[name = string("key_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_41_pad_0 = const()[name = string("key_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_41_dilations_0 = const()[name = string("key_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_41_groups_0 = const()[name = string("key_states_41_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1623_to_fp16 = const()[name = string("op_1623_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350454400)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_41_cast_fp16 = conv(dilations = key_states_41_dilations_0, groups = key_states_41_groups_0, pad = key_states_41_pad_0, pad_type = key_states_41_pad_type_0, strides = key_states_41_strides_0, weight = var_1623_to_fp16, x = var_1662_cast_fp16_0)[name = string("key_states_41_cast_fp16")]; |
| string value_states_41_pad_type_0 = const()[name = string("value_states_41_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_41_strides_0 = const()[name = string("value_states_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_41_pad_0 = const()[name = string("value_states_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_41_dilations_0 = const()[name = string("value_states_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_41_groups_0 = const()[name = string("value_states_41_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1624_to_fp16 = const()[name = string("op_1624_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350978752)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_41_cast_fp16 = conv(dilations = value_states_41_dilations_0, groups = value_states_41_groups_0, pad = value_states_41_pad_0, pad_type = value_states_41_pad_type_0, strides = value_states_41_strides_0, weight = var_1624_to_fp16, x = var_1662_cast_fp16_0)[name = string("value_states_41_cast_fp16")]; |
| tensor<int32, [4]> concat_40x = const()[name = string("concat_40x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_105_cast_fp16 = reshape(shape = concat_40x, x = query_states_41_cast_fp16)[name = string("x_105_cast_fp16")]; |
| tensor<int32, [4]> concat_41x = const()[name = string("concat_41x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_107_cast_fp16 = reshape(shape = concat_41x, x = key_states_41_cast_fp16)[name = string("x_107_cast_fp16")]; |
| tensor<int32, [4]> concat_42x = const()[name = string("concat_42x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_43_cast_fp16 = reshape(shape = concat_42x, x = value_states_41_cast_fp16)[name = string("value_states_43_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1688_cast_fp16 = mul(x = x_105_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1688_cast_fp16")]; |
| tensor<int32, [2]> var_1689_split_sizes_0 = const()[name = string("op_1689_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1689_axis_0 = const()[name = string("op_1689_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_1689_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_1689_cast_fp16_1 = split(axis = var_1689_axis_0, split_sizes = var_1689_split_sizes_0, x = x_105_cast_fp16)[name = string("op_1689_cast_fp16")]; |
| bool var_1692_interleave_0 = const()[name = string("op_1692_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_1692_cast_fp16 = concat(axis = var_1627, interleave = var_1692_interleave_0, values = (var_1689_cast_fp16_1, var_1689_cast_fp16_0))[name = string("op_1692_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1693_cast_fp16 = mul(x = var_1692_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1693_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_43_cast_fp16 = add(x = var_1688_cast_fp16, y = var_1693_cast_fp16)[name = string("query_states_43_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1695_cast_fp16 = mul(x = x_107_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1695_cast_fp16")]; |
| tensor<int32, [2]> var_1696_split_sizes_0 = const()[name = string("op_1696_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1696_axis_0 = const()[name = string("op_1696_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_1696_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_1696_cast_fp16_1 = split(axis = var_1696_axis_0, split_sizes = var_1696_split_sizes_0, x = x_107_cast_fp16)[name = string("op_1696_cast_fp16")]; |
| bool var_1699_interleave_0 = const()[name = string("op_1699_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_1699_cast_fp16 = concat(axis = var_1627, interleave = var_1699_interleave_0, values = (var_1696_cast_fp16_1, var_1696_cast_fp16_0))[name = string("op_1699_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1700_cast_fp16 = mul(x = var_1699_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1700_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_43_cast_fp16 = add(x = var_1695_cast_fp16, y = var_1700_cast_fp16)[name = string("key_states_43_cast_fp16")]; |
| tensor<int32, [2]> var_1702_split_sizes_0 = const()[name = string("op_1702_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_1702_axis_0 = const()[name = string("op_1702_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1702_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_1702_cast_fp16_1 = split(axis = var_1702_axis_0, split_sizes = var_1702_split_sizes_0, x = query_states_43_cast_fp16)[name = string("op_1702_cast_fp16")]; |
| tensor<int32, [2]> var_1704_split_sizes_0 = const()[name = string("op_1704_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1704_axis_0 = const()[name = string("op_1704_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1704_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1704_cast_fp16_1 = split(axis = var_1704_axis_0, split_sizes = var_1704_split_sizes_0, x = key_states_43_cast_fp16)[name = string("op_1704_cast_fp16")]; |
| tensor<int32, [2]> var_1706_split_sizes_0 = const()[name = string("op_1706_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1706_axis_0 = const()[name = string("op_1706_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1706_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1706_cast_fp16_1 = split(axis = var_1706_axis_0, split_sizes = var_1706_split_sizes_0, x = value_states_43_cast_fp16)[name = string("op_1706_cast_fp16")]; |
| bool attn_weights_121_transpose_x_1 = const()[name = string("attn_weights_121_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_121_transpose_y_1 = const()[name = string("attn_weights_121_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_121_cast_fp16 = matmul(transpose_x = attn_weights_121_transpose_x_1, transpose_y = attn_weights_121_transpose_y_1, x = var_1704_cast_fp16_0, y = var_1702_cast_fp16_0)[name = string("attn_weights_121_cast_fp16")]; |
| fp16 var_1710_to_fp16 = const()[name = string("op_1710_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_123_cast_fp16 = mul(x = attn_weights_121_cast_fp16, y = var_1710_to_fp16)[name = string("attn_weights_123_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_125_cast_fp16 = softmax(axis = var_1627, x = attn_weights_123_cast_fp16)[name = string("attn_weights_125_cast_fp16")]; |
| bool var_1713_transpose_x_0 = const()[name = string("op_1713_transpose_x_0"), val = bool(false)]; |
| bool var_1713_transpose_y_0 = const()[name = string("op_1713_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1713_cast_fp16 = matmul(transpose_x = var_1713_transpose_x_0, transpose_y = var_1713_transpose_y_0, x = var_1706_cast_fp16_0, y = attn_weights_125_cast_fp16)[name = string("op_1713_cast_fp16")]; |
| bool attn_weights_127_transpose_x_1 = const()[name = string("attn_weights_127_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_127_transpose_y_1 = const()[name = string("attn_weights_127_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_127_cast_fp16 = matmul(transpose_x = attn_weights_127_transpose_x_1, transpose_y = attn_weights_127_transpose_y_1, x = var_1704_cast_fp16_1, y = var_1702_cast_fp16_1)[name = string("attn_weights_127_cast_fp16")]; |
| fp16 var_1716_to_fp16 = const()[name = string("op_1716_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_129_cast_fp16 = mul(x = attn_weights_127_cast_fp16, y = var_1716_to_fp16)[name = string("attn_weights_129_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_131_cast_fp16 = softmax(axis = var_1627, x = attn_weights_129_cast_fp16)[name = string("attn_weights_131_cast_fp16")]; |
| bool attn_out_21_transpose_x_0 = const()[name = string("attn_out_21_transpose_x_0"), val = bool(false)]; |
| bool attn_out_21_transpose_y_0 = const()[name = string("attn_out_21_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_21_cast_fp16 = matmul(transpose_x = attn_out_21_transpose_x_0, transpose_y = attn_out_21_transpose_y_0, x = var_1706_cast_fp16_1, y = attn_weights_131_cast_fp16)[name = string("attn_out_21_cast_fp16")]; |
| bool attn_output_21_interleave_0 = const()[name = string("attn_output_21_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_21_cast_fp16 = concat(axis = var_1630, interleave = attn_output_21_interleave_0, values = (var_1713_cast_fp16, attn_out_21_cast_fp16))[name = string("attn_output_21_cast_fp16")]; |
| tensor<int32, [4]> concat_43x = const()[name = string("concat_43x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_109_cast_fp16 = reshape(shape = concat_43x, x = attn_output_21_cast_fp16)[name = string("x_109_cast_fp16")]; |
| string hidden_states_105_pad_type_0 = const()[name = string("hidden_states_105_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_105_strides_0 = const()[name = string("hidden_states_105_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_105_pad_0 = const()[name = string("hidden_states_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_105_dilations_0 = const()[name = string("hidden_states_105_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_105_groups_0 = const()[name = string("hidden_states_105_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_1629_to_fp16 = const()[name = string("op_1629_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351503104)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_105_cast_fp16 = conv(dilations = hidden_states_105_dilations_0, groups = hidden_states_105_groups_0, pad = hidden_states_105_pad_0, pad_type = hidden_states_105_pad_type_0, strides = hidden_states_105_strides_0, weight = var_1629_to_fp16, x = x_109_cast_fp16)[name = string("hidden_states_105_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_107_cast_fp16 = add(x = hidden_states_101_cast_fp16, y = hidden_states_105_cast_fp16)[name = string("hidden_states_107_cast_fp16")]; |
| fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1731_cast_fp16 = mul(x = hidden_states_107_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_1731_cast_fp16")]; |
| bool doubled_85_interleave_0 = const()[name = string("doubled_85_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_85_cast_fp16 = concat(axis = var_1630, interleave = doubled_85_interleave_0, values = (hidden_states_107_cast_fp16, var_1731_cast_fp16))[name = string("doubled_85_cast_fp16")]; |
| tensor<int32, [1]> out_43_axes_0 = const()[name = string("out_43_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_43_gamma_0_to_fp16 = const()[name = string("out_43_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355697472)))]; |
| fp16 var_1741_to_fp16 = const()[name = string("op_1741_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1741_to_fp16, gamma = out_43_gamma_0_to_fp16, x = doubled_85_cast_fp16)[name = string("out_43_cast_fp16")]; |
| tensor<int32, [2]> var_1752_split_sizes_0 = const()[name = string("op_1752_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1752_axis_0 = const()[name = string("op_1752_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1752_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1752_cast_fp16_1 = split(axis = var_1752_axis_0, split_sizes = var_1752_split_sizes_0, x = out_43_cast_fp16)[name = string("op_1752_cast_fp16")]; |
| string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1619_to_fp16 = const()[name = string("op_1619_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355701632)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = var_1619_to_fp16, x = var_1752_cast_fp16_0)[name = string("input_21_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1760_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("op_1760_cast_fp16")]; |
| string var_1765_pad_type_0 = const()[name = string("op_1765_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1765_strides_0 = const()[name = string("op_1765_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1765_pad_0 = const()[name = string("op_1765_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1765_dilations_0 = const()[name = string("op_1765_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1765_groups_0 = const()[name = string("op_1765_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364090304)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1765_cast_fp16 = conv(dilations = var_1765_dilations_0, groups = var_1765_groups_0, pad = var_1765_pad_0, pad_type = var_1765_pad_type_0, strides = var_1765_strides_0, weight = var_1620_to_fp16, x = var_1752_cast_fp16_0)[name = string("op_1765_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_113_cast_fp16 = mul(x = var_1760_cast_fp16, y = var_1765_cast_fp16)[name = string("x_113_cast_fp16")]; |
| string hidden_states_109_pad_type_0 = const()[name = string("hidden_states_109_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_109_strides_0 = const()[name = string("hidden_states_109_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_109_pad_0 = const()[name = string("hidden_states_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_109_dilations_0 = const()[name = string("hidden_states_109_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_109_groups_0 = const()[name = string("hidden_states_109_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_1621_to_fp16 = const()[name = string("op_1621_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372478976)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_109_cast_fp16 = conv(dilations = hidden_states_109_dilations_0, groups = hidden_states_109_groups_0, pad = hidden_states_109_pad_0, pad_type = hidden_states_109_pad_type_0, strides = hidden_states_109_strides_0, weight = var_1621_to_fp16, x = x_113_cast_fp16)[name = string("hidden_states_109_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_111_cast_fp16 = add(x = hidden_states_107_cast_fp16, y = hidden_states_109_cast_fp16)[name = string("hidden_states_111_cast_fp16")]; |
| int32 var_1781 = const()[name = string("op_1781"), val = int32(-2)]; |
| int32 var_1784 = const()[name = string("op_1784"), val = int32(1)]; |
| fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1795_cast_fp16 = mul(x = hidden_states_111_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_1795_cast_fp16")]; |
| bool doubled_89_interleave_0 = const()[name = string("doubled_89_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_89_cast_fp16 = concat(axis = var_1784, interleave = doubled_89_interleave_0, values = (hidden_states_111_cast_fp16, var_1795_cast_fp16))[name = string("doubled_89_cast_fp16")]; |
| tensor<int32, [1]> out_45_axes_0 = const()[name = string("out_45_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_45_gamma_0_to_fp16 = const()[name = string("out_45_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380867648)))]; |
| fp16 var_1805_to_fp16 = const()[name = string("op_1805_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1805_to_fp16, gamma = out_45_gamma_0_to_fp16, x = doubled_89_cast_fp16)[name = string("out_45_cast_fp16")]; |
| tensor<int32, [2]> var_1816_split_sizes_0 = const()[name = string("op_1816_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1816_axis_0 = const()[name = string("op_1816_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1816_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1816_cast_fp16_1 = split(axis = var_1816_axis_0, split_sizes = var_1816_split_sizes_0, x = out_45_cast_fp16)[name = string("op_1816_cast_fp16")]; |
| string query_states_45_pad_type_0 = const()[name = string("query_states_45_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> query_states_45_strides_0 = const()[name = string("query_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> query_states_45_pad_0 = const()[name = string("query_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> query_states_45_dilations_0 = const()[name = string("query_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 query_states_45_groups_0 = const()[name = string("query_states_45_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1776_to_fp16 = const()[name = string("op_1776_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380871808)))]; |
| tensor<fp16, [?, 2048, 1, 5]> query_states_45_cast_fp16 = conv(dilations = query_states_45_dilations_0, groups = query_states_45_groups_0, pad = query_states_45_pad_0, pad_type = query_states_45_pad_type_0, strides = query_states_45_strides_0, weight = var_1776_to_fp16, x = var_1816_cast_fp16_0)[name = string("query_states_45_cast_fp16")]; |
| string key_states_45_pad_type_0 = const()[name = string("key_states_45_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> key_states_45_strides_0 = const()[name = string("key_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> key_states_45_pad_0 = const()[name = string("key_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> key_states_45_dilations_0 = const()[name = string("key_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 key_states_45_groups_0 = const()[name = string("key_states_45_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1777_to_fp16 = const()[name = string("op_1777_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385066176)))]; |
| tensor<fp16, [?, 256, 1, 5]> key_states_45_cast_fp16 = conv(dilations = key_states_45_dilations_0, groups = key_states_45_groups_0, pad = key_states_45_pad_0, pad_type = key_states_45_pad_type_0, strides = key_states_45_strides_0, weight = var_1777_to_fp16, x = var_1816_cast_fp16_0)[name = string("key_states_45_cast_fp16")]; |
| string value_states_45_pad_type_0 = const()[name = string("value_states_45_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> value_states_45_strides_0 = const()[name = string("value_states_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> value_states_45_pad_0 = const()[name = string("value_states_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> value_states_45_dilations_0 = const()[name = string("value_states_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 value_states_45_groups_0 = const()[name = string("value_states_45_groups_0"), val = int32(1)]; |
| tensor<fp16, [256, 1024, 1, 1]> var_1778_to_fp16 = const()[name = string("op_1778_to_fp16"), val = tensor<fp16, [256, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385590528)))]; |
| tensor<fp16, [?, 256, 1, 5]> value_states_45_cast_fp16 = conv(dilations = value_states_45_dilations_0, groups = value_states_45_groups_0, pad = value_states_45_pad_0, pad_type = value_states_45_pad_type_0, strides = value_states_45_strides_0, weight = var_1778_to_fp16, x = var_1816_cast_fp16_0)[name = string("value_states_45_cast_fp16")]; |
| tensor<int32, [4]> concat_44x = const()[name = string("concat_44x"), val = tensor<int32, [4]>([-1, 16, 128, 5])]; |
| tensor<fp16, [?, 16, 128, 5]> x_115_cast_fp16 = reshape(shape = concat_44x, x = query_states_45_cast_fp16)[name = string("x_115_cast_fp16")]; |
| tensor<int32, [4]> concat_45x = const()[name = string("concat_45x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> x_117_cast_fp16 = reshape(shape = concat_45x, x = key_states_45_cast_fp16)[name = string("x_117_cast_fp16")]; |
| tensor<int32, [4]> concat_46x = const()[name = string("concat_46x"), val = tensor<int32, [4]>([-1, 2, 128, 5])]; |
| tensor<fp16, [?, 2, 128, 5]> value_states_cast_fp16 = reshape(shape = concat_46x, x = value_states_45_cast_fp16)[name = string("value_states_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1842_cast_fp16 = mul(x = x_115_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1842_cast_fp16")]; |
| tensor<int32, [2]> var_1843_split_sizes_0 = const()[name = string("op_1843_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1843_axis_0 = const()[name = string("op_1843_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 16, 64, 5]> var_1843_cast_fp16_0, tensor<fp16, [?, 16, 64, 5]> var_1843_cast_fp16_1 = split(axis = var_1843_axis_0, split_sizes = var_1843_split_sizes_0, x = x_115_cast_fp16)[name = string("op_1843_cast_fp16")]; |
| bool var_1846_interleave_0 = const()[name = string("op_1846_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> var_1846_cast_fp16 = concat(axis = var_1781, interleave = var_1846_interleave_0, values = (var_1843_cast_fp16_1, var_1843_cast_fp16_0))[name = string("op_1846_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> var_1847_cast_fp16 = mul(x = var_1846_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1847_cast_fp16")]; |
| tensor<fp16, [?, 16, 128, 5]> query_states_cast_fp16 = add(x = var_1842_cast_fp16, y = var_1847_cast_fp16)[name = string("query_states_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1849_cast_fp16 = mul(x = x_117_cast_fp16, y = rope_cos_to_fp16)[name = string("op_1849_cast_fp16")]; |
| tensor<int32, [2]> var_1850_split_sizes_0 = const()[name = string("op_1850_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_1850_axis_0 = const()[name = string("op_1850_axis_0"), val = int32(-2)]; |
| tensor<fp16, [?, 2, 64, 5]> var_1850_cast_fp16_0, tensor<fp16, [?, 2, 64, 5]> var_1850_cast_fp16_1 = split(axis = var_1850_axis_0, split_sizes = var_1850_split_sizes_0, x = x_117_cast_fp16)[name = string("op_1850_cast_fp16")]; |
| bool var_1853_interleave_0 = const()[name = string("op_1853_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2, 128, 5]> var_1853_cast_fp16 = concat(axis = var_1781, interleave = var_1853_interleave_0, values = (var_1850_cast_fp16_1, var_1850_cast_fp16_0))[name = string("op_1853_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> var_1854_cast_fp16 = mul(x = var_1853_cast_fp16, y = rope_sin_to_fp16)[name = string("op_1854_cast_fp16")]; |
| tensor<fp16, [?, 2, 128, 5]> key_states_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1854_cast_fp16)[name = string("key_states_cast_fp16")]; |
| tensor<int32, [2]> var_1856_split_sizes_0 = const()[name = string("op_1856_split_sizes_0"), val = tensor<int32, [2]>([8, 8])]; |
| int32 var_1856_axis_0 = const()[name = string("op_1856_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1856_cast_fp16_0, tensor<fp16, [?, 8, 128, 5]> var_1856_cast_fp16_1 = split(axis = var_1856_axis_0, split_sizes = var_1856_split_sizes_0, x = query_states_cast_fp16)[name = string("op_1856_cast_fp16")]; |
| tensor<int32, [2]> var_1858_split_sizes_0 = const()[name = string("op_1858_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1858_axis_0 = const()[name = string("op_1858_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1858_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1858_cast_fp16_1 = split(axis = var_1858_axis_0, split_sizes = var_1858_split_sizes_0, x = key_states_cast_fp16)[name = string("op_1858_cast_fp16")]; |
| tensor<int32, [2]> var_1860_split_sizes_0 = const()[name = string("op_1860_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1860_axis_0 = const()[name = string("op_1860_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1, 128, 5]> var_1860_cast_fp16_0, tensor<fp16, [?, 1, 128, 5]> var_1860_cast_fp16_1 = split(axis = var_1860_axis_0, split_sizes = var_1860_split_sizes_0, x = value_states_cast_fp16)[name = string("op_1860_cast_fp16")]; |
| bool attn_weights_133_transpose_x_1 = const()[name = string("attn_weights_133_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_133_transpose_y_1 = const()[name = string("attn_weights_133_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_133_cast_fp16 = matmul(transpose_x = attn_weights_133_transpose_x_1, transpose_y = attn_weights_133_transpose_y_1, x = var_1858_cast_fp16_0, y = var_1856_cast_fp16_0)[name = string("attn_weights_133_cast_fp16")]; |
| fp16 var_1864_to_fp16 = const()[name = string("op_1864_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_135_cast_fp16 = mul(x = attn_weights_133_cast_fp16, y = var_1864_to_fp16)[name = string("attn_weights_135_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_137_cast_fp16 = softmax(axis = var_1781, x = attn_weights_135_cast_fp16)[name = string("attn_weights_137_cast_fp16")]; |
| bool var_1867_transpose_x_0 = const()[name = string("op_1867_transpose_x_0"), val = bool(false)]; |
| bool var_1867_transpose_y_0 = const()[name = string("op_1867_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> var_1867_cast_fp16 = matmul(transpose_x = var_1867_transpose_x_0, transpose_y = var_1867_transpose_y_0, x = var_1860_cast_fp16_0, y = attn_weights_137_cast_fp16)[name = string("op_1867_cast_fp16")]; |
| bool attn_weights_139_transpose_x_1 = const()[name = string("attn_weights_139_transpose_x_1"), val = bool(true)]; |
| bool attn_weights_139_transpose_y_1 = const()[name = string("attn_weights_139_transpose_y_1"), val = bool(false)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_139_cast_fp16 = matmul(transpose_x = attn_weights_139_transpose_x_1, transpose_y = attn_weights_139_transpose_y_1, x = var_1858_cast_fp16_1, y = var_1856_cast_fp16_1)[name = string("attn_weights_139_cast_fp16")]; |
| fp16 var_1870_to_fp16 = const()[name = string("op_1870_to_fp16"), val = fp16(0x1.6ap-4)]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_141_cast_fp16 = mul(x = attn_weights_139_cast_fp16, y = var_1870_to_fp16)[name = string("attn_weights_141_cast_fp16")]; |
| tensor<fp16, [?, 8, 5, 5]> attn_weights_cast_fp16 = softmax(axis = var_1781, x = attn_weights_141_cast_fp16)[name = string("attn_weights_cast_fp16")]; |
| bool attn_out_transpose_x_0 = const()[name = string("attn_out_transpose_x_0"), val = bool(false)]; |
| bool attn_out_transpose_y_0 = const()[name = string("attn_out_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [?, 8, 128, 5]> attn_out_cast_fp16 = matmul(transpose_x = attn_out_transpose_x_0, transpose_y = attn_out_transpose_y_0, x = var_1860_cast_fp16_1, y = attn_weights_cast_fp16)[name = string("attn_out_cast_fp16")]; |
| bool attn_output_interleave_0 = const()[name = string("attn_output_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 16, 128, 5]> attn_output_cast_fp16 = concat(axis = var_1784, interleave = attn_output_interleave_0, values = (var_1867_cast_fp16, attn_out_cast_fp16))[name = string("attn_output_cast_fp16")]; |
| tensor<int32, [4]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [4]>([-1, 2048, 1, 5])]; |
| tensor<fp16, [?, 2048, 1, 5]> x_119_cast_fp16 = reshape(shape = concat_47x, x = attn_output_cast_fp16)[name = string("x_119_cast_fp16")]; |
| string hidden_states_115_pad_type_0 = const()[name = string("hidden_states_115_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_115_strides_0 = const()[name = string("hidden_states_115_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_115_pad_0 = const()[name = string("hidden_states_115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_115_dilations_0 = const()[name = string("hidden_states_115_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_115_groups_0 = const()[name = string("hidden_states_115_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 2048, 1, 1]> var_1783_to_fp16 = const()[name = string("op_1783_to_fp16"), val = tensor<fp16, [1024, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386114880)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_115_cast_fp16 = conv(dilations = hidden_states_115_dilations_0, groups = hidden_states_115_groups_0, pad = hidden_states_115_pad_0, pad_type = hidden_states_115_pad_type_0, strides = hidden_states_115_strides_0, weight = var_1783_to_fp16, x = x_119_cast_fp16)[name = string("hidden_states_115_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_117_cast_fp16 = add(x = hidden_states_111_cast_fp16, y = hidden_states_115_cast_fp16)[name = string("hidden_states_117_cast_fp16")]; |
| fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1885_cast_fp16 = mul(x = hidden_states_117_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_1885_cast_fp16")]; |
| bool doubled_93_interleave_0 = const()[name = string("doubled_93_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_93_cast_fp16 = concat(axis = var_1784, interleave = doubled_93_interleave_0, values = (hidden_states_117_cast_fp16, var_1885_cast_fp16))[name = string("doubled_93_cast_fp16")]; |
| tensor<int32, [1]> out_47_axes_0 = const()[name = string("out_47_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_47_gamma_0_to_fp16 = const()[name = string("out_47_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390309248)))]; |
| fp16 var_1895_to_fp16 = const()[name = string("op_1895_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1895_to_fp16, gamma = out_47_gamma_0_to_fp16, x = doubled_93_cast_fp16)[name = string("out_47_cast_fp16")]; |
| tensor<int32, [2]> var_1906_split_sizes_0 = const()[name = string("op_1906_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1906_axis_0 = const()[name = string("op_1906_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1906_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1906_cast_fp16_1 = split(axis = var_1906_axis_0, split_sizes = var_1906_split_sizes_0, x = out_47_cast_fp16)[name = string("op_1906_cast_fp16")]; |
| string input_pad_type_0 = const()[name = string("input_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_strides_0 = const()[name = string("input_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_pad_0 = const()[name = string("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_dilations_0 = const()[name = string("input_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_groups_0 = const()[name = string("input_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1773_to_fp16 = const()[name = string("op_1773_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390313408)))]; |
| tensor<fp16, [?, 4096, 1, 5]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = var_1773_to_fp16, x = var_1906_cast_fp16_0)[name = string("input_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1914_cast_fp16 = silu(x = input_cast_fp16)[name = string("op_1914_cast_fp16")]; |
| string var_1919_pad_type_0 = const()[name = string("op_1919_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1919_strides_0 = const()[name = string("op_1919_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1919_pad_0 = const()[name = string("op_1919_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1919_dilations_0 = const()[name = string("op_1919_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1919_groups_0 = const()[name = string("op_1919_groups_0"), val = int32(1)]; |
| tensor<fp16, [4096, 1024, 1, 1]> var_1774_to_fp16 = const()[name = string("op_1774_to_fp16"), val = tensor<fp16, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398702080)))]; |
| tensor<fp16, [?, 4096, 1, 5]> var_1919_cast_fp16 = conv(dilations = var_1919_dilations_0, groups = var_1919_groups_0, pad = var_1919_pad_0, pad_type = var_1919_pad_type_0, strides = var_1919_strides_0, weight = var_1774_to_fp16, x = var_1906_cast_fp16_0)[name = string("op_1919_cast_fp16")]; |
| tensor<fp16, [?, 4096, 1, 5]> x_123_cast_fp16 = mul(x = var_1914_cast_fp16, y = var_1919_cast_fp16)[name = string("x_123_cast_fp16")]; |
| string hidden_states_119_pad_type_0 = const()[name = string("hidden_states_119_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> hidden_states_119_strides_0 = const()[name = string("hidden_states_119_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> hidden_states_119_pad_0 = const()[name = string("hidden_states_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> hidden_states_119_dilations_0 = const()[name = string("hidden_states_119_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 hidden_states_119_groups_0 = const()[name = string("hidden_states_119_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 4096, 1, 1]> var_1775_to_fp16 = const()[name = string("op_1775_to_fp16"), val = tensor<fp16, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407090752)))]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_119_cast_fp16 = conv(dilations = hidden_states_119_dilations_0, groups = hidden_states_119_groups_0, pad = hidden_states_119_pad_0, pad_type = hidden_states_119_pad_type_0, strides = hidden_states_119_strides_0, weight = var_1775_to_fp16, x = x_123_cast_fp16)[name = string("hidden_states_119_cast_fp16")]; |
| tensor<fp16, [?, 1024, 1, 5]> hidden_states_cast_fp16 = add(x = hidden_states_117_cast_fp16, y = hidden_states_119_cast_fp16)[name = string("hidden_states_cast_fp16")]; |
| int32 var_1933 = const()[name = string("op_1933"), val = int32(1)]; |
| fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1935_cast_fp16 = mul(x = hidden_states_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_1935_cast_fp16")]; |
| bool doubled_97_interleave_0 = const()[name = string("doubled_97_interleave_0"), val = bool(false)]; |
| tensor<fp16, [?, 2048, 1, 5]> doubled_97_cast_fp16 = concat(axis = var_1933, interleave = doubled_97_interleave_0, values = (hidden_states_cast_fp16, var_1935_cast_fp16))[name = string("doubled_97_cast_fp16")]; |
| tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [2048]> out_gamma_0_to_fp16 = const()[name = string("out_gamma_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415479424)))]; |
| fp16 var_1945_to_fp16 = const()[name = string("op_1945_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [?, 2048, 1, 5]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1945_to_fp16, gamma = out_gamma_0_to_fp16, x = doubled_97_cast_fp16)[name = string("out_cast_fp16")]; |
| tensor<int32, [2]> var_1956_split_sizes_0 = const()[name = string("op_1956_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1956_axis_0 = const()[name = string("op_1956_axis_0"), val = int32(1)]; |
| tensor<fp16, [?, 1024, 1, 5]> var_1956_cast_fp16_0, tensor<fp16, [?, 1024, 1, 5]> var_1956_cast_fp16_1 = split(axis = var_1956_axis_0, split_sizes = var_1956_split_sizes_0, x = out_cast_fp16)[name = string("op_1956_cast_fp16")]; |
| tensor<int32, [4]> x_begin_0 = const()[name = string("x_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> x_end_0 = const()[name = string("x_end_0"), val = tensor<int32, [4]>([0, 1024, 1, 1])]; |
| tensor<bool, [4]> x_end_mask_0 = const()[name = string("x_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])]; |
| tensor<fp16, [?, 1024, 1, 1]> x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, x = var_1956_cast_fp16_0)[name = string("x_cast_fp16")]; |
| string var_1974_pad_type_0 = const()[name = string("op_1974_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_1974_strides_0 = const()[name = string("op_1974_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_1974_pad_0 = const()[name = string("op_1974_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_1974_dilations_0 = const()[name = string("op_1974_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1974_groups_0 = const()[name = string("op_1974_groups_0"), val = int32(1)]; |
| tensor<fp16, [2048, 1024, 1, 1]> var_1969_to_fp16 = const()[name = string("op_1969_to_fp16"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415483584)))]; |
| tensor<fp16, [2048]> var_1968_to_fp16 = const()[name = string("op_1968_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419677952)))]; |
| tensor<fp16, [?, 2048, 1, 1]> var_1974_cast_fp16 = conv(bias = var_1968_to_fp16, dilations = var_1974_dilations_0, groups = var_1974_groups_0, pad = var_1974_pad_0, pad_type = var_1974_pad_type_0, strides = var_1974_strides_0, weight = var_1969_to_fp16, x = x_cast_fp16)[name = string("op_1974_cast_fp16")]; |
| string var_1974_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_1974_cast_fp16_to_fp32_dtype_0"), val = string("fp32")]; |
| tensor<fp32, [?, 2048, 1, 1]> lm_embed = cast(dtype = var_1974_cast_fp16_to_fp32_dtype_0, x = var_1974_cast_fp16)[name = string("cast_73")]; |
| } -> (lm_embed); |
| } |
